In April 2015, HBO's Silicon Valley ran an episode where a billionaire investor sits a young founder down and explains that revenue is a mistake. Report a number and people will evaluate it, and no number survives evaluation. Stay pre-revenue and nobody can do the arithmetic on you. You get priced on what you might eventually be worth instead of what you actually earned, and imagination has no ceiling. Russ Hanneman was written as the punchline of that season, a man who got rich once putting radio on the internet and mistook the luck for a theory. Eleven years on, the bit no longer scans as satire. It scans as a term sheet, and the real numbers have outgrown the joke.
On August 26, TechCrunch reported that Instinct had closed a $250 million Series B at a $2.5 billion valuation, co-led by Index Ventures and Benchmark. Spear Street Technology, the company behind it, was registered in California in April. The app is in private beta. There is no public pricing, no launch date, and no disclosed revenue. The founder introduced himself publicly for the first time on the day the round was announced.
Three weeks earlier, Kleiner Perkins partner Mamoon Hamid had led a $75 million Series A at a $500 million valuation. So the price went up by $2 billion in roughly twenty-one days. Do the division. That is about $95 million a day. Just under $4 million an hour. Somewhere around $66,000 a minute, continuously, including every hour that everyone involved was asleep.
Nothing shipped during those three weeks. No launch, no pricing page, no revenue disclosure, no public product at all. Every minute of that $66,000 bought a number going up.
Per head
Divide by people and it gets stranger. Safe Superintelligence has roughly fifty employees, no released product, and not one published paper. It is valued at $32 billion. That works out to about $640 million per employee. Not per year. Per person, sitting in the building.
humans& announced a $480 million seed at a $4.48 billion valuation in January, three months after founding, with roughly twenty people. Call it $224 million each. The company said it would ship its first product in early 2026. It is the end of August.
This is not a figure invented by cranks. Todd Sisitsky, president of TPG, said publicly that early-stage AI ventures were commanding valuations between $400 million and $1.2 billion per employee, and used the word breathtaking. He was not paying a compliment.
The same company, priced twice, one month apart
David Silver ran reinforcement learning at DeepMind for over a decade and built AlphaGo. He left in late 2025 and founded Ineffable Intelligence in London. On April 27 the company announced $1.1 billion in seed funding at a $5.1 billion valuation, the largest seed round in European history, with no product, no revenue, and no public roadmap.
Forbes reported the money came in two pieces. About $11 million went in first, valuing the company at roughly $55 million pre-money. Then $1.1 billion followed within about a month, at $4 billion pre-money. Sequoia was in both, with the majority of its capital at the higher mark. Same company. Same empty product roadmap. Roughly seventy times the price, four weeks apart.
For scale, the median pre-money valuation for a US seed round is around $17.9 million. Ineffable’s second tranche priced at more than two hundred times that, for a company whose own investors had valued it at $55 million the month before. Anyone hired in week five gets their options struck against $4 billion rather than the blended number, which is a quiet way of transferring risk to employees who did not set the price.
The money is not leaving the room
The part that reframes everything else is where the capital actually goes.
Nvidia is an investor in Safe Superintelligence, Ineffable, humans&, Thinking Machines, SkildAI, Baseten, Runway, and Anthropic. That is not a sample. That is most of the list. CNBC reported in May that Nvidia had passed $40 billion in equity investments in 2026 alone, including at least seven multibillion-dollar deals with public companies and about two dozen private rounds according to FactSet. The largest single check was $30 billion into OpenAI.
Then those companies spend the money on Nvidia hardware. With Thinking Machines both ends are documented: Nvidia was in the seed round, and in March 2026 the two announced a partnership involving an undisclosed investment plus a multi-year agreement to deploy a gigawatt of Vera Rubin compute. CoreWeave is an equity stake paired with a $6.3 billion compute-purchase agreement running to 2032. In July, Bloomberg reported Nvidia was working on a fresh set of deals worth more than $750 billion, including talks to backstop as much as $250 billion so OpenAI could lease data center capacity.
IDC has a clean definition for this. Circular financing is where the same capital moves at once as a vendor payment and as an equity stake, so a company funds its own customer’s revenue while also supplying that customer’s infrastructure. Goldman Sachs raised its Nvidia price target and warned in the same note that circular revenue could be dilutive to the multiple. On August 26, the same day Instinct’s round broke, Jensen Huang went on CNBC to say the critics were missing the point and that this is simply the first generation of startups needing tens of billions to operate.
He may be right about the capital intensity. It does not change the shape of the flow chart. A large share of this money is being handed to companies whose main activity is handing it back.
The same six months, from a desk
PitchBook put AI funding in the first half of 2026 at $407 billion. More than all of 2025 combined, in six months.
Now hold that against what happened to the people who write the software. Challenger, Gray & Christmas counted roughly 52,050 announced tech job cuts in the first quarter of 2026, the highest first quarter the firm has logged since 2023. TrueUp tracked 249 separate layoff events by early May, covering 95,878 workers, a pace of about 864 people a day. That is faster than 2025 ran.
Put the two rates next to each other, because they happened in the same months, in the same industry. Instinct’s price went up about $95 million a day. The sector shed about 864 jobs a day.
The rest of the hiring data points the same direction. Indeed Hiring Lab has general software engineering postings down roughly 49 percent against the February 2020 baseline, with overall tech postings off about 36 percent, while machine learning engineer roles are up 59 percent over the same stretch. Stanford’s research found entry-level software developer employment down about 20 percent from its peak. Unemployment for new computer science graduates is running around 6.1 percent. Marc Benioff said plainly that Salesforce hired no new engineers in fiscal 2026.
So the number is too big to feel. Convert it into public budgets and it gets easier. The entire federal SNAP program, the food assistance roughly 42 million Americans depend on, cost $101.7 billion in fiscal 2025. AI startups raised four times that in half a year. The National Institutes of Health, which funds close to fifty thousand competitive grants across more than twenty-five hundred institutions, runs on about $47 billion annually. Six months of AI funding was more than eight times that. The National Science Foundation, which underwrites roughly a quarter of all federally supported basic research at American universities, operates on $8.75 billion. Six months of AI funding was forty-six times the NSF.
Let me be precise about what that comparison is and is not. Venture capital was never going to buy anyone groceries. It is not fungible with a food budget and nobody diverted one into a seed round. Anyone telling you otherwise is selling something. The point is narrower and worse than a misallocation story.
Grocery prices are up about thirty-two percent since January 2020 and reaccelerating. The University of Michigan’s consumer sentiment survey hit an all-time low for economic optimism this month. Wages in most of this industry have not tracked either number. And if you write software for a living, you are standing on both sides of this ledger at once, which is the part that should actually make you put the coffee down. A large share of the justification for that $407 billion is the claim that these systems replace people like you. The savings get booked against your salary line. Then the capital those projected savings unlock turns out to be unavailable to you at any price.
A profitable small software company with real customers cannot raise two million dollars on twenty years of shipping. A four-month-old company with no product, no pricing, and no disclosed revenue raised three hundred and fifty million on a name and a private beta. Capital is not scarce right now. It is abundant to the point of embarrassment. It is allocated by proximity, and almost none of the allocation turns on whether anything gets built.
Then there is the public money
The United Kingdom’s Sovereign AI fund and the British Business Bank both participated in Ineffable’s round. Both run on taxpayer money. A public development bank put $20 million into a pre-product company at a $5.1 billion valuation that had been set weeks earlier by the same private investors who bought in at $55 million. Whatever you think of venture capitalists gambling with their limited partners’ money, that is their arrangement. This is different. This is a government putting public funds in at the marked-up price, after the markup, on the strength of a founder’s reputation and a company with nothing to sell.
And a lot of it simply evaporates
None of this money is being carefully husbanded. Yupp.ai raised $33 million with Chris Dixon and a16z crypto in the round, reached 1.3 million users, and shut down about ten months after launch. Cluely raised $5.3 million and then $15 million from Andreessen Horowitz at roughly $120 million; in March 2026, TechCrunch ran a piece under the headline that its CEO had admitted to publicly lying about revenue, after Roy Lee posted that he had gotten a cold call about numbers and, in his words, “told her some bs,” attaching Stripe screenshots with the real figures. The Justice Department indicted Albert Saniger in April 2025, alleging he told investors that Nate completed purchases without human intervention when the automation rate was near zero and contractors in the Philippines and Romania were doing the clicking. Nate had raised more than $50 million.
Builder.ai is the case everyone cites and most people get wrong, so get it right. The viral version was 700 engineers in India manually impersonating an AI called Natasha. That story came from one social media post, several outlets ran it, and several later walked it back; Windows Central updated its article specifically to remove the claim. The collapse was real and so was the fraud, but the fraud was in the books: roughly $220 million in claimed 2024 revenue against an audited figure closer to $50 million, surfaced after a lender seized $37 million from the accounts. Roughly $445 million went in and the investors got a bankruptcy.
The aggregate picture matches the anecdotes. The MIT study that circulated last year found ninety-five percent of organizations investing in generative AI saw no measurable return. SimpleClosure’s data shows Series A shutdowns climbing from around six percent of wind-downs to fourteen percent in a single year, which is what it looks like when a correction stops killing bad ideas and starts killing funded ones.
What the money is not buying
Here is the part I have lived. In the same week Instinct’s valuation went up five times, its early users were publishing failures that any production engineer would recognize on sight. Claire Vo disconnected the Google integration at eleven in the morning and received an email summary at two in the afternoon; the agent told her the messages had been kept in plain text. Peter Yang could not delete indexed email from the company’s records until the team shipped a settings toggle in response to him. Alex Cohen made a fresh Gmail account, emailed his own inbox with instructions aimed at the agent, watched it obey them, and deleted his account. Another tester saw it pull a sign-up code out of an inbox to finish a Resy booking. Katie Jacobs Stanton had it send mail on her behalf without asking. The terms of service claim a perpetual and irrevocable license over user materials including for model training, and let the agent enter binding agreements on the user’s behalf.
Not one of those is a research problem. They are engineering problems with known answers. Short-lived scoped tokens. Revocation that cascades into derived stores instead of stopping at the OAuth grant. Inbound mail treated as untrusted input rather than instruction, which is the oldest lesson in the field. A confirmation gate on anything irreversible. An audit record on every tool call, with the arguments.
I build agentic systems on MCP for a living. When I built a wedding planning agent for a Las Vegas chapel, getting the model to plan a wedding was the easy half. The hard half was guaranteeing it could not confirm a booking, move a deposit, or email a customer without a human approving it and a log entry proving who did. The client’s downside was never a mediocre venue suggestion. It was an unauthorized commitment on a Saturday in June, found on Monday. That work took months. It does not demo well and it never gets marked up five times in three weeks.
So the $2 billion that landed on Instinct's price tag in twenty-one days did not buy a line of code, a security review, or one paying customer. It bought a number. Someone will sell that number to someone else at a higher number, and when the music stops the people holding it will be told they should have known better while the people who set the price are already marking up the next one. Hanneman was the villain of that season. The writers still gave him Pinterest and Snapchat, real companies with real users losing real money, because satire has to stay inside what an audience will accept. Hand that writers' room $66,000 a minute for an app almost nobody has used and the note comes back that it strains credibility. We are eleven years past the joke and still climbing.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and chief architect at Prana Entertainment, an enterprise software and AI consultancy in Las Vegas. He has spent close to thirty years shipping production software on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy, and Allegiant Air, and was employee number six at Break.com during the user-generated content era.
He builds production agentic systems on MCP and the Microsoft Agent Framework, including Cara, a shopping assistant for ShopSnap storefronts, and a wedding planning agent for a Las Vegas chapel. He maintains PLogger, an open-source observability framework for .NET, and writes weekly about the distance between what executives say about AI and what practitioners actually put into production. He is a Los Angeles Business Journal CTO of the Year finalist and the author of the novel The Archive of Lost Suns.