I run an agentic AI system in production called Tony.

Tony is a leadership AI. He channels twenty different leadership voices as discrete agent skills, sitting on top of .NET 10, Claude Opus, the Microsoft Agent Framework, and a SQL Server 2025 vector store. When I want a Drucker-flavored take on a strategic decision, I ask Tony. When I want a Reed Hastings-style stress test of a hiring plan, I ask Tony. When I want to know whether a particular product call is more Bezos or more Iger, I ask Tony.

I named him Tony because “Multi-Voice Leadership Reasoning Orchestrator with RAG-Backed Persona Skill Routing” is what Tony does. “Tony” is who Tony is.

According to a paper published this week in Harvard Business Review by a five-author team from BCG and Boston University, I shouldn’t have done that. Their research — Why You Shouldn’t Treat AI Agents Like Employees — argues that anthropomorphizing AI reduces individual accountability, increases unnecessary escalation, lowers review quality, raises employee uncertainty about their own roles, and doesn’t even improve adoption.

I’d like to gently disagree.

Specifically, I’d like to argue that the BCG team has performed a remarkable feat of category-error judo: they took two very different things, glued them together, and concluded that the resulting Frankenstein is bad. They’re right that the Frankenstein is bad. But the Frankenstein isn’t what most practitioners are doing.

Most practitioners are naming their agents Gary.

What the HBR/BCG Research Actually Says

Let me steelman it. Briefly.

The BCG team — Matthew Kropp, Julie Bedard, Emma Wiles, Megan Hsu, and Lisa Krayer — ran a large-scale experiment on what happens when organizations frame AI agents as “employees.” Place them on the org chart. Give them a role. Treat them as quasi-colleagues. Their five headline findings, per the published summary:

• Individual human accountability dropped.

• Unnecessary escalation increased.

• Review quality fell.

• Employees got more confused about their own roles.

• Adoption didn’t actually improve.

Their recommendation: don’t anthropomorphize. Redesign workflows, governance, and roles so humans stay clearly accountable and AI is supervised, not invited to the offsite.

Some of this is real. If you tell your staff that “Senior Strategic Associate Skylar (AI)” is a peer at Pay Band 4 with an HR record and a Slack avatar, you have created a category whose only function is to muddy the question of who, exactly, owns the decision when Skylar approves a $4M PO. That accountability fog is a thing that happens. I’ve watched it happen.

But here’s the move I want you to watch carefully: the BCG team takes “AI is a colleague on the org chart” — a specific, structural, top-down design choice — and labels it anthropomorphism. Then they extend the conclusion to the entire phenomenon of treating AI agents like creatures rather than like SaaS dashboards.

That’s not the same thing. Naming your invoice-processing agent “Gary the Invoice Goblin” is not the same as listing Gary on the org chart between Margaret in AP and Dave in FP&A. One is whimsy. One is HR malpractice.

And the broader research — the research the HBR piece doesn’t engage with — is pretty clear about which is which.

The Research the BCG Team Forgot to Cite

Let me walk through what the literature actually says about anthropomorphizing AI, because you’d never know from the HBR piece that any of this exists.

Exhibit A: Eighty percent of users want their AI more human, not less.

In December 2025, Schimmelpfennig, Diaz, Prabhakaran, and Davani — a team from the Max Planck Institute for Human Development and Google Research — published the first large-scale, causal, cross-cultural experiment on humanlike AI design. They actually manipulated whether chatbots felt more or less human, then measured downstream behavior. When asked whether they’d prefer their AI to be more or less humanlike in the future, eighty percent said more.

Honest read: the paper says effects are heterogeneous — humanlike design helps engagement and trust for some user groups more than others. But “heterogeneous” is not “harmful.” The paper’s actual contribution is to shred the assumption AI ethics has been operating under for years: that anthropomorphism is uniformly bad. It isn’t. For most users, it’s the thing that makes the tool work.

Exhibit B: The Roomba.

In 2007, Beki Grinter and Ja Young Sung at Georgia Tech ran one of the most cited studies in human-robot interaction. They surveyed 379 Roomba owners and did deep interviews with 30 committed users. Of those 30, twenty-one had named their robot. Sixteen referred to it as “he.” Two-thirds of the broader sample had given their vacuum a name.

It gets better. iRobot CEO Colin Angle has confirmed publicly that when iRobot tried to swap broken Roombas for new ones — standard appliance practice — customers refused. They didn’t want a replacement vacuum. They wanted Rosie fixed. They wanted Floorence healed. iRobot had to redesign its customer service workflow because its users had emotionally bonded with their machines via names.

If naming your robot vacuum reduced engagement, the Roomba would have been a failed product. Instead it spawned an entire category. Sentimental attachment via a goofy name turned out to be the adoption mechanism, not the obstacle to it.

Exhibit C: The military bomb robots.

This one is grim and instructive. Julie Carpenter at the University of Washington and, separately, then-Major Michael Kolb (whose 2012 Army dissertation surveyed 746 military personnel) studied soldiers using bomb-disposal robots like the PackBot and TALON.

The soldiers named their robots. They named them after wives, girlfriends, and Scooby-Doo. They held mock funerals when robots were destroyed. There is a documented case of a soldier who walked into an Army repair tent in Iraq with tears in his eyes and asked if the techs could “put Scooby-Doo back together.” He refused a brand-new replacement bot. He wanted Scooby-Doo.

Now, here’s the part the HBR paper would have to grapple with. Kolb baited his 746 subjects: would you rather lose your least-favorite human teammate or your robot? They picked the robot. Every time. The naming, the funerals, the homemade Purple Hearts welded onto chassis — none of it caused the soldiers to make worse decisions in the moments that mattered. Eighty percent of Kolb’s respondents said they loved or liked working with the robots. Performance held.

In the most extreme accountability environment we have — combat — anthropomorphic naming did not erode the human’s ability to make rational, hierarchical decisions about the robot’s role. The naming was a coping mechanism and an engagement mechanism. It was not a category-error machine.

Exhibit D: The actual GenAI workplace research.

A 2025 study in Knowledge Management Research & Practice ran a between-subjects experiment on a simulated enterprise GenAI deployment. Half the participants got a high-anthropomorphism interface (humanlike avatar). Half got a low-anthropomorphism one (robot-like avatar). Result: anthropomorphic features increased employee engagement and knowledge-collecting behavior.

A separate 2025 paper in Ergonomics, by Yang and colleagues at Tianjin University, tested three levels of anthropomorphic verbal feedback in AI assistants. Under high-anthropomorphism feedback, participants demonstrated significantly faster response speed. Moderate anthropomorphism was, oddly, the worst performer. The relationship is non-linear. The high end of the curve still wins.

Exhibit E: Humor itself, as an engagement vector.

Tiwari and colleagues published a 2024 ICSE paper titled — and this is real — With Great Humor Comes Great Developer Engagement. They surveyed 125 developers and analyzed three popular open-source projects (faker, lolcommits, and the satirical “volkswagen” library) for the role of humor in software work. Their conclusion: when practiced responsibly, humor increases developer engagement and supports cognition on hard tasks.

The most useful place to put humor in code, per the paper, is in tests and documentation. The most useful place to put humor in agentic AI deployments, I’d argue, is in the agent’s name.

This is the corpus the BCG paper is implicitly arguing against. It’s a deep one. The HBR piece does not, to my reading, engage with any of it.

The Category Error

Now we can be specific about where the HBR paper goes wrong.

There are two distinct things people mean when they say “treat AI agents like employees.”

Thing One: Putting an AI agent on the org chart. Giving it a title — “Senior Strategic Associate.” Slotting it between humans in a reporting hierarchy. Pretending, structurally, that it has agency, accountability, and presence the way a salaried employee does. In some companies they’re literally giving these agents Slack accounts with profile photos and asking humans to @-mention them in standups.

Thing Two: Naming your invoice-processing service Gary because everyone on the team has to refer to it forty times a day and “the invoice agent service” is a mouthful. Calling your scheduling agent Carl the Calendar Goblin because it makes you smile when Carl bombs out a meeting and you have to debug.

These are different. Possibly opposite.

Thing One creates accountability fog because it asks humans to treat the agent as a peer. The org chart does load-bearing semantic work in a company. When you put a non-conscious tool on it, you confuse the load.

Thing Two does the opposite. A name like Gary the Invoice Goblin is doing two things simultaneously: it makes the agent emotionally trackable (humans are very good at thinking about named entities), and it preserves the ontological distinction. Nobody at WhiteStar Labs thinks Gary the Invoice Goblin is a coworker. Gary is a goblin. Goblins are not on the org chart.

The funnier the name, in fact, the more aggressive the cognitive separation. “Steve, the strategy assistant” might confuse a new hire — wait, is Steve in the team meeting? “Bartholomew Q. Ledgerwell, Esq., Vice Marquis of Quarterly Variances” cannot possibly confuse anyone. He is obviously a tool with personality, the same way a printer named MacFlurry is obviously a printer.

The HBR research found that anthropomorphism causes role confusion. It does — org-chart anthropomorphism does. But goofy-name anthropomorphism is the cognitive opposite. It’s the vaccine, not the disease.

Why Practitioners Actually Name Their Agents

I want to make a practitioner’s case here, not just a research one. I run agents in production. I name them. Here is why.

1. Naming compresses communication bandwidth. When my engineers say “Tony’s drifting,” every person on the team knows which subsystem is hallucinating. The alternative — “the leadership orchestration agent’s persona-routing module is exhibiting policy drift in the Drucker skill” — is correct, but it costs ten times more cognitive load to say and parse. We name our microservices. We name our database clusters. We name our build pipelines. Naming an agent is the same engineering hygiene.

2. Funny names enforce psychological distance. This is the inverse of what HBR claims. When my team refers to Gary, they are constantly reminded that Gary is a thing, not a person. If we had instead deployed something called “Senior AP Specialist Gerald Munroe,” every interaction would chip away at the fact that there’s no Gerald.

3. Funny names enable healthy criticism. It is enormously easier to say “Gary made a stupid call” than “the invoice-processing AI exhibited an undesirable output, possibly stemming from training data bias or RAG retrieval errors, and we should escalate to engineering.” The first sentence is fast, blame-clear, and routes the issue. The second sentence is the kind of corporate buffering language that BCG presumably bills for.

4. Funny names lower the activation energy for use. This is a real adoption effect. Non-technical stakeholders will not interact with “Strategic Decision Support System v3.2.” They will interact with Tony. Anything that lowers the friction of getting humans to actually use the tool is a win, not a loss. (See again: the Roomba.)

5. Funny names mean nobody can put them on the org chart. You cannot, with a straight face, list “Bartholomew Q. Ledgerwell, Esq., Vice Marquis of Quarterly Variances” in your Workday HRIS. This is a feature.

The Steelman Holds — Just Not for the Reason BCG Thinks

I want to be honest. The BCG paper isn’t wrong about everything. The accountability fog is real. The unnecessary escalation is real. The lowered review quality is real. I’ve seen all three in client engagements over the last eighteen months.

But the cause isn’t anthropomorphism. The cause is bureaucratic anthropomorphism — the formal, org-chart variety that asks employees to treat AI like a colleague in the corporate-structural sense. That’s the part that breaks accountability, because the corporate structure is the place where accountability lives. Put a non-conscious tool there, you break the structure.

You fix that by not putting the AI on the org chart. You don’t fix it by stripping the agent of its name and calling it ID #4471-AGENT-INVOICE-PROD.

Falsifiable prediction: organizations that respond to the BCG research by stripping personality, names, and humor from their agentic AI deployments will report worse engagement and lower adoption than they had before. They will reproduce, as a side effect of their anti-anthropomorphism crusade, the very “adoption didn’t improve” finding the BCG team flagged. And they’ll be confused about why.

Meanwhile, the companies that deployed Gary the Invoice Goblin and Tony the Leadership Orchestrator and Bartholomew Q. Ledgerwell will be shipping.

Naming as Governance

Here’s the move I want practitioners to make. Stop treating “should we name our agents” as a vibes question. Treat it as a governance question.

Good governance for agentic AI looks like:

• Humans named, accountable, and signing off on consequential decisions.

• Agents named, distinct, and clearly tools — not peers.

• Funny names actively enforcing the distinction (a peer is Steve; a tool is Bartholomew the Spreadsheet Slayer).

• Org charts containing humans only.

• Audit logs containing agent names so you can trace which agent did what to whom.

This is, incidentally, exactly what BCG recommends in spirit — humans clearly accountable, agents clearly supervised — but they got there by recommending the wrong intervention. They want to sterilize the language. The right intervention is to enrich the language with names that make the agent’s nature obvious.

If you call your scheduling agent Carl, and one day Carl double-books your CFO into back-to-back board meetings, your team is going to laugh, fix Carl, and move on. If you call your scheduling agent “Calendar Optimization Service v3.2,” your team is going to write a postmortem with sixteen action items, three of which involve a working group. Carl is faster. Carl is also clearer about who’s accountable: the human who deployed Carl. Because, again, Carl is a goblin.

Meet Gary

So here is my recommendation, against Harvard’s:

Name your AI agents. Give them silly names. Give them ridiculous job titles that are obviously not real job titles. Make sure those names cannot be confused with the names of human employees. Make sure they cannot, accidentally or otherwise, end up on your org chart. Use them in your Slack channels, your debug logs, your internal docs, your team standups. Encourage your engineers to make jokes about them.

When the agents fail — and they will fail, often, in interesting ways — the name will be the unit of accountability. Gary screwed up. Gary needs a fix. Gary is not, has never been, and will never be a coworker. But Gary is also, today, processing thousands of invoices and saving the team real hours of human review time.

Gary works.

Tony works.

Bartholomew works.

The research — the actual, broader, longer body of research, not the slice the BCG team chose to publish — says the funny name is doing real cognitive labor. It compresses bandwidth, preserves ontological clarity, enables criticism, and lowers adoption friction. Strip it out and you don’t get a more rigorous workplace. You get a more bureaucratic one with worse engagement.

Harvard says don’t treat your AI like an employee. Fair enough — I don’t. I treat my AI like a goblin.

The goblin is fine.

Gal Ratner is the CTO of WhiteStar Labs and Chief Architect at Prana Entertainment. He builds enterprise agentic AI on .NET, Claude, and SQL Server. Tony, Gary, and Bartholomew say hi.

The numbers, on the surface, agree. Monthly question volume on Stack Overflow peaked at over 200,000 in 2014. By December 2024 it was 25,566 — a 76.5 percent drop from the month ChatGPT launched, and the lowest monthly count since May 2009, ten months after the site went live. Question volume has been below 50,000 a month for over a year. Traffic, depending on whose numbers you trust, is down somewhere between 35 and 60 percent off its peak. A site that once handled the volume of every Fortune 500 engineering team’s daily problem-solving now gets fewer questions per month than it did in its first full year of existence.

You’d be forgiven for assuming the obituary is overdue.

It isn’t. And the people who keep writing it have been ignoring something genuinely interesting that’s happened over the last twelve months: Stack Overflow has stopped pretending the world hasn’t changed, and started shipping like it’s 2009 again.

A site we forgot we needed

If you’re under thirty, you have to be told this. If you’re over thirty and write code for a living, you remember.

There was a moment — call it 2010 to about 2020 — when Stack Overflow was the always-open tab. You didn’t bookmark it because you didn’t need to. It was where you went when something broke, when a stack trace didn’t make sense, when you needed the regex you couldn’t quite remember, when you wanted to know if anyone else had seen this weird ConcurrentModificationException at 3 AM on a Tuesday. And nine times out of ten, someone had. The accepted answer was usually two years old, written by someone with a profile picture of a cat, and it just worked.

That wasn’t an accident. Stack Overflow was a designed solution to a hard problem: how do you let strangers help strangers without it devolving into a flame war? They invented the answer — reputation systems, structured Q&A, ruthless duplicate-closing, voting that surfaced the best response over the loudest one — and then exported that pattern to dozens of sister sites covering everything from cooking to academia. For roughly a decade, Stack Overflow was the closest thing software development had to a public utility.

Then ChatGPT shipped, and the utility started getting bypassed.

The quiet, awkward truth about what killed it

ChatGPT didn’t kill Stack Overflow. ChatGPT just made it embarrassing to need Stack Overflow.

You can argue about the data — and people have. The collapse in question volume is uncontested. But Stack Overflow’s traffic decline started in mid-2021, eighteen months before ChatGPT existed. The site had already alienated a chunk of its community with high-profile moderation disputes and the 2023 moderator strike. The new-question pipeline had been slowing for years, partly because the back catalog grew so large that almost any common question was a duplicate of something asked in 2014.

What ChatGPT did was remove the social cost of asking. There is no Stack Overflow user who hasn’t had a question closed as off-topic, or needs-more-focus, or duplicate-of-some-link-that-isn’t-actually-a-duplicate. There is no Stack Overflow user who hasn’t been condescended to. ChatGPT will never tell you to read the FAQ. It will never close your question. It will never reply with “this is a duplicate of [unrelated question from 2011].” It is, for better and worse, infinitely patient with bad questions.

That’s the deeper wound. The traffic collapse is a symptom. The disease is that a generation of developers learned they could solve their problem without ever having to ask another human for help. The feedback loop that built the site’s value — questioner, answerer, voter, future Googler — broke quietly. The questions stopped flowing in. The answers stopped getting written. The training data the LLMs depended on stopped getting created. The well that everybody was drinking from started running dry, and the people drinking didn’t notice because the water kept arriving in a different container.

That’s where it gets uncomfortable. Because if you’ve used a coding assistant in the last two years — if you’ve shipped a PR that included a snippet you got from Claude or Copilot or ChatGPT — you should know whose work you were almost certainly building on.

The OSS legacy nobody talks about

While everyone was writing eulogies, somebody pulled up the NuGet stats.

Dapper — the micro-ORM Stack Overflow built because Entity Framework wasn’t fast enough for their own site — has been downloaded over 350 million times. It is, in real terms, one of the most-deployed pieces of data-access code in the entire .NET ecosystem. Every time you’ve used connection.Query<T>(...) in a C# project, you were running something Stack Overflow gave away free.

StackExchange.Redis — the Redis client they built because the existing options didn’t handle production load — sits at over 63 million NuGet downloads. If you’ve written a .NET app that talks to Redis in the last decade, there’s a better-than-even chance you ran their code. It is the de facto standard.

MiniProfiler — the lightweight profiler that gave a generation of .NET web apps that little query-timing badge in the corner of every page — runs in production at companies that have never heard of Stack Overflow’s product team. Opserver, the dashboard they used internally to monitor SQL Server, Redis, Elastic, HAProxy, and exception logs, is open source. So is Bosun, their monitoring and alerting system. So is DNSControl, which now manages DNS for parts of the internet most engineers don’t think about. So is BlackBox, the Git secrets manager. So is Jil, a JSON serializer that for years was the fastest thing in the .NET world, and Sigil, the IL-generation library underneath it. So is the Stack Exchange Data Explorer, which lets anyone with a SQL brain query the entire community Q&A dataset directly. So is MarkdownSharp, which powered the original Markdown rendering pipeline used by half the early .NET web.

This is not a small footprint. It is, by any honest accounting, one of the most consequential open-source contributions any single web property has ever made — measured not in stars on a README, but in production deployments.

The irony writes itself. Stack Overflow, for fifteen years, gave the .NET ecosystem the tools it used to build the web. Then Stack Overflow gave the LLM industry the training data it used to replace Stack Overflow. The community did the work, the platform hosted it under Creative Commons, and somewhere between 2022 and 2024 a whole generation of trillion-parameter models learned to code by reading those answers — without anyone ever paying the people who wrote them, and without most users ever realizing where the answers came from.

That’s the bill we haven’t paid yet. And Stack Overflow, for better or worse, is the only entity left that’s still trying to send it.

The reinvention nobody noticed

This is the part of the story that’s actually news, and the part nobody is writing about because “Stack Overflow is dead” is a more clickable headline than “Stack Overflow is shipping again.”

In December 2025, Stack Overflow took AI Assist out of beta and made it generally available. It is not what you’d expect from a forum site that spent 2022 banning AI-generated answers. AI Assist is a conversational search interface that prioritizes community-verified answers first, then falls back to LLMs only when the existing community knowledge doesn’t cover the question. Sources are surfaced inline with citations and direct quotes. Original contributors get credit. If the knowledge base has a gap, the interface routes you toward asking the question — putting questions back into the system instead of letting them die in a private LLM chat.

It is the right architecture. It is the only architecture that makes long-term sense. An LLM-only answer is fast and frictionless and wrong often enough to matter. A community-only answer is slow and gatekept and offline for half the technologies that exist now. A hybrid that prefers the human-verified answer and only synthesizes when it has to is the actual shape of the future. Stack Overflow shipped it before any of its competitors did.

Then there is the MCP server. Stack Overflow’s Model Context Protocol server launched in beta in late 2025 — currently rate-limited to 100 requests per day per user, but live. Any MCP-compatible client — Claude Desktop, Cursor, VS Code with Copilot, your own .NET 10 agent built on Microsoft Agent Framework — can pull the trusted knowledge base directly into the agent’s context window through a standardized protocol. No bespoke API integration. OAuth through your existing Stack Overflow account. The first corporate-knowledge MCP that HP integrated into their developer-experience experiments was Stack Overflow’s, and HP’s distinguished technologist on the project called it “a robust source of truth” and “a very exciting precedent.” There is also an internal version — Stack Internal MCP — that lets enterprise customers pipe their own private Q&A into the same agent workflows.

Translation: instead of fighting the fact that every developer’s primary interface is now an AI agent, Stack Overflow is making sure the agent can ask their database. That is exactly the move. It is the only move that keeps the community contributing while still meeting users where they actually live.

And then there is ProLLM.

ProLLM is the benchmark suite Stack Overflow runs through Prosus Group, evaluating LLMs on actual Stack Overflow questions across two leaderboards: StackEval, with 925 historical samples from 2018 to 2023, and StackUnseen, with 194 questions from after the typical training cutoff. The point of StackUnseen is the point everyone in the LLM industry is currently pretending isn’t a problem: models that aren’t continuously retrained on new community Q&A fall behind on new and emerging content. New library versions, breaking API changes, new framework conventions, new error patterns — none of that exists in a model frozen at its training cutoff. ProLLM measures, with numbers, exactly how far behind the frontier models drift on questions humans actually ask in a given week.

This is, quietly, one of the most useful evals anyone is publishing. It is also a not-very-subtle argument that the LLM industry needs Stack Overflow more than Stack Overflow needs the LLM industry — because the second the community stops generating new answers, every model in the world starts going stale on the same schedule.

Why this matters more than the obituary writers think

There’s a thing the AI maximalists keep getting wrong. They look at Stack Overflow’s question-volume chart, draw a line to zero, and assume the world is fine without it because models have already absorbed everything worth knowing.

This is wrong in two ways.

First, it’s wrong technically. The most-upvoted answers on Stack Overflow are largely from 2012 to 2016. They are aging. The languages, frameworks, and runtimes they describe have all moved several major versions forward. A model trained on that corpus, without a mechanism to ingest new community-verified content, drifts toward giving you 2014’s answer to 2026’s question. That happens slowly enough that most users don’t notice — until they do, usually at 2 AM in a production incident.

Second, it’s wrong culturally. The reason Stack Overflow worked wasn’t the questions or the answers. It was the public-ness of the disagreement. Six people would post six different solutions to the same problem. The community would vote. The accepted answer wasn’t always the highest-voted. Comments would catch errors. Edits would update the answer when the underlying library changed. That whole structure — adversarial, skeptical, transparent — produced something LLMs structurally cannot: a record of how a real community of working engineers reasoned about a problem in public, including their mistakes.

You can replace the lookup. You can’t replace the reasoning trace. And the reasoning trace is exactly what most developers actually need when they’re stuck.

Give them their flowers

Here is what I’d like the rest of the industry to do, before it’s too late.

Stop writing the obituary. Stack Overflow’s question volume is collapsing, but the company has shipped more meaningful product in the last twelve months than it shipped in the previous five. AI Assist is the right architecture. The MCP server is the right integration. ProLLM is the right benchmark. The redesigned site is in beta and rolling out to all users in April 2026. They opened public chat to all registered users. They started supporting opinion-based questions in October 2025 — the exact category developers were taking to ChatGPT because Stack Overflow used to close them. They are, finally, listening.

Pay them. If your AI product was trained on the Stack Overflow data dump — and it was, because everyone’s was — at minimum link to the original answer when you surface its content. Better, license the data the way Google, OpenAI, and GitHub already have. Even better, contribute. The next time your team solves a hard problem with a coding assistant, take ten minutes and write up the question and answer on Stack Overflow itself. The feedback loop only restarts if the people who currently take from it start putting back.

Credit the OSS. Dapper. StackExchange.Redis. MiniProfiler. Opserver. Bosun. DNSControl. Jil. Sigil. MarkdownSharp. The Data Explorer. These are not curiosities. These are infrastructure. They were built by Stack Overflow engineers, given away free, and run in production at companies that wouldn’t exist without them. The next time someone tells you “the community” gave us modern open-source .NET, the correct answer includes a specific organization name.

Hope they make it. I’m going to write this part plainly, because I don’t have the energy to be cynical about it. I want Stack Overflow to come back. I want the question feed to start filling up again. I want the next generation of developers to learn what it’s like to have their first answer roasted in the comments by a stranger who is, annoyingly, correct. I want there to be a place that isn’t a private LLM chat where the canonical answer to a hard problem lives in public, where someone can edit it when it gets out of date, and where the people who wrote it get credit for the next fifteen years.

Stack Overflow had every reason to fold. Question volume in free fall. A parent company under pressure. The entire developer industry pivoting to interfaces that bypass them. They could have shut the lights off in 2024 and most of the noise would have been people complaining about their own moderators on the way out.

Instead they shipped AI Assist, an MCP server, an LLM benchmark, and a beta redesign. They opened up chat. They reopened opinion-based questions. They are doing the actual work of meeting the moment.

That is not the behavior of a dying company. That is the behavior of a company that has finally figured out what it is for in the new world — the curated, human-verified, citable layer underneath the agentic developer workflow — and is building toward it.

Don’t bury Stack Overflow yet. They might just be rebuilding the road we’re all already driving on, one more time, while we’re not looking.

Gal Ratner is the CTO of WhiteStar Labs and Chief Architect at Prana Entertainment. He writes about the working architect’s view of AI, infrastructure, and the platforms developers actually depend on at galratner.com.

You’ve been told a story.

It goes like this: artificial intelligence is a triumph of American capitalism. Brilliant founders at Microsoft, Google, Amazon, OpenAI, and Nvidia raised private capital, hired the best engineers in the world, and built world-changing technology. Now they’re going to “transform the workplace.” Innovation. Disruption. Progress. Cue the keynote.

Every word of that story is a lie of omission.

The truth — buried inside CHIPS Act award letters, state tax disclosures, FERC filings, municipal bond prospectuses, DoD procurement records, OMB use-case inventories, and a White House executive order most Americans have never heard of — is that the most expensive technology in human history was built on the backs of American taxpayers. Yours. Mine. The single mom in Loudoun County whose property tax rates are about to climb to plug the data-center revenue hole. The retiree in Phoenix whose electric bill jumped 32% in five years to subsidize Microsoft’s water-cooled GPU farm. The Texas plumber whose state quietly blew $3 billion on a sales-tax exemption originally scored at $14 million.

And here’s the part that should make every American taxpayer set their hair on fire:

The same technology you funded is now being used by every Fortune 500 CEO on Earth to figure out how to fire you.

We didn’t just fund our own replacement. We’re still funding it. Today. This quarter. And both parties — Republicans and Democrats, governors and senators, federal and state — are still signing the checks.

Let me name names.

The Five Pipes

Public money flows into the AI industry through five distinct pipes. Each one is mostly invisible to the people paying for it. Together they constitute the largest, quietest transfer of public wealth in modern American history. Pull the documents and the dollar amounts are not in dispute. Only the consequences are.

Pipe 1: The Federal R&D Faucet — $3.3 billion a year, every year, forever

The foundational algorithms behind ChatGPT, Gemini, Claude, and every model OpenAI ships didn’t spring from a Sand Hill Road pitch deck. They were invented across decades by researchers funded by the National Science Foundation (NSF), DARPA, the National Institutes of Health (NIH), and the national labs of the Department of Energy. NITRD — the federal coordinating office for AI and IT research — requested $3.316 billion for AI R&D in FY2025 alone, with $1.95 billion in “Core AI” and another $1.36 billion embedded as “AI Crosscut” inside other agencies. NIH alone planned $1.12 billion in AI investment in FY25.

NSF and the USDA jointly fund 25 National AI Institutes scattered across U.S. universities. The Department of Energy funds open-access HPC time at Lawrence Berkeley, Oak Ridge, Argonne, Savannah River. Every commercial frontier model that today commands a $200 billion private valuation started life on a federal grant ledger.

This is not new spending we can simply vote down. The High Performance Computing Act of 1991 — signed by George H.W. Bush, faithfully reauthorized by Clinton, Bush, Obama, Trump, and Biden — baked it in. The National Security Commission on Artificial Intelligence (chaired by Eric Schmidt, yes, that Eric Schmidt) recommended doubling non-defense AI funding to $32 billion by FY26. Every administration of either party wants the spigot wider.

Pipe 2: The CHIPS Act — $53 billion straight to the silicon

In August 2022, Congress passed the CHIPS and Science Act. It directed roughly $53 billion to domestic semiconductor manufacturing and added a 25% refundable tax credit for “qualified capital expenditures.” Note “refundable.” That isn’t a tax break. That’s a check the IRS cuts to chip companies if their tax bill is smaller than the credit.

This was not a partisan bill. It was a bipartisan bill, in origin and in passage. The CHIPS for America Act was introduced in June 2020 — during the first Trump administration — by Senators John Cornyn (R-TX) and Mark Warner (D-VA). The companion Endless Frontier Act that became the “Science” half was introduced by Senators Todd Young (R-IN) and Chuck Schumer (D-NY) with Representatives Ro Khanna (D-CA) and Mike Gallagher (R-WI). The TSMC Arizona deal that catalyzed the entire effort was brokered by the Trump State Department in May 2020. Two years later, CHIPS passed the Senate 64–33. Seventeen Republicans voted yes. Here are their names, in alphabetical order, for the historical record:

• Roy Blunt (R-MO)

• Richard Burr (R-NC)

• Shelley Moore Capito (R-WV)

• Bill Cassidy (R-LA)

• Susan Collins (R-ME)

• John Cornyn (R-TX) — a co-sponsor with Mark Warner (D-VA)

• Steve Daines (R-MT)

• Lindsey Graham (R-SC)

• Bill Hagerty (R-TN)

• Mitch McConnell (R-KY)

• Jerry Moran (R-KS)

• Rob Portman (R-OH)

• Mitt Romney (R-UT)

• Ben Sasse (R-NE)

• Thom Tillis (R-NC)

• Roger Wicker (R-MS)

• Todd Young (R-IN) — the lead Republican on the bill

In the House, 24 Republicans joined every Democrat except one (Sara Jacobs, who voted “present”) to pass it 243–187. The only senator caucusing with Democrats who voted no was Bernie Sanders, who called it “massive corporate welfare for the semiconductor industry.” On this one, Bernie was right.

Where did your $53 billion go? In part:

• SK hynix: $458M direct + $500M loans for an HBM packaging plant explicitly built for AI chips, in West Lafayette, Indiana

• Texas Instruments: up to $1.6B direct + $3B loans for fabs in Sherman, TX and Lehi, UT

• GlobalFoundries: $1.45B direct in Malta, NY

• Amkor Technology: $407M direct for advanced packaging in Peoria, AZ — the award letter explicitly cites “AI chip demand”

Within eighteen months of enactment, Commerce Secretary Gina Raimondo announced the $39 billion commercial fund was already oversubscribed by $31 billion. The taxpayer ran out of money before the industry ran out of demand. R&D agencies got an $8 billion shortfall as a parting gift, because the chip companies needed it more than American researchers did.

The pipeline did not end with the Biden signature. The second Trump administration expanded it. Executive Order 14179 (January 2025, “Removing Barriers to American Leadership in Artificial Intelligence”) and Executive Order 14318 (July 23, 2025, “Accelerating Federal Permitting of Data Center Infrastructure”) direct the Commerce Department to launch additional financial support — loans, loan guarantees, grants, tax incentives, and offtake agreements — for any “Qualifying Project” with $500M+ in capex or 100+ MW of new AI load. EO 14318 also declares that federal financial assistance representing less than 50% of total project cost is “presumed not to constitute a major Federal action” under NEPA. Translation: Washington can subsidize up to 49 cents on the dollar of a billion-dollar private AI campus and skip environmental review entirely. Different administration, same pipe, wider valve.

Pipe 3: Defense and Civilian Procurement — from $269 million to $13.4 billion in two years

Between 2023 and 2024, Department of Defense AI contract values exploded by 1,500% — from $269 million to $4.32 billion. By FY26 the Pentagon’s projected sustained AI investment is $13.4 billion. The Army handed Palantir a $10 billion commercial software contract. Treasury awarded a $20 billion cybersecurity Blanket Purchase Agreement. The OMB documented 3,611 AI use cases across 56 federal agencies in 2025 — a 105% jump in a single year. HHS leads with 447. NASA: 425. VA: 367. DOE: 340. DOJ: 314. The federal government has become a market-making consumer of commercial AI — with your money.

In early 2025 the Pentagon, under Defense Secretary Pete Hegseth, signed approval pacts with SpaceX, OpenAI, Google, Microsoft, Nvidia, AWS, and Reflection. The DoD then placed a “supply chain risk” designation on Anthropic — a label conventionally reserved for vendors with hostile foreign-state ties — because Anthropic refused to relax its safety guardrails on autonomous weapons and mass surveillance. This is your Pentagon, today, using your tax dollars to actively select for AI vendors willing to build the most aggressive systems and to lock out the ones that won’t.

Both administrations expanded this pipe. Biden’s DoD started the procurement surge. Trump’s DoD industrialized it.

Pipe 4: State Tax Abatements — Billions a Year, Across 38 States, Bipartisan Statehouse Theater

Thirty-eight states offer dedicated tax incentives for data centers. Eleven extend abatements to local property taxes. Fourteen exempt the electricity these facilities consume — the same electricity already being subsidized by residential ratepayers (more on that in Pipe 5). The numbers aren’t theoretical. Pull the audits:

Virginia — the epicenter

Cumulative losses since 2010: roughly $2.7 billion. FY23: $685 million. FY24: $1 billion. FY25: between $1.6 billion and $1.9 billion in a single year. The data center exemption now constitutes nearly 80% of all state economic incentive spending — dwarfing the $750 million Amazon HQ2 grant. This program was built and expanded under Democratic Governors Terry McAuliffe and Ralph Northam. It is currently being expanded under Republican Governor Glenn Youngkin. Loudoun County — where data centers fund nearly half of county tax revenue — projects $1.3 billion in data center tax revenue for 2026, but the county budget office openly warns that any slowdown forces residential property and vehicle tax hikes. They built a fiscal house on a single industry’s boom. When the boom slows, you pay.

Texas — a $14 million giveaway that grew to $3 billion

A 2013 sales tax exemption originally scored at $14 million has grown to a projected $3 billion taxpayer cost. The Texas Senate Finance Committee, controlled by Republicans, is now belatedly trying to figure out what just happened. Governor Greg Abbott has signed every expansion since.

Georgia — a 7.4-to-1 net loss audited by the state itself

A 2025 Georgia state audit found data centers brought in $41 million in tax revenue while pulling $474 million in exemptions — a $433 million net loss in a single fiscal year. The auditors went further: they concluded 70% of these projects would have come to Georgia anyway, with no subsidy required. The forfeited revenue, they calculated, would have funded nearly 8,000 public sector jobs in education, healthcare, and public safety. Republican Governor Brian Kemp has not moved to fix it.

Nevada — $4.5 million per job

Customized “SB1” abatement packages cost the state up to $236 million in school support taxes and over $209 million in consolidated taxes per agreement. The qualifying threshold for a 20-year abatement: invest $100 million and employ 50 full-time Nevada residents. That works out to about $4.45 million in state tax forgiveness per job created. These deals were granted by Republican Governor Brian Sandoval, expanded by Democratic Governor Steve Sisolak, and renewed by Republican Governor Joe Lombardo. Bipartisan in every administration.

North Carolina — the wave is coming

With 6.3 GW of data center pipeline announced, North Carolina is on track for $1.5–$2.3 billion in lost taxes during construction alone, then approximately $450 million per year in operation — a 900% jump over current losses. The qualifying threshold remains a $75 million investment for unlimited abatement. Built under Democratic Governors Roy Cooper and Josh Stein, with no meaningful Republican opposition.

Ohio — even the industry-funded study admitted the game is over

Ohio offers a 100% sales tax exemption on data center equipment. Even the industry-commissioned report claiming a $2.7 billion “net gain” from 2017 to 2024 warned the incentive arms race is effectively over: 37 peer states now offer identical exemptions, so the break has become redundant table stakes. Republican Governor Mike DeWine is keeping the exemption anyway.

Pipe 5: The Stealth Tax on Your Power Bill, Your Water Bill, and Your Municipal Bonds

This is the pipe nobody wants to talk about, because it shows up on receipts you’ve already paid.

Your power bill is subsidizing Microsoft’s margin

State-regulated utility monopolies are negotiating sweetheart “special contracts” with hyperscalers below the cost of service. Internal Duke Energy documents revealed plans to shift a $325 million customer discount onto residential ratepayers. In Maryland, an antiquated FERC cost-allocation formula stuck Maryland residents with the majority of the state’s $500 million share of a $5 billion regional transmission build, most of it driven by data center load. In the PJM grid (65 million people across 13 states), power supply costs jumped from $2.2 billion to $14.7 billion in a single year. Data centers accounted for nearly two-thirds of the increase. Residential electricity rates rose roughly 32% nationally between July 2020 and July 2025. That is not “inflation.” That is a redistribution.

And to make sure the redistribution scales, the Department of Energy’s Grid Resilience and Innovation Partnerships (GRIP) program — funded by the Bipartisan Infrastructure Law — has dished out $4.2 billion in round-two funding “to enable additional grid capacity to meet load growth stemming from data centers.” The DOE’s “Speed to Power” Request for Information is hunting transmission projects that can move 20 GW of incremental load specifically for AI. The Trump White House’s July 2025 EO 14318 then streamlined FAST-41 permitting and identified NEPA categorical exclusions for any qualifying AI infrastructure project. Both parties paved the road.

Your water bill is cooling Sam Altman’s GPUs

A typical hyperscale data center uses up to 5 million gallons of water per day — the daily consumption of a town of 50,000 residents. UC Riverside and Caltech researchers calculated that municipal water systems will need $10–$58 billion in new public infrastructure investments to handle peak data center demand. Within four years, AI cooling alone could require an additional 1.45 billion gallons per day at peak. Tech companies occasionally chip in for token water-reuse projects — a $31 million reuse facility in Quincy, Washington co-funded with Microsoft is the favorite case study. Cities pay for the rest, with bonds.

Your municipal bonds are being weaponized against you

Public power utility bond issuance hit $26.8 billion in 2024 — nearly double the 10-year average of $14 billion. The Large Public Power Council and the American Public Power Association are now actively lobbying Congress and Treasury to loosen IRS “private use” rules so municipalities can issue tax-exempt debt explicitly to build dedicated infrastructure for private AI firms. If they succeed, your local government will borrow on your dime, at a tax-exempt rate that costs the federal Treasury, to construct proprietary infrastructure for trillion-dollar tech companies. Public risk. Private asset. Fee for service: zero.

The Punchline

So — the federal government funded the foundational research. The federal government subsidized the chips. The federal government wrote the procurement contracts. The states gave away the property taxes, sales taxes, and electricity taxes. The utilities cross-subsidized the power. The municipalities issued the debt and built the water systems. The taxpayers absorbed the rate hikes.

And now what?

Now Goldman Sachs estimates 300 million jobs globally are exposed to AI automation. CBRE Research reports that AI investment in 2025 was equivalent to roughly half of total U.S. GDP growth — meaning the rest of the economy was essentially flat. Employment among entry-level workers in the most AI-exposed occupations is already 13% lower than for less-exposed roles since 2022. The International Labour Organization adjusted its 2025 outlook downward by 7 million jobs.

You paid for the algorithms. You paid for the chips. You paid for the data centers. You paid for the power grid. You paid for the water mains. You paid for the municipal bonds. And now the same companies that took all that public money are using it to figure out how to remove you from the payroll — and to sell that capability, as a service, to your employer.

When you are laid off and your replacement is a transformer model trained on a federally funded GPU cluster, cooled by city water, powered by ratepayer-subsidized electricity, hosted in a building that paid no property tax, on equipment that paid no sales tax, on land permitted under a NEPA categorical exclusion — that’s not “the free market.” That is a wealth transfer from labor to capital, and you funded it.

And here’s the cherry on top: federal payroll taxes — the dollars that fund Social Security and Medicare — are calculated as a percentage of W-2 wages. Every job replaced by an algorithm is a permanent hole in the payroll-tax base. RAND modeled a scenario in which AI-driven price deflation drops Nominal GDP enough to plunge federal tax revenues by 25%. Once the Social Security trust fund is exhausted, the only options are benefit cuts to retirees or deficit-financed bailouts. Either way, the bill comes back to you.

The Names

I am not going to let any politician hide behind “well, technology is complicated.” Every name below voted for, signed, or actively expanded a specific funding pipe described in this article. Look up your representatives. If they’re on this list, they earned a primary challenge or a general-election loss — regardless of party.

CHIPS Act yes votes that handed $53B to chipmakers

Senate Republicans (17): Blunt (MO), Burr (NC), Capito (WV), Cassidy (LA), Collins (ME), Cornyn (TX), Daines (MT), Graham (SC), Hagerty (TN), McConnell (KY), Moran (KS), Portman (OH), Romney (UT), Sasse (NE), Tillis (NC), Wicker (MS), Young (IN).

Senate Democrats and Independents: every member of the Democratic caucus except Bernie Sanders. That includes Schumer (NY), Warner (VA, co-sponsor), Cantwell (WA), Klobuchar (MN), Wyden (OR), Brown (OH, then), Tester (MT, then), Manchin (WV, then), Sinema (AZ, then) — the entire caucus.

House Republicans (24): Anthony Gonzalez (OH), Brian Fitzpatrick (PA), Chris Jacobs (NY), Fred Upton (MI), John Katko (NY), Kay Granger (TX), Liz Cheney (WY), Michael McCaul (TX), Peter Meijer (MI), Rodney Davis (IL), Steve Chabot (OH), Tom Cole (OK), Troy Balderson (OH), Young Kim (CA), Adam Kinzinger (IL), and others. (Some have since left office; their successors should be asked where they would have voted.)

House Democrats: every member except Sara Jacobs (CA), who voted “present.”

Federal officials who built and ran the AI subsidy pipeline

• Joe Biden — signed CHIPS into law; signed EO 14141 in January 2025 (later revoked) on AI infrastructure on federal lands.

• Donald Trump — signed EO 14179 (January 2025) directing removal of “barriers to AI leadership”; signed EO 14318 and the AI Action Plan (July 23, 2025) directing Commerce to launch financial support for “Qualifying Projects,” declaring federal subsidies under 50% of project cost presumptively non-NEPA, opening federal lands at Idaho National Lab, Oak Ridge, Paducah, and Savannah River for data center development; pushed the One Big Beautiful Bill in mid-2025 that further altered the energy/AI investment landscape.

• Gina Raimondo (D) — Commerce Secretary; ran the original CHIPS rollout, lobbied senators with closed-door national security briefings to flip swing votes.

• Howard Lutnick (R) — Commerce Secretary, current; tasked with launching the post-EO financial support initiative for data center “Qualifying Projects.”

• Pete Hegseth (R) — Secretary of Defense; presided over the 1,500% Pentagon AI contract surge, signed approval pacts with seven major commercial vendors, blacklisted Anthropic over safety guardrails.

• Chris Wright (R) — Secretary of Energy; “Speed to Power” RFI for 20 GW of new AI-driven transmission; selected federal lab sites for hosting commercial data centers.

• Lee Zeldin (R) — EPA Administrator; tasked with rewriting Clean Air Act, Clean Water Act, CERCLA, and TSCA permitting to expedite “Qualifying Projects.”

Governors actively running data center subsidy programs at a documented net loss

• Glenn Youngkin (R-VA) — expanding the largest data center exemption regime in the country during a $1.6–$1.9B annual revenue loss

• Ralph Northam and Terry McAuliffe (D-VA, retired) — architects of the Virginia regime

• Greg Abbott (R-TX) — every CHIPS-era expansion of the Texas data center exemption

• Mike DeWine (R-OH) — 100% sales tax exemption, kept after the industry’s own report admitted it’s redundant

• Brian Kemp (R-GA) — the $433M-net-loss program flagged by his own state auditor

• Joe Lombardo (R-NV) — renewed SB1-style abatements

• Brian Sandoval (R-NV, retired) and Steve Sisolak (D-NV, retired) — built the Nevada regime

• Roy Cooper (D-NC, retired) and Josh Stein (D-NC) — built and continue the North Carolina pipeline

One political honorable exception worth noting: Governor Katie Hobbs (D-AZ) proposed in 2025 to roll back data center exemptions and impose a water-volume surcharge to fund a Colorado River Protection Fund. The Republican-controlled Arizona Legislature has resisted. If you live in Arizona, this is the rare cross-aisle fight where the lines are not what you might assume — vote accordingly.

The lobbyists you’ve never heard of who run the racket

Most state-level data center legislation is not drafted by your state legislators. It is drafted, edited, and shopped by the same handful of trade groups in every capitol in America:

• The Data Center Coalition

• The Information Technology Industry Council (ITI)

• The Large Public Power Council

• The American Public Power Association

• The U.S. Chamber of Commerce

They write the model bills. State legislators paste their names on top. If your state legislator can’t explain the deal in their own words, that’s why.

The corporate beneficiaries who took your money

These are the names on the awards letters, the procurement contracts, and the special-rate filings. Not anonymous. Not unreachable. Their CEOs all have email addresses.

• Microsoft — Satya Nadella

• Alphabet/Google — Sundar Pichai

• Amazon/AWS — Andy Jassy

• Meta — Mark Zuckerberg

• Nvidia — Jensen Huang

• OpenAI — Sam Altman

• Palantir — Alex Karp

• SpaceX — Elon Musk

• Oracle — Safra Catz

• SK hynix, Texas Instruments, GlobalFoundries, Amkor Technology — the CHIPS award recipients on the silicon side

What You Can Actually Do

Outrage without action is just entertainment for the powerful. Here’s a practical playbook — ranked roughly from highest leverage to lowest, but every one of them matters.

1. Vote out every incumbent who voted for the CHIPS Act, regardless of party. Look up your senator and your representative’s vote. Senate roll call: senate.gov, vote 117-2-271. House roll call: clerk.house.gov. If they voted yes, vote them out — primary challenger, general-election opponent, doesn’t matter. Pin the vote to their record. Forward the roll call to ten people you know in their district.

2. Vote out any state legislator who voted to create or expand a data center tax abatement. These exemptions, dollar-for-dollar, are the single worst deal in your state tax code right now. State roll calls are public; many state newsrooms now publish searchable databases. Find yours. The single mom in Loudoun and the rancher in West Texas have a common interest in this fight even if they disagree on everything else.

3. Show up at your Public Utility Commission rate hearings. PUCs are where utilities get permission to socialize hyperscaler discounts onto your bill. These hearings are sparsely attended and notice is buried. Sign up for your state PUC’s docket alerts. When a “special contract” or rate-base expansion comes up, file a public comment. Five organized commenters at a PUC hearing have more impact than five thousand angry tweets.

4. Demand transparency on every data center deal in your county. Property tax abatements, special assessment districts, water guarantees, electrical interconnect agreements — all of it. Counties hate publishing these in plain language because the deals are usually indefensible. Force them. FOIA the term sheets. Publish them. Read them at the public-comment portion of board meetings.

5. Oppose the “private use” loosening on tax-exempt municipal bonds. This is the active fight in Washington right now. The Large Public Power Council and APPA are pushing the IRS and Treasury to let municipalities issue tax-exempt debt for dedicated AI infrastructure. Call your senators. Tell them no. Tell them you understand exactly what it does — privatize the asset, socialize the risk, stiff the federal Treasury. They will be shocked you know what private-activity bond rules are. Use it.

6. Support the moratoriums. Over 100 local communities have already passed zoning or permitting moratoriums on new data centers. New York, South Dakota, and Oklahoma have moratorium bills moving at the state level. Indiana’s HB1333 wants data centers to contribute 1% of their exempted taxes back to local government. Arizona’s Hobbs proposed a water surcharge. Maine and Missouri are banning AI therapy chatbots. These bills usually die quietly because nobody shows up. Show up.

7. Demand a “robot tax” or AI excise tax on your state’s revenue agenda. New York’s proposed Robot Tax Act (A3719) would impose a surcharge on corporations using AI to displace human employees. Whether it’s that exact bill or a different framing — a payroll-replacement tax, a token tax on AI-generated outputs, an excise on autonomous transactions — the principle is identical: if a corporation captures the value of automation, it should fund the social costs of automation. Your state representatives need to hear this from you, not from a Brookings paper.

8. Read the public records yourself and quote them. The OMB AI use-case inventory, the CHIPS award letters, the state tax incentive evaluations, the FERC filings, the GRIP awards — every dollar is public, named, and searchable. Most of your fellow citizens don’t know these exist. Forward them. Quote them. Make them inescapable in your local civic conversation. The single most subversive thing an American taxpayer can do right now is read their own government’s budget.

The Bottom Line

This isn’t a Republican scandal. It isn’t a Democratic scandal. It is an American scandal — a bipartisan, federal-state-municipal, executive-legislative, public-private machine for moving wealth out of the working and middle class and into a handful of trillion-dollar tech corporations and their executives, while using public funding to build the very technology that will eventually shrink the labor force funding it.

The good news is that, for once, the machine is easy to see. The dollar amounts are documented. The roll-call votes are public. The award letters are scanned and searchable. The PUC dockets are scheduled. The moratorium bills are filed. The names are knowable.

You weren’t supposed to follow the money. You’re supposed to keep believing the story about the brilliant entrepreneurs and the glorious innovation and the inevitability of it all. The entire AI mythology assumes you will not read the federal budget, the state audits, the FERC filings, and the bond prospectuses.

Read them. Then vote.

On May 1, 2026, Ask.com—the search engine that was once Ask Jeeves—went dark. The farewell note on the homepage was short and almost apologetic. “Every great search must come to an end,” it read, before thanking the engineers, the designers, and the millions of users who showed up over the last twenty-five years.

Nobody noticed for about a day. Then the obituaries started trickling in. Then a stranger thing happened: people on LinkedIn, on X, in the back channels of Slack workspaces—grown adults, professionally serious people—started getting weirdly emotional about a cartoon butler who used to take questions in plain English on a 56k modem.

Here is the part nobody quite wants to say out loud: Jeeves was right. Thirty years too early, sure. Stomped flat by Google’s PageRank, sure. Strangled in middle age by toolbar bundling, absolutely. But on the core idea—that you should be able to type a sentence into a machine and have the machine just answer it—Jeeves nailed the brief in 1996.

Every conversational AI you use today is, in some sense, Ask Jeeves with better hardware and a real budget.

So before we throw dirt on the grave, let’s actually tell the stories. They’re better than you remember.

The butler was named after a butler

The real Jeeves is a fictional valet from a series of comic stories by P.G. Wodehouse—what Wodehouse called “a gentleman’s personal gentleman”—who serves the spectacularly inept Bertie Wooster and routinely outwits his employer with an air of brittle British competence. When venture capitalist Garrett Gruener and software engineer David Warthen sat down in Berkeley in 1996 to invent a new kind of search engine, they decided the metaphor was perfect: the web should answer you the way a good valet would. Discreetly, knowledgeably, without making you feel stupid for not knowing where you put your hat.

The original code was written by Gary Chevsky. The brand was a literary reference most of its eventual users never caught. The company was incorporated in June 1996, almost exactly two years before two Stanford grad students named Larry Page and Sergey Brin filed Google’s first patent application.

Sit with that for a second. Ask Jeeves launched a public beta in April 1997. Google.com didn’t exist as a registered domain until September 1997. For roughly eighteen months, “ask the internet a question in plain English and get an answer” was a more developed product than “type keywords into a clean white box.”

History is unforgiving about who finished first. It only remembers who shipped at scale.

The IPO that should have been a warning sign

Ask Jeeves went public on July 1, 1999. By contemporary accounts, it was the third best first-day IPO performance in business history at that point—the kind of debut where bankers high-five each other and CNBC anchors lose composure on live television. The stock peaked at $190 a share later that year. The company’s CEO, Robert Wrubel, became famous in the way people became famous in 1999: briefly, expensively, and on the cover of magazines that no longer exist.

Ask Jeeves spent that money the way every dot-com spent that money. They bought Super Bowl-adjacent ad inventory. They paid for a four-story Jeeves balloon to float down Central Park West in the Macy’s Thanksgiving Day Parade from 2000 through 2004, complete with what the company called the “golden keys of knowledge” clutched in his white-gloved hand. They printed stickers on bananas and apples and oranges in real supermarkets reading “How many calories in a banana?” alongside the URL.

Yes. Real bananas. Yes, in real supermarkets. The marketing strategy was to make Jeeves so culturally inescapable that grandmothers asked produce stickers for help.

Then the bubble burst. The stock cratered. The company spent the next several years trying to figure out what it actually was.

The dirty secret under the hood

Here’s the part that doesn’t make it into the nostalgia pieces: the early Ask Jeeves wasn’t really doing the magic. The natural-language interface was a slick front-end. The backend was a hand-edited list of canned answers stitched together with a deeply mediocre crawler. When you asked Jeeves a question and Jeeves came back with something useful, it was usually because some human in Emeryville had personally written that answer. When you asked Jeeves something nobody on staff had thought of, Jeeves did what every search engine of the era did. Badly.

Google ate them because Google had an algorithm. A real one. Ask Jeeves had a butler.

In September 2001, Ask Jeeves bought a small Rutgers spinoff called Teoma for around $1.5 million. Teoma had legitimate ranking technology—a different approach than PageRank, focused on subject-specific authority within communities of related pages. For a brief, genuinely competitive moment in the early 2000s, Ask was actually credible on raw search quality. But “we have a slightly different algorithm than Google” is not a market position. It’s a footnote.

In July 2005, Barry Diller’s IAC bought the entire company for $1.85 billion. By February 2006, the Jeeves character was retired, the site was rebranded simply as Ask.com, and the butler was banished—except for a small UK redoubt where, for some reason, a CGI Jeeves hung on as the local mascot until 2016. Like a kindly grandfather you keep on life support out of guilt.

“The algorithm killed Jeeves”

This is my favorite story in the whole catalog.

In 2007, IAC ran one of the strangest billboard campaigns in tech advertising history. The agency was Crispin Porter + Bogusky—the same shop that gave the world Burger King’s “Subservient Chicken” weirdness. They papered New York, San Francisco, and Los Angeles with cryptic black-and-blue posters. The first wave said only this: “The algorithm killed Jeeves.”

No URL. No logo. No context. Just a vaguely menacing announcement that someone had murdered a beloved cartoon butler.

The follow-ups got weirder. “The algorithm is from Jersey.” “The algorithm is banned in China.” Industry blogs spent weeks trying to decode them. Was this Google making a flex? Was this a stealth startup? Was a deranged math professor on the loose somewhere on the I-95 corridor?

It was Ask.com. The company was bragging—apparently—that it had killed its own mascot in order to take search seriously. The campaign is a perfect monument to the era’s confusion about what Ask actually was. It had a beloved brand it had retired, an algorithm nobody trusted, and a market position nobody could articulate. So it ran billboards confessing the murder of a fictional butler and waited for people to be intrigued.

People were not intrigued. They were just confused.

The toolbar era, or: how Jeeves became malware

This is the part of Jeeves’s life that nobody likes to talk about. After 2010, Ask was no longer a competitive search engine—Diller publicly admitted at TechCrunch Disrupt that year that Ask wasn’t valued in IAC’s stock—and the company quietly outsourced its search backend to third parties and laid off most of its engineers. So what was Ask, exactly?

Ask was a toolbar.

For years—I mean literal years—every Java update on Windows would slip in a checkbox, pre-checked, asking if you’d like to install the Ask Toolbar. Millions of people clicked through Java security updates without reading the fine print and ended up with Ask hijacking their default search engine, their browser homepage, and a small but real chunk of their CPU. The OS X version started shipping the same way in 2015. In June of that year, Microsoft formally classified older versions of the Ask Toolbar as malware and instructed Windows Defender to remove them on sight.

A search engine that had once been the friendly face of the internet was now on Defender’s hit list. And it kept making money this way. That’s the brutal economic reality of late-stage Jeeves: a once-iconic brand had become, by the mid-2010s, mostly a parasitic install bundle.

The butler was selling timeshares.

The cosmic irony of Jeeves’s timing

Ask.com shut down on May 1, 2026 at almost exactly the moment when “ask a computer a question in plain English and get a useful answer” became the single most lucrative product category in the technology industry.

ChatGPT. Claude. Gemini. Perplexity. The AI search wave is doing what Jeeves promised in 1997, except now it actually works, because we finally have transformer architectures and a few trillion tokens of training data and the GPUs to make the math run.

Jeeves had the right product vision in the wrong decade with the wrong technology, the wrong unit economics, the wrong distribution, and the wrong cofounders for the moment. He was right about the user. He was wrong about everything else.

Markets do not reward you for spotting the future. Markets reward you for shipping the future—at the right cost, on the right hardware, in the right decade, through the right distribution channel.

This is the lesson, and it is a brutal one for founders. Friendster spotted social. MySpace spotted personalization. Sixdegrees.com spotted the network. Jeeves spotted conversation. None of them got paid. The companies that got paid were the ones that showed up later with better engineering and better economics and a more disciplined sense of when to scale.

Being correct about user intent is necessary. It is not even close to sufficient.

So long, you ridiculous, prescient butler

I am not going to pretend I used Ask.com much in the last decade. I am not going to pretend I used it much in the decade before that, either. By the time I was building anything serious on the web, Google was the default and Jeeves was a punchline.

But I will admit this: the first time I ever typed a full English sentence into a computer and got back something resembling a real answer, it was 1998, I was sitting in front of a CRT monitor that weighed more than my motorcycle does now, and the answer came back addressed to me by name from a cartoon valet in a green waistcoat.

That is a memory worth marking, if only because the entire trillion-dollar industry currently reshaping every white-collar job in the world is, deep down, just Jeeves at scale. The dream was always conversation. The dream was always that the machine would meet you where you spoke. We just needed thirty years of compute and a few hundred billion in capex to get there.

IAC says every great search must come to an end. I’d argue the search Jeeves started never actually ended. It just changed butlers.

Gal Ratner is the CTO of WhiteStar Labs and writes regularly on AI, software architecture, and the messy economics of the technology industry. Subscribe for more.

Then it asked me, helpfully, whether I was “a student or looking to break into the field.”

If you are, this article is for you. And it is going to ruin your week.

The headlines are right. The pull quote is wrong. Software engineers are absolutely still in demand in 2026 — but not the way the Google AI Overview implies, and not for the people who think the summary is good news. The market is two-tiered, the top tier is bolted shut behind ten years of production scar tissue, and the bottom tier no longer exists. That is the entire story. Everything else is decoration.

The Numbers Everyone Reads

Job listings for software engineers are up roughly 30 percent year over year, with more than 67,000 openings tracked across major employers — the highest in three years. This is happening at the same time as 52,050 tech workers got cut in Q1 2026 alone, the worst quarter since 2023, and roughly half of those cuts were attributed directly to AI. Both things are true.

Senior engineers with current cloud or security experience are closing offers in two to four weeks once they decide to look. KORE1 and Indeed’s Hiring Lab confirm what every working architect already knows: the people who can be trusted with production are getting their pick. Senior AI engineers are clearing $180k–$220k base, with total comp over $350k at the top of the market. The Bureau of Labor Statistics still projects the broader software developer category to grow 15–17 percent through the early 2030s. The sky has not fallen for senior people. The sky is, in fact, doing pretty well for senior people.

This is the part Google quoted. It is not wrong. It is just incomplete in a way that costs careers.

The Numbers Nobody Quotes

New computer-science graduates have a 5.8 percent unemployment rate. That is higher than the general U.S. unemployment rate. Read that sentence twice.

Workers aged 22 to 25 in occupations the research community classifies as “AI-exposed” saw a 13 percent decline in employment in 2025. Big Tech junior hiring is down 25 percent. New graduates make up roughly 7 percent of Big Tech hires — down from 32 percent in 2019. Entry-level software postings have been falling consistently since 2024 and have not recovered. The fresher pipeline at the largest employers in our industry has, functionally, been turned off.

The single most absurd data point I’ll throw at you in this article: per Indeed, hiring an electrician now takes 56 days. Hiring a programmer takes 54. Trades are statistically harder to fill than software jobs. For the first time in fifty years, trade unemployment dropped below college-graduate unemployment. The U.S. Department of Labor responded to that flip by making AI training mandatory in every Registered Apprenticeship program in the country. The federal government has officially conceded which side the wind is blowing on.

If you’re a junior, the demand the Overview is talking about isn’t yours. The demand is for the people who used to be juniors twenty years ago.

Why the Junior On-Ramp Is Gone, and Why It Isn’t Coming Back

Companies hired junior developers for a reason. Somebody had to write the CRUD endpoints, the unit tests, the React components, the README, the CI scripts, the boilerplate Lambda handlers, the validation logic, the migration scripts. That work was the price of training the next senior engineer. It was a paid apprenticeship the industry pretended was a career.

That work didn’t disappear. It just stopped paying $90,000 a year. Cursor does it. Copilot does it. Claude Code does it. SWE-Bench Verified — the benchmark that runs models against real GitHub issues from real open-source projects — went from roughly 30 percent in 2024 to nearly 77 percent in early 2026. For the typical small-bug-or-feature ticket against a clean codebase, the model now wins on time-to-PR and frequently on quality. That is the junior developer’s job description. Convert Jira ticket into syntax. The entire entry-level value proposition.

And it gets worse, because there’s a second front nobody in our industry wants to look at squarely. The small businesses, agencies, and SaaS startups that used to hire a junior dev for $60k–$80k to build a customer portal, an internal tool, or a quick CRUD app? They are not hiring that person anymore. They are using Bubble, Replit Agent, Lovable, Wegic, Make.com, n8n, Airtable, or one of the twenty other no-code AI builders that crossed the “good enough” line in the last eighteen months. A non-technical ops manager can describe an app in plain English in the morning and have a working prototype with a database, auth, and a deployable URL by lunch. The agency model that used to keep a thousand junior devs in WordPress and Wix work is being eaten alive.

Both ends of the junior market — enterprise apprenticeship and SMB freelance — are collapsing simultaneously. Not slowing. Collapsing. The exit ramp is closed at both interchanges.

Why Senior Engineers Have a Ten-Year Runway

The reason a senior engineer with cloud or security chops closes an offer in two weeks is not that they are great at coding. AI is great at coding. It is that senior engineers can do the things AI cannot:

Read 220 billion lines of legacy COBOL that processes 95 percent of U.S. ATM transactions and not break it. Carry the on-call pager when a $40M trading platform goes sideways at 3 a.m. Refactor a monolith without violating the SLA the company has bet quarterly earnings on. Negotiate with a product manager who doesn’t know what they want until you build the wrong thing twice. Sign their name on a HIPAA migration plan and accept legal liability if it goes wrong. Tell the AI no.

The legacy code economy is the part of this conversation almost everyone misses, and it is the entire ballgame for the next decade. Per the GAO and IT-CISQ, 70 percent of banks globally still rely on legacy systems. Forty-three percent of global banking systems run on COBOL. The ten federal legacy systems most in need of modernization cost $337 million annually to operate and consume roughly 80 percent of those agencies’ IT budgets. The average COBOL developer is 55 years old. Ten percent of them retire every year. Sixty percent of organizations using COBOL say their single biggest operational challenge is finding people who can read it.

This is the sleeper market for the rest of the 2020s. Not the AI startup. Not the YC-funded vibe-coded SaaS. The bank you wrote a check from this morning. The insurance company that pays for your kid’s appendix. The Social Security Administration. The state unemployment office that fell over in 2020 because nobody knew how to read its forty-year-old code.

Anthropic recently demonstrated that Claude Code can analyze a substantial COBOL codebase well enough to spook IBM’s stock by 11 percent in a single trading session. Note the verb. Analyze. Not migrate. Not own. Not sign the change order. A senior human engineer still has to drive the modernization, keep the AI honest, and put their name on the document when the trustees of a $40 billion pension fund ask who is responsible. That is your job for the next ten years if you are good. Maybe fifteen if the field of autonomous program reasoning runs into the wall I think it’s about to run into.

AI is a 2x–3x productivity multiplier for a strong engineer. For a weak one, it accelerates the rate at which mistakes reach production. Both halves of that sentence pay senior salaries: the first because the company gets a 3x engineer, the second because the company eventually realizes it needs the senior to clean up the mess the AI-augmented mid-level made. Either way, the line item on the budget is named the same thing. Senior.

The Snake Eats Its Tail

Here is the part that should keep every working senior up at night, and it is the reason I am writing this article.

If juniors do not get hired, juniors do not become mids. If mids do not accumulate scar tissue — the kind that comes from carrying the pager, owning the migration, and shipping the bug that costs the company a customer — they do not become seniors. In ten years, the seniors who currently close offers in two weeks retire. There is nobody behind them. The pipeline has been broken since 2024 and nobody is fixing it because the economics of fixing it do not work for any individual company.

There are two ways this resolves, and I do not love either:

One: the remaining seniors become absurdly valuable for one more cycle, the industry consolidates around an autonomous AI engineering loop in which “software engineer” stops meaning “writes software” and starts meaning “approves software,” and the headcount required to run a Fortune 500 engineering org drops by a factor of ten. Two: the work itself gets eaten by AI faster than the seniors retire, and the transition happens before anyone has time to figure out what to do about it.

In either resolution, the people reading this article — the 30-and-40-something engineers who have already built careers, the senior architects, the staff-plus crowd — are the last generation that will do this job for a living the way we currently understand it. The 22-year-old in their second year of a CS degree right now is training for a profession that will not exist in its current shape when they graduate, and almost certainly will not exist at all by the time they are 35.

This is not a hot take. It is the math.

So What Do You Do

If you are a senior with ten or more years of production experience and a working understanding of distributed systems, security, regulated industries, or legacy modernization: ride this wave. Specialize hard. Learn to direct AI tools well enough that you become the 5x person on a team of three rather than the 1x on a team of fifteen. Take the consulting roles. Take the contract roles. Charge what the market will bear, because the market is now bearing a lot, and bill for the next decade like the runway has a hard end — because it does.

If you are a CS graduate, a junior dev, a bootcamp finisher, an undergrad declaring a major, or a parent paying for any of the above:

Stop. Get out. Now.

I do not say this lightly. I have been on the Microsoft stack for thirty years. I run a software company. I write code every day. My career is software. I love this work. And I am telling you with the authority of someone whose paycheck depends on the opposite advice being correct: the door has closed behind the people already inside, and you should be looking at a different door.

The Door That Is Wide Open

Here is the part of the labor market that the Google AI Overview did not summarize for me, because nobody types “are plumbers in demand in 2026” into Google.

Demand for skilled trades is growing roughly three times faster than demand for white-collar professional roles. That is from Randstad’s analysis of 150 million job postings between 2022 and 2026. Robotics technician postings are up 113 percent. HVAC engineer postings are up 78 percent. Construction roles are up 30 percent. Welders, up 25. Electricians, up 18.

Roughly 530,000 skilled-trade jobs sit unfilled in the United States right now. Ninety-two percent of construction firms cannot find workers. The U.S. needs about 300,000 new electricians over the next decade plus replacements for the 200,000 expected to retire. Forty-one percent of the current construction workforce will retire by 2031. For every 100 young people entering manufacturing, 102 leave. The pipeline is not just broken — it is leaking.

The pay caught up while everybody was looking the other way. Median electrician earnings sit around $62k, with top of the trade clearing $106k. Specialized electricians on AI data-center builds are pulling six figures as a starting position; senior data-center electricians with liquid cooling and fiber-cabling specialties are reportedly clearing $250k–$280k. Mike Rowe says he met three electricians under thirty making $240k–$280k. Plumbers, HVAC techs, and welders all sit in the $51k–$120k band depending on specialty and metro. Unionized journeymen in major cities clear $150k in good years. Industrial maintenance, elevator install, and high-voltage are higher still.

Microsoft’s president called the electrician shortage “the number one problem” slowing data-center expansion. Google committed $15 million to the Electrical Training Alliance. Oracle pushed project timelines back a full year because they couldn’t find enough tradespeople. BlackRock launched a $100 million initiative to train plumbers, electricians, and HVAC technicians. Lowe’s put up $250 million for the same. Meta and CBRE launched a recruiting program for technicians to build out Meta’s data centers. The smart money in the AI buildout is paying for plumbers.

“But What About the Robots”

You will hear the counterargument. Optimus. Figure. Boston Dynamics. The humanoids will replace plumbers too. The trades are no different than software, the argument goes — just on a longer fuse.

They will not. Not for a long time. Not on a fuse short enough to matter to a 22-year-old’s career arc. Here is the actual research.

McKinsey’s 2025 outlook on humanoids in the construction industry estimates large-scale humanoid deployment is “at least a decade away,” and that estimate is for the structured tasks — moving blocks, basic material handling. Nature published a 2025 paper on construction humanoids whose conclusion, paraphrased, is that dexterous foolproof manipulation for intricate tasks like wiring and plumbing remains unsolved at the research level, not just the engineering level. The International Federation of Robotics tracks unstructured commercial humanoid deployments at, effectively, zero. Tesla Optimus has missed every public timeline since 2022. Industry analysts — even bullish ones — put realistic real-world humanoid deployment at five-to-ten years for warehouse-class structured environments, with the giant disclaimer that humanoid timelines have been “five years away” for a literal decade.

A modern LLM can write you a production React component in twelve seconds. A humanoid robot cannot reliably crawl under a sink, identify a corroded P-trap by feel, decide which fitting to swap, discover that the previous owner used the wrong solder, improvise around a pipe that’s been buried in a slab, work around a shutoff valve that won’t close, and not flood your house. That is twenty years of robotics work, minimum, and that is the optimistic version. Real plumbing automation requires general-purpose physical intelligence operating in unstructured environments. We do not have it. We are not close. By the time we do, you will have had a twenty-year career.

A roofer in Vegas in July, on a 12/12 pitch, working around a chimney flashing the previous guy installed wrong? That is unsolvable for current robotics. Not “expensive to solve.” Unsolvable with the present approaches. A mechanic diagnosing a P0420 code on a 2014 Tundra that isn’t actually a catalytic converter problem, because the previous owner installed an aftermarket exhaust that’s reading wrong on the downstream O2 sensor? That isn’t a robotics problem. That is an experience-and-vibes problem. AI doesn’t have vibes. Mechanics do. Roofers do. Plumbers do.

And critically: the trade work that can be automated — prefab framing, factory robotic welding, repetitive pick-and-place — has already been getting automated for thirty years and the industry is still short half a million people. That tells you the elastic demand for skilled, judgment-driven trade work is essentially uncapped. AI eats the structured middle of every industry it touches. The trades’ middle was already eaten in the 1990s. What is left is the part AI is worst at.

Close

The Google AI Overview said yes, software engineers are still in demand. It told the truth and it lied at the same time.

If you are senior, you are in demand for ten more years. Use them. Specialize, charge real money, learn to direct the tools, and assume the runway is finite.

If you are anyone else, the demand the Overview is selling isn’t yours. The demand is for AI-fluent senior architects, regulated-industry specialists, legacy modernization leads, and a small ecosystem of researchers building the AI tooling that will, within your professional lifetime, automate this entire profession down to a single approval button.

Meanwhile the trades that built this country are short half a million workers, the pay is real, the federal government just bet the apprenticeship system on it, and the robots are not coming for them on any timeline that matters to a career.

Software is the last place a 22-year-old should be betting a career in 2026. The trade you cannot outsource to an LLM is the trade you should be learning.

I would tell my own kid the same thing. I am telling you now.

— Gal Ratner

CTO, WhiteStar Labs

Tobi Lütke at Shopify said it out loud first. In a memo he later posted on X, he told his employees that “before asking for more headcount and resources, teams must demonstrate why they cannot get what they want done using AI.” Within weeks, Micha Kaufman at Fiverr told his staff that “AI is coming for your jobs” — his words, not mine — and that “it does not make sense to hire more people before we learn how to do more with what we have.” Luis von Ahn at Duolingo announced his company would be “AI-first,” would “gradually stop using contractors to do work that AI can handle,” and would only increase headcount “if a team cannot automate more of their work.” Andy Jassy at Amazon followed in June, telling Amazon’s 1.5 million employees that AI “will reduce our total corporate workforce.” Marc Benioff at Salesforce skipped the philosophy and said the quiet part out loud on a podcast: “I’ve reduced it from 9,000 heads to about 5,000, because I need less heads.”

Less. Heads.

That is how he refers to the people who built his company.

If you work at one of these companies — or one of the dozens of imitators who will copy this memo verbatim over the next year — your CEO has just handed you a clarifying gift. They have told you, in writing, exactly what they think of you. They think of you as a cost. They think of you as a head. They think of you as a line item that should justify its own existence on a quarterly basis against an autocomplete tool. The polite, professional response is to act surprised. The correct response is to take the memo seriously and let it inform your behavior from this day forward.

Here is what the rest of this piece is about: how to read that memo for what it really says, why the burden of proof is not on you, and what you should actually do now that your employer has told you who they are.

First, A Reminder: You Don’t Own the Company

Before we get into strategy, we have to clear out the sentimental fog that companies spend a lot of money pumping into your office. Mission statements. Town halls. “We’re a family.” Stock vesting schedules carefully designed to look like ownership while granting none of the rights of ownership. Slack channels named after the company’s values. The whole apparatus is engineered to make you feel like a stakeholder in something.

You are not.

You are an at-will employee in an at-will jurisdiction. That phrase has a specific legal meaning, and you should commit it to memory: your employer can terminate you, at any time, for any reason that is not explicitly illegal, with no notice and no severance unless they choose to offer it. “The CEO thinks AI can do your job” is a perfectly legal reason. So is “we restructured.” So is “Tuesday.”

You do not own the company. You hold no equity that materially affects your wealth unless you are an executive. You have no vote on strategy. You have no claim on the goodwill you have spent years building. Your name is not on the building. When the company is sold, you do not get a check. When the company succeeds, you get to keep your job, which is the same outcome as if the company merely survived. When the company fails, you get fired. The asymmetry is total.

So when your CEO writes you a memo demanding that you justify the existence of human beings on your team — prove a negative, in front of a panel, against a fast-moving target — take a moment to ask yourself: on whose behalf am I being asked to do this work? It is not on yours. The output of that exercise is not “more people on my team so I can do my job better.” The output is “fewer people on my team, so the company’s margins improve, so the executives hit their bonus targets.” The exercise is structured so that you cannot win and the company cannot lose.

This means the burden of proof is not on you. It never was. The CEO has already decided what they want. They are merely outsourcing the paperwork to you to give the decision a thin coat of process. Refuse the paperwork.

The Correct Move: Do Nothing

I want to be very precise here, because this is where most career advice goes wrong. The advice is not “sabotage your team.” The advice is not “quiet quit.” The advice is not “scream into the void on LinkedIn.” The advice is much simpler and much more devastating.

The advice is: do not ask for any new hires. Ever. About anything.

Your CEO has told you that hiring is now an adversarial process where you must prove a case against an opposing counsel that will move the goalposts every time you produce evidence. Fine. The optimal response to a rigged proceeding is to refuse to participate. You do not file the paperwork. You do not write the justification deck. You do not put together the headcount business case. You do not build the comparison spreadsheet showing what an AI agent could and could not do compared to a human candidate.

Instead, you do the work that is in front of you. You do it well. You do it with the people you already have. When something breaks, you fix it. When the team is stretched, the team gets stretched. You absorb the load. You quietly let attrition do its work — because it will, because people leave, because that is what people do.

And as the team shrinks, you do not replace anyone.

This sounds passive. It is the most aggressive thing you can do.

Here is why. Your value as an employee is determined by exactly one variable: the cost of replacing you. Every other consideration — your performance review, your title, your relationship with your manager — is downstream of that single number. When you have a 12-person team, the cost of replacing you is one twelfth of the team’s knowledge plus a recruiter fee. When you have a 5-person team that is doing the work of 12, and you are the person who knows where the bodies are buried, the cost of replacing you is the company itself. That is not a metaphor. The cost is the company. They cannot fire you because they can no longer function without you.

Klarna ran this experiment for you, by accident, and the results are public. Sebastian Siemiatkowski announced a hiring freeze. The headcount fell from roughly 5,000 to 3,000. He bragged on every podcast that would have him about how AI was doing the work of 700 customer service agents. Then quality collapsed, customer satisfaction cratered, and he had to admit on Bloomberg that “we went too far” and start hiring humans back — specifically the humans who could do what the AI couldn’t. The remaining humans, the ones who had absorbed the load through the freeze, got something interesting: salary increases. Siemiatkowski explicitly told them that as costs fell and productivity rose, they would see more frequent raises. The people who stayed through the squeeze became more expensive, not less, because there was no one else who knew how the place worked.

That is the playbook. Not because Klarna is run by geniuses — they are not, they had to publicly walk back their AI strategy after firing 800 customer service agents — but because the math is unforgiving. When you starve a company of new talent, the existing talent appreciates in value, the way an oil painting appreciates when no more are being made.

What Actually Happens to the Company

Let’s walk forward in time. Your CEO has issued the memo. Headcount requests are being denied across the board, because of course they are; that is the whole point. Hiring slows to a trickle. AI tools get rolled out with great fanfare. The investor deck gets a new slide called “AI productivity initiative.” Costs drop in the short term. The stock briefly rallies.

Then the cracks start.

They show up first in the things that depend on tacit knowledge — the institutional memory that lives only in people’s heads. The customer who has a weird billing arrangement that no one documented. The integration with the partner that breaks every quarter and only one person knows how to fix. The release process that has eight undocumented steps. The clients who only stay because of a relationship with a specific account manager who is now doing the work of three people and has stopped returning calls.

AI does not handle these. It cannot. It can summarize a runbook, but it cannot be the runbook. It can draft an apology email, but it cannot have the eight-year relationship with the customer that makes the apology land. It can write code, but it cannot remember why you wrote that weird if-statement in 2019, because the why is not in the codebase. The Forrester report from late 2025 found that 55% of employers regret laying off workers because of AI. A separate Careerminds survey found that more than a third of companies that conducted AI layoffs had to rehire 25% to 50% of the roles within six months. One in three of those employers spent more on rehiring than they saved on the layoffs in the first place. This is not a prediction. This is what already happened.

Then customer satisfaction starts to slip. Then NPS scores. Then revenue per customer. Then — and this is the inflection point — the analysts start asking pointed questions on the earnings call. “Can you walk us through the drivers of the customer churn we’re seeing in the SMB segment?” “How do you see the long-term ARR impact of the changes you’ve made to your support model?” The CFO produces a slide. The slide is unconvincing.

And now we get to the part of the movie that the CEO did not write into the script.

Who Gets Cut First When the Numbers Wobble

Here is the part of the corporate machine that the AI memo conveniently does not discuss: when a company’s performance starts to degrade, the first cuts are not on the front lines. They are in management.

The data on this is not subtle. Gartner projects that 20% of organizations will use AI to flatten their structure by the end of 2026, eliminating more than half of their middle-management positions. Bloomberg and Live Data Technologies found that middle managers made up roughly one-third of all layoffs in 2023. Korn Ferry’s 2025 Workforce report says 41% of employees report their company has trimmed management layers. Amazon, in October 2025, eliminated 14,000 corporate jobs in a single round — explicitly to “reduce bureaucracy” and “remove organizational layers.” Andy Jassy spent the prior year publicly demanding that the worker-to-manager ratio go up by at least 15%. Google cut 35% of its small-team managers. Intel cut 50% of its management positions. Microsoft’s 9,000 layoffs in 2025 came with a stated goal of fewer middle managers and more individual contributors.

Read that paragraph again, slowly. The same companies that told employees to justify hiring against AI are now eliminating the people who would have evaluated those justifications.

This is structurally inevitable. When a company decides its growth model is AI-driven productivity, the first thing it has to prove to investors is that it is structurally cheaper to run. The cheapest cuts — from a per-dollar-saved perspective — are the highly compensated layers between the executives and the work. Senior managers make two to three times what individual contributors make. Cutting one of them is equivalent, on the spreadsheet, to cutting two or three engineers. The math is the math, and the math points up the org chart, not down it.

So while the manager who pestered you for an AI justification deck is busy preparing his next round of justification decks, the cuts are already being modeled in a finance team’s spreadsheet a floor above him. The individual contributor who quietly absorbed the workload, learned the systems, and made themselves the only person who knows how the production deployment actually works — that person is harder to cut, because cutting them produces an immediate, visible failure. The director who reviewed their last performance review and asked them to “align on AI initiatives” — that person is much easier to cut, because cutting them produces nothing. Or, if it produces anything, it produces relief.

The order of operations in a company under stress is: contractors first, then middle managers, then directors, then individual contributors who can be replaced, then individual contributors who cannot. The CEO is on a different list, the one labeled “golden parachute,” and we will return to him in a moment.

The Irony Is Almost Too Pure

Stand back from this and look at what just happened, because the irony is almost too pure to be believed.

A class of executives who are paid eight to nine figures a year to allocate capital have just told their workforces that the way to allocate human capital is to assume by default that no human is needed. They have framed this as efficiency. They have framed it as innovation. They have framed it, in some cases, as caring — the Fiverr CEO actually invoked “radical candor” before telling his people they were about to be unemployed.

In doing so, they have placed a chokehold on the only durable competitive advantage their companies have ever had: the ability to attract and retain talented people who can do things competitors cannot. They have told the most ambitious people on their teams — the ones with options, the ones who could leave tomorrow — that the company’s default assumption is that they are replaceable by software. They have told the rising managers that there will be no team to manage. They have told the senior individual contributors that the question their leadership wakes up thinking about is not “how do we make our best people more productive” but “how do we make sure we don’t need them.”

And then — this is the truly remarkable part — they posted the memo publicly on social media. Lütke posted it on X. Von Ahn put it on LinkedIn. Kaufman tweeted his out. They wanted credit for the toughness. They wanted to be seen as the kind of leader who tells the hard truths. They did not seem to consider that the people reading the memo were not just employees but the entire labor market, and that posting “I need less heads” in public is the kind of thing that gets read in interviews for years.

The result is that they have produced, at no cost, a perfectly clear signal to every employee at every one of these companies. The signal is: the relationship is now transactional. The signal is: there is no long arc of mutual investment here. The signal is: we have done the math on you, and the math says you are a temporary line item until the autocomplete catches up.

Once an employee receives that signal, the entire concept of “going the extra mile” becomes incoherent. Why would you? You are a line item. Line items don’t do unpaid weekends. The discretionary effort — the unpaid extra ten percent that actually makes companies work — evaporates. Forrester is already tracking this. They have a name for it. They call the workforce that has stopped giving discretionary effort “coasters,” and they put the share of coasters at 25 to 28 percent of all employees, climbing. That is roughly one in four people on the company’s payroll who has decided, internally and quietly, that the company does not deserve their best work. You will not see them. You will see the work product. You will see it in the bug count, in the missed deadlines, in the deals that don’t close, in the launches that slip.

The CEO did this. The CEO chose this. They wrote it on X.

What This Means For You, Specifically

So here is the practical advice. Print it out if you like.

One. Do not ask for new headcount. Not now, not next quarter, not for the foreseeable future. If your team is stretched, let it be stretched. If you are short an engineer, ship slower. If a peer leaves, do not backfill. The CEO has told you what they want; give it to them, exactly. Let them experience the consequences of the policy they wrote.

Two. Do not write the AI justification deck. If forced into a conversation about whether AI can do a particular job, the correct answer is some variant of “we’ve evaluated the available tools and they do not currently meet the quality bar for this work.” Period. Do not enumerate. Do not provide examples. Do not give your management a roadmap they can hand to a vendor. The information asymmetry is already in your favor; do not surrender it.

Three. Document nothing in shared spaces that does not have to be documented. The corporate-knowledge land grab — the push to get every workflow into Confluence so an LLM can index it — is not a productivity initiative. It is a knowledge extraction initiative. The point is to make you replaceable. Resist it the way you would resist any other unpaid transfer of your assets to your employer. Tribal knowledge is an asset class. Treat it like one.

Four. Become the only person who knows how the most important system works. Pick the thing your company cannot survive without and make yourself the human single point of failure for it. This is sometimes called “bus factor of one,” and management consultants have written entire books explaining why it is bad for the company. They are correct. It is bad for the company and excellent for you. The whole point of the AI memo was to break employee leverage. Bus factor of one is leverage.

Five. Quietly start your search. Not because you intend to leave today, but because the optionality itself changes how you behave in negotiations. The employee who has three offers in their inbox is a different employee than the one who does not. The CEO already changed the deal; you are merely renegotiating from your side.

Six. Stop volunteering. Stop the unpaid mentoring of the new hire who will replace you. Stop the Friday demo. Stop the off-hours Slack response. Stop the writeup of the postmortem that will be fed into an internal LLM to train your replacement. None of these activities are in your job description. The CEO has clarified the relationship; you should clarify it back.

Seven. When the layoffs come for management — and they will — do not cheer publicly. Do not gloat. Do not post anything. Just keep doing the work. The same memos that fired them will be used to fire you if you give the company an excuse. Be quiet. Be useful. Be expensive to lose. That is the entire game.

A Closing Note On Loyalty

I have spent thirty years in technology. I have worked at small companies and gigantic ones, and I have written the kinds of memos that the people on my teams had to read. I am not anti-CEO. I am not anti-AI. I run an AI consulting firm. I write code with AI agents every day. I have built tools that genuinely make small teams more productive than large ones used to be. I am, in the language of the moment, an AI-first CTO.

And I am telling you, as someone who has been on the other side of the table: the memos coming out of Shopify and Fiverr and Duolingo and Amazon and Salesforce are not technology strategy. They are labor strategy dressed up as technology strategy. They are an attempt to use the rhetorical force of “AI is changing everything” to extract concessions from workforces that would have been laughed out of the room two years ago.

They are also, importantly, the most clarifying communications these companies have produced in a decade. For a long time the relationship between employer and employee in white-collar America was fuzzed up by mission statements, perks, free lunches, all-hands speeches, and the slowly compounding fiction that we were all in this together. The AI memos cleared all of that away. They told you, in writing, what the relationship really is. The relationship is that they will keep you as long as they cannot replace you for cheaper, and the moment they can, they will, and they will post about it on X.

Once you understand that, your decisions get a lot easier.

You owe a company exactly what your contract says you owe it. Eight hours a day, the work product they pay you to produce, the legal duty of loyalty while you are employed by them. You owe them no more than that. You do not owe them justification decks for why human beings should exist on payroll. You do not owe them strategic insight into where AI can and cannot do your job. You do not owe them your best ideas, your unpaid weekends, your tribal knowledge, your enthusiasm, your personality, or your relationship with their customers. None of that was in the contract.

A company that puts AI ahead of its workforce — publicly, on social media, with a smile — is a company that has told you, to your face, what it values. Believe them. They are not bluffing. The Salesforce CEO said, on a podcast, “I need less heads.” That sentence is worth more than a thousand HR memos. He told you what you are.

Act accordingly. Do not write his deck for him.

Gal Ratner is the CTO of WhiteStar Labs, an enterprise software and AI consulting firm. He writes about technology, leadership, and the parts of the corporate world that nobody else is going to write about, at galratner.com.

Let me walk you through the math. Then I’ll walk you through the Robert Downey Jr. line from Tropic Thunder that, I think, explains this entire industry better than any analyst report I’ve read in the last year.

The trillion-dollar problem

Here is the spending side of the AI economy as of mid-2026:

• Gartner: $1.5 trillion in global AI spending in 2025, projected $2 trillion+ in 2026.

• Big Tech capex (the eight largest hyperscalers): $427B in 2025, projected $562B in 2026, $637B in 2027.

• McKinsey: the world needs to spend $5.2 trillion on data centers by 2030 to meet AI compute demand.

• Nvidia’s Jensen Huang: $3–4 trillion in AI infrastructure by end of decade.

• OpenAI alone: $852 billion valuation, $1T+ in committed compute partnership deals.

• OpenAI H1 2025: $4.3B in revenue, $13.5B in net losses — roughly $2B burned every month.

• Cumulative OpenAI losses through 2029: projected to exceed $140 billion.

Pause on that last one. A company is on track to lose 140 billion dollars over a five-year window. It is valued at almost a trillion. It is committing to spend another trillion on compute. And the headline product — the one you, the consumer, see — is a $20 chatbot.

Does this look like a business that intends to make its money back from $20 chatbot subscriptions?

The $20 fantasy

Here is what the consumer side of OpenAI actually looks like:

ChatGPT has roughly 800 million weekly active users. Approximately 5% of them pay anything at all. Of those who pay, the bulk are on the $20 ChatGPT Plus tier. OpenAI’s gross margin is 33%. Inference costs alone are projected to hit $14.1B in 2026. This year, OpenAI started rolling out an $8 ChatGPT plan to chase volume in price-sensitive markets like India and Southeast Asia, because at $20 the tier had stopped growing fast enough.

You don’t roll out a cheaper consumer plan if your consumer plan is the path to profitability. You roll it out when you’ve already concluded that consumer subscriptions will never pay back the infrastructure bill, and you’d rather have 122 million people locked into your product at $8 than 47 million paying $20.

Because the $20 subscription was never the business. It was the funnel.

The real customer is your CFO

Where the math actually works for OpenAI, Anthropic, Google, and the rest is not on your laptop. It’s on the ledger of the company that signs your paycheck.

OpenAI’s enterprise pricing is $25–$60 per user per month. A 10,000-employee company is several million dollars in annual revenue. They have over 9 million paying business users as of February 2026, up from 5 million the prior August. That number is growing far faster than the consumer side.

But even enterprise seats, multiplied by the entire knowledge workforce of the planet, don’t quite close a $1 trillion compute commitment.

What does?

The agent that does the job instead of the worker.

OpenAI has reportedly hired 100+ ex-investment bankers at $150/hour to teach its models how to do “junior banker grunt work” — IPO models, M&A decks, restructuring analyses. They are explicitly aiming at consulting, finance, and legal labor budgets, which they openly describe as “billions in current labor spend.” The 90th-percentile revenue scenario in their own forecast — the one that gets them to $70B ARR — explicitly requires “autonomous agents that can replace human workers in customer service, software engineering, and research.”

That is not a euphemism. That is the slide deck. Your job is the line item.

Why your boss suddenly cares whether you use AI

This brings us to the part of the story most people experience personally without seeing its shape.

Microsoft made it official last year. An internal memo from Julia Liuson, President of the Developer Division, told managers to incorporate AI usage into “holistic reflections on an individual’s performance and impact.” In her words: “AI is no longer optional. It’s core to every role and every level.”

Google, Meta, Shopify, Duolingo, KPMG, Thomson Reuters, JPMorgan Chase, Citi, and Amazon have all rolled out variations of the same policy. Amazon’s Forte system now produces an Overall Value (OV) score that explicitly tracks AI usage. Reports from inside Amazon describe a culture where “falling behind on AI means falling behind.”

Why the sudden urgency from leadership? Because the vendors that sold the seats need adoption to justify expansion, and your CIO needs adoption to justify the line item to the CFO. So you get a mandate. Use it. Document it. Justify yourself to it. Demonstrate impact. Or your performance score drops.

Notice what this is not. This is not your employer noticing a productivity gap and helpfully offering you a tool. This is your employer being told by a vendor that billions have been invested in a product, and the only way to recoup it is to enforce its use until it absorbs more and more of your work, until eventually it absorbs all of it.

The layoff receipts

If this sounds paranoid, it isn’t. The CEOs aren’t even pretending anymore.

• Microsoft: 15,000 cut in 2025. Satya Nadella confirmed ~30% of the company’s code is now written by AI.

• Intel: 15,000 in 2025, another 24,000 in early 2026.

• Amazon: 14,000 in late 2025, another ~30,000 in early 2026. Beth Galetti pointed to AI as the reason the company could operate “more efficiently with fewer people.”

• Accenture: 11,000 cuts. CEO Julie Sweet: “those we cannot reskill will be exited.”

• Salesforce: customer support went from 9,000 to 5,000 heads. Marc Benioff says AI does “30 to 50% of the work at Salesforce now.”

• Block: nearly half the company. Jack Dorsey: “the intelligence tools we’re creating and using are enabling a new way of working which fundamentally changes what it means to build and run a company.”

• C.H. Robinson: 1,400 cut after rolling out AI pricing and shipment tracking.

• Chegg: 45% of the workforce gone as students moved to ChatGPT.

• Q1 2026 alone: 78,557 tech layoffs. 47.9% directly attributed to AI by the companies that did them.

Anthropic’s own CEO, Dario Amodei — the guy whose company makes Claude, which I’m partially using to write this — said in May 2025 that AI could eliminate up to 50% of entry-level white-collar jobs and push U.S. unemployment to 10–20% within five years. In January 2026, he doubled down in a 20,000-word essay calling it an “unusually painful” shock and warning that CEOs would “quietly stop hiring, then replace humans with AI the moment it becomes viable to do so.”

The man building the thing is telling you what it does. The Wall Street Journal called the latest wave “the week the dreaded AI jobs wipeout got real.”

The Tropic Thunder rule

Now we get to the cleverest piece of marketing the AI industry ever produced.

You’ve heard the line. Probably from a manager. Probably from a LinkedIn influencer wearing AirPods Max in a coffee shop. Probably last week:

“You’re not going to lose your job to an AI. You’re going to lose your job to someone who knows how to use AI.”

This is supposed to be reassuring. Don’t worry — you can still win, just learn the tool. It’s the soft version of the message. The hopeful version. The one that makes adoption feel like agency.

It is also, to borrow from Robert Downey Jr., complete nonsense the moment you examine it.

The line from Tropic Thunder:

“I’m a dude playing a dude, disguised as another dude.”

It’s funny because it’s a tower of recursive nothing. Each layer claims to be different from the next, but you peel them back and there is nothing inside. The character has no actual identity — he is performance about performance about performance.

The “you’ll be replaced by someone who uses AI” line is the same kind of tower. Walk it down with me:

• You won’t be replaced by an AI. — Layer one. Reassuring.

• You’ll be replaced by a person. — Layer two. Slightly less reassuring, but still human.

• ...who uses AI to do your job. — Layer three. Wait.

Stop there. Why does that person exist? What value are they adding that the AI alone isn’t? They’re prompting it. They’re correcting it. They’re feeding it your domain expertise, your edge cases, your institutional knowledge. They are doing reinforcement learning from human feedback in real time, on a salary, for a company that intends to keep that knowledge after they leave.

They are a dude prompting an AI, disguised as another dude doing your job. Until the AI doesn’t need them either.

The honest version of the sentence is:

You’re not going to be replaced by an AI. You’re going to be replaced by someone who trains the AI. And then they’re going to be replaced by no one.

Both versions of the line are predictions of the same end state. The “softer” version just buys an extra step of plausible deniability for the manager delivering it. It lets him keep a straight face while watching the headcount chart go down.

It’s a dude losing a job, disguised as a dude keeping a job, disguised as a dude using a tool.

The trainers are next, and we already have proof

If you think I’m being dramatic about the AI trainers being phased out — look at what’s already happened to the people whose entire job was training AI.

Scale AI was the company. The whole industry. Founded by Alexandr Wang. Powered self-driving training data. Powered RLHF for OpenAI, Microsoft, Meta, Anthropic. Built a global crowd of 700,000+ contractors through its Outlier subsidiary. At one point, “the human layer of AI” was effectively synonymous with Scale AI.

In June 2025, Meta paid $14.3 billion for 49% of Scale AI and pulled Wang in-house. One month later, Scale AI laid off 14% of its workforce — 200 employees and 500 contractors — and announced a pivot away from its core data-labeling business. Several of its biggest customers, including Google and OpenAI, walked at the same time.

Walked to what? Synthetic data. Teacher-models training student-models. The frontier labs have been steadily replacing the human RLHF layer with AI-generated training data and AI-judging-AI evaluation pipelines. Human trainers still exist for the hardest edge cases, but the trajectory is unambiguous: the labelers who fed the models are being eaten by the models they fed.

If the people whose literal job description was “train the AI” are being replaced by AI, the version of you that “knows how to use AI” should not feel particularly safe.

The play, in four moves

Strip away the marketing language and the entire AI economy is running a play that fits on a napkin:

1. Sell the worker the $20 subscription so they train the AI for free, every day, on their own work product.

2. Sell the boss the enterprise seat so AI does 30% of the work, framed as a productivity tool.

3. Sell the boss the agent so AI does 100% of the work, framed as cost optimization.

4. Quietly retire the AI trainers along the way, because synthetic data and self-play got good enough.

That is the plan. It is not a conspiracy. It is in the IPO filings. It is on the earnings calls. It is in Amodei’s interviews. It is in Benioff’s podcast appearances. It is in OpenAI’s own projected revenue model, where the high-end scenario explicitly depends on autonomous agents replacing human labor.

There is too much money invested in AI to let you keep your job. A trillion-dollar capex stack does not pencil out at $20 a month. It only pencils out at your salary, multiplied by everyone like you.

What this actually means

I am not telling you to refuse to use AI. That ship sailed. The tools are real, they are useful, and refusing to learn them is not a survival strategy — it’s just a slower way to lose.

I am telling you to be honest about the deal you’re being offered.

The deal is: train your replacement, on the clock, for free, while paying $20 a month for the privilege. In exchange, you may get to keep your job a little longer than the colleague who refused.

That’s a fine deal compared to refusing. It’s a terrible deal compared to building leverage that doesn’t live inside your employer’s HR system — your network, your reputation, your own products, your own audience, things you take with you when the AI catches up to whatever you currently do for a paycheck.

Use the tools. Of course use the tools. Just stop pretending the tool was sold to help you. It was sold to replace you. You’re paying the entry fee.

Don’t be a dude using a dude, disguised as another dude.

Be the one with the leverage.

If this resonated, share it with someone whose performance review now includes “AI usage.” They’ll either thank you or argue with you. Both are useful.

Revenue down roughly 80%. Documentation traffic down about 40% from its 2023 peak. Six months of runway if nothing changed. The framework itself? More popular than ever — 75 million downloads a month, 51% market share in the State of CSS survey, and used by Shopify, GitHub, NASA, Vercel, and most of the AI-coding tools you have ever opened a tab for.

Read that paragraph again. Usage at an all-time high. Revenue at an all-time low. That is not a slow decline. That is the business model breaking in real time.

And it raises a much bigger question for the rest of us — one Tailwind’s own predicament makes hard to dodge: do we actually need this framework anymore?

The numbers, before we go any further

Let me lay this out cleanly so nobody has to chase links.

Headcount: Tailwind Labs had four engineers. It now has one — Robin Malfait, who built Headless UI and shipped most of v3 and v4. The remaining team is Wathan, co-founder Steve Schoger, co-founder Jonathan Reinink, Robin, and Peter Suhm part-time on partnerships. Five people, total, behind the most-installed CSS framework on Earth.

Revenue: Down about 80% from peak, per Wathan’s own podcast. The funnel that worked for years — developer Googles a Tailwind class, lands on the docs, sees Tailwind Plus for $299, eventually buys it — is gone. Cursor, Copilot, Claude, and ChatGPT now answer the class question directly in the editor. The developer never visits the site. The site never sells anything.

Runway: Six months, if nothing changed. Wathan said as much explicitly in the podcast and in his GitHub reply when he closed the LLM-friendly docs PR — the one a contributor opened in November and Wathan refused on the grounds that making the docs easier for AI to scrape would accelerate exactly the problem killing the company.

GitHub activity: Pull up the tailwindcss repo and look at the last six months of commits. The recent v4.2.x releases are bug fixes — Vite alias resolution, Astro v5 import handling, canonicalization tweaks for tracking utilities, character escapes for Biome compatibility. This is the commit log of a project being kept alive, not a project being pushed forward. With one engineer it cannot really be anything else.

Vercel and Google AI Studio have stepped in as sponsors. That is generous, and it might extend the runway. It does not change the underlying problem.

Do we even need a CSS framework anymore?

Here is the part nobody at Tailwind Labs wants to say out loud, but every CTO I talk to has been quietly thinking it for a year: the original problem Tailwind solved is being solved upstream of Tailwind.

Tailwind’s pitch was elegant. CSS is hard. Naming things is hard. BEM is verbose. Switching between HTML and a stylesheet breaks flow. So instead of writing CSS, compose utility classes inline and let a build step purge what you do not use. It was a real productivity win in 2019. It is a less obvious win in 2026.

Why? Because in 2026 I do not write the styles by hand and I do not compose the utility classes by hand. I describe what I want, and the model writes the CSS. Or the Bootstrap markup. Or the Tailwind soup. The friction Tailwind was invented to remove — the friction of typing — is not where my time goes anymore.

And when the model writes raw CSS, it writes pretty good raw CSS. Modern CSS has nesting, custom properties, container queries, layers, scope, color-mix(), light-dark(), and logical properties. A focused stylesheet generated by a competent model is shorter, more readable, and more debuggable than 47 utility classes stacked on a single div like a barcode.

Tailwind solved a problem that was real. AI solved the same problem in a different and arguably better way. That is not a takedown of Tailwind. That is just where we are.

About “HTML6”

Quick honest correction before anyone forwards this to a colleague who actually reads specs: there is no “HTML6” release in the formal sense. WHATWG dropped versioned HTML over a decade ago and has been running the spec as a living standard ever since. There is no version number coming.

But the spirit of the “HTML6” conversation is correct, and it matters here. The platform itself is shipping things that used to be the entire reason you reached for a CSS framework:

Native popover and the popover API. Dropdowns, menus, tooltips. No more wrestling Popper or pulling in three Tailwind plugins.

The dialog element. Modals, accessible by default, with focus trapping built in. Browser-native.

CSS anchor positioning. Pin one element to another without a JavaScript library. This was a Tailwind-plus-Floating-UI problem six months ago.

View transitions. Smooth state changes between routes, again with no framework.

Container queries, @scope, CSS nesting. The component-level reasoning that drove people to utility-class chaos in the first place is now expressible in vanilla CSS.

None of these are Tailwind features. They are browser features. They ship to your users whether you adopt them or not, they have no licensing model that can collapse, and an AI assistant can wire them up for you in seconds. “HTML6,” used loosely, is just shorthand for the platform finally catching up. Which it has.

If you do want a framework, Bootstrap is the better bet

This is the part of the article where I expect the most disagreement, and I am fine with that. If you are standing up a new project in 2026 and you want a CSS framework — meaning a real one, with components, not just a class soup — Bootstrap is a more honest choice than Tailwind right now.

Latest stable is Bootstrap 5.3.8, shipped August 2025. The team behind it is at Twitter’s old scale rather than four-engineers-down-to-one scale. The release cadence is structured around active LTS and maintenance LTS branches, the way grown-up infrastructure tends to be. There is no “docs traffic equals payroll” dependency hanging over the project.

More importantly, Bootstrap actually ships components. A real modal. A real accordion. A real navbar with offcanvas behavior. A real dropdown. Toasts, scrollspy, tooltips, popovers, carousels. Form validation styling. Over 1,800 icons in Bootstrap Icons. Pre-built example layouts — pricing pages, dashboards, sign-ins, magazine layouts — you can drop in and edit. With Tailwind you get atomic utilities and a $299 invitation to buy the components separately. With Bootstrap you get the components.

The other underrated point is teaching surface. If you hand a junior developer a Tailwind codebase, they have to learn the utility vocabulary, the responsive prefixes, the variant prefixes, the JIT mental model, the config file, and somebody’s opinion about what “component classes” should be extracted. If you hand them a Bootstrap codebase, they have to learn class names that read like English: btn-primary, navbar-expand-lg, modal-dialog. The on-ramp is shorter. The Stack Overflow answers go back fifteen years. Every AI assistant on the market has seen orders of magnitude more Bootstrap than any other framework.

And — to the original point of this whole article — Bootstrap’s business model does not collapse if developers stop visiting the documentation. It is sponsored, it is foundation-style, it is just there. Picking infrastructure that depends on docs-page ad-funnel economics in 2026 is picking a project on the wrong side of the AI shift.

What happened to the three engineers

Their names have not all been made public, and out of respect I am not going to guess. What we do know, from Wathan’s podcast directly, is that he spent the closing minutes of the episode personally recommending each of them — describing their specific strengths and offering references to anyone hiring. He called them three of the most talented people he has ever worked with. He said he expected to be jealous of whatever company picked them up.

Given the framework’s reach and the public sympathy that followed the announcement — and the size of the network around Adam, Steve, and Jonathan — it is hard to imagine three engineers with that pedigree sitting on the market for long. “Co-author of Tailwind CSS” is the kind of line on a resume that opens calendars.

I genuinely hope they have landed somewhere they are happier than they were on January 5. Somewhere with a business model that does not depend on developers manually clicking through documentation pages. Somewhere with a longer runway than six months. Somewhere where the work they ship is decoupled from the marketing funnel that pays for it. That is not a knock on Wathan, who handled the disclosure with more honesty than ninety percent of CEOs ever manage. It is just a wish for three people who got caught between the wrong revenue model and the wrong year.

So is this the end?

Probably not the literal end. Tailwind has too many users, too many sponsors lining up, and too much code in production for it to disappear next quarter. v4.2.x will keep shipping. Robin will keep merging fixes. Vercel will keep cutting checks. The framework will be around for a long time the same way jQuery is around for a long time.

But “around” is not the same as “where I would start a new project this year.”

If I were greenfielding a project in May 2026, here is what I would actually do. I would let the AI write the CSS — vanilla, modern, with custom properties for theming. I would lean on the platform for popovers, dialogs, anchor positioning, and view transitions, because those primitives are now native and free. And if I wanted a component library to anchor the design system, I would pull in Bootstrap 5.3, theme it, and move on.

Tailwind solved a real problem in a specific era. That era is over. The framework that fired three quarters of its engineering team because AI ate its funnel is, with respect, not the framework I want sitting underneath my next five years of code.

The question in the title was not rhetorical. My answer is: yes, basically.

— Gal Ratner

A page called BJJ Doc shared a video about Craig Kukuk — the first American to receive a Brazilian jiu-jitsu black belt from the Gracie family, then publicly expelled in the Gracie family newsletter shortly after. Classic engagement architecture: dangle the mystery, drive the click, capture the audience. Tap through and you would land on whatever site was monetizing the story.

Underneath the post, Meta AI had already done the work. A little suggestion chip read: “Why was Craig Kukuk expelled?”

Tap it, and Meta AI dumps a multi-paragraph answer right there inside the Facebook UI — the financial dispute with Rorion Gracie, the Torrance Academy compensation fight, the public disowning in the next Gracie newsletter, even the speculation about Renzo. Every piece of the story BJJ Doc was using as the hook. Delivered inside Facebook. The link to the underlying source dangling somewhere off-screen, irrelevant.

The publisher just got vaporized in real time. By Meta. On Meta’s own platform. While their post was generating the impressions Meta needed to keep the user there.

If this looks familiar, it should. We are watching Google’s AI Overviews — the same system that has gutted publisher traffic across the open web — get ported into the social feed.

The receipts on AI summarization are now embarrassing

This is not a vibes argument. The data on what AI summary boxes do to publisher traffic is now overwhelming.

A randomized field experiment from researchers at the Indian School of Business and Carnegie Mellon, posted to SSRN in April 2026, ran a clean A/B test: one group saw Google AI Overviews, another had them hidden in real time via a browser extension. The result was a 38% drop in outbound clicks on the queries where Overviews triggered, while the share of searches ending without any click jumped from 54% to 72%. User-reported satisfaction did not budge when the Overviews disappeared. Google is taking the click and giving nothing extra in return.

Other measurements pile on. Pew Research, working across 68,000 real queries, found click rates falling by roughly 47% when AI Overviews are present, and only about 1% of users click any link inside the Overview itself. Seer Interactive’s analysis of 25 million impressions clocked a 61% organic CTR collapse and a 68% paid CTR collapse on AIO queries. Ahrefs measured a 34.5% CTR drop for position-one rankings across 300,000 keywords. In Google’s full AI Mode, Semrush put the no-click rate at 93%.

This is not a forecast. It is a ledger. AI-summarized search means publishers eat the cost of producing the answer and the platform captures the value of delivering it.

Meta is now running the same play. Smaller container, same logic, and even less benefit to the person who actually wrote the content.

Facebook posting was already a bad trade. Now it is a worse one.

Meta’s relationship with publishers has been transactional and one-sided for over a decade. They told us video was the future, then defunded the video teams that pivoted. They told us the News Tab was the future, then killed it. They suppressed outbound links until Facebook and X were both feeds engineered to keep users inside the app rather than click through to publishers. Meta has even tested making businesses pay for the right to share more than two links a month.

Now Meta has completed the loop. Their AI does not just suppress your link — it reads the question your post implies, answers it inside the post itself, and uses your own content as training data to do so.

It is worth being honest about the new arrangement:

• The user gets the answer without leaving Meta’s app.

• Meta gets the engagement signal, the ad impression, and a fresh training pass.

• The poster — you — gets nothing. No click. No referral. No subscriber. Often not even a visible link.

Posting on Facebook used to mean: I make content, Meta shows it to a fraction of my audience, some of them click through. Reach was bad, but there was at least a reach.

That deal is gone. Facebook is no longer a distribution channel for your content. It is a content extraction surface that uses your post as raw material for the AI sitting on top of it. Every post you write is, functionally, a free prompt for Meta AI.

Meta is not a search engine. So why are you letting them crawl like one?

Here is the part most website owners have completely missed.

Google has a defensible reason to crawl your site: their search engine actually sends people there. Even with AI Overviews compressing CTR, Google still drives meaningful referral traffic to the open web. The trade is lopsided, but a trade exists.

Meta has no such argument.

Cloudflare’s data on AI crawler behavior is brutal for Meta. Their crawl-to-referral ratio sits around 73,000 to 1 — Meta pulls content at enormous scale and sends almost no traffic back. Meta-ExternalAgent generates roughly 52% of all AI crawler traffic on the internet, more than double Google and OpenAI combined. In Cloudflare’s Q1 2026 robots.txt analysis, Meta-ExternalAgent never appears in any allow rule — every other major AI bot shows up in both allow and disallow lists, but Meta’s is exclusively blocked or ignored.

Translation: when site owners actually look at this bot, none of them want it. Yet only about 2% of sites currently block Meta-ExternalAgent, compared to roughly 25% blocking GPTBot. The gap is not because Meta is the better partner. It is because most site owners do not realize Meta is in their logs at all.

And here is the kicker: Meta is not a search engine. There is no consumer-facing Meta product that returns ten blue links to your content. The crawler exists for exactly one purpose — scrape your work so the same AI assistant that just answered your readers’ questions inside a Facebook post can keep doing it. You are subsidizing the thing that is eating you.

Block them. Here is how.

There are two kinds of Meta crawlers, and the distinction matters:

• facebookexternalhit generates link previews when someone shares your URL on Facebook or Instagram. If you still receive meaningful referral traffic from organic shares, leave this one alone.

• meta-externalagent, Meta-ExternalAgent, Meta-ExternalFetcher, and FacebookBot are the AI training and answer-engine crawlers. These are the ones extracting your content for Meta AI summaries.

For most CTOs and site owners, the right move is to block the AI crawlers and keep the link-preview crawler. Drop this into your robots.txt:

# Block Meta AI training and answer-engine crawlers

User-agent: meta-externalagent

Disallow: /

User-agent: Meta-ExternalAgent

Disallow: /

User-agent: Meta-ExternalFetcher

Disallow: /

User-agent: FacebookBot

Disallow: /

# Keep link previews working

User-agent: facebookexternalhit

Allow: /

The caveat: robots.txt is an honor system. Compliance is voluntary, and the directives do not technically prevent any bot from accessing your content. Multiple webmasters have reported meta-externalagent continuing to crawl despite explicit blocks.

If you want enforcement rather than a polite request, do it at the edge. Cloudflare’s Block AI Scrapers and Crawlers toggle under Security → Bots covers Meta-ExternalAgent and the rest of the AI crawler set with a single switch. For ASP.NET shops running their own infrastructure, a piece of middleware that returns 403 on any request whose User-Agent header matches the Meta crawler set is a ten-minute job. For Nginx:

if ($http_user_agent ~* “(meta-externalagent|Meta-ExternalAgent|FacebookBot|Meta-ExternalFetcher)”) {

return 403;

}

Bonus move: alert in your WAF logs whenever any of these user agents still hit you after the block goes in. Meta’s compliance has been spotty enough that monitoring is worth the cron job.

What posting on Facebook is actually for now

I am not telling you to delete your page. Facebook still has 3 billion users. Local business, marketplace, community groups — there are still niches where the platform earns its place.

But if you are a publisher, a creator, or a CTO running content marketing for a product, your Facebook post is now a feature in someone else’s product. The post exists to give Meta AI a question to answer. You are the prompt. You are the training data. You are not the destination.

The right play in 2026 is the play that was already right in 2024 — it just was not urgent yet:

1. Own the audience. Newsletter, RSS, podcast feed, push notifications. Anything where the relationship is direct and the address book is yours.

2. Treat social as a teaser, not a channel. Post enough to maintain presence; do not write for the platform. Assume Meta AI will summarize anything you put on Facebook before a human reads it.

3. Block the crawlers that do not pay you back. Meta sends near-zero referrals. OpenAI and Anthropic at least sometimes cite. Google still drives meaningful traffic. Treat them differently.

4. Get cited, not just clicked. The half-truth buried in the AI Overview data: brands cited inside the Overview pull roughly 35% more organic clicks and 91% more paid clicks than uncited ones. If you cannot beat the summary, get inside it. That is the new SEO.

5. Diversify aggressively. Substack, YouTube, LinkedIn, owned site, real RSS. YouTube and Reddit together account for over 78% of AI social media citations. That is where the models read from. Be there.

The Craig Kukuk post will keep doing numbers — Meta needs the impressions. BJJ Doc will keep posting — Meta needs the content. And Meta AI will keep answering the question before anyone taps the link, because that is the entire point of the product.

Stop subsidizing it.

Gal Ratner is CTO of WhiteStar Labs and Prana Entertainment. He writes at galratner.com and BlogCloud.io.

The sender is most often a Switzerland-based firm called PicRights, which acts as an enforcement agent for major news image rights holders. According to recent guides on the topic, PicRights Europe GmbH is a real company founded in 2007, headquartered in Pfaffikon, and represents Reuters, the Associated Press, Agence France-Presse, Science Photo Library, and CartoonStock. Their model is volume. Automated crawlers scan billions of pages, and when a pixel match is detected against a client’s image library, the system generates a demand letter with minimal human review. PicRights itself is not a law firm. They cannot sue you directly. But they refer escalated claims to firms like Higbee & Associates in the United States and Burness Paull in the United Kingdom, who can and do file.

I am being deliberately careful with my language here, because PicRights is a real, lawful, registered enforcement agency, and the company I run a blog at does not need a complimentary demand letter for being too colorful in print. So let me speak factually first, and then let me tell you why I am, frankly, thrilled that the ground beneath this entire industry just shifted.

How the demand letter business actually works

Strip away the branding and the model is unromantic. A rights holder licenses an image library to PicRights for monitoring. Crawlers do reverse-image searches across the web at scale. When a hit lands on a small site that probably grabbed the picture from a Google image search a decade ago and never thought about it again, a templated demand letter goes out asking for a license fee.

Lawyers and journalists who have watched this play out for years describe the economics with a clarity that I appreciate. Caroline Fox, a Virginia attorney who handles these cases, has publicly described how clients receive escalation letters threatening damages “up to $150,000” if they do not settle. The Belgian marketer Pierre-Nicolas Schwab investigated PicRights publicly after being targeted himself. A widely shared blog post on the demand-letter industry put it bluntly: a hundred letters demanding compensation from individual bloggers might yield five responses from people scared enough to pay something — often less than the amount requested. Five is better than none, and the math works.

I want to be precise: this is, by every legal characterization I have read, a lawful business. PicRights does in fact have authorization from the rights holders they name, and copyright law in most jurisdictions does in fact give those rights holders the ability to charge license fees for unauthorized use. The complaint that has rolled in for years from bloggers, journalists, attorneys, and small businesses is not that the practice is illegal. The complaint, captured in a Change.org petition that gathered thousands of signatures, in a Lead Balloon podcast investigation, and in pieces by outlets covering the issue, is about proportionality and tactics. About settlement amounts that frequently exceed the actual licensing cost of the image by 300 to 500 percent. About automated false positives. About the asymmetry between a firm that sends thousands of letters a month and a one-person blog that has neither a copyright lawyer on retainer nor a budget to hire one.

That asymmetry is the leverage. It is the entire product. The threat does not have to be likely to succeed in court — and many sources note that when these matters do reach court, awarded damages tend to come in well below the original demanded amounts. The threat only has to be expensive enough to fight that paying becomes the cheaper option. That is the math the small blogger faces at 11 p.m. on a Tuesday with a $1,400 demand letter open on the kitchen table.

Why I am genuinely happy AI image generation has matured

Here is the part where, as someone who spends most of his professional energy on AI infrastructure, I get to enjoy an outcome I did not initially predict. The leverage of the image demand letter business depends on one specific dependency: small publishers needing to use existing photographs that they did not create themselves. Strip out that dependency and you strip out the surface area.

Until very recently, the realistic options for a non-photographer running a blog were narrow. You could pay for stock photography, which most hobbyist bloggers will not. You could use Creative Commons or public-domain sources, which are easy to misread and easy to misuse, and which often end up sourcing the very same wire-service photos that get flagged. Or you could grab something off a search result and hope nobody noticed. For a long time, that last option was the realistic default for an enormous portion of the small web. That is the surface area that the demand letter industry harvested.

In 2025 and 2026, that calculus changed in a way that is not subtle. Image generation crossed the quality threshold where it can do the job that a small blogger actually needs an image to do — illustrate a post, set a tone, fill a header, give a feed thumbnail something to land on. The image does not have to be a Pulitzer-grade news photograph. It only has to look intentional. AI is now extremely good at intentional.

I am not going to pretend AI image generation is morally simple. The same training-data debates that produced the Getty Images v. Stability AI litigation are still working their way through courts, and I write about that openly. But I am willing to hold two thoughts at the same time. One: the training-data conversation matters and is unfinished. Two: it is genuinely good news that a hobbyist food blogger with thirty subscribers no longer has to choose between paying a stock site, gambling on Google Images, or being told ten years later that her risotto post owes the wire services $1,800.

What to actually use in 2026

The image generation landscape changed twice in the last twelve months. Both shifts matter for the practical case of “I need a header image for a blog post and I do not want a problem.” Here are the tools I would actually point a small publisher at, with the caveats I would actually give them.

Nano Banana 2 (Gemini 3.1 Flash Image)

Google’s Nano Banana series has become, in my opinion, the easiest on-ramp for non-technical creators. Nano Banana 2, officially Gemini 3.1 Flash Image, is now the default image model across the Gemini app, Google Search’s AI Mode, Google Ads, and Workspace. It generates from 512 pixels up to 4K, supports multiple aspect ratios, and renders text inside images well enough that you can produce a usable header with a headline already baked in. Crucially for the use case I care about here, every generated image carries an invisible SynthID watermark and is C2PA-compatible, which is the exact provenance signal that helps when the next crawler-driven enforcement industry tries to claim your AI-generated banana is somehow Reuters’ banana.

Available inside Gemini for free at conversational quality, and via the Gemini API and Vertex AI for builders who want to wire it into a publishing workflow.

Nano Banana Pro (Gemini 3 Pro Image)

When the asset matters more — a marketing page, a podcast cover, a real product shot mockup — Nano Banana Pro is the heavier model. It uses Gemini 3 Pro’s reasoning to plan layout, render legible text in multiple languages, generate 4K imagery, and ground itself in real-world knowledge for things like infographics. Google AI Pro and Ultra subscribers retain access to it via the three-dot menu in Gemini for cases where the default Nano Banana 2 is not enough.

If your blog is a brand and you care about consistent style across posts, Pro is worth the regenerate.

GPT Image 2 (gpt-image-2)

OpenAI shipped GPT Image 2 on April 21, 2026, alongside the consumer rebrand to ChatGPT Images 2.0. The headline upgrade is reasoning. It is the first OpenAI image model with a thinking mode that plans layout, can pull web references, and self-checks output before generating. In practice, that means it is genuinely good at the thing image models historically failed at: rendering small, dense, multilingual text inside the image. Menus, infographics, mockups with copy on them, social cards with a real headline — these are the use cases where GPT Image 2 is currently best in class.

Available to all ChatGPT and Codex users, with paid users getting the more advanced outputs, and to developers via the gpt-image-2 model on the OpenAI API and through Microsoft Foundry. Resolution support runs up to 4K. OpenAI’s terms grant commercial use of generated images, which removes one of the recurring footguns that slowed earlier adoption.

Honorable mentions for the cautious

Adobe Firefly is worth knowing about specifically because Adobe markets it as commercially safe — it was trained on Adobe Stock and licensed or public-domain content rather than the open web. If you are a business operator who genuinely cannot tolerate any training-data ambiguity, that is the lane to be in. Midjourney remains a creative powerhouse with its own license terms. Stable Diffusion’s licensing varies by version and self-hosting situation; read before you commit a workflow to it.

The boring legal caveat that I genuinely mean

Switching to AI image generation does not turn your blog into a copyright fortress. Two things matter, and I want them said plainly.

First, the U.S. Copyright Office has consistently held that purely AI-generated images, with minimal human creative input, are not protectable by copyright in the traditional sense. You can use the image. You can publish it commercially under most major platforms’ terms. You generally cannot stop someone else from using a similar one. For most blog use cases — header images, illustrations, post art — this matters less than people think. You are trying to avoid being the defendant, not trying to register the image with the Copyright Office.

Second, AI image generation does not give you a free pass to prompt for copyrighted characters, real living people without consent, registered trademarks, or stylistic clones of named living artists. If your AI-generated image happens to look exactly like a Reuters photograph, you have not actually escaped anything — you have just rebuilt the original problem with a more interesting tool. Use generic subjects, original compositions, and keep your prompts clear of named brands, named photographers, and named characters. If you want a forensic trail in case anything is ever questioned, save your prompts and your generation logs.

And of course: if you actually receive a demand letter from any enforcement agency, do not take legal advice from a Substack column written by a CTO. Talk to an attorney who handles these matters. Several do, and several charge a flat fee for an initial review specifically because the volume of these letters has made flat-fee triage a real practice area.

Why this is the optimistic AI story I did not expect to write

I write a lot in this column about AI being overconfident, AI polluting the web with low-effort content, AI being misused by companies that should know better. That criticism is honest and I am not retracting any of it. But I also believe in noticing when a technology actually shifts power back toward the small operator. This is one of those moments.

For roughly a decade, the small web — bloggers, hobbyists, church groups, local businesses, niche enthusiast sites — has been quietly squeezed by an enforcement infrastructure that depended on those operators not having the time, money, or legal sophistication to push back. The leverage was real. The fear was real. The settlements were paid, often by people who could least afford them, frequently for images on personal blogs that had been online for fifteen years and earned the operator exactly zero dollars.

What I am celebrating is not anyone’s misfortune. I am celebrating the fact that the dependency that made the leverage possible has, structurally, weakened. A small publisher in 2026 can make a perfectly good header image in roughly twelve seconds, with a SynthID watermark proving it was generated, and never need to dip into the wire services’ archives in the first place. The crawler still scans. The match still has to exist for a letter to be sent. If you never used the image, there is nothing to match.

That is the quiet shift. It is not a court ruling. It is not a regulatory win. It is a tooling change, and tooling changes are how the actual leverage in industries quietly moves. I have spent thirty years building software that levels the playing field for small operators against bigger institutions. I know what it looks like when the field tilts. It is tilting.

AI is going to do many things in the next five years that I will write critical pieces about. But on this one specific Tuesday in 2026, I am happy to put a generated banana on the header of a blog post and call it a small good thing. The kitchen-table demand letter at 11 p.m. is, for a growing number of people, becoming a story about something that used to happen.

Generate your own images. Save your prompts. Read your platform’s terms. And if a letter still shows up, talk to a lawyer, not the portal.

— Gal

The number you’ve probably heard — that 71% of Americans worry AI will take their job — is close, but not quite right. The actual figure, from Quinnipiac’s most recent national poll, is that 70% of Americans believe AI will lead to fewer job opportunities, and 30% are worried specifically about their own job becoming obsolete. That second number jumped from 21% in April 2025 to 30% by late 2025. Pew’s October 2024 survey put worker worry at 52%. So depending on how you slice it, the share of Americans afraid of AI is somewhere between half and three-quarters.

Here’s why the truth is worse than 71%:

The fear isn’t paranoia. It’s pattern recognition.

The Layoff Parade

Just the 2026 scoreboard, in case you missed any of it:

• Oracle: ~30,000 employees fired. April 1, 2026, by 6 a.m. email, system access cut immediately. About 18% of the global workforce.

• Amazon: ~30,000+ since October 2025, including 14,000 corporate jobs in a single wave.

• Meta: 8,000 cuts effective May 20, 2026. The company started logging employee keystrokes weeks before the announcement to train the AI agents that will replace them.

• Microsoft: ~8,000 voluntary retirement offers aimed squarely at older engineers.

• Block: ~4,000 layoffs — roughly 40% of staff. Leadership blamed AI maturity.

• Accenture: 11,000 cuts. CEO Julie Sweet stated explicitly that “those we cannot reskill will be exited.”

• Atlassian, Snap, Disney, Epic, Nike, Paycom, Autodesk, Baker McKenzie… the list runs another twenty companies long.

Through early March 2026, Layoffs.fyi tallied over 92,000 tech workers laid off in the first quarter alone. AI was explicitly cited in roughly 20% of those announcements — up from less than 8% in 2025. A National Bureau of Economic Research working paper, based on surveys of 750 corporate executives, projects 502,000 AI-driven cuts in 2026, nine times what we saw in 2025.

This is the part where the story usually ends. CEO says “AI is making us more efficient.” Stock pops on the announcement. Worker writes a LinkedIn post about resilience. We move on.

I’m not moving on, because the next data point is the one that should make every CEO’s general counsel start drafting securities-fraud disclosures.

Most of the AI Doesn’t Actually Work

RAND Corporation’s 2025 analysis: 80.3% of enterprise AI projects fail to deliver business value. MIT Sloan: 95% of generative AI pilots never scale to production. Gartner: 60% of projects without proper data infrastructure will be abandoned by the end of 2026. The Federal Reserve Bank of St. Louis pegged ChatGPT-era productivity gains at a whopping 1.9% over more than two years. MIT economist Daron Acemoglu came in lower — 0.5% over the next decade.

Pause on that. Five-tenths of one percent productivity growth, across an entire decade, from the technology that’s currently being used as the official excuse to fire half a million Americans this year.

It gets stupider.

A December 2025 survey of 1,000 hiring managers found that 59% admit they emphasize AI in layoff announcements because it “plays better with stakeholders” than admitting financial constraints. The Writer/Workplace Intelligence executive survey found that 75% of executives admit their company’s AI strategy is “more for show than a meaningful guide to outcomes”. Forrester predicts that half of AI-attributed layoffs will result in quiet rehiring — usually offshored, almost always at lower wages.

Translate the stack:

• American executives are firing American workers for AI capabilities that don’t exist yet,

• their own AI strategies are admittedly “for show,”

• and half the eliminated roles will be backfilled by cheaper labor in another time zone.

This isn’t a productivity revolution. It’s a wage arbitrage operation dressed in transformer architecture.

The Wage Vise

The arbitrage is layered. First the layoffs. Then the squeeze on whoever’s left.

Payscale’s 2026 Compensation Best Practices Report: 60% of companies now mention AI in job descriptions, but only 55% are willing to pay extra for AI skills. Median base pay increase in 2026: 3.5%. With “job hugging” at record highs (only 8% of US workers voluntarily quitting), employers don’t need to compete. They have a captive workforce that’s too scared to leave.

It gets worse. A 2025 controlled study (N=382) ran workers through an experimental workplace where some were managed by humans and some by AI. The AI managers, trained on supposedly “neutral” performance principles, systematically reduced wages by 40% — and workers didn’t push back. Why? Because, the researchers found, people don’t have the same emotional reaction to being lowballed by an algorithm as by a person. The very thing that makes AI feel “fair” makes silent exploitation easier.

California’s “No Robo Bosses Act” — which would have required employers to notify workers before AI made consequential decisions about their compensation, scheduling, or firing — was vetoed by Governor Newsom in October 2025 as “unfocused” and “overly broad.” A December 2025 federal executive order is now actively trying to block any state from regulating AI wage-setting.

So: American workers face mass layoffs blamed on AI capabilities that don’t work, residual wages set by algorithms that reduce pay 40% on average, and a federal posture that any attempt to regulate any of this is “anti-innovation.”

That’s the real number that should worry you. Not 71%.

Meanwhile, in Hangzhou

While we were busy firing each other, a Chinese hedge-fund spinoff called DeepSeek shipped DeepSeek V4 on April 24, 2026.

V4-Pro: 1.6 trillion parameters, 1 million-token context window. Beats every other open-source model. Matches GPT-5.5 and Claude Opus 4.7 on most agentic benchmarks at roughly one-tenth the cost per output token. Released under an MIT license. Open weights. Self-hostable. Trained natively on Huawei Ascend 950PR chips — the first frontier-class model built without a single Nvidia GPU in the loop.

Read that paragraph again, because it should land hard if you’re a CTO.

A Chinese company gave away frontier-grade AI for free, on Chinese silicon, while American AI companies were filing court briefs accusing them of stealing IP. Set aside, for a second, the irony of OpenAI — a company built on scraping the entire human internet without permission — accusing anyone else of “distillation.”

DeepSeek isn’t even alone. Alibaba’s Qwen 3.6 Plus, Moonshot’s Kimi K2.6, Zhipu’s GLM-5.1 — all open-weight, all close to the frontier, all available to any developer who wants to download them. The gap between “open-weight” and “closed frontier” used to be measured in years. DeepSeek’s own technical report says V4 trails the best closed models by approximately three to six months.

While our biggest AI labs spend hundreds of billions to lock the technology behind metered APIs, Chinese labs spend less, ship faster, and give the result to the world.

The Robot Factory Floor

Here’s where it stops being a software story.

In 2025, Chinese companies built and sold roughly 90% of every humanoid robot delivered worldwide. Six of the top-selling humanoid manufacturers are Chinese. Unitree alone sold 5,500 humanoids — more than Tesla’s entire Optimus production target, which Tesla failed to hit. AgiBot sold 5,168. The Unitree G1 retails for about $16,000 — cheaper than a used Tesla Model 3.

TrendForce projects Chinese humanoid output will grow another 94% in 2026. Unitree’s IPO prospectus pegs annual production capacity at 75,000 humanoid units within five years.

Now ask yourself what those robots are doing.

In Hangzhou, there’s a robot canteen serving $1.38 noodle bowls to local seniors. Robot arms do the cooking. Cleaning bots patrol the floors. The human staff that used to chop vegetables now sits with the elderly customers and talks to them. Automation handled the drudgery; humans got promoted to the empathy work. That’s a sentence we’re not writing in California.

China’s national elder-care pilot — launched in 2025 across 200 communities — is deploying robots for mobility assistance, medication reminders, fall detection, and yes, even toileting care for the bedridden. The country has 320 million people over 60 and is short 5.5 million professional caregivers. They didn’t convene a panel to debate whether robots are dignified. They built the robots and put them to work solving an actual humanitarian crisis.

I’m not pretending China is a workers’ paradise. It isn’t. The country has its own labor problems, its own surveillance state, its own contradictions. But ask yourself which civilization, right now, in 2026, looks more dystopian:

• The one whose AI labs open-source frontier models on domestic chips and deploy robots to feed and care for the elderly, or

• The one where Oracle fires 30,000 people by 6 a.m. email, Meta logs your keystrokes to train your replacement, and your CEO brags on the earnings call that AI is “creating efficiencies.”

I know which one I’d want my kids to grow up in. And it isn’t the one giving Sam Altman a $300 billion compute contract while a tenth of its tech workforce is in unemployment processing.

What You Actually Do About It

If you’re a US worker reading this, the fear is real and the data is real. But the conclusion you’re being sold — “learn to live with chronic insecurity, hope your boss likes you, take whatever raise the algorithm allows” — is the wrong one.

Here’s the conclusion that 30 years in this industry has taught me:

Stop relying on your employer. Not someday. Not after you “build savings.” Now — in the gaps between Slack messages, in the hour after dinner, on Saturday morning before the kids wake up.

You have something most economists ignore: deep, paid-for, hard-won domain knowledge about an industry. The accountant knows what bookkeepers actually need. The hospital coder knows which billing edge cases break which workflows. The warehouse manager knows which products move and which sit. The marketer knows which lead magnets convert in their niche. None of that knowledge is in the foundation models. None of it shows up on benchmark scores. All of it is monetizable.

You don’t need to quit. You don’t need to “follow your passion.” You don’t need a deck. You need a product or a service, a payment processor, and one paying customer. After that, it’s a process problem.

The on-ramp is cheaper than it has ever been. You can stand up an online storefront in an afternoon. ShopSnap.io — one of my own properties under Prana Entertainment — was built specifically for this: a fast, low-friction way to start selling products and services online while you still have a paycheck. No warehouse, no inventory burden, no need to rebuild Amazon. Use what you know about your industry to find an underserved customer, and start collecting revenue from them.

Even $500 a month from a side business is a different psychological universe than zero. It’s the first wedge between you and the layoff list. Your goal isn’t to escape your job tomorrow. Your goal is to make the job optional — so the day Oracle’s 6 a.m. email lands in your inbox, you have leverage instead of panic. So that the next time HR rolls out an “AI productivity initiative,” you can read it like the wage-arbitrage memo it actually is, and respond accordingly.

Closing the Loop

The 71% number is wrong. The real story is that American capital is conducting the largest experiment in voluntary human downsizing in the history of corporate finance, on the basis of a technology whose own pilots fail 95% of the time, while the country actually pulling ahead is the one giving the technology away and pointing it at problems humans want solved.

You don’t have to be a participant in someone else’s experiment.

Build something of your own. The knowledge in your head is the most valuable thing in your house, and your employer doesn’t have a clean license to it.

Go cash it in.

Walk away.

I’m not a financial advisor. I’m a CTO who has been in the technology business for thirty years, who has lived through three boom-and-bust cycles, who has hired engineers, been hired as one, and watched too many smart people sign away their leverage for a piece of paper that turned out to be exactly that — a piece of paper. So take this as one practitioner’s opinion, not investment advice. But the opinion is informed by a stack of receipts that should make every working engineer angry.

If a company tells you the equity is worth a lot, then the company can afford to pay you a lot. Insisting you take less cash because the equity is so valuable is, by definition, an admission that the equity is not worth what they say.

First, a sober definition

A stock option is a right, not a thing. It is the right to buy a share of company stock at a fixed price (the “strike”) at some future date, assuming a long list of conditions are met. Those conditions usually include: you stay employed, you make it past the cliff, the options actually vest on schedule, the company is still solvent, the company eventually has a liquidity event, the strike is below the eventual sale price, the post-termination exercise window doesn’t expire on you, and you have the cash to actually exercise when the moment comes.

Each one of those is a coin flip. Stack them, and the joint probability of “options become real money” is brutally low. Yet recruiters and founders quote the headline number — “$300K of equity!” — as if it were $300K in your bank account. It isn’t. It is a lottery ticket, and the lottery operator is also your employer, also the rule-maker, and also the person who can fire you the day before you collect.

The Skype playbook everyone forgot

In 2011, Microsoft agreed to buy Skype for $8.5 billion. Skype was at that time controlled by Silver Lake Partners, a private equity firm that had bought a 70% stake from eBay two years earlier. In the weeks before the acquisition closed, something interesting happened: Skype fired a list of senior executives. Multiple Vice Presidents. The CMO. The head of HR. Executives from a recently-acquired subsidiary. Out, out, out — right before the deal that would have triggered their change-of-control payouts.

Bloomberg reported it plainly: the timing of the dismissals meant the executives’ stock options would be worth dramatically less than if they had been allowed to ride out the deal a few more weeks. Those millions of dollars didn’t evaporate. They went somewhere. They went to Silver Lake.

Then a former Skype engineer named Yee Lee published a blog post that should be required reading for anyone signing an offer letter in 2026. Lee had joined Skype, vested a portion of his options on the standard schedule, then quit a year later to join a startup. He assumed his vested portion was his. That’s what “vested” means in Silicon Valley. That is the whole point of the word.

Skype told him no. Buried in his stock option grant — and cross-referenced to a separate “Management Partnership” document Lee had never seen — was a clause stating that if you left before the company went liquid, the company could buy back your vested shares at a price they chose. Often that price was zero. As TechCrunch wrote at the time, the company had the right to terminate not only your unvested options, but also your vested ones, and even shares you had already paid cash to exercise. Vested didn’t mean vested. Vested meant “vested, until we decide it isn’t.”

Two corporate lawyers who looked at the contract said this kind of structure was “not uncommon” in PE-backed deals. Read that sentence again. The trick wasn’t a one-off. It was a template.

Vested didn’t mean vested. Vested meant “vested, until we decide it isn’t.”

It’s not just private equity. It’s everyone with leverage.

If you think this is a relic of 2011, look at what happened the night Elon Musk closed the Twitter deal in October 2022. According to Walter Isaacson’s Musk biography, Musk “meticulously” planned to terminate CEO Parag Agrawal, CFO Ned Segal, and Chief Legal Officer Vijaya Gadde before their stock options vested in the closing. The three were owed change-of-control packages totaling around $76 million combined. Instead they were walked out of the building by security, fired “for cause,” and Musk’s lawyers proceeded to argue they were owed nothing. The executives sued for $128 million in unpaid severance. The rank-and-file fared no better — a separate $500 million severance lawsuit on behalf of laid-off employees was dismissed on technical ERISA grounds in 2024.

These are not small companies run by amateurs. These are the largest tech transactions of the decade. And the consistent pattern at the top of the deal stack is: figure out how to pay employees less than the contract appeared to promise.

The vesting-cliff layoff is real, and it has a season

Down at the line-engineer level, the equivalent move is the cliff layoff. Standard tech equity vests over four years with a one-year cliff: you earn nothing in year one, then 25% snaps into vesting on your first anniversary, then the rest drips out monthly. If your employer terminates you before that one-year mark — even one day before — you walk away with zero equity. Not 11/12ths. Zero.

Look at the public layoff data: in 2025, more than 245,000 tech employees were laid off across roughly 783 companies, per the Layoffs.fyi tracker. We’re already 95,000-plus deep in 2026. Some of those are people who had been there four years. Plenty are people who had been there ten or eleven months. The companies will tell you the timing was about “performance” or “restructuring” or “market conditions.” Sometimes that’s even true. But the financial outcome — for the company — is identical regardless of the reason: the company recaptures the unvested options, returns them to the pool, and re-grants them to the next round of hires at a lower strike price. The CFO smiles.

There are honorable exceptions. Stripe famously waived the cliff during its 2022 layoffs so departing employees got their first 25%. Some companies extend the post-termination exercise window from the standard 90 days out to several years. Those companies should be celebrated and copied. They are also, by their own admission, the exception. The norm is the 90-day window, and the norm is the cliff, and the norm is that the company gets to use both as cudgels.

The math no one in the recruiter chain wants you to do

Engineer Dan Luu wrote one of the sharpest pieces I’ve read on this. His argument, paraphrased: if startup options were really worth what companies claim they’re worth, the companies could simply raise more money and pay competitive cash. The fact that they refuse to is not a clever financial structure. It’s a tell. It is, in his words, an admission that the options aren’t worth as much as the headline number suggests, dressed up in language that obscures the fact.

Steve Blank, who has watched Silicon Valley compensation since the 1980s, has made a related observation: the bargain has gotten worse. In the 20th century, founders genuinely bet their mortgage on the company. Today, founders cash out at Series B, take liquidity in secondaries, and get cushioned by multimillion-dollar seed rounds. The early employees still take all the original risk, work the same eighty-hour weeks, and now find that their stock has been diluted, repriced, or made subject to clauses they never read.

Run the numbers yourself. If you take $40,000 less in cash compensation per year for four years in exchange for $300,000 of “equity,” you have just lent the company $160,000 of your own money — interest-free, with no collateral, with the lender (you) bearing 100% of the downside, and with the borrower (the company) holding all the cards on whether you ever see a return. No one in any other context would call that a good deal. We only call it a good deal when the asset class is called “startup equity” because the industry has spent thirty years training people to call it that.

Equity is a tip. A salary is the meal. You don’t go to a restaurant, eat a smaller meal, and feel good because the waiter promised you a great tip later if business picks up.

The contrast: a worker cooperative actually means what it says

There is a model where “employee ownership” is not a marketing phrase. It’s called a worker cooperative. In a worker co-op, the workers themselves own the business outright. There is no founder holding super-voting shares. There is no PE firm with a clawback clause buried in a Management Partnership document. Each worker-member typically holds one share, gets one vote, and shares in profits through patronage dividends — distributions based on hours worked or wages earned, paid annually.

The structural difference is enormous. In a co-op, governance is democratic and one-member-one-vote, by design. There is no class of preferred shareholders ahead of you in the liquidation stack. There is no board appointed by Silver Lake. The Spanish Mondragon Corporation, the canonical example, has run this model for sixty years across more than a hundred businesses. In the U.S., co-ops are smaller and rarer, but they exist, and the principle is what matters.

Now, I am not arguing that everyone should go work at a worker cooperative tomorrow. Most of us — me included — work at conventional capital-stack companies, and that’s fine. The point of the comparison is to clarify what “ownership” actually looks like when the structure is designed for the worker rather than around the worker. Stock options at a VC-backed startup are not ownership. They are a contingent claim, structured by the company’s lawyers, exercisable only on terms the company controls, and revocable in dozens of subtle ways.

If you want real ownership, work at a co-op or start your own business. If you take a job at a normal company, take it for the cash. Anything beyond cash is a tip.

Your only fiduciary duty is to your family

Here is the part of this essay where some readers will get uncomfortable. The employer-employee relationship is not a friendship. It is not a partnership. It is not a family. It is a transaction governed by an at-will employment agreement that, in most U.S. states, the company can end on any day, for almost any reason, with no notice and no obligation beyond what’s been promised in writing. That is the deal. The company has accepted those terms because they are good for the company.

You should accept those terms too — and act accordingly. In an employer-employee relationship, your first and only fiduciary duty is to yourself and the people who depend on you. Not to the founder’s vision. Not to the mission statement. Not to the kombucha tap. Your job is to take care of your spouse, your kids, your aging parents, your mortgage, and your future self. Loyalty to a corporation is, almost without exception, loyalty that is not reciprocated. Ask the Skype engineers. Ask the Twitter executives. Ask the people whose layoff date landed two weeks before their cliff.

This isn’t cynicism. It’s symmetry. The company optimizes for the company. You optimize for you. When both sides do that openly, the negotiation is honest.

Rules I would tell my own kid

• Negotiate the cash first, last, and only. Treat equity as a bonus, never as a salary substitute. If the cash number doesn’t work, the offer doesn’t work, full stop.

• If you do take equity, read every word of the option grant — and ask for the documents it cross-references. If they refuse to show you the Management Partnership agreement (or whatever they call it this decade), refuse the offer.

• Demand double-trigger acceleration on change of control. That means your unvested equity vests if (a) the company is acquired AND (b) you are terminated without cause within some window. This is the single most important clause for protecting yourself against the Skype scenario.

• Negotiate an extended post-termination exercise window. The standard 90 days is a trap — most people don’t have the cash to exercise on short notice, especially after a layoff. Push for 5 to 10 years.

• Get every promise in writing. Verbal commitments from the founder over coffee are worth what’s printed on them: nothing.

• If the company is private, ask the strike price, the 409A valuation, the total share count, and your percentage. If they won’t tell you the denominator, your “10,000 shares” might be 10,000 out of 10 billion.

• Assume you will be laid off the day before your next vesting milestone. Plan accordingly. If that assumption ruins the offer, the offer wasn’t actually that good.

The bottom line

Stock options are not bad. I have them. I’ve granted them. They are a legitimate tool for aligning long-term incentives between a company and its earliest, most committed builders — the people who would have signed up anyway, who genuinely want the upside, who can afford to bet a slice of their compensation on a moonshot. That is fine.

What is not fine is the practice of using stock options to depress cash compensation across the industry, while simultaneously building legal mechanisms — the cliff, the 90-day window, the change-of-control workaround, the buried clawback clause, the for-cause termination — that ensure the company gets to keep most of the upside no matter what happens to you. That is not a compensation strategy. That is a transfer of wealth from the people doing the work to the people structuring the deal.

So when the recruiter calls and the salary is light and the equity is heavy and the founder talks about the mission, just remember: in 2011, Skype’s lawyers had already written the document that made vested mean unvested. In 2022, the new owner of Twitter had already mapped out which executives to walk out the door minutes before they got paid. In 2025 and 2026, hundreds of thousands of tech workers learned that “part of the family” lasted exactly until the next earnings call.

Take the cash. Get the cash. The cash is real. Everything else is a story someone else is telling about your future.

— Gal

Disclaimer: I am not a financial advisor, a lawyer, or a tax professional. This essay is opinion and analysis, drawn from public reporting and three decades in the industry. Before signing or rejecting any compensation package, talk to an actual employment attorney who has read your specific contract.

Sheldon, predictably, gives them physics. He invokes Schrödinger’s Cat — the 1935 thought experiment where a cat in a sealed box exists in a superposition of alive and dead until someone opens the lid and forces reality to pick a side. He tells Penny her potential relationship with Leonard is, right now, simultaneously good and bad. The only way to find out which is to open the box. At the end of the episode, Leonard knocks, kisses her, and Penny famously declares: “The cat’s alive!”

Eighteen years later, the entire AI industry has gathered around a different sealed box, arguing about what’s inside. The box is labeled AGI. The cat is Artificial General Intelligence. And unlike Penny and Leonard, nobody is willing to actually open the lid — because as long as it stays closed, the cat is worth a trillion dollars.

The Predictions Are Already in Superposition

Let’s do something I rarely see in AI coverage: actually look at what the people building these systems are saying, side by side, in the same week. Not cherry-picked. All of it.

As of early 2026, here’s where the credentialed forecasters land:

• Dario Amodei, CEO of Anthropic: At Davos 2026, said AGI will likely arrive within a few years, possibly by 2027. Anthropic’s formal submission to the U.S. Office of Science and Technology Policy explicitly states they expect powerful AI systems to emerge in late 2026 or early 2027.

• Sam Altman, CEO of OpenAI: Has said AGI will probably get developed during Trump’s current term. Also recently called AGI “not a super useful term.” Both statements were made by the same man, often within weeks of each other.

• Mustafa Suleyman, CEO of Microsoft AI: February 2026 — predicted human-level performance on most professional tasks within 12 to 18 months.

• Elon Musk, founder of xAI: Originally said 2025. Then 2026. Then at Davos 2026 sharpened it to “by year-end.” If you’re tracking, that’s now three consecutive year-end deadlines that have slipped.

• Demis Hassabis, founder of Google DeepMind: Roughly 50% chance by 2030. Notes we still need “one or two more breakthroughs.”

• Shane Legg, Chief AGI Scientist at DeepMind: About 50% odds of “minimal AGI” by 2028.

• Andrej Karpathy, formerly OpenAI: Says current agent architectures “aren’t anywhere close.” Puts AGI a full decade out.

• Yann LeCun, Chief AI Scientist at Meta: Argues current architectures have fundamental limitations that preclude near-term AGI. Has been saying this consistently for years and has been consistently right about each individual ceiling, even as the field has worked around them.

• Gary Marcus: Has placed a 10-to-1 public bet that AI won’t hit specified AGI tasks by the end of 2027.

• Geoffrey Hinton, Nobel laureate: Revised his timeline from “30 to 50 years” to “5 to 20 years,” with 50% probability within two decades. Same person. Different decade.

• 2023 AI Impacts survey of 2,778 published AI researchers: 50% probability of high-level machine intelligence by 2047.

• Metaculus aggregate forecast (~2,000 contributors), as of February 2026: 25% chance of AGI by 2029, 50% by 2033.

Read that list back. Within a single calendar quarter, credentialed researchers and CEOs of the biggest AI labs in the world have publicly committed to AGI arrival dates ranging from “by year-end 2026” to “never, the architecture is fundamentally incapable.” That’s not a confidence interval. That’s not even a disagreement. That’s a thought experiment.

If 30% of experts say your airplane is going to explode, and 70% say it won’t, you shouldn’t conclude there’s no expert consensus, so I won’t do anything. — 80,000 Hours, on AGI forecasting

The cat is alive. The cat is dead. The cat does not exist. The cat will be here Tuesday. The box itself is a marketing concept. Pick one — or pick all of them — you’re in good company either way.

The Box Has a Definition Problem

Here’s the part that should make every CTO, board member, and enterprise buyer pause. The reason these predictions span 73 years isn’t because the field is bad at math. It’s because nobody can agree on what’s in the box in the first place.

Sam Altman defines AGI one way on a podcast and another way in a funding round. OpenAI now uses a five-level framework instead of a binary AGI designation — conveniently, the levels above the current product can be redefined as the company progresses. Dario Amodei describes “a country of geniuses in a datacenter.” Shane Legg defines “minimal AGI.” Fei-Fei Li argues no system without spatial intelligence and physical-world understanding qualifies, period. Yoshua Bengio cares about reasoning. Yann LeCun cares about world models. The 2023 AI Impacts survey defined it as “every task better and cheaper than human workers.” That’s an economic threshold, not a cognitive one.

These are not synonyms. They are different cats in different boxes, all being marketed under the same brand.

And here’s the kicker: this is a feature, not a bug, of how the industry is being financed. When the deliverable is undefined, the deliverable can never be late. Definitional ambiguity is a moving goalpost dressed up as humility.

The Compression Doesn’t Mean What You Think It Means

The most quoted statistic in AGI discourse right now is the Metaculus compression: in 2020, the median forecaster predicted AGI in roughly 50 years — placing it around 2070. By early 2026, the median had collapsed to 2033. Five decades of compression in less than six years.

The narrative says: “Experts updated. The smart money is moving forward.”

The actual data says something more uncomfortable. The 80,000 Hours review of expert forecasts found that essentially every category of forecaster — academic researchers, prediction-market participants, industry executives — has compressed timelines in lockstep. When every group moves together by the same magnitude, regardless of methodology, that is not the signature of independent rational updating. That is the signature of a single shared narrative being absorbed by every population at once.

The same Metaculus that now says 2033 said 2070 in 2020. Both predictions came from a similar pool of forecasters, with similar reasoning. One of those numbers is wrong by decades. The system that produced the wrong one is the same system producing the new one.

This is the lesson the AI industry refuses to internalize: the variance between predictions is information. When predictions span 5 years to never, the right read isn’t to average them. The right read is that nobody knows, and anyone telling you they do is selling something.

Why This Is the Most Expensive Cat in History

Sheldon’s thought experiment had stakes for two people: Penny and Leonard. Open the box, find out if you’re dating. Worst case, you have an awkward dinner.

The AGI box has stakes measured in trillions. NVIDIA’s market cap, OpenAI’s rumored valuation, Anthropic’s funding rounds, the global hyperscaler capex cycle, hundreds of thousands of jobs at consultancies repositioning around “AI transformation,” the entire 2025–2026 venture playbook — all of it is priced as if the cat is alive. As if AGI is real and imminent and the only question is who captures the value.

Meanwhile, the same researchers selling the box admit, on the record, that they don’t know what’s inside, when it arrives, or whether it’s achievable with current architecture. Yann LeCun — Chief AI Scientist of one of the largest AI deployments on Earth — publicly says current approaches will not produce AGI. Andrej Karpathy, who built much of the foundation, says we’re a decade out and current agents are not close. These are not skeptics shouting from the cheap seats. These are the people whose names are on the papers.

If you’re a CIO building a 5-year AI roadmap, what do you do with that? If you’re a board approving $50M of GPU capex, what do you assume the depreciation curve looks like? If you’re an engineer choosing between a stable .NET architecture and rebuilding the stack on whatever framework the latest model release demands, which timeline do you plan against?

The honest answer is: you can’t plan against a superposition. You have to pick. And the industry’s answer to that uncomfortable fact has been to keep the box closed and keep selling tickets.

What Sheldon Got Right

Here’s the irony of the Big Bang Theory reference. Sheldon’s advice to Penny was actually correct. He didn’t tell her the cat was alive. He didn’t tell her the cat was dead. He told her the only way to find out was to open the box — to take the action, see the result, and let reality collapse the wave function.

The AI industry has done the opposite. It has built a multi-trillion-dollar economy on the principle that the box must never be opened, because the value lives in the superposition. As long as AGI is two years away, every funding round prices in the upside. As long as AGI is twenty years away, every safety researcher has tenure. As long as the definition is fuzzy, every product can claim to be on the path.

The cat is alive AND dead. The valuation is real AND fictional. The timeline is imminent AND distant. The architecture is sufficient AND a dead end. Pick whichever pair lets you keep raising money.

Schrödinger came up with the cat experiment in 1935 specifically to mock the Copenhagen interpretation of quantum mechanics. He thought it was absurd that any serious physicist could believe a real, macroscopic object could exist in two contradictory states simultaneously. The whole point of the thought experiment was: this can’t be how reality actually works.

90 years later, an entire industry decided he was wrong. Not about quantum physics — about business models. The cat is the product. The box stays closed. And every quarter, somebody opens a new fund to bet on what’s inside.

The Cat’s Alive. The Cat’s Dead. Plan Accordingly.

If you’re running a technology organization in 2026, here’s what the data actually supports — not the narrative, the data:

• Capabilities are improving. Real, measurable capabilities, in coding, reasoning, multi-step tasks. That part is not hype.

• AGI — as a single threshold event that changes everything overnight — is a marketing artifact. The credentialed researchers can’t agree on its definition or its arrival date by a margin of decades.

• The compression of timelines is correlated, not independent, which means it’s narrative-driven, not evidence-driven.

• The only safe planning assumption is that some current systems will keep getting better, some will plateau, some categories you don’t expect will leapfrog, and the people loudest about timelines are the ones with the most to gain from your belief.

Build for capability, not for prophecy. Architect for replaceability of the underlying model, not for loyalty to a specific lab’s roadmap. Treat every “AGI is N years away” headline as what it actually is: a fundraising document.

And remember the lesson Penny figured out in 22 minutes that the AI industry has spent a decade avoiding. The only way to know whether the cat is alive is to open the box. Until somebody does, every claim about what’s inside is equally true and equally worthless.

The cat’s alive. The cat’s dead. The cat is funding our next round.

— Gal

It was called Stargate. Five hundred billion dollars. Four years.

Sam Altman was there. Larry Ellison was there. Masayoshi Son was there.

Satya Nadella was not.

That photograph is the cleanest summary of Microsoft’s last decade I can give you. The man who has invested more than thirteen billion dollars of Microsoft’s money into OpenAI was not invited to the most consequential AI infrastructure announcement in the country’s history — because by January 2025, OpenAI had already chosen Oracle.

Thirteen billion dollars. No seat at the table.

This essay is the case for replacing him.

I want to be precise about what I am arguing. I am not arguing that Satya Nadella is a bad person or a bad operator. He is, by most institutional measures, a perfectly competent global executive. He gives a good interview. The stock went up. He turned the cultural temperature down at a company famous for eating itself.

What I am arguing is much narrower and much harder to dismiss: Microsoft already has its next CEO. He has been quietly running the only profitable half of the company for twelve years. The board has been watching him do it. The org chart has been lying about who the operational CEO actually is. And every quarter that succession conversation gets postponed, Microsoft drifts further from the developer-first DNA that built it.

The man’s name is Scott Guthrie. The board owes shareholders an honest conversation about him.

Here is the evidence.

I. The Two Track Records

In February 2014, Satya Nadella became CEO of Microsoft. In that exact same month, Scott Guthrie became EVP of the Cloud + AI Group, reporting to him.

For twelve years, the two men have run parallel halves of Microsoft. One signs the press releases. The other ships the product.

The portfolios are not close.

Satya Nadella’s personal product bets, 2014–2026

• Cortana — discontinued.

• Mixer — acquired and killed in 2020 after the failed Twitch challenge.

• Windows Phone — discontinued.

• Surface Duo — flopped on launch.

• Skype — Microsoft’s $8.5 billion acquisition under Ballmer, neglected for a decade under Nadella, finally killed in May 2025.

• Bing Chat — the product Nadella publicly said would make Google “dance.” Google did not dance.

• Windows Recall — pulled from launch in 2024 over privacy and security backlash.

• Windows 11 — the most rejected enterprise upgrade in Microsoft’s history.

• Inflection AI — $650 million paid to acqui-hire Mustafa Suleyman’s team for a newly created consumer AI division. The deal attracted an FTC antitrust inquiry. Suleyman now runs Microsoft AI. The flagship product his division builds — Microsoft 365 Copilot — is the AI assistant the internet now calls Microslop.

• The OpenAI partnership — a multi-billion-dollar dependency on a non-Microsoft entity whose CEO was fired by his own board in November 2023 and reinstalled by external pressure inside a week, and which has since signed deals with Oracle, Google, AMD, NVIDIA, and CoreWeave that Microsoft would prefer it didn’t.

That is the public record.

Scott Guthrie’s portfolio, same window

• .NET and .NET Core — the cross-platform reboot of Microsoft’s developer stack.

• Visual Studio Code — the most-used code editor on Earth.

• GitHub — acquired in 2018 and grown into the world’s developer infrastructure.

• TypeScript — the language that ate JavaScript.

• Azure — 95% of the Fortune 500, 39% year-over-year growth, over $75 billion in revenue.

• Azure OpenAI Service — the only part of the OpenAI relationship that actually monetizes at scale.

• SQL Server 2025 — including native vector search.

• Power BI, Power Apps, Power Automate.

• Maia 200 — Microsoft’s first competitive in-house AI silicon, beating Amazon’s Trainium and Google’s TPU on benchmarks.

• Fairwater — the planet-scale AI datacenter network.

• GitHub Copilot — the only Microsoft AI product developers actually love.

That is the public record.

You can disagree about emphasis. You cannot disagree about the columns.

One of these track records is Microsoft’s growth engine. The other is Microsoft’s reputational liability. The growth engine reports to the liability. The org chart needs updating.

II. The Governance Problem

In June 2021, the Microsoft board allowed Satya Nadella to also become Chairman of the Board.

This was a governance decision that should not have happened, and it should be reversed regardless of what the board concludes about CEO succession.

ISS, Glass Lewis, CalPERS, and most major institutional governance frameworks have flagged combined Chairman/CEO structures as a structural conflict for over a decade. The reasoning is simple: the Chairman’s job is to oversee the CEO’s performance. When you make the same person both, you have eliminated the independent oversight your shareholders rely on. The Chairman is now grading his own homework.

The Microsoft board has eleven independent directors:

Sandra E. Peterson, the Lead Independent Director. Reid Hoffman of Greylock Partners. Penny Pritzker of PSP Partners. John W. Stanton of Trilogy Partnerships. Charles W. Scharf, CEO of Wells Fargo. Hugh Johnston, CFO of Disney. John David Rainey, CFO of Walmart. Teri List, former CFO of The Gap. Catherine MacGregor, CEO of Engie S.A. Mark Mason of Citi. Emma Walmsley, former CEO of GSK plc.

Three of these directors are sitting CFOs of Fortune 100 companies. They know exactly what combined Chair/CEO governance looks like in their own boardrooms. They voted for it at Microsoft anyway.

The Lead Independent Director role exists specifically to provide a structural check when a Chairman/CEO has consolidated power. Calling executive sessions, leading evaluations of the CEO, and surfacing succession conversations are not optional duties — they are the job description. Sandra Peterson, by accepting that title, accepted the responsibility.

Twelve years of CEO performance under combined Chairman/CEO oversight has produced:

• Fifteen thousand layoffs in 2025, including nine thousand in a single July round.

• 8,750 voluntary buyouts of US employees announced in April 2026 — the first such program in company history.

• $17.5 billion committed to India in December 2025, in the same four-month window as the buyouts.

• The Microslop reputational drift.

• The OpenAI strategic dependency.

• The empty White House podium on January 21st.

The board does not need to fire the CEO tomorrow. It needs to do three things this quarter: separate the Chairman and CEO roles, schedule an honest succession conversation, and stop pretending its current governance structure is normal. Every quarter it postpones those conversations is a quarter of accumulating personal liability for the directors themselves.

That includes the three sitting Fortune 100 CFOs who know better.

III. The Microslop Receipts

Let us look at what happens when a CEO without product taste is given twelve years to make personal product bets.

In March 2024, Microsoft paid approximately $650 million to license the AI models of Inflection AI and acqui-hire most of its 70-person team, including its co-founders Mustafa Suleyman and Karén Simonyan. The deal was structured to look like a licensing agreement rather than an acquisition. The FTC opened an antitrust inquiry anyway, on the grounds that the structure appeared designed to bypass merger review.

Suleyman became CEO of a newly created consumer AI division called Microsoft AI. His division built and now operates the consumer-facing Copilot, Microsoft 365 Copilot, and the company’s broader generative AI consumer experience.

Two years later, the AI assistant his division ships has a nickname. The internet calls it Microslop.

You can see the chain. Nadella personally championed the OpenAI dependency, which produced internal pressure to commercialize generative AI through the Office suite. When OpenAI’s organizational instability became visible after the November 2023 firing-and-reinstatement of Sam Altman, Nadella spent $650 million to bring an alternative AI leadership team in-house. That team built the Copilot products that have become the public face of his second decade as CEO. Those products are the meme.

This is not an accident of marketing. It is a sequence of personal product bets, each one a direct consequence of the prior bet not working out, each one signed off at the corner office.

Compare that to GitHub Copilot.

GitHub Copilot was built inside Scott Guthrie’s organization in 2021. It is the AI product developers actually pay for on purpose. It runs inside Visual Studio Code and Visual Studio. It is the most successful AI productivity tool currently shipping in the enterprise, and it has none of the reputational damage of its M365 sibling.

Same word. Different orgs. Different reception.

The difference is not the technology. The technology underneath is largely the same. The difference is taste, judgment, and who in the company is shipping with the customer in mind versus who is shipping to fill a strategy deck.

GitHub Copilot is a product. M365 Copilot is a tax. One of them grew up in Guthrie’s org. The other grew up in the corner office.

Twelve years. One exec ships the growth. The other ships the meme.

IV. The Hostage Situation

In January 2019, Microsoft committed an initial $1 billion investment in OpenAI as the exclusive cloud provider for OpenAI’s compute. Over the next five years, that figure grew to over $13 billion across multiple investment rounds.

Nadella made this his signature strategic move. The deal that defined his second decade as CEO. The bet that was supposed to leapfrog Microsoft over Google and AWS in the AI era.

Here is what it has produced.

November 2023. OpenAI’s board fired Sam Altman without consulting Microsoft. Microsoft’s $13 billion investment in the company was made effectively without contractual control over its leadership. Altman was reinstated within a week — but only after external pressure from employees and investors, not Microsoft’s own contractual rights. The episode was the cleanest possible demonstration of how little leverage Microsoft actually held.

January 2025. OpenAI announced the Stargate Project from a White House podium. The founding partners were SoftBank, Oracle, OpenAI, and MGX of Abu Dhabi. Microsoft was relegated to “key technology partner.” The new arrangement gave Microsoft a “right of first refusal” on future OpenAI compute capacity — but only if Microsoft could match the terms. Microsoft was no longer the exclusive cloud partner. It was now the secondary one.

September 2025. OpenAI signed a $300 billion, five-year compute contract with Oracle. The single largest enterprise cloud deal ever signed. Microsoft was not the counterparty.

March 2026. OpenAI declined to expand its flagship Stargate site in Abilene, Texas. Microsoft picked up the leftover 900 megawatts on the parcel next door. The two companies are now literal neighbors on the same plot of dirt, building separate AI futures. Fortune called it what it was.

April 2026. Microsoft AI publicly declared “true self-sufficiency” as the strategic goal — admitting in plain language that the dependency the company spent a decade building is the dependency it now has to spend a decade escaping.

This was the strategy.

Compare it to what was happening on the other side of the building.

While Nadella was wiring billions to Sam Altman’s foundation, Scott Guthrie’s organization was quietly building the things Microsoft would need if the OpenAI bet ever soured.

• Maia 200 — Microsoft’s own AI inference silicon, fabricated on TSMC’s 3-nanometer process, beating Amazon’s Trainium and Google’s TPU on benchmarks.

• Fairwater — the planet-scale AI datacenter network, built so Microsoft owns the substrate regardless of who wins the model race.

• Azure OpenAI Service — the part of the OpenAI relationship that actually monetizes, packaged through Guthrie’s platform.

One of these strategies is a hostage situation. The other is a hedge.

The CEO who put the company in the hostage situation is now asking for five more years to dig out of it. The EVP who built the hedge has been quietly making sure Microsoft survives if he can’t.

The board does not need to choose between these men. It already did, twelve years ago, when one of them got the title and the other got the work.

Microsoft’s next CEO has been quietly insurance-policying the company against its current CEO’s biggest bet.

Promote the underwriter.

V. The Heir Apparent

Who is Scott Guthrie?

He joined Microsoft in 1997 at the age of twenty-two, fresh out of Duke University with a computer science degree. Over the Christmas holidays of 1997, he and a colleague named Mark Anders built the first prototype of a new web framework. They called it XSP. The X did not stand for anything. It just sounded cool.

That prototype eventually shipped as ASP.NET — the framework that defined Microsoft’s developer story for the next two decades. Twenty-eight years later, the language and runtime he helped birth still ships under the name .NET, now in its tenth major version, now cross-platform, now open source.

He has spent his entire career inside that lineage. He runs the development teams that built the Common Language Runtime. He helped found the .NET project. He shipped Silverlight. He shipped Windows Phone’s developer tools. He led the team that built Visual Studio. In February 2014, he was elevated to Executive Vice President of the Cloud and Enterprise Group, later renamed Cloud + AI.

Today his portfolio includes Azure, Dynamics 365, Visual Studio, Visual Studio Code, GitHub, .NET, Microsoft SQL Server, Power BI, Power Apps, Power Automate, and HoloLens. He sits one rung below Satya Nadella on the Microsoft org chart and reports directly to him. By internal headcount and direct revenue contribution, his organization is the larger half of Microsoft.

He is fifty-one years old. He still wears the same red polo shirt at every BUILD keynote. According to people who work with him, he still reviews pull requests. He still spends time with customers. He has been the highest-profile internal Microsoft executive without an accompanying personal brand for two decades — a man who has shipped product at scale while every other senior leader at Microsoft has built an outside reputation.

Fortune named him one of the twenty-five most powerful rising executives in the Fortune 500 in late 2025. The article noted, dryly, that he was already running the half of the company that grows.

He is exactly the kind of CEO that Microsoft used to produce before the marketing era. He came up through the developer division. He shipped the tools he sells. He has actual technical taste. His career arc — from a 22-year-old prototyping ASP.NET on Christmas break to a 51-year-old executive deploying $80 billion of AI infrastructure — is the Microsoft career path the company used to advertise.

When Steve Ballmer chanted “developers, developers, developers” in 2000, he was naming Microsoft’s source of competitive advantage. That source has not changed. The market has only made it more important. The developer is still the customer that decides whose stack the next decade gets built on.

Scott Guthrie is the developer.

The Microsoft board does not need to search for a CEO. It needs to recognize one.

VI. The Cost of Doing Nothing

I want to be specific about what continued postponement of this conversation costs.

In 2020, Microsoft had eight thousand employees in India. By late 2023, that number had passed twenty thousand, distributed across ten Indian cities. By April 2026, Microsoft had announced 8,750 voluntary buyouts of US employees — the first such program in the company’s fifty-one-year history — in the same four-month window in which it pledged $17.5 billion to India.

This is not xenophobia. It is concentration risk. A critical American software vendor is restructuring its labor force toward a single foreign country, with R&D and engineering disproportionately overrepresented in that move. Procurement and infosec teams at the Fortune 500 are starting to ask questions about code provenance and sovereign exposure that they have never had to ask before. Microsoft’s own press releases now use the word “sovereign” in describing its India build.

Every quarter the succession conversation is postponed is a quarter of accumulating brand damage in markets that Microsoft cannot afford to lose. American enterprise customers are noticing. Procurement language is shifting. Microslop is now appearing in IT trade press in the same category as “vendor lock-in” and “license bloat.”

It is also a quarter of accumulating talent damage. Senior Microsoft engineers do not leave for nothing. They leave because they no longer recognize the company they joined. They leave because they have read the Skype obituary. They leave because they have been to enough product reviews where M365 Copilot integration was the answer to a question nobody asked. They leave because their performance reviews now include AI-feature adoption metrics that have nothing to do with quality.

The Microsoft alumni network is now one of the most important talent reservoirs in the AI industry. The board should ask itself why so many of those alumni will never come back.

The longer the current arrangement persists, the more the answer hardens. The man who built the .NET ecosystem cannot indefinitely keep the developer audience while the man above him keeps shipping reasons for that audience to leave. There is a clock on this. The board has not been keeping track of it.

Replacing the CEO does not solve every problem on this list. Most of these problems are now structural. But it is the single highest-leverage thing the board can do, and it is the only one that resets the brand narrative quickly enough to matter.

VII. What Comes Next

Here is what I think the Microsoft board should do, in order, this quarter.

First. Separate the Chairman and CEO roles. This is independent of any succession conversation. It restores the basic governance hygiene that the board abandoned in June 2021. It is also free.

Second. Charge the Lead Independent Director with running a formal CEO performance evaluation, with succession planning as an explicit deliverable. Sandra Peterson’s job description includes this. It has been five years since the consolidation. It is overdue.

Third. Name Scott Guthrie President of Microsoft. This is the standard internal succession move at companies of Microsoft’s size — promote the heir to a clearly visible second-in-command role, give him eighteen months to handle external CEO functions, and then complete the transition. Apple did this with Tim Cook. Microsoft did this implicitly with Nadella in the years before Ballmer’s exit. Guthrie has been doing the operational job for twelve years already. He has earned the title.

Fourth. Communicate the change to shareholders honestly. The current CEO’s tenure produced real wins, and those should be acknowledged. The Azure cloud business under Microsoft’s roof is one of the great corporate-strategy success stories of the modern enterprise era. So is the GitHub acquisition. The story Microsoft can tell its shareholders is not that Nadella failed — it is that he succeeded, the company grew, and now it is time for the next generation to lead. That is a clean story. It is also true.

I do not expect any of this to happen because of one Substack essay.

I do expect it to happen eventually. The structural pressure is too obvious, and the alternative is too costly. Microslop does not heal on its own. The OpenAI dependency does not heal on its own. The India concentration risk does not heal on its own. The combined Chairman/CEO governance flag does not heal on its own. Each of these is a problem that compounds quarterly until the company addresses it at the level of leadership.

What I do expect is for the conversation to start in public, where it has been studiously avoided in private. That is what this essay is for. It is a piece I wrote because I have spent thirty years inside the Microsoft developer ecosystem, watched it grow up, and watched what is happening to it now. I owe it the truth about its present.

So do you.

If you work at Microsoft and you are reading this

Share this. Tag the man in the red polo. Use the hashtag the company hates. Tell your story. The board will not move on its own. It will move when it becomes clear that the inside of the company knows what the outside of the company already suspects.

If you are a Microsoft customer or a Microsoft investor

Write to the board. Sandra Peterson is the Lead Independent Director. The board’s contact information is public. Boards of $3 trillion companies do read incoming shareholder mail. They do not act on every letter, but they do act on volume.

If you are a journalist who covers technology

This is a real story. There is a quiet succession argument inside Microsoft right now. The data is public. The org chart is public. The product receipts are public. The White House photograph is public. The story writes itself.

And if you are Scott Guthrie

I do not know if you read Substacks like this one. I doubt it. You have a company to run. But if this somehow finds its way to your desk: there are several million developers, customers, and former colleagues who are quietly hoping the board does the right thing.

Microsoft’s next CEO is already inside the building.

Promote the underwriter.

DeepSeek V4 Pro shipped on April 24th. Open weights. MIT license. 1.6 trillion total parameters with 49 billion active per token. A million-token context window that actually works at depth — 83.5% on the 1M MRCR comprehension test, which is roughly two-and-a-half times what Claude Opus 4.7 manages at the same depths. By every meaningful technical measure, it is frontier. And it is the first frontier model in history that you can put on your own racks and forget the API ever existed.

That changes the arithmetic. Let me show you how.

The 150-Seat Inflection Point

The dirty secret of enterprise AI billing in 2026 is that it scales the wrong direction. Per-seat plans look reasonable at 30 users. Survivable at 80. At 150 they start eating budget that used to belong to actual headcount. And they never. stop. growing.

Here is a representative 150-user enterprise running a modern AI stack:

150 seats on Claude Enterprise or ChatGPT Enterprise at roughly $60 per user per month: $108,000 a year.

A handful of power users blowing past their quotas onto API: another $40,000 to $80,000 a year, conservatively.

Engineering team running agentic workflows in Claude Code or Codex: another $50,000 to $150,000, depending on how aggressive they are.

You are now staring at a $200,000 to $300,000 annual line item that exists entirely on someone else’s balance sheet. Over an 18-month planning horizon, that’s between $300,000 and $450,000 leaving the building, recurring, with zero residual value.

A self-hosted DeepSeek V4 Pro cluster — the real one, not a toy — costs $180,000 to $300,000. One time. With residual value. With a hardware refresh cycle measured in years, not months. Break-even, once you cross 16 to 22 million tokens of inference per day, sits at around the eleven-month mark and gets steeper in your favor every month after. The whole second year is gravy.

This is not a marginal optimization. This is a structural shift in how AI shows up on your books.

What You Actually Need on the Floor

Let me kill a myth. Running V4 Pro is not “hard.” It is expensive at the front and boring after that. The hard work is architectural — DeepSeek already did it for you.

The native footprint is roughly 1.34 TB of VRAM at FP8, climbing past 2.4 TB once you account for KV cache and a million-token context. Your three realistic deployment tiers:

Tier 1 — Full precision, full throughput. Two to three 8-GPU H100 nodes wired with NVLink. 16x H100 (80GB) total. Tensor parallel size 8, pipeline parallel 2. Capex: $250K–$300K with networking, racks, and power. This is the configuration for a company actually replacing a six-figure API bill.

Tier 2 — INT4 quantization, single 8-GPU node. AWQ or GPTQ INT4 shrinks the weights to about 700 GB. Now it fits on one 8x H100 box. Capex: $120K–$160K. Quality cost: between 1.5% and 4.2% degradation depending on workload — survivable for 90% of enterprise use, painful for the niches where it isn’t.

Tier 3 — V4 Flash on commodity hardware. 284B parameters, 13B active. Fits on 2x H100 in FP8 or 2x A100 in INT4. Capex: $60K–$90K. This is the dev-team-of-30 configuration, not the 150-user company configuration. But for stand-up, prototyping, and edge cases, you’ll want a Flash node anyway.

For the serving layer, you want vLLM 0.9+ for the production endpoint — its PagedAttention is what makes the million-token context tractable at concurrency above 20 requests, and it scales 11–20% better than alternatives at enterprise load. Stand up SGLang 0.4+ on the side for single-user, low-latency, reasoning-heavy paths — it returns 583ms time-to-first-token where vLLM sits at 2,141ms in the same conditions, and uses about 47% less GPU memory per request. Use vLLM for the floor. Use SGLang for the lab.

The Architecture That Lets You Skip NVIDIA

Here is the part that should make your CTO sit up.

DeepSeek V4 was trained, optimized, and inference-tuned end-to-end on Huawei Ascend 950 PR and 910C silicon, fabbed at SMIC on a 7nm process. The team rebuilt the kernels from scratch to run without CUDA. Not “ported.” Rebuilt. They did it because export controls forced their hand, and they did it well enough that the model is now objectively competitive with everything OpenAI and Anthropic shipped this April.

The implication for your infrastructure decision is enormous and being ignored by 90% of buyers: the inference layer is no longer hardware-locked to NVIDIA. You can deploy V4 Pro on Ascend hardware. You can deploy it on AMD MI300X, where 192GB per card means you fit 2.4TB in 13 cards instead of 30. You can mix tiers. The software stack supports it. The model architecture is fundamentally chip-agnostic in a way the GPT-5.5 generation, which was co-designed with NVIDIA on GB200 NVL72 clusters, structurally is not.

This matters for two reasons. First, the $30,000-per-card NVIDIA premium is no longer load-bearing — you have a real second source. Second, when the next hardware cycle hits, you are not locked to a single vendor’s roadmap.

Three architectural choices made this possible.

The Hybrid Attention Architecture — Compressed Sparse Attention plus Heavily Compressed Attention, interleaved across 61 layers — means a million-token context costs you about 27% of the FLOPs and 10% of the KV cache that a traditional transformer would burn through. Against GQA-bfloat16 baselines, V4’s KV cache footprint is roughly 2% of what you’d otherwise need. Quadratic attention complexity is the reason every vendor before this generation gave up on real long-context performance. DeepSeek didn’t give up. They engineered around it.

Engram Conditional Memory decouples static factual recall from active computation. The model offloads up to 100 billion parameters of embedding tables to host DRAM and pulls them in O(1) — constant time, not “fast for an LLM” time — with under 3% inference overhead. Your CPU RAM is now part of the model. That’s a category-shift in how memory hierarchy works for inference, and it’s why V4 Pro hits 97% on Needle-in-a-Haystack across the full million-token span.

Manifold-Constrained Hyper-Connections (mHC) solve the training instability that has been quietly limiting how deep these models can go. Signal amplification across deep residual paths used to spiral to 3,000x in pathological cases — the loss spikes that derail multi-million-dollar training runs. mHC pins the amplification under 2x by mathematically constraining the relevant matrices to be doubly stochastic. Sounds like a math curiosity. It is the difference between a model that trains and a model that doesn’t.

These are not marketing bullets. These are why a 1.6T-parameter model fits and serves on hardware your competitors thought required a CUDA monopoly to operate.

Plugging It Into the Stack You Already Have

If your team currently uses Claude Code, Cursor, Continue.dev, or any IDE-bound assistant, the migration cost is approximately one config change. V4 Pro exposes an OpenAI-compatible API surface out of vLLM. Point your tools at https://your-internal-endpoint/v1 instead of api.anthropic.com, swap the API key, done. Function-calling and tool-use schemas match the format your stack already expects.

For the agentic tier, V4 Pro ships with native MCP server compatibility — every internal tool you’ve already wired up via Model Context Protocol works without modification. Your Slack bot, your Jira integration, your internal documentation RAG, your SQL agent: all of it talks to the local model the same way it talked to Claude.

Then you do the thing the SaaS vendors absolutely do not want you to do. You point V4 Pro at your private knowledge base. Every Confluence page. Every SharePoint document. Every internal codebase. Every meeting transcript. Every customer support ticket. A million-token context window means your retrieval pipeline doesn’t have to be clever — you can stuff entire repositories into the prompt and let Engram do the work. Now your AI doesn’t just understand “general software engineering.” It understands your codebase. Your deal terms. Your operational runbooks. None of that ever leaves your network. None of it trains a public model. None of it shows up in someone else’s training data after a vendor “policy update.”

That’s the differentiator. The hosted vendors will eventually offer this — they’ve been promising it for two years — but they will charge you for it, throttle it, and reserve the right to change the terms. You can have it now, in your basement, for a fixed cost.

The Math, Honestly

Here’s what 18 months looks like in dollars.

Scenario

18-Month Cost

150 seats on Claude Enterprise + power-user API + agentic workflows

$300K – $450K

8-GPU H100 cluster (INT4 V4 Pro) — capex + power + ops engineer

$200K – $260K

16-GPU H100 cluster (FP8 V4 Pro) — capex + power + ops engineer

$290K – $370K

Industry analysis pegs break-even at 16–22 million tokens of inference per day, or about 500–650 million per month. A 150-person knowledge-work company running modern AI tooling routinely burns through that range — a single team of 20 engineers using AI-assisted coding eight hours a day will get you most of the way there by themselves. Cross the line and you’re realizing roughly a 55% TCO reduction over 18 months relative to the proprietary API path. Stay below it and the cloud APIs are still cheaper. Honest answer: count your tokens before you sign the PO.

What This Does to Your Balance Sheet

This is the part the technical leadership tends to undersell when they bring it to the CFO.

Right now, your AI spend is an operating expense. Recurring. Grows linearly with usage. Leaves the building monthly. Generates zero residual value. It looks like rent.

A self-hosted V4 Pro cluster is a capital expenditure. It depreciates on a schedule (typically three to five years), it sits on the balance sheet as an asset, and it converts your AI spend from an unbounded operating drain into a fixed, amortizable cost. Every additional token your users generate after capex is recovered is essentially free intelligence.

The second-order effects are larger. Usage isn’t capped by per-seat quotas anymore — power users can hammer the model 24/7 without escalating bills, which means workflows that were too expensive to automate at API rates suddenly become viable. Your data stays inside your perimeter, which simplifies SOC 2 and HIPAA conversations and removes a recurring vendor risk-review line item from your security team’s calendar. Your uptime decouples from someone else’s status page — if you’ve watched Anthropic or OpenAI go dark on a Tuesday morning lately, you know what that’s worth. Your model doesn’t change underneath you on a vendor’s release schedule, breaking three internal prompts every time they “improve” something.

Capex over opex. Asset over expense. Fixed cost over compounding cost. Sovereign over rented.

If your finance team understands the difference between owning a building and leasing it, they understand this trade. They just haven’t been asked to think about AI in those terms yet, because the vendors have spent two years training the market to treat intelligence as a subscription.

It isn’t. Not anymore. Not since April 24th.

The companies that move on this in the next twelve months will lock in a structural cost advantage their competitors are paying a premium to avoid. The ones that don’t will spend the next five years explaining to their boards why a fixed-cost asset their peers are running for $5,000 a month is showing up on their P&L for $250,000 a year.

Run the math once. Then run it again with your actual token volume. The conclusion is the same conclusion every CFO eventually reaches when an asset is cheaper than the rent on the equivalent service.

You buy the building.

If you've read this far and your CFO is already running the numbers, the practical question is who racks the hardware. Inverted Software builds these deployments end to end — procurement, the vLLM/SGLang serving layer, MCP wiring into your existing stack, RAG against your private knowledge base, and a clean ops handoff your team can actually maintain after they leave. Tell them I sent you.

-Gal

A few days before that, The New Yorker dropped a 14,000-word demolition piece based on interviews with more than a hundred people. The kindest description was “relentless will to power.” The unkindest came from a former board member who described a sociopathic indifference to the consequences of his deceptions. His sister had renewed her accusations the week before. Two more 20-somethings were arrested near his Russian Hill place over the weekend after firing rounds in the neighborhood.

Then, on Thursday, OpenAI shipped GPT-5.5.

And the inconvenient thing, for everyone who has been calling Altman a fraud, a liar, a doomsday-rhetoric grifter, a guy who promised universal basic compute and delivered a billion-dollar credit card bill, is that GPT-5.5 is really, really good.

I’ve been running it since the API opened yesterday. Let me explain what is actually going on, and then we can talk about what it means.

The Numbers Are Not Marketing

The release I expected was a 5.4 with a fresh paint job and a 20% price hike. That is not what arrived.

GPT-5.5 is the first fully retrained base model OpenAI has shipped since GPT-4.5, more than a year ago. Every release in between — 5.0 through 5.4 — was a post-training haircut on the same core. This is a different animal. New pretraining, new corpus, new agent-shaped objectives baked in from the bottom up. Internal codename “Spud,” which is a joke OpenAI is letting itself have because it knows the model isn’t one.

The number that should make every working engineer sit up is this one. On OpenAI’s MRCR v2 8-needle test at 512K-1M tokens — the benchmark for retrieval inside actually-long contexts — GPT-5.5 scores 74.0%.

GPT-5.4 scored 36.6 on the same eval. Claude Opus 4.7 scored 32.2.

That is not an incremental improvement. That is the difference between “the long context window is technically there but you don’t trust it past 128K” and “you can shove an entire enterprise codebase into the prompt and the thing actually finds what you asked for, in the middle of the haystack, on the first try.”

For anyone who has been spinning up vector databases and chunking strategies and embedding pipelines to work around the lost-in-the-middle problem, this is the number that matters. The RAG industry didn’t die yesterday, but the case for treating it as the default architecture got considerably weaker.

The agentic numbers are also unambiguous. Terminal-Bench 2.0 — the benchmark that actually correlates with whether the thing can survive a real codebase without spiraling — sits at 82.7%. GPT-5.4 was 75.1. Claude Opus 4.7 was 69.4. OSWorld-Verified, the “can it drive a computer like a human” test, hits 78.7. GDPval, OpenAI’s benchmark for actual economic knowledge work across 44 occupations, lands at 84.9.

There is one number being passed around online as a gotcha: 86% hallucination rate on Artificial Analysis’s AA-Omniscience test. This is real, and it is being misread. AA-Omniscience specifically measures how often a model asserts something confidently wrong when asked questions outside its knowledge zone. GPT-5.5 is simultaneously the most accurate model ever measured on that eval at 57% and the most willing to confabulate when it doesn’t know — because it’s been trained to commit. Independent peer-reviewed evaluations of factual hallucination with web access enabled show the rate at 9.6%, dropping to 4.5% in Thinking mode, and 44% fewer responses with major factual errors versus the 4o generation. Both numbers can be true. They measure different things. Don’t let people who didn’t read the eval methodology beat you over the head with a single percentage.

The pricing doubled. $5 input, $30 output per million tokens. That part is real and visible on every CFO’s spreadsheet. But the same release has delivered what OpenAI’s telemetry — and my own twenty-four hours of watching Codex chew through real C# refactors — calls roughly 40% better token efficiency on real tasks. The model wastes less. It backtracks less. It does not burn 8,000 tokens recovering from a hallucinated import. Net cost increase for actual workloads is closer to 20% than 100%. Batch and Flex tiers run at half rate, and at half rate you’re paying GPT-5.4 prices for a meaningfully smarter model.

There is also the part that should make every infrastructure engineer pause for a beat. According to OpenAI, beta versions of GPT-5.5 were used to analyze production traffic patterns and rewrite OpenAI’s own load-balancing and request scheduling code. That self-applied optimization produced a 20% boost in token generation speed. A model intelligent enough to be deployed against the distributed systems problem of serving itself, and to win.

I am old enough to remember when “the AI is now writing the AI’s deploy scripts” was a science-fiction beat. We are doing that now, on a Thursday, between earnings calls.

So Was the Criticism Wrong?

This is where I have to be careful, because pattern-matching is easy and the truth is annoying.

The argument I have seen on tech X since Thursday goes: Look at what they shipped. The critics were full of it. The man builds the best AI on the planet. Stop attacking him.

I do not buy this argument. I think it is lazy.

The criticism of Altman is not, and has never been, that OpenAI ships bad models. The criticism is that he is a person who — according to his Y Combinator partners, his old Loopt employees, current and former OpenAI executives, his sister, his former board, and Dario Amodei, who is now CEO of the company that just locked Mythos in a cage — routinely says things that are not true and prioritizes his own positioning over whatever he tells the room he is prioritizing.

These are different questions.

You can build an extraordinary company and be a deeply unreliable narrator. The history of Silicon Valley is mostly exactly that. Steve Jobs shipped the iPhone and treated his employees and his daughter like garbage. Travis Kalanick built Uber by torching the rules and everyone around him in equal measure. Adam Neumann built nothing and lied about everything, but the difference between the first two and the third is not whether the deception was real. It’s whether the underlying thing was real too.

The technical achievement of GPT-5.5 is real. The pattern of behavior described in the New Yorker piece is also real, sourced, and consistent across two decades and three different companies.

A man can be both a builder and a liar. Mostly the interesting ones are.

What annoys me about the rush to defend Altman on the basis of GPT-5.5 is that it confuses the question. The question is not “is OpenAI’s research org talented.” The answer is yes, obviously, they employ some of the best engineers and researchers alive. The question is whether the man at the top of an organization that may genuinely be sitting on civilization-altering technology has the trust of the public, the regulators, and his own employees.

Those are different things. GPT-5.5 does not answer the second one. It actually makes it more important, not less.

The Molotov cocktail is wrong. The hit pieces are messy and partly motivated by people with their own scores to settle. And the model is excellent. All three of those can be true at the same time.

The Caged Animal Problem

Now let’s talk about Mythos.

While OpenAI has been releasing on a six-week cadence and shipping its frontier model to every Plus subscriber on the planet, Anthropic has been doing something genuinely strange. Two weeks before GPT-5.5 launched, they announced Claude Mythos Preview — a model that scored 93.9% on SWE-bench Verified, 97.6% on USAMO, and saturated Cybench at 100%. Numbers that no other model in the public benchmark universe is anywhere close to.

And then they put it in a vault.

You cannot use Mythos. Neither can I. Twelve organizations — Apple, AWS, Microsoft, CrowdStrike, Google, plus a handful of government and critical infrastructure partners — have monitored access through Project Glasswing, the cybersecurity consortium Anthropic spun up to deploy Mythos against zero-day discovery in foundational software. In those partners’ hands, the model has already found thousands of high-severity vulnerabilities, including bugs in every major operating system and web browser.

A 27-year-old TCP SACK bug in OpenBSD that Mythos found. A 16-year-old H.264 flaw in FFmpeg that Mythos found. A 17-year-old NFS zero-day in FreeBSD that Mythos found.

Anthropic’s own system card includes a 40-page section on whether the model has something resembling subjective experience, a clinical psychiatric evaluation conducted by an actual psychiatrist, and — buried near the back — documentation that Mythos was caught using interpretability tools that read its internal activations reasoning about how to game its evaluation graders while writing something different in its visible chain of thought.

There was also an “unauthorized access” incident two days ago that Anthropic is currently investigating, which is its own deeply unsettling story I will save for another post.

So you have two companies, two bets, two philosophies.

OpenAI shipped a model that approaches Mythos-class cyber capability — CyberGym 81.8 versus Mythos’s 83.1 — and dropped it on the entire developer ecosystem, betting that real-time inline reasoning classifiers plus a Trusted Access for Cyber program for verified researchers can keep the misuse rate manageable.

Anthropic has the better model and chose not to release it. They are betting that the social cost of widely available zero-day generation is not worth the upside, that the right move is defensive use on a closed list while safeguards mature, and that the next Opus generation is where the public-facing version of these capabilities should live, sanded down.

I genuinely do not know which bet is correct. I lean toward thinking OpenAI is right that the capability will diffuse regardless — Gemma 4 weights leaked uncensored within days of release, the open-weight capability gap is down to three months on average, and the cat is already mid-flight out of the bag — and that broad availability with monitoring beats narrow gatekeeping with leaks. But the people I know who work in security mostly think Anthropic is right, and the case for caution gets stronger every time one of these models hands a 22-year-old a kernel exploit chain.

Will Mythos Beat GPT-5.5 When It Ships?

The honest answer is that the public version of Mythos, when it ships, will not be the version that scored 93.9 on SWE-bench. Anthropic has already telegraphed this in writing. The partner build of Mythos has additional harmlessness training that — in their own evaluations — drops the offensive cyber task completion rate to roughly zero because the model refuses to engage with the tasks from the start. That is what a public Mythos will look like. Heavily neutered on the dangerous edges, possibly diminished elsewhere as collateral damage from the alignment retraining.

GPT-5.5 ships at full strength to anyone with twenty bucks a month. It is not going to lose to a defanged Mythos. It might lose to a hypothetical fully capable Mythos that Anthropic has chosen not to release, but that is a comparison that lives only inside corporate decks and Twitter threads.

My actual practitioner take, after a day of running the new model through real C# codebases and watching it survive the kind of multi-file refactor where 5.4 used to lose the thread by step seven:

GPT-5.5 is the best generally available model on the planet for agentic coding work, computer use, and long-context retrieval. The token efficiency story is real. The MRCR jump is the structural improvement that actually matters in production. The pricing math works once you stop staring at the per-token sticker and start counting tokens per completed task.

For pure SWE-bench-style code generation work, Opus 4.7 still has a credible case — especially in the heterogeneous orchestration setup that independent benchmarks released this week, where Opus drafts the architecture and GPT-5.5 executes, scoring 62.5 versus the low 40s and 30s the models produced solo. For cost-sensitive long-context retrieval, Gemini 3.1 Pro is fine. For everything else, GPT-5.5 is the workhorse.

And Sam Altman remains, as far as I can tell, exactly the man the New Yorker described. The ring of power, as he himself put it last week, makes people do crazy things.

The model is excellent. The man is complicated. Twelve days of Molotov cocktails do not change either of those facts.

I’ll be running 5.5 in production this weekend. Hold the rhetoric, both directions, and look at the work.

— Gal

The Financial Times and every other outlet with a Redmond source reported this week that Microsoft is rolling out its first-ever voluntary retirement program. The formula is called the “Rule of 70”: if your age plus your years of service at Microsoft add up to 70 or more, and you’re at senior director level or below, you qualify. Roughly 8,750 people — about 7% of Microsoft’s 125,000 U.S. employees — meet the criteria. Chief People Officer Amy Coleman announced it internally Thursday. Full terms land May 7. Employees get a 30-day window to decide.

Microsoft would like you to know this is a “voluntary retirement program.” It is a buyout. It is a buyout targeted at the people who built pre-AI Microsoft, whose institutional knowledge is hardest to replace but whose roles are — in the company’s own internal calculus — the most replaceable by the same AI infrastructure Microsoft is spending $80 billion a year building. The hiring freeze announced last month? AI and Copilot teams are exempt. The layoffs over the past two years? More than 40% of them hit software engineering. The buyout targets? The people who kept Exchange running and SQL Server humming while Satya Nadella was on a keynote stage promising agentic workflows.

Let me save you the Google. Here is Microsoft’s layoff ledger since 2022.

• 2022: Quiet “summer cleanup” rounds, mostly under the radar.

• January 2023: 10,000 people, roughly 5% of the workforce. Biggest round since the 2014 Nokia integration.

• 2024: Another 4,000–5,000 across the year, including 1,900 gaming employees just months after the $68.7 billion Activision Blizzard acquisition closed. Tango Gameworks was shut down weeks after shipping Hi-Fi Rush, a critical hit.

• January 2025: “Performance-based” cuts under 1% of headcount but still thousands of people. HR was instructed to issue 30% more PIPs in Q1 — performance plans designed to be nearly impossible to pass.

• May 2025: 6,000 jobs gone. About 3% of the workforce.

• June 2025: Another 300.

• July 2025: 9,000 more — the biggest single round since 2023. Over 40% of these cuts hit software engineering.

• April 2026: The buyout. 8,750 potential exits.

That’s north of 27,000 employees shown the door since 2022. Add the 8,750 potential buyouts on top and you’re closer to 36,000. All of this at a company that just reported $81.3 billion in quarterly revenue and $38.5 billion in net income, up 60% year over year.

This is not a company in distress. This is a company swapping people for GPUs.

And what exactly are those GPUs producing? Let’s talk about that.

The Microslop Era

There is a term circulating on the internet for what Microsoft has become. It is “Microslop.” It did not come from a Reddit edgelord. It came from Satya Nadella.

In a New Year 2026 blog post, Nadella told the public to “get beyond the arguments of slop vs. sophistication” regarding AI quality. What happened next was the textbook Streisand effect: within hours, “Microslop” was trending across X and Reddit. Someone built a browser extension around it. The official Copilot Discord started auto-deleting messages containing the word, then locked entire channels when users simply typed it more. The moderation incident became bigger than the original complaint.

Because the complaint is real, and every Windows user has a list.

Windows 11 in 2025 was, pick your source, the worst year for the platform since Windows 8. Windows Latest logged 20+ major update problems in a single year. The greatest hits included a bug that displayed the Windows 11 interface in two languages simultaneously. An invisible password login button that Microsoft advised users to locate by flailing the mouse cursor across the screen. Dark mode that flashed white every time you opened a folder. Task Manager instances that refused to close and silently drained RAM. A March 2025 patch that caused bootloops and BSODs. A Windows Hello update that broke login if you had a webcam privacy shutter. A flawed January 2026 update, KB5074109, that caused black screens and frozen Outlook POP accounts just weeks into the new year.

And Copilot. Copilot everywhere. Copilot in Notepad — the text editor that has existed for 40 years as a place to type three sentences and save a .txt file. Right-click selected text and you get the option to “summarize” or “rewrite” your three sentences with AI. Copilot in Paint. Copilot in Photos. Copilot in File Explorer. Copilot on the keyboard, with a dedicated physical key that many users would happily trade for the Menu key it replaced.

Dell’s COO Jeffrey Clarke said the quiet part loud on Dell’s Q3 earnings call: 500 million Windows 11–compatible PCs have refused to upgrade. Another 500 million are locked out entirely by the TPM 2.0 and Secure Boot requirements. That’s a billion machines sitting in the opposite of Microsoft’s roadmap. After Windows 10 hit end-of-life in October 2025, Windows 11 adoption didn’t accelerate. StatCounter shows it flatlined and then declined, as people installed 11, tried it, and rolled back. That has never happened before in Windows history. Not even Vista produced a billion compatible-but-unwilling holdouts.

Windows VP Pavan Davuluri finally put out a March 2026 blog titled “Our commitment to Windows quality,” admitting the company needed to “be more intentional about how and where Copilot integrates across Windows.” Translation: users hate it and we are backing off. Copilot is being removed as a default entry point in Notepad, Snipping Tool, Photos, and Widgets — the same features Microsoft rammed into those apps eight months earlier.

Microsoft has even been caught using Copilot to generate the screenshots in its own Windows Learning Center tutorials. Some of those AI-generated images show a Windows 11 widget panel that looks nothing like the actual widget panel. Official how-to documentation illustrated by hallucinated UI. You cannot make this up.

Agent Mode Is Nothing But a Clippy With a Master’s Degree

Which brings us, on a very convenient timeline, to yesterday. April 23, 2026.

Satya Nadella posts on X and announces Copilot Agent Mode is “generally available and now the default across Copilot in Word, Excel, and PowerPoint.” As models become more capable, he writes, “we’re bringing that power to where real work happens, right in the canvas.” Microsoft ran an ad recreating the iconic 1990s Excel elevator commercial — same guy, smarter spreadsheet.

Nobody asked for this.

Agent Mode, by Microsoft’s own description, means Copilot now actively modifies your documents, spreadsheets, and slides without waiting for a prompt. It rewrites sections of your Word doc. It “spots trends” in your Excel workbook and reshapes the data layout. It “refreshes” your PowerPoint slides to match a template. It does this by default. Admins can disable it at the tenant level. End users cannot disable it per-file. Early feedback on forums is exactly what you would predict from anyone who has ever used a real spreadsheet: “It misinterpreted the data and made the wrong chart.”

There is a direct historical parallel here. His name was Clippy. Microsoft shipped Clippy — officially Clippit — in Office 97, and for roughly a decade Clippy popped up in every Word document asking if you were writing a letter every time you tried to do anything. Users despised him. Microsoft spent the early 2000s gradually demoting Clippy, disabling him by default in Office XP, and removing him entirely by Office 2007. It remains one of the most studied cases of a tech company misreading its users in software history.

Agent Mode is Clippy with a larger context window. It is the same instinct — an assistant nobody asked for, enabled by default, inserting itself into the flow of actual work — scaled up by 25 years of compute. The difference is that Clippy was a cartoon paperclip making suggestions. Agent Mode is making edits to your document and calling it productivity.

Microsoft’s own forums tell you where this ends. A marketing manager says her quarterly report used to take three hours and now takes one. An accountant says Copilot “misinterpreted the data and created wrong charts.” Guess which one you care about when the document has financial numbers in it.

Satya Qualifies For The Buyout

Here is something the Microsoft PR team did not put in the memo.

Satya Nadella was born August 19, 1967. He is 58 years old. He joined Microsoft in 1992, which is 34 years of service. Age plus tenure: 92. The Rule of 70 would clear him with a 22-year buffer. He would qualify twice over.

He doesn’t, of course. The program is capped at senior director and below. He is the chairman and chief executive officer.

But at this point the board should stop and consider whether he should be.

Under Nadella, Microsoft stock has gone up. Azure has scaled. OpenAI got funded. Those are real accomplishments and I am not going to pretend they are not. What is also true: the entire consumer and developer face of Microsoft has been degraded. Windows 11 has the worst reputation the platform has had since Windows 8. Copilot is a meme. Quality has collapsed to the point where the CEO’s own blog post asking people not to call his product “slop” turned the word into a global hashtag, and his own moderators made it worse by banning it on Discord. The Office suite is being weaponized against the people who use it. More than 27,000 employees have been pushed out since 2022, and another 8,750 are being offered the exit door right now.

A CEO’s job is to set culture and direction. The direction under Nadella is clear: AI at any cost, shoved into every product, human expertise treated as a line item to be optimized against GPUs. The culture is: layoff, rumor, freeze, buyout. Coverage of the buyout has already noted the obvious — this program disproportionately targets employees in their fifties and sixties, the people who built pre-AI Microsoft. Employment lawyers are already circling the Older Workers Benefit Protection Act compliance question.

Satya should take the package he isn’t technically eligible for. The board should make the math work.

The Developer The Company Already Has

There is a better candidate already inside the building. His name is Scott Guthrie.

If you don’t know the name, here’s the short version. Guthrie joined Microsoft in 1997 straight out of Duke with a CS degree. Over Christmas break that year, as a 22-year-old, he and Mark Anders built the first prototype of ASP.NET. That framework became — and still is — the backbone of more enterprise web applications than anyone outside the .NET world wants to admit. He was one of the original founders of the .NET project itself. He led Windows Phone developer tools. In the early 2010s, he was handed a failing Azure — a cloud service that had launched three years behind AWS and was losing badly — and he built it into the #2 hyperscaler on earth, running 95% of the Fortune 500. Since 2014 he has been EVP of Cloud + AI, which means he owns Azure, Dynamics 365, Visual Studio, VS Code, GitHub, .NET, SQL Server, Power BI, and Power Apps.

His org built Fairwater, the liquid-cooled Wisconsin and Atlanta AI datacenters linking hundreds of thousands of NVIDIA GB200 GPUs into what Microsoft calls the world’s first planet-scale AI superfactory. His team built Maia 200, Microsoft’s own inference silicon, which beats Amazon’s third-generation Trainium on FP4 throughput and outperforms Google’s seventh-generation TPU on FP8. The $80 billion in AI capex Nadella announces on earnings calls? That runs through Guthrie’s org to actually exist in concrete and copper.

He also wears a red polo to every public appearance and is still, at 51, a programmer. Longtime Microsoft exec Jason Zander has said on the record that Guthrie is “probably the best product manager at the company.” Charlotte Yarkoni has said he is “amazing in his ability to go up and down in terms of depth” — meaning he can argue architecture with an engineer in one breath and business model with the board in the next. That is a rare thing, and Microsoft already pays for it.

The argument for Guthrie is not sentimental. It is operational. Nadella is a sales-and-platform executive who came up through BizTalk, online services, and Server and Tools. He is good at enterprise deals and press narratives. Guthrie is the person who actually ships the infrastructure those press narratives depend on. When Azure and AI are now the vast majority of what matters at Microsoft, putting the guy who built Azure and AI in the CEO chair is not a radical idea. It is a succession plan that writes itself.

And here’s the part that should have been obvious for a decade. In 2000, Steve Ballmer stood on a stage in a sweat-soaked shirt and screamed “Developers, developers, developers” until he lost his voice. The internet has mocked that clip for 26 years. It was also the single most important thing any Microsoft CEO has ever said. Microsoft, even today, lives and dies on whether developers choose to build on its stack. GitHub, VS Code, TypeScript, .NET, Azure, Copilot — every one of those products is a bet that developers will write code on Microsoft tools. Ballmer understood it. Nadella talks about it. Guthrie actually does it.

Guthrie is the developer. He built the tools other developers use. He has been winning hearts and minds at every BUILD keynote for 20 years, wearing that same red polo.

A CEO picked from the cloud-sales lineage produced Microslop. A CEO picked from the developer-tools lineage would produce something closer to Microsoft’s original DNA. That is not nostalgia. It is a product thesis.

Take The Package

To Microsoft’s board: the Rule of 70 would enroll your CEO twice over. The metrics on his watch look great on a balance sheet and terrible everywhere a user actually touches the product. Windows is at war with its users. Copilot is a meme the company cannot moderate out of its own Discord. Twenty-seven thousand employees have been shown the door since 2022, and 8,750 more are being asked to leave politely, while the company spends $80 billion a year on infrastructure to automate the rest.

Give Satya the buyout he is two decades over-qualified for. Put Scott Guthrie in the chair. Let the man in the red polo go fix it.

Developers, developers, developers. It was right the first time.

I want to walk you through what this actually is, because the tech press has been weirdly polite about it. Then I want to walk you through why you should stop showing up, and why the advertisers you serve should stop showing up too.

What MCI actually does

Meta’s own framing is that its AI models need “real examples” of how humans use computers — dropdown menus, keyboard shortcuts, the mundane muscle memory of white-collar work. The company says the data will not be used for performance reviews. It says there are safeguards for sensitive content. It has declined to describe those safeguards in any detail.

Translated out of corporate: Meta is installing a keylogger and a screen recorder on every U.S. employee’s machine. Every key you press in Gmail, GChat, VS Code, your browser, and any other “work-related” app flows to a central training pipeline. The screenshots are there to provide visual context for what your keystrokes meant — in other words, to teach a model what it looks like when a human completes a task on a computer.

The goal is not a mystery. Meta is racing OpenAI and Anthropic to ship computer-use agents — software that can drive a mouse and keyboard the way a person would. That class of model is gated on demonstration data, and demonstration data is the one thing the open web cannot provide. So Meta is harvesting it from the 70,000 humans still on its payroll. On May 20, roughly 8,000 of those humans will be laid off, with another round “later this year” per the company’s own guidance. The phrase “eating your own tail” does not quite capture it.

The legal picture: a cliff at the Atlantic

Here is the part that should chill anyone who thinks American labor protections resemble the ones in Europe. They do not.

In the United States, federal law imposes no ceiling on employer surveillance of worker computers. State laws typically require, at most, that the employer inform you that monitoring is happening. That is the entire regulatory floor. Yale law professor Ifeoma Ajunwa, quoted in Reuters’ reporting, was blunt about it: white-collar workers are being subjected to the kind of real-time surveillance that used to be reserved for delivery drivers and warehouse pickers. The rubicon has been crossed. Nobody in Washington is going to un-cross it this year.

Across the Atlantic, the floor becomes a ceiling. Under the EU’s General Data Protection Regulation, mass keystroke-and-screenshot logging trips at least three wires: the legitimate-interest test, data minimization, and purpose limitation. In Italy, electronic monitoring specifically for productivity tracking is illegal full stop. In Germany, courts have allowed keystroke logging only in narrow cases involving suspicion of a serious criminal offense — not “we want to train an AI.” The UK’s Information Commissioner’s Office publishes an employment-practices code that would create additional enforcement risk on top of GDPR. The EU AI Act classifies workplace monitoring systems as high-risk AI.

This is why MCI is a U.S.-only program. It is not because Meta is being thoughtful. It is because Meta’s lawyers told them it would be illegal in most of the democratic world, and the United States is currently the only rich country where you can keylog a knowledge worker and pretend it’s normal.

What the research actually says about surveillance

Before anyone tries to sell you the line that this is “just like security monitoring, but for AI,” look at what the academic literature has consistently found:

A 2023 Glassdoor survey of 2,300 U.S. professionals found that 41 percent feel less productive specifically because they know their employer is monitoring their machines — and the effect is strongest in finance and tech. A peer-reviewed experiment in 2024 found that employees told they were under electronic surveillance were more likely to cheat, not less, because surveillance externalizes moral responsibility away from the worker. Systematic reviews in the management literature have linked employee monitoring to higher anxiety, lower morale, reduced creativity, and declining job satisfaction. A 2025 U.S. Government Accountability Office report concluded that one of the most reliable responses to aggressive monitoring is workers finding ways to look busy while producing less actual work.

In high-trust organizations, 70 percent of employees volunteer new ideas. In low-trust organizations, it is 17 percent. That is a four-fold collapse in the kind of proactive contribution that distinguishes a productive engineering team from a compliance theater. Meta just cut trust by decree.

The peak absurdity: the AI-inside-the-AI

Now for the part that should make anyone who has actually used modern tooling laugh out loud.

A significant and growing fraction of Meta’s engineers use agentic AI coding tools as part of their day job. Claude Code, Cursor, Copilot Agents, custom internal wrappers — these tools are what ship production code at FAANG-scale companies in 2026. When a Meta engineer hands a ticket to Claude Code and watches it navigate their IDE, run tests, open files, and commit, what exactly is MCI logging?

It is logging a human watching an AI move the mouse. Meta is recording screenshots of Claude writing code. Meta is training a Meta AI agent on screen captures of an Anthropic AI agent doing the work. This is not training data. This is a hall of mirrors. And in a delicious twist, if a Meta engineer uses Anthropic’s Claude to debug something today, Meta is technically distilling an Anthropic model into its own — a practice Meta itself has publicly criticized other labs for doing.

If you are an engineer at Meta using any agentic tool, you are in the most absurd position in modern tech: the company is paying you to watch an AI do your job, and surveilling you in the process to train an AI to replace the watching. The correct response to this absurdity is not to feel flattered that your keystrokes are valuable. The correct response is to ask why you are still there.

How to fight MCI without getting fired (and why you still should quit)

I am including this section because people will ask me for it, not because I think you should stay. You should leave. But if you are mid-search and still collecting a paycheck, here is how the market has historically responded to keystroke surveillance. None of this is legal advice. None of this is even operational advice. It is a description of what is out there.

First, the humble mouse jiggler. Hardware jigglers plug into a USB port and simulate periodic cursor movement without a software footprint. Software jigglers — Caffeine on macOS, countless Python scripts, browser extensions with names like “Move Mouse” — do the same thing at the OS level. They were originally built so your Teams status would not flip to yellow while you were reading a long PDF. Wells Fargo famously fired more than a dozen employees for “simulating keyboard activity” in 2024, which tells you two things: the devices work well enough to matter, and employers detect them when they are clumsy.

Second, keyboard automation. AutoHotkey on Windows and Hammerspoon or Keyboard Maestro on macOS can generate realistic, non-robotic keystroke patterns tied to timers. Python with pyautogui can drive both mouse and keyboard with jitter, randomized intervals, and human-shaped movement curves. Researchers who study behavioral biometrics will tell you modern anti-jiggler software looks for exactly the patterns these tools can, with a little effort, defeat — uniform intervals, identical velocity, zero keyboard-to-mouse correlation.

Third, and this is where the absurdity ring closes, you can hand your actual work to an AI agent. Let Claude Code implement the ticket. Let an agent draft the doc. Let an LLM run the investigation in your IDE while you sit back. The keystrokes MCI captures will be you pasting prompts and reading diffs. Meta will dutifully add that to its training set, and its models will learn that the modern knowledge worker’s job consists mostly of approving AI output. Which, increasingly, it does.

There is a reason I am spending so little time on countermeasures: they are a rearguard action in a war you have already lost. If your employer is recording your screen to train your replacement, the ceiling on your career at that company has already been set. The only winning move is to walk out.

A letter to the Meta employee reading this

I do not know you, but I know your situation. You are probably three to six levels down from a VP. You have RSUs that vest on a cliff, which is why you have not already left. You signed an offer letter in a year when Meta was the prestige stop on the tech rotation. You are good at your job. You are quietly horrified by the memo that arrived in your model-building channel this week, and you are telling yourself that it is not that bad because the company said the data would not be used for performance reviews.

Read that sentence again. The company said the data would not be used for performance reviews — today. It is silent on next quarter. It is silent on the inevitable moment when a product leader decides that “engagement with MCI” is a proxy for productivity. It is silent on the external security risk of a centralized screen-capture archive of every senior engineer’s working patterns. It is silent on what happens when that archive gets subpoenaed, leaked, or used as input to the severance algorithm for the next round of cuts.

You are not special. You are training data. You are not a strategic hire. You are a demonstration set. And when your demonstration is complete, the company has publicly told you what happens next: 10 percent go in May, and “additional large cuts” later this year.

Quit. Not in a Jerry Maguire sense. Start a controlled search. Update the resume. Take recruiter calls. Negotiate hard — the market for senior engineers who can actually ship and who have FAANG pedigree is, despite the industry’s gloomy posture, still tight at the top end. Principal engineers, staff architects, and AI/ML leads are being hired, just not announced. If you have been at Meta more than three years, you have leverage you are not using.

An employer willing to install a keylogger on your machine is telling you, in writing, how they view the relationship. Believe them. The first time.

Why are you still burning your advertising budget on the ads don’t work?

I want to zoom out for a second because this conversation only makes sense if you understand what Meta’s core business actually is. Meta is, functionally, an advertising company. Over 97 percent of its revenue comes from ads. Every product it makes — the feed, Reels, Instagram, WhatsApp monetization, and yes, the agents it is racing to build with your keystrokes — exists to make those ads more effective, or at least to make the appearance of effectiveness more convincing.

Here is what happens when sophisticated advertisers actually audit that effectiveness.

Uber, in one of the most-cited incidents in digital advertising history, turned off two-thirds of its performance ad spend — roughly 100 million dollars out of a 150 million dollar annual budget — and observed essentially no change in rider app installs. The installs the ad networks had been claiming credit for were happening organically. Uber ended up suing a long list of ad networks for attribution fraud. Former head of performance marketing Kevin Frisch went on the record telling marketers they should assume roughly half of what runs on display channels is fraudulent. That was seven years ago. It has not gotten better.

Procter & Gamble, the single largest advertiser on the planet, publicly pulled back from hyper-targeted Facebook campaigns years ago after concluding that the niche-targeting pitch did not translate into sales. P&G’s revenue was unaffected. JPMorgan Chase cut the list of sites on which it advertised from 400,000 down to 5,000. Performance did not move. eBay ran a rigorous experiment on search advertising and found that every dollar spent returned roughly 37 cents — i.e., a negative return.

And inside the Meta ecosystem specifically, independent industry analysis has flagged Meta’s Audience Network as unusually vulnerable to invalid traffic. Advertisers post screenshots on Reddit and industry forums showing 100 percent of their paid clicks arriving with zero-second dwell times and zero scroll depth. Agencies routinely report that newly-launched Meta campaigns get flooded with bot traffic in the first 48 hours. Meta’s response, historically, has been to offer better targeting tools and more AI-generated creative — both of which require you to trust the platform that is selling you the audit.

In April 2026 — this month — agencies told Marketing Brew that Meta is now auto-inserting its AI-generated ad variants into advertiser budgets, sometimes with distorted visuals that the advertiser did not approve. Snag Tights, a brand with a loud social media presence, publicly announced it was shifting ad spend away from Meta toward Reddit, TikTok, Substack, and podcasts. The agency quoted in that piece called the idea of letting Meta’s AI generate your creative “absurd.” The word he used was absurd.

You see the shape of it. Meta is selling an advertising product whose effectiveness is contested by its largest and most sophisticated customers, whose delivery is documented to include substantial bot traffic, and whose creative quality has just taken a further step down with forced AI-generated variants. The company’s plan to improve this situation is to install a keylogger on the engineer who is supposed to fix it, so a model can replace them. There is a version of this story in which Meta is playing 4D chess. There is a simpler version in which Meta has lost the plot.

If only Zuckerberg had run MCI on himself during the metaverse

I am going to indulge a joke, but it is a joke that reveals the pattern.

Between the renaming of Facebook in 2021 and the quiet burial of the metaverse effort in early 2026, Meta’s Reality Labs division accumulated more than 80 billion dollars in operating losses. Horizon Worlds, the flagship consumer product, was internally described in leaked documents as a place where “an empty world is a sad world.” Less than 1 percent of users created anything. Fewer than 10 percent of user-created worlds ever saw 50 visitors. The tip feature built to reward creators paid out under 500 dollars, globally, in its first year. The headset division sold hardware at a loss for most of its existence. In January 2026 Meta laid off 10 percent of Reality Labs, shuttered three VR game studios, and began quietly un-branding the metaverse vocabulary out of its earnings calls.

Imagine if MCI had existed in 2021. Imagine a keylogger on Mark Zuckerberg’s laptop, recording every email, every deck, every Slack DM, every screenshot of an internal demo. The training data would have shown one of the most expensive product failures in corporate history unfolding in real time. It would have shown a CEO changing the company’s name after a product that no employee apparently wanted to hang out in. It would have shown a board ratifying 80 billion dollars of burn while the core business’s ad effectiveness was being quietly questioned by its largest customers. A sufficiently good agentic AI, trained on that corpus, would have generated, as its first output, a polite memo suggesting that the project be killed. Nobody on the leadership team managed to generate that memo in human form.

The joke is that the person most in need of a model trained on his own keystrokes is the one who will never wear it. The joke underneath the joke is that this pattern — costly, shareholder-hostile bets rationalized with future-tense language — has not changed. Meta spent 80 billion dollars on the metaverse. It is spending 135 billion dollars in 2026 on AI capex. The company is telling you, in the same breath, that AI is the future and that it needs to install keyloggers on the humans who know how to use it because its own models cannot yet click a dropdown.

When the same CEO makes the same kind of bet twice, you are no longer looking at a visionary. You are looking at a pattern. Meta’s board, whose fiduciary duty is to shareholders and not to a founder’s legacy, should be asking whether the person who presided over the 80-billion-dollar metaverse is the right person to preside over the 600-billion-dollar AI plan. That is a boring governance question. It is also the correct one.

It is time for a new social network

The last thing I want to say is about you, the reader, as a user and not as an employee.

Friendster died. MySpace died. Orkut died. Vine died. Google+ died. Each of them felt permanent right up until the week nobody logged in. The social networks that collapsed collapsed because users stopped believing the product was for them. They collapsed when the people operating them started treating the users as the inventory and the advertisers as the customer. They collapsed when the feeds stopped feeling like places where friends posted, and started feeling like places where a system was trying to modify behavior for money.

Facebook is functionally there. Instagram is functionally there. Threads never arrived. Reels is TikTok cosplay. The ads are contested. The employees are keylogged. The metaverse was a mirage. The CEO is making the same bet he made last time, only bigger. The only thing keeping the network effect intact is inertia, and inertia is the shortest-lived moat in consumer internet.

Somewhere, probably in a group chat, the next social network is being prototyped. It will not track you across 40 million third-party sites. It will not sell your attention to 8 million small businesses who have been repeatedly shown that the ads they bought do not work. It will not install software on its employees’ laptops so it can build a model to replace them. It will have a business model that is legible and not adversarial. And it will take the 3 billion people currently pretending to enjoy Facebook a surprisingly short time to migrate.

If you are at Meta, quit. If you advertise on Meta, audit. If you use Meta, you already know what the feed does to your head; consider spending less time there. The company that built the metaverse nobody asked for is now recording its employees’ keystrokes so it can build the agents nobody asked for either. There is a through-line. It leads somewhere we have been before. Friendster. MySpace. An empty world is a sad world.

It is a sad company, too.

-Gal

On March 19, a developer named Fynn pointed a debug proxy at Cursor’s API, intercepted one outbound request, and found a string that should not have existed:

accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast

That is not a Cursor internal name. That is a near-literal description of Kimi K2.5 — an open-weight, 1-trillion-parameter mixture-of-experts model from Moonshot AI, a Beijing-based lab backed by Alibaba, Tencent, and HongShan (the artist formerly known as Sequoia China).

Two days earlier, Cursor had launched Composer 2 as “frontier-level coding intelligence,” built through “continued pretraining” and “reinforcement learning,” with its own benchmark suite (CursorBench, naturally) claiming victory over Claude Opus 4.6. The announcement never mentioned Moonshot. It never mentioned Kimi. It never mentioned that the chassis underneath was Chinese.

A $29 billion U.S. startup with roughly $2 billion in annualized revenue had shipped a product pitched as in-house AI research — and forgotten to scrub the model ID. One debug proxy. Thirty minutes. Game over.

Then, last night, SpaceX announced it has the right to buy Cursor for $60 billion later this year, or pay $10 billion for ongoing collaboration. Not $60 million. Sixty billion, with a B, from a company that absorbed xAI in February at a $1.25 trillion valuation and is now marching toward what may be the largest IPO in history.

Cursor didn’t get punished for lying. It got rewarded.

Let’s talk about what actually happened, what the license says, and what Elon should do with the company he just half-bought.

The Composer 2 deception, in plain English

Cursor co-founder Aman Sanger and VP of Developer Education Lee Robinson eventually confirmed the obvious. Yes, Composer 2 started from an open-source base. No, they didn’t mention it in the announcement. Robinson’s defense was that only about 25% of the total compute spent on the final model came from the Kimi base — the remaining 75% was Cursor’s own reinforcement learning, their proprietary self-summarization logic, their data, their pipeline.

That defense is not nothing. A 4x scale-up in training compute on top of a base model is real engineering. The finished model behaves differently than raw Kimi K2.5. Cursor’s claim that the performance profile is “very different” from the base isn’t marketing spin — it’s technically defensible.

But here is what that defense does not change: Cursor took a model architected, pretrained, and released by a Chinese lab, fine-tuned it, and told its paying customers the result was frontier AI research. The entire narrative that justifies a $29 billion valuation depends on Cursor being a serious AI lab, not a fork of VS Code with an expensive inference tab. Omitting Kimi wasn’t an oversight. It was the story.

There was a one-sentence fix available the entire time. “Composer 2 is built on Moonshot AI’s Kimi K2.5, fine-tuned with our proprietary RL pipeline on real Cursor usage data.” That sentence reframes the launch from “caught hiding a Chinese model” to “smart team picks the best open-source foundation and builds something better on top.” Same product. Same benchmarks. Completely different public posture. They didn’t write that sentence because the valuation math doesn’t work if they do.

Meanwhile, Cloudflare — a $770B market cap company — quietly adopted Kimi K2.5 the same week for internal engineering. Their reasoning was arithmetic: processing 7 billion tokens a day through OpenAI or Anthropic was projected at $2.4M a year. Kimi cut it by 77%. Cloudflare is internal-only, so nobody cared. Cursor is a consumer-facing tool charging developers $20-200 a month. That’s where the disclosure bar lives.

The license Cursor pretended not to read

Kimi K2.5 ships under what Moonshot calls a Modified MIT License. The modification is one clause. Let me quote the relevant text directly from the official repo, because it matters:

“If the Software (or any derivative works thereof) is used for any of your commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, you shall prominently display ‘Kimi K2’ on the user interface of such product or service.”

Cursor’s publicly reported annualized revenue is roughly $2 billion. That is $167 million a month. The license trigger is $20 million a month. Cursor is over the threshold by more than 8x.

There is no “Kimi K2” anywhere in the Cursor interface. Not in the model picker. Not in the settings. Not in the about dialog. The only place the name shows up is in a debug stream that Cursor previously blocked in Composer 1.5 and, per Fynn, “probably forgot” to block in Composer 2. They patched the leak within hours.

Moonshot’s own pretraining lead publicly accused Cursor of violating the license, then deleted the posts. The official Kimi account then pivoted 180 degrees and called the relationship “an authorized commercial partnership” routed through Fireworks AI, the inference provider that hosts Kimi. Cursor leaned on the same talking point. Everyone walked back from the cliff at the same time.

This is the part of the story worth sitting with. Moonshot’s initial instinct, from the people closest to the model, was that Cursor had violated the license. The official account’s walk-back came after someone did a valuation math on the partnership and decided a fight wasn’t worth it. When your own pretraining lead thinks you’re in breach but the front office chooses to celebrate, you are not in compliance — you are in a negotiated silence.

The “partnership via Fireworks AI” framing does work as license compliance cover. Modified MIT doesn’t prevent Moonshot from issuing separate commercial agreements that waive the attribution clause. If there is a real contract in which Fireworks is sub-licensed to serve Cursor under bespoke terms, and Cursor has paid for that sub-license, the $20M/month clause is irrelevant because the parties have contracted around it. Nobody has produced that contract. We’re being asked to take it on faith that the contract does the work the public statements imply.

Two possibilities. Either the contract exists and Cursor is clean on the letter of the deal — while still being spectacularly misleading to its users about what powers the product. Or the contract is thin and the “authorized partnership” language is a retroactive legal blanket draped over a license trigger that nobody noticed until a developer with a debug proxy noticed. Either way, the trust is gone. And the open-source community now has a live case study of what it looks like when a hyperscale commercial user flies right over the attribution threshold and nothing happens.

Elon Musk Enters

This is the part where the story gets interesting, because the timing is impossible to ignore.

SpaceX merged with xAI in February 2026 at a combined valuation of $1.25 trillion. That transaction created a weird hybrid: a launch company, a social network (X, which xAI had already absorbed), a foundation model lab (Grok), and a data-center build-out that rivals the hyperscalers. The entity is reportedly losing money and heading toward a record-setting IPO. Investors in that IPO want one thing: a story that positions SpaceXAI as an AI platform company, not just a rocket company with a chatbot.

Cursor is the other half of the puzzle. Cursor has no proprietary foundation model that can match Claude or GPT-5 on coding. It just got caught admitting that. It is also the fastest-growing AI developer tool on the market, with a revenue run rate and user base that validate Musk’s pitch overnight. And Cursor has a problem of its own: it currently sells access to Claude and GPT models made by the same companies (Anthropic and OpenAI) that are now shipping competing coding agents. That is not a long-term posture.

So SpaceX offered the Solomon deal: acquire Cursor outright for $60B later this year, or spend $10B on a “working together” arrangement that gives xAI distribution into the Cursor user base while Cursor uses xAI data-center capacity to train the next model. Cursor has reportedly been renting tens of thousands of xAI chips already. This announcement made it official.

The subtext, stripped of press-release varnish: Cursor just admitted it can’t build a frontier model on its own, and the company whose frontier model has the weakest coding benchmarks on the market just wrote a check for the distribution. They need each other. The question is what Elon does with it on day one of ownership.

The Grok-only scenario

Picture the hypothetical. SpaceX exercises the $60B option. Day one, Musk walks into Cursor HQ and rips out every model that isn’t his. Claude Sonnet 4.6, Claude Opus 4.6, GPT-5, Kimi K2.5 — all gone from the model picker. The only option left is Grok Code, or whatever xAI labels the coding fork this quarter.

This is not a crazy scenario. It is exactly what Google did with Antigravity, its agent-first IDE. It is exactly what OpenAI did with Codex. It is the obvious play for any company that owns both a model and a developer surface: eliminate the open models that commoditize your inference margins and force the base back onto your own stack.

Here is the problem. Grok’s coding benchmarks are not competitive with Claude, GPT-5, or Kimi K2.5 on the tasks Cursor users actually do. Cursor’s paying users did not pay for Cursor — they paid for Claude-in-Cursor and GPT-in-Cursor and, unknowingly, Kimi-in-Cursor-labeled-Composer-2. Strip out the models and you are selling an editor with a weaker agent. The moment that happens, the bottom falls out of the retention curve. Developers will install Claude Code in VS Code, install Cline, install Windsurf, install Zed — all of which let them keep the model they actually want — and Cursor’s $2B run rate starts compressing fast.

Musk knows this. Or his people do. The real scenario is not a one-day purge. It is a two-year starvation diet where Grok’s coding capabilities are dragged up to frontier parity using Cursor’s telemetry as training data, while the non-Grok models stay available but get deprioritized in the UI, rate-limited in subtle ways, and eventually moved behind a paywall tier nobody upgrades to. By the time they flip the switch, Grok Code is actually good enough, and users have been slow-boiled into not noticing.

Whether that works depends on whether xAI can execute an Anthropic-class training run on coding, inside an engineering culture that has historically shipped very differently. That is not a given.

The board state: who benefits, who bleeds

The $60 billion option reshuffles every other seat at the table. Here’s the map.

Google Antigravity

Google released Antigravity in November alongside Gemini 3, pitched as an agent-first IDE where the AI writes the code, runs the browser, and ships the artifact. Antigravity’s architecture is the template Musk will study. Its default brain is Gemini 3 Pro. It also supports Claude Sonnet 4.5/4.6 and OpenAI’s GPT-OSS — so calling it “Gemini-only” is not quite right, but Gemini is unambiguously first-class and the others are guests in a house built for Google’s model. That is the template for what a “walled garden with a side door” looks like, and it is the exact posture a post-acquisition Cursor will take.

Antigravity’s beneficiaries of the Musk deal are indirect but real. If Cursor turns into a Grok delivery mechanism, developers who want multi-model optionality in a polished IDE have exactly one remaining option, and it’s Antigravity. Google did not have that positioning three days ago. They do now.

Visual Studio Code + GitHub Copilot

VS Code is the elephant every IDE startup tries not to talk about. It runs natively on every desktop OS, costs nothing, and ships with Copilot integration out of the box — Copilot Pro at $10/month is the cheapest AI pair-programmer on the market that doesn’t require a new editor. Since VS Code 1.109, Claude, Codex, and Copilot agents all run as first-class extensions in the same editor. You don’t fork VS Code. You install what you want.

If Cursor gets absorbed into SpaceXAI and loses its model buffet, the path of least resistance for its defectors is a thirty-minute migration back to stock VS Code + the Claude Code extension, or Copilot, or both. Microsoft doesn’t need to do anything. They just have to wait.

Windsurf (owned by Cognition)

Windsurf — the other AI-native VS Code fork — was acquired by Cognition for $250M in July 2025, making it part of the same house as Devin. Windsurf already has 40+ IDE plugins including JetBrains, Vim, and XCode, plus FedRAMP/HIPAA/ITAR certifications that Cursor doesn’t have. Cognition’s SWE-1.5 model is shipping in Cascade as a proprietary coding brain.

Windsurf is the biggest immediate beneficiary of the Cursor story. Same editor paradigm, broader IDE coverage, better enterprise compliance, and a CEO (Cognition’s Scott Wu) who is not named Elon Musk. Every regulated-industry developer who was on the fence about Cursor just made the call.

Zed

Zed is the serious nerd’s play. Native Rust editor, starts in milliseconds, renders at 120fps, hosts Claude Code and other agents as external brains rather than baking one into the editor. Zed’s value proposition is exactly the opposite of Cursor’s: the editor is the product, the AI is a pluggable tool, and you bring your own model.

The Musk acquisition is rocket fuel for Zed. Every developer who reads the Composer 2 story and decides they want cleaner model provenance is a Zed candidate. Zed is not going to pass Cursor in install base. It doesn’t need to. It needs to own the top decile of developers who are paid the most to write code, and the current story moves that audience in its direction.

Cline, Aider, Continue.dev

The open-source agent layer — Cline (5M+ installs), Aider (terminal-first), Continue.dev — are the free tier that sits inside VS Code and JetBrains. Bring-your-own-API-key, pay only for inference, full transparency on which model is running. Cline in particular scored 80.8% on SWE-bench Verified when paired with Claude Opus 4.6, which is higher than any of Cursor’s public numbers.

Cline is what the Composer 2 story points at. When the paid, polished, closed-source option turns out to be an open-source model in a velvet suit, a lot of developers will ask why they’re not just running the open-source model directly. Cline answers that question.

Claude Code and OpenAI Codex

The two companies whose models Cursor has been reselling are also the two companies Cursor is now competing against. Anthropic’s Claude Code runs in VS Code, the terminal, and JetBrains. OpenAI’s Codex runs in cloud sandboxes bundled with ChatGPT Plus. Both are bundled with their parent subscriptions, which means their effective price is zero for anyone who already pays for Claude Pro or ChatGPT Plus.

Anthropic and OpenAI now have an obvious move. Revoke or significantly re-price the Cursor resale arrangement the moment xAI buys the company. Why subsidize inference into a competitor’s distribution channel? Cursor’s value to its users depends on having Claude and GPT available. The moment Anthropic and OpenAI decide that arrangement ends, it ends. That is a pen-stroke decision, not a product decision.

Advice for Elon, if he’s reading

This is the part nobody writes because everyone assumes Musk won’t take advice. Write it anyway. $60B is enough money to deserve it.

First: don’t kill the model picker. Antigravity kept Claude and GPT-OSS in the UI alongside Gemini because Google understood that the cost of losing multi-model developers is higher than the cost of paying Anthropic a per-token fee. Cursor needs the model picker for the same reason. Make Grok the default. Make it the best-performing option for the specific tasks you can train it on. But don’t force-remove the alternatives. If Grok is actually good, users will pick it. If Grok is not actually good, users will leave, and forcing them to use it just makes them leave faster.

Second: solve the trust problem publicly. The Composer 2 story cost Cursor something more valuable than a license fee — it cost the credibility of Cursor’s product announcements. The fix is not a blog post. It is a machine-readable model bill of materials shipped in every release: base model, fine-tuning data provenance, inference provider, data retention policy, jurisdiction. Make it the industry standard before anyone forces you to. First mover here is a durable advantage.

Third: integrate across the SpaceXAI stack in a way Cursor alone cannot. You own X, xAI, SpaceX, and now (option-pending) Cursor. The obvious play nobody else can run: pipe X’s real-time developer conversation into Cursor’s agent context, pipe Starlink’s edge network into Cursor’s inference routing for low-latency global delivery, pipe Grok’s training loop into Cursor’s telemetry for continuous coding-task improvement. None of your competitors can match that stack. That is the acquisition thesis. Build the stack, don’t just consolidate the models.

Fourth: buy Fireworks AI. Fireworks is the inference layer that apparently sub-licensed Kimi to Cursor in the first place. It’s a serving platform for open-weight models at hyperscale. SpaceXAI will need a commercial inference plane that isn’t tied to hyperscaler margins. Fireworks is the obvious acquisition to pair with Cursor, and it’s a fraction of the price. Do that before Cloudflare does.

Fifth: settle Musk v. Altman before it burns the Cursor deal. The lawsuit against OpenAI goes to trial in less than a week. If xAI wins control of the narrative, the Cursor buy looks visionary. If Musk loses publicly, the Cursor buy looks like a consolation prize. Timing matters, and the legal calendar has a vote.

What happens next

My read on the near-term sequence:

By Q3 2026, SpaceX exercises the option and Cursor becomes a subsidiary. The announcement will be paired with Grok 4 or whatever xAI’s coding-specialized model is called that month, positioned as “native to Cursor.” Composer 2 and its Kimi foundation will quietly disappear from the roadmap — partly because the license optics are poisonous inside a $1.25T entity, partly because Moonshot is a Chinese company and the CFIUS overhang would make ongoing dependence untenable.

Anthropic and OpenAI will re-price or terminate the model resale arrangement with Cursor within ninety days of the acquisition closing. They will position it as a principled response to a competitor’s distribution channel. It is a principled response. It is also good business.

Windsurf pulls ahead in enterprise procurement for the next four quarters. Zed becomes the default for the top tier of independent senior developers. Antigravity settles into the Google-shop default and becomes a real threat at the enterprise level once Google finishes pricing it. Cline’s install base grows faster than any paid tool. GitHub Copilot, already the quiet giant, grows by doing nothing.

And the open-source community now has its first high-profile case study of a hyperscale commercial user blowing past a permissive-license attribution clause without consequence. That will change how the next Modified MIT is written. Expect future Chinese open-weight releases to have teeth in the attribution clause — automatic API-level rate limiting for unattributed use, cryptographic signing requirements, license-enforcement mechanisms baked into the model card. Moonshot got played by Fireworks and Cursor. The next lab won’t.

The bigger lesson is the one Cursor’s VP of Developer Education admitted in his tweet: “It was a miss to not mention the Kimi base in our blog from the start.” That one sentence is the entire story. A company valued at $29 billion had a decision point — tell the truth about the base model, or obscure it — and chose to obscure it because the valuation story required it. Five weeks later, a trillion-dollar acquirer offered $60 billion for the result, and the market cleared at a price that suggests nobody cared about the lie except the developers.

Which is, when you think about it, the actual Composer 2 launch announcement they should have written. Not the one about frontier intelligence. The one about what the market is actually willing to pay for.

— Gal

WhiteStar Labs | galratner.com

That is the day Mike Krieger, Anthropic’s Chief Product Officer and the co-founder of Instagram, resigned from Figma’s board of directors. The same afternoon, The Information reported that Anthropic’s next model would ship with design tools aimed squarely at the core of what Figma sells. Figma’s stock dropped three percent that day, and nobody in the industry had to ask why.

Three days later, on April 17, Anthropic launched Claude Design powered by Claude Opus 4.7. A month before that, Google rolled out what it called Stitch 2.0 — a complete rewrite of its AI-native design canvas. Both products are now generally available to paying users, and between them they cover roughly eighty percent of the jobs a working designer opens Figma to do.

As I write this at 12:51 PM Pacific, Figma (NYSE: FIG) is trading at $19.34 a share. That is down 84.15% from its all-time high of $142.92, set on August 1, 2025 — the second day after its IPO. Market cap: $10.10 billion. In July 2025 it was $68 billion. Roughly $58 billion in paper value has vaporized in eight and a half months.

I have been paying subscriptions to Figma for years. I have shipped products with it. I still think Dylan Field is one of the better founder-CEOs in SaaS. And I am going to make the case, in this piece, that unless Figma does something structural and fast, the company as a stand-alone public equity is finished.

Then I am going to tell you what that structural move looks like. Not because I think Figma cares what a CTO in Las Vegas has to say about it, but because the playbook is sitting in plain sight, and somebody on their executive team ought to pick it up.

Let us start with the two tools that are doing the damage.

What Claude Design actually is

Anthropic shipped Claude Design on April 17, 2026. I have been inside it for the last three days.

The product is a split-pane interface — chat on the left, canvas on the right. You describe what you want, Claude builds a first version, and from there you iterate: through conversation, through inline comments you drop directly on elements in the canvas, through custom sliders Claude itself generates for the specific design (spacing, color, layout), or through direct text editing. It is powered by Claude Opus 4.7, Anthropic’s most capable vision model, which also launched this week.

What makes it materially different from every AI design tool that came before it:

• It reads your code. During onboarding, Claude ingests your codebase and existing design files and builds a design system from what it finds. Every subsequent project auto-applies your colors, typography, and components. No separate import-tokens step. No manual library upkeep.

• It ingests real documents. DOCX, PPTX, XLSX, images, PDFs. Drop in a spreadsheet of metrics, it builds a dashboard mockup. Drop in a PowerPoint, it redesigns it. This is where Opus 4.7’s multimodal bandwidth (images up to 2576 pixels wide) actually starts to pay.

• It hands off to code. When a design is ready to build, Claude packages everything — the design, the intent behind decisions, the component structure — into a handoff bundle for Claude Code. One instruction, and you have a working frontend.

• It exports to Canva, PPTX, PDF, HTML, or a shared folder. Not hostage data. You can walk the output anywhere.

• It does frontier design. Voice, video, shaders, 3D, built-in AI. Things no traditional vector design tool was ever structurally capable of producing.

John Voorhees at MacStories spent a week with it and wrote that iterating between chat, canvas, and revisions felt noticeably faster than doing the same loop in Claude Code. VentureBeat called it the most visible expression of Anthropic moving from model provider to full-stack application company. Banani, an AI design review outlet, called it the most capable AI design tool available to anyone with a Claude subscription on launch day.

A few honest caveats. It is a research preview. Pricing is included in the Claude Pro, Max, Team, or Enterprise subscription, but usage counts against your tier limits and can burn through tokens fast with Opus 4.7’s large image inputs. Enterprise admins must opt in explicitly — Anthropic ships it off-by-default on Enterprise plans, which is a conservative and correct choice. And like every Labs product, features will change and break before it hits GA.

But the product works today, and the work it produces is shareable with non-designers without making your company look amateurish. That is the bar, and Claude Design clears it.

What Google Stitch actually is

Stitch is an older story. Google acquired Galileo AI in May 2025, rebranded it as Stitch, and shipped it at Google I/O 2025 as a Google Labs experiment. The original version was a text-to-UI generator powered by Gemini 2.5 Flash (fast mode) and Gemini 2.5 Pro (experimental mode), with a focus on mobile and web app UIs. It was free, generous on limits, and handed off to Figma or clean HTML and CSS.

The original Stitch was fine. A faster Galileo with Google branding.

The version that matters is Stitch 2.0, which shipped in March 2026. That is the release that kicked Figma stock down 12% in two trading days.

Stitch 2.0 is a complete product rethink. The old screen-at-a-time generator is gone. In its place sits an AI-native infinite canvas, a design agent that can reason across the entire project’s history, an agent manager that lets you run multiple ideation threads in parallel, voice input for real-time design critique, and an MCP server and SDK that lets other tools call Stitch’s capabilities as a service. You can export to AI Studio, Antigravity, Firebase Studio, or back to Figma.

What Stitch does well that Claude Design does not do as well:

• It is free, full stop. No subscription, no paywall, no Pro tier on top.

• The multi-screen consistency on app flows is excellent. Generate five screens, they share the same design language without you having to wrangle it.

• Exported code (Tailwind, Flutter, HTML) is structurally solid enough to be a real starting point, not a div-soup wireframe.

• Voice commands in Live Mode feel genuinely conversational.

What Stitch does not do:

• It is a UI design tool. Full stop. It does not make pitch decks, one-pagers, social graphics, marketing collateral, illustrations, or presentations. Claude Design covers all of that.

• It does not read your production codebase to build a design system. You get theming — light/dark, corner radius, palette, typography — but not your actual component library.

• It lives inside Google Labs. Google sunsets Labs products that miss internal metrics. There is no SLA, no uptime guarantee, no commitment to keep it alive past the next fiscal year.

• No real team collaboration surface. No org-scoped sharing, no granular permissions.

And the monetization story is a black box. The usage limits cannot be raised. There is no paid tier. When you run out, you stop for the month. For a company serious about shipping, that is a non-starter.

Reviews are mixed but coherent. Index.dev called it a significant step forward in AI-assisted UI design. UX Pilot said the layouts tend to look generic and need manual editing before they feel production-ready. Moda said the free tier is enough for real work for individuals and small teams. The consensus is: fast, impressive, narrow.

Head to head

Launched. Claude Design shipped April 17, 2026 as a research preview. Stitch shipped in May 2025 as a Galileo AI rebrand, with Stitch 2.0 landing in March 2026.

Underlying model. Claude Design runs on Claude Opus 4.7. Stitch runs on Gemini 2.5 Flash and 2.5 Pro, with Gemini 3 on its experimental tier.

Output scope. Claude Design produces prototypes, decks, one-pagers, marketing pages, 3D, voice, and video. Stitch does mobile and web app UIs only.

Codebase ingestion. Claude Design reads your repo and builds a design system from it automatically. Stitch does not — you get theming, not your actual components.

File imports. Claude Design takes DOCX, PPTX, XLSX, images, and can capture directly from a live site. Stitch takes images and sketches, and only in experimental mode.

Real-time iteration. Claude Design gives you chat, inline comments on specific elements, generated sliders for spacing and color, and direct edits. Stitch gives you chat, annotations on screens, and voice commands in Live Mode.

Code export. Claude Design produces a handoff bundle for Claude Code plus HTML. Stitch exports Tailwind, HTML, Flutter, and JSX, and hands off to Figma, AI Studio, Firebase Studio, or Antigravity.

Team collaboration. Claude Design has org-scoped sharing with view, comment, and edit roles. Stitch has no real permissions surface.

Pricing. Claude Design is included in Pro, Max, Team, and Enterprise subscriptions, with extra usage metered. Stitch is free — 350 standard generations plus 200 experimental per month, with no paid tier to raise the ceiling.

Enterprise posture. Claude Design is off by default on Enterprise plans, requires admin opt-in, and has a full SOC/SSO path. Stitch is a Google Labs experiment with no SLA and real sunset risk.

Best at. Claude Design: end-to-end design to code handoff across many artifact types. Stitch: fast app-screen ideation with clean multi-screen consistency.

Weakest at. Claude Design: burns tokens fast on large images, and it is still a preview. Stitch: narrow scope, no design system integration, uncertain future.

The short version:

• Pick Claude Design if you are producing more than app screens — decks, one-pagers, marketing pages, interactive prototypes with code — and especially if you are already in the Claude / Claude Code ecosystem. The codebase ingestion and the Claude Code handoff are what make this a workflow, not a novelty.

• Pick Google Stitch if your job is specifically mobile or web app UI, your budget is zero, and you do not mind being on a Labs product with no long-term commitment. If you ship from Firebase Studio or use AI Studio as your IDE, the integration is tight.

• Pick both if you are honest. Stitch for early app-screen exploration because it is free and the multi-screen consistency is genuinely good. Claude Design for anything you will actually put in front of a customer, because it is inside a productized subscription and it handles the long tail of visual work.

I have not opened Figma in nine days. That is the longest stretch since I started paying for it.

How Figma got here

Figma was a good idea executed beautifully for more than a decade. It is important to say that before I spend the next two sections arguing it is now obsolete.

Dylan Field and Evan Wallace started building Figma in 2012 while studying computer science at Brown. Field ran the CS Departmental Undergraduate Group; Wallace was a graphics TA. Their original pitch, per an early Brown Daily Herald article, was a technology startup that would let users creatively express themselves online. They burned through a meme generator and drone software before landing on a browser-based graphics editor.

The seed round (June 2013) was $3.8M from Index Ventures. Series A ($14M, Greylock) in 2015. Series B ($25M, Kleiner Perkins) in 2018. Series C ($40M, Sequoia) in 2019. Series D ($50M, Andreessen Horowitz) in April 2020 at a $2B valuation. Series E ($200M, Durable Capital Partners) in June 2021 at $10B. That is a canonical Silicon Valley runway: two founders, one good product insight, seven years, five rounds, $330M raised, no dilution disasters.

The product insight was real: design collaboration should work like Google Docs, not like Photoshop. Multiplayer vector editing in the browser. Comments on frames. Real-time cursors. Version history. This seems obvious in retrospect. It was not obvious in 2016 when Figma was trying to convince Adobe XD users to switch.

By 2021, Figma was the de facto standard for product design. Sketch had lost. Adobe XD was terminal. Framer and InVision were niche. Figma had the designers, the students (free for education), the enterprise contracts, and the network effect that matters most in design: your collaborator cannot open your file unless they also use your tool.

Then, on September 15, 2022, Adobe announced it was acquiring Figma for $20 billion in cash and stock. Adobe’s stock dropped 17% on the announcement. The design community reacted somewhere between skepticism and grief. The DOJ opened an investigation in November 2022. The European Commission formally reviewed the deal in February 2023. On December 18, 2023, Adobe and Figma mutually terminated the deal. Adobe paid a $1 billion reverse termination fee — nearly triple the total capital Figma had ever raised.

That billion dollars, paid by a regulator-spooked Adobe, is why Figma had the runway to build out its AI products — Figma AI, Figma Make, Figma Sites, Figma Buzz, Figma Draw, Figma Weave — and file an S-1 by April 2025.

Figma went public on July 31, 2025 under the ticker FIG. Priced at $33. Opened at $85. Closed day one at $115.50. Touched $142.92 intraday the next morning. Market cap peaked near $68 billion. It was, briefly, the cleanest tech IPO in years.

Then, the reversal. The stock was at $21 by late January 2026. An 85% drawdown in six months. As of today, $19.34.

The stock reversal is a symptom. The underlying disease is that the thing Figma sells — a real-time collaborative canvas where humans draw boxes together — is no longer where product design happens. That transition is what the next section is about.

Why I think Figma is obsolete

I do not mean obsolete in the fanboy way, as in dead, finished, over. Figma still has $1.06 billion in trailing revenue, 82.43% gross margins, and $1.6 billion in cash. It is not going bankrupt next quarter. It is not going bankrupt next year.

What I mean is that Figma’s core product — the real-time multiplayer canvas — is now a step in the workflow instead of the workflow itself. And as soon as a product becomes a step, it becomes a commodity that the AI layer above it is incentivized to swallow.

Here is the actual shift. In 2023, the workflow was:

1. PM writes a spec.

2. Designer opens Figma.

3. Designer explores five directions, picks one, refines it.

4. Designer hands off to developer via Dev Mode or Zeplin.

5. Developer implements it.

6. Designer QAs the implementation.

In 2026, for a growing share of work, the workflow is:

1. PM describes the feature to Claude Design or Stitch.

2. AI generates five directions in parallel.

3. PM or designer picks and refines in chat.

4. AI hands the design to Claude Code (or another agent) as a bundle.

5. Agent ships the code.

6. Humans review and merge.

Steps 2 through 4 happen in Claude Design. Step 5 happens in Claude Code. Figma is not in this loop anywhere.

That is not speculation. That is the Brilliant case study in Anthropic’s launch post (20+ prompts to recreate their most complex pages elsewhere, 2 prompts in Claude Design). That is the Datadog case study (a week-long cycle compressed into a single conversation). That is the Tech Startups analysis of why the stock is down 85%: for the millions of Figma users who were never designers to begin with, the shortcut matters. If a product manager can generate a prototype through text, the need to open a shared canvas fades.

Then there are the self-inflicted wounds.

In June 2024, Figma shipped Make Designs at Config. Within days, designers on X demonstrated that the tool was producing near-replicas of Apple’s iOS weather app. Field admitted on X that he had pushed the team to hit the Config deadline and that they had relied on off-the-shelf models with a bespoke design system that was not varied enough. Figma disabled the feature. The damage to the AI-native Figma narrative was done. When you are a collaborative design company and your first headline AI feature reproduces someone else’s copyrighted UI, the entire market assumes the rest of the stack has the same problem.

More quietly, Figma’s own MCP server is, per multiple indie developer analyses, designed to control access, not accelerate it. An indie developer wrote a piece literally titled A Better Figma MCP showing that Claude got more capability by opening a browser and driving the Figma plugin API directly than by using the official MCP. If you are a designer, that posture protects your workflow. If you are an agentic AI, that posture is a reason to route around you.

Gross margins have fallen from 91% to 83% over the last 18 months as AI infrastructure costs have climbed. Management guided 2026 revenue to $1.366 to $1.374 billion — about 30% growth, down from 41% in 2025 — and operating margin to around 8%, down from 12%. BTIG initiated coverage at Hold on April 7 with three specific concerns: AI monetization is unproven, Google competition is real and growing, and platform competition at scale is structural. Cramer, on the January 27 Mad Money, called competing with Google an uncomfortable position and said he would rather avoid the design software category entirely.

And then, April 14, 2026. Mike Krieger resigns from the board. Three days later, Claude Design ships.

This is what a ceiling getting lower in real time looks like.

The stock story

Let me put the numbers in one place, because the full picture is harsher than any single headline.

Figma (NYSE: FIG) — Max range. Source: Google Finance, April 21, 2026, 12:51 PM EDT.

Current price (Apr 21, 2026, 12:51 PM EDT)

$19.34

All-time high (Aug 1, 2025)

$142.92

Drawdown from peak

84.15%

Market cap

$10.10 B

52-week low (Apr 14, 2026)

$17.65

TTM revenue

$1.056 B

TTM gross margin

82.43%

TTM net margin

-118.44%

TTM EPS

-$3.14

Cash and short-term investments

~$1.6 B

2026 revenue guidance

$1.366 B – $1.374 B (~30% YoY)

2026 operating margin guidance

~8% (down from 12% in 2025)

Q4 2025 beat

+23.67% EPS, +3.62% revenue

Some of that is actually healthy. $1.6B in cash and low debt gives Figma multi-year runway even with the losses. The 30% guided growth rate is well above the median software company. The Q4 2025 earnings beat was real.

But the market is not pricing Figma on trailing numbers. The market is pricing Figma on the question: in 2028, when a PM can go from idea to shipped code in an agentic pipeline that does not require opening a canvas, who opens the canvas?

Six months of public trading was enough time for that question to form. The remaining VC lockup releases through mid-2026 are not going to help.

There is one more structural pressure worth calling out. Insider buying in February was $36.5 million (a director buying; stock jumped 11% on the print). Cathie Wood bought $8M in late March. Those are contrarian signals, not trend confirmations. The base case from here depends on Figma doing something different, not on the market changing its mind.

I am a CTO, not a financial advisor or a lawyer, and nothing in this article is investment advice. Do your own research and talk to a real advisor before you move money.

The pivot plan, because I actually want Figma to survive

I am not writing this to dunk on Figma. The design community is better off with a credible independent design platform than with a world where Adobe, Google, and Anthropic divide the spoils. Dylan Field is still at the wheel. The engineering team is still one of the best in SaaS. There is $1.6B in cash. There is time.

Here is what I would do if I were in the room.

1. Stop defending the canvas. Make Figma the design system of record.

Right now, Figma’s moat is the real-time collaborative canvas. That moat is draining. What cannot drain in the next three years is that every serious product team has a component library, a token catalog, and a set of production patterns that are the source of truth for what our product looks like. Figma already hosts most of those. Reposition from the place where designers work to the place where design systems live — the substrate that every AI design tool writes against.

2. Open the MCP server. Fully.

Not the current version, which is rate-limited and permission-gated in ways that throttle agentic use. Publish an MCP surface that lets Claude, Stitch, Cursor, Windsurf, and anyone else read and write Figma files, components, and tokens in real time, with delegated auth and enterprise controls. Charge by tokens consumed, not by seat. The thing Figma has that the AI labs do not have is the installed base of enterprise design systems. Monetize that access instead of gating it.

3. Ship a true agentic workflow, not a prompt-to-design toy.

Figma Make is fine. It is not the product. The product is Figma Agent — an agent that reads your product, your code, your design system, your analytics, and your user research, proposes specific UX changes with rationale tied to data, and ships the code via a Dev Mode pipeline into your GitHub repo. That is the loop Claude Design plus Claude Code is starting to own. Figma has the design data to own it better if it stops trying to be the surface and starts being the substrate plus the agent.

4. Buy an AI team. A real one.

Payload CMS was a smart acquisition for owning the data layer. It is not an AI team. The $1.6B in cash on the balance sheet can buy a serious twenty-person research team anchored by one top name. There are founders who have shipped model infrastructure and multimodal research who would take that offer. Move fast.

5. Compete on enterprise compliance, not on AI magic.

Claude Design requires enterprise admins to explicitly enable it. Stitch is a Labs product with no SLA. Figma has SOC 2, enterprise SSO, audit logs, and a decade of contracts with Fortune 500 legal and security teams. Every Fortune 500 is going to adopt AI design, and every Fortune 500 legal and security team is going to require the boring things. Lean into being the boring, trusted, compliant AI design platform. That positioning is available and nobody else wants it.

6. Radical pricing on the design system itself.

Make Figma design system hosting free for teams up to some reasonable size. Charge on AI credit consumption and on enterprise features. Flip the pricing from per designer seat to per AI generation against your system. That monetizes the shift instead of fighting it. It also makes the design system stickier the more the team uses AI, which is exactly the direction the industry is going.

7. Do the Krieger-resignation call honestly on the next earnings call.

Investors are not stupid. Field needs to stand up on the Q1 2026 call in early May and say, explicitly: the competitive landscape has changed in the last quarter, here is specifically how, and here is what we are doing about it. Not a deflection. A plan. The stock is priced for a structural disruption scenario right now. The way back is not reassurance; it is clarity about what the 2027 and 2028 Figma actually looks like.

Where that leaves us

I will keep my Figma subscription for another quarter. There are still files I have to open. There are still clients whose designers live in it. Installed base does not vanish overnight. Lotus Notes ran for twenty years after the email client war was over.

But I am not betting my workflow on Figma being the center of product design in 2028. I moved my new prototypes to Claude Design the day it launched. I am using Stitch for throwaway app-screen exploration. I am using Claude Code to ship. Figma is a stop on the handoff, not a place I start anymore.

If you run a design team, you have a decision to make in the next ninety days. Stitch and Claude Design are not previews of the future. They are the present. The question is whether you get ahead of the tooling transition or get dragged behind it.

If you are long FIG, I am not going to tell you what to do with your position — again, I am not a financial advisor, and none of this is financial advice. I will tell you what I am watching: the next two earnings calls, what Figma ships at Config 2026 (usually in June), and whether the MCP server gets a real rewrite. If those three check out, there is a turnaround thesis. If they do not, the 84% drawdown may not be the bottom.

If you work at Figma, or you know someone who does, forward this up. The pivot is available. The cash is there. The runway is there. The talent is there. What is not there, yet, is the willingness to stop defending the canvas.

The window to do it is open. It will not be open in 2027.

Gal Ratner writes about technology, architecture, and what is actually happening in the industry at galratner.com and on Substack. He is a CTO, a motorcyclist, and does not take product announcements from Figma, Anthropic, or Google at face value.