The AI boom's easy money is gone — and only the labs can't stop paying

There was no earnings miss, no reversal of conviction, no announcement. The easy-money phase of the AI boom ended the way most administrative eras do: in a series of small instruments nobody called a turning point. The first arrived when Uber, which had mandated AI coding agents as an absolute necessity for all 30,000 employees ^[1], simply ran out of money for them — its entire annual AI budget gone within a few months ^[2]. Since then the dials have turned, one after another, widening outward across the chain. Accenture capped Claude at $29 per employee per month and cut off its heaviest users ^[3]. Microsoft reportedly restricted which third-party coding tools its engineers could use ^[2]. Alibaba banned Claude Code outright, classifying it as high-risk software after finding hidden tracking code ^[4]. The big banks put their AI agents on read-only, their security systems unable to tell a legitimate request from fraud ^[5]. Then the money itself moved: the neoclouds — CoreWeave, Nebius, Nscale — found themselves shut out of the high-yield bond market ^[6]. OpenAI cut its raise from a $30 billion target to $11 billion, with a potential retail offering to wealthy individuals to bridge the gap ^[7]. Anthropic walked its IPO target from $2 trillion toward $1.5 trillion, on a perceived revenue slowdown ^[8]. None of these caps is a verdict on whether the technology works. Accenture is the tell. Its own evidence showed heavy users delivering outsized value — a website built in three days for a few hundred dollars of tokens against a $60,000 outside quote ^[3]. The cap came anyway. It was an accounting decision — one that did not wait for the value question to be settled either way ^[3]. But there is one full stop in this litany. Everyone in it can turn a dial back down. The one spender assumed to be the biggest risk-taker cannot. Anthropic has signed for $518 billion of compute over the next decade, largely non-cancelable, alongside a $60 billion Broadcom equipment-lease package ^[9]. A $63 billion chip-leasing deal with Broadcom required the chipmaker's own chief executive to intervene personally before banks would sign ^[6]. That is what the tightening looks like — the gates are up across the chain. Anthropic signed papers that make tightening irrelevant to its obligations. Anthropic's disclosures put the mechanism in the plainest words available.

If our actual spend falls short, we must pay Google the difference. — Anthropic

That is not a hedge against a slowdown; it is a fixed cost a slowdown cannot reduce. And the slowdown stress is live: bulls project year-end annualized revenue above $110 billion, while SemiAnalysis puts it nearer $74 billion as customers like Meta shift to internal tools ^[9]. Whatever demand does, the bill is already owed. None of this is retreat. Azure grew 48% on rising enterprise-agent adoption ^[10]; hyperscaler cloud revenue rose 54% and S&P 500 profits 33% ^[11]; Anthropic's own quarter was $16.8 billion, up 45 to 50 percent sequentially ^[9]. The caps govern the rate of spend, not its direction. Which leaves one pocket with no gatekeeper above it. Households bought a record $840 billion of US equities in the first half of the year, at a record 55% of their financial assets — and Goldman flags exactly that demand as the vulnerable part of the market, just as 2027's public equity supply is set to be the largest since 2000 ^[12]. OpenAI's retail bridge is reaching for the same pocket in miniature ^[7]. The full vote comes in mid-November, when Anthropic prices its IPO after walking the target down on a perceived revenue slowdown, with analysts calling it the critical test of investor appetite for capital-intensive AI ^[8].

Instead of proceeding headlong with its IPO, the company should focus its energies on addressing these unprecedented risks. — Mark Levine

Every other dial in this chain has already been turned down. The household investor holds the last one, and it sits in their own hand.