SAN FRANCISCO — The numbers that have defined the AI boom looked a little different last week. Not the headline figures — those keep climbing. Anthropic just closed a $65 billion round at a $965 billion valuation, leapfrogging OpenAI. SpaceX is preparing a public listing that could seek $75 billion at a $1.75 trillion valuation. The money is still pouring in. But something else is shifting, quietly, in the spreadsheets of the companies that were supposed to make all of this pay off.
Uber burned through its entire AI budget for 2026 in four months. Microsoft canceled Claude Code licenses in key product divisions. The company’s chief technology officer at Uber told staff the firm was going back to the drawing board after discovering it had no clear way to link AI spending to any meaningful improvement in useful product features. According to Fortune, the practice of pushing employees to consume as many AI tokens as possible — “tokenmaxxing” in the industry’s own argot — is already fading. Nobody, it turns out, is comfortable accidentally burning half a billion tokens in a month.
Gary Marcus, the NYU professor emeritus and longtime AI skeptic, was among the first to name what was happening. The enterprise, he wrote this week, is undergoing a correction away from AI overuse. The largest accounts with OpenAI and Anthropic subscriptions had engineers thrown at them to prevent churn. The moment that kept those engineers busy — the brief and extraordinary period when companies told workers to use AI for everything, cost be damned — may already have peaked. “Tokens got burned for millions of dollars without any real significant ROI to show for it,” one tech executive put it in a social post Marcus cited. The phrase reads less like a complaint than an epitaph.
None of this has slowed the money. Three quarters of all venture capital raised over the past year went into five companies, according to Niko Bonatsos of Verdict Capital, speaking at TechCrunch’s StrictlyVC event in Athens last week. In 17 years in Silicon Valley, he said, he had never witnessed more groupthink. The observation carried real weight: Bonatsos helped build General Catalyst’s early-stage franchise before leaving to start his own firm. He was the first backer of what became Cursor, the AI coding tool Elon Musk reportedly holds an option to acquire for $60 billion. He is not a person given to reflexive pessimism about AI investing. And yet.
“If you’re not in one of those two buckets” — AI-native founder, or American dynamism play — “it’s really tough,” Bonatsos told TechCrunch’s Connie Loizos. “Today, if you’re a 40-year-old tenured professor at Stanford not building something in AI, no one wants to meet you.” The joke landed. It also described, without much exaggeration, the actual conditions on the ground in San Francisco in 2026.
Andreas Stavropoulos of Threshold Ventures, who sat alongside Bonatsos at the Athens event, acknowledged that a correction was likely. The promise and optimism surrounding AI, he said, remained “significantly ahead” of the near-term ability to show results. That gap — between what investors believe AI will eventually produce and what enterprises can actually account for today — is the precise tension now pulling at the seams of the boom. It is not a new tension. But the tokenmaxxing retreat suggests it has moved from theoretical to operational.
The approaching IPO wave makes that tension harder to defer. Anthropic, which has spent the past several months drawing scrutiny for its decisions around humanizing AI behavior, is now also the most valuable AI startup in the world by private-market valuation. OpenAI sits just behind it. SpaceX, whose roadshow reportedly begins around June 4, would be the first of the three to test public-market appetite. What price investors assign to Elon Musk’s rocket company will establish, de facto, a comparable for every AI listing that follows it through 2027.
The bull case is not imaginary. Ben Blume of Atomico, the third investor at the Athens conversation, made it with precision: SpaceX is a one-of-one company, he said, and its listing will capture a widespread imagination in a way that draws in investors who have never had direct financial access to the space sector before. More broadly, the liquidity events that mega-IPOs generate flow back into the next generation of companies. The compounding effect of the Google IPO in 2004 is still traceable in the current wave of AI founders. These are real mechanisms, not just promotional talking points.
But the FT data circulating among skeptics paints a different picture of what has happened so far. Even under best-case assumptions, analysts have estimated AI return on investment at deeply negative levels for most of the major technology companies — Microsoft around negative nine percent, Google around negative fifteen, Meta around negative twenty-eight, Oracle around negative thirty-five. Amazon barely comes out positive. None of these figures are audited, and all carry the methodological rough edges that any honest analyst would acknowledge. The fact that they are circulating widely, however, suggests a shift in the mood. When the spreadsheet starts to matter more than the vision, the easy phase of a boom has ended.
CNBC reported this week that hundreds of startups carrying premium valuations from the 2021 venture surge are now facing a quiet reckoning, as the AI boom resets expectations across entire categories of enterprise software. The companies that built workflow-driven SaaS products in that era, which embedded themselves in employee processes and often charged by the user, face a particular problem: the assumptions underlying their business models were not disrupted by a competitor offering a better price. They were disrupted by a technology that offers to eliminate the workflow entirely. Samir Kaul of Khosla Ventures put it plainly to CNBC: “The next generation of entrepreneurs, their coding language is spoken English.”
What the tokenmaxxing pullback actually measures is not the death of AI investment. It is the end of the first phase — the phase in which cost discipline was subordinated entirely to the imperative of being seen as an AI adopter. That phase created a lot of short-term revenue for Anthropic and OpenAI. It did not create an equivalent amount of value for the companies doing the adopting. As budgets tighten and procurement teams start asking harder questions about what exactly was produced, the pressure on the labs to demonstrate sustained enterprise value will compound.
Whether the IPOs scheduled for this year will precede or follow that reckoning remains the central open question. The Ken, citing analysts including Abhishek Pathak of Motilal Oswal, framed it as the end of AI’s venture era proper — a moment when the commercial logic of the foundation model business finally has to stand on its own, without the subsidy of tokenmaxxing campaigns and first-mover enterprise experiments. The labs that raised at trillion-dollar private valuations will need to explain, to public investors, how they sustain those valuations in a market that is learning, somewhat belatedly, to count.
—Inputs from Sputnik.
