The correction can reprice the labs. It can't cancel what they've already signed.

Everyone arguing about the AI sell-off agrees on one thing: mid-November. Jim Cramer won't call a bottom in data-center stocks until several conditions line up, and one of them is a successful Anthropic listing — his test of whether this spending chain can ever be a profitable business ^[1]. The two loudest voices on whether the boom is a bubble at all, one calling it a credit-fueled oversupply and the other insisting enterprise demand is real, disagree on nearly everything except the trigger they're watching: the labs' IPOs ^[2]. The verdict is supposed to arrive on a date. So it's worth reading the document that date is meant to produce. In the fine print of Anthropic's offering materials sits the sentence that decides what a November verdict can actually reach.

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

That is take-or-pay: a promise to pay for compute whether or not you end up using it. And the sentence sits on top of a stack. Anthropic has signed roughly $518 billion of compute commitments over the next decade, much of it non-cancelable, on top of about $44.6 billion already agreed with the cloud provider Nscale and a $60 billion equipment-lease package with Broadcom ^[3][4]. Against that, the disclosures show $42 billion of net losses in 2025 ^[3]. None of this reverses if a hostile market shows up. A falling share price does not unwind a signed contract. One rung up the same ladder, the people building the machines are funding this year's bill from debt rather than profit. Amazon is spending roughly $220 billion on AI infrastructure this year while running free cash flow — the cash left after all spending — negative at minus $7.6 billion, and carrying $119.1 billion of long-term debt ^[5]. The next dollar is also getting dearer: Anthropic recently offered Blackstone yields above 10% to fund data centers, up from 8.5% four months earlier ^[6]. More expensive is not the same as unavailable, and the contracts carry no price-escape hatch. Now watch what the two biggest labs did as the market turned on them. In May, OpenAI was preparing an IPO for late 2026 ^[7]. In June it filed a confidential registration but began floating a delay to 2027 ^[8]. By mid-September it was fielding a $1.2 trillion proposal with no formal talks, and its own chief executive called a listing ill-advised ^[9]. This year was ruled out ^[10]. Then, on October 19, it was racing to close a $30 billion private round at more than $1.4 trillion — timed, the reporting says, to lock in long-only and sovereign-wealth money before Anthropic's disclosures print the first side-by-side comparison of the two firms ^[11]. Five months, one direction: away from the public market, into private money. That market hasn't stopped pricing the labs — it just reaches them sideways. CoreWeave, the public cloud provider, hit a one-year low on debt-financing worries, and the decline dragged OpenAI's secondary-market valuation — what its private shares trade for among investors — below $1.5 trillion ^[12]. Anthropic arrives with backstops in place: its listing was pushed from October to November, reportedly so it could show two consecutive quarters of positive free cash flow, and Alphabet, holding about 14%, and Amazon are expected to support the valuation through the process ^[10][3]. None of this is an exit. It's a sorting. The same week OpenAI's mark fell below $1.5 trillion, Anthropic's rose 2.4% above $2 trillion ^[12]. And the managers who sat out the trade on return-on-investment skepticism haven't been vindicated: 70% of active managers underperformed through early October simply for being underweight AI ^[13]. The market is choosing which names can absorb the spending, not voting on whether the spending happens. So where can the correction actually cut? Look at what hasn't been signed. Anthropic has roughly $44.6 billion already committed to Nscale — and another $45 billion still under negotiation ^[4]. That un-signed piece is the first big chunk of the buildout a hostile price could genuinely stop, and it's the measure of the correction's reach. Everything behind it is locked. Everything in front of it is still up for grabs. The date everyone agreed to wait for is real, and it will move shares. But read the fine print and the test shrinks. A mid-November listing can punish the labs' investors, squeeze the suppliers, and freeze whatever hasn't been committed yet. What it cannot do is cancel the spending that has been. The bill arrives whether or not the demand shows up — that's what the sentence says — which leaves the market's verdict grading the buildout only at its uncommitted edges.