Three weeks ago, Nvidia posted the best quarter in its history and the market yawned. Since then, roughly three trillion dollars of technology market value has gone missing, CoreWeave has fallen by a third, and Anthropic has filed the paperwork to sell $100 billion of itself to the public. Nobody rang a bell. There was no single headline. Just a slow, grinding realisation that the money side of AI and the product side of AI had drifted very far apart.
That gap is now visible in a public document. Anthropic's S-1, filed on 26 October 2026, is 330 pages long and about 90 of them are risk factors. You can read it as a thriller about rogue models and existential risk, and plenty of people have. Read it as an operator and it says something more useful and more uncomfortable: the subsidised era of AI has a balance sheet, and the balance sheet has a date on it.
What actually broke: the financing loop
Start with how a GPU gets paid for, because the mechanism matters more than the ticker.
Nvidia sells the chip. You cannot afford the chip. Nvidia invests in you so you can buy the chip, and you borrow against the chip you have not yet bought to cover the rest. Normally no bank takes depreciating silicon as collateral. It works here because Nvidia has effectively promised a floor: if nobody rents the capacity, the capacity gets rented anyway. Nvidia's money buys Nvidia chips that get rented back to Nvidia, which lifts Nvidia's stock, which funds more investments in more companies that borrow more money to buy more Nvidia chips.
Wall Street calls this circular financing. It holds up beautifully as long as three things stay true at once: the chips rent for more than they depreciate, the labs keep buying, and somebody keeps lending. The third one stopped being true in the second week of October, when it became clear that CoreWeave, carrying about $52 billion of debt and lease obligations, was paying close to 15% on new money. Credit default swaps on that debt hit 1,115 basis points, roughly double where they had been. A company paying 15% to fund assets that lose value every quarter is not a company with a business model. It is a company with a schedule.
Then Elon Musk said out loud what the bond market had been saying quietly: debt markets have largely shut for companies whose only revenue is AI data centres, and he expects more than 200 neoclouds to die. Gartner's estimate is gentler and, frankly, scarier: about 80% of neocloud providers cease independent operations by 2030.
Notice what did not happen. Nobody said the chips are useless. Nobody said inference stopped working. What broke was the financing layer wrapped around the compute, and that layer was propping up the price.
Read the S-1 like an operator, not an investor
The interesting number in Anthropic's filing is not the valuation, which would put it above 492 of the S&P 500's 500 companies. It is the commitments.
The filing itemises future compute and infrastructure spending in the range of $542 billion to $567 billion, with tallies that include negotiated capacity crossing $750 billion. Some of it is astonishingly specific: roughly $3.1 billion a month to SpaceX for access to Nvidia GPUs once planned capacity is delivered, up to about $120.5 billion by 2029, plus at least $114.5 billion to Google through mid-2033 and $110 billion to Amazon through 2036. The vast majority of it is non-cancellable or payable whether or not the compute gets used.
Now put that next to the operating picture: a net loss in the tens of billions for the first half of 2026, extraordinary revenue growth, gross margins that look healthy only because compute is currently priced below its replacement cost, and five customers accounting for roughly a quarter of revenue.
Here is the part most commentary skips. A company with non-cancellable $3.1 billion monthly obligations cannot run a permanent discount. Subsidised tokens are a customer acquisition cost, and customer acquisition costs get retired. Every founder building on top of these APIs is, whether they have noticed or not, standing on a price that is scheduled to change.
The subsidy has an expiry date
This is the practical consequence of October, and it has nothing to do with whether you believe in AGI.
For two years, the dominant economic fact of building software has been that intelligence was underpriced. Frontier tokens arrived at a fraction of their cost to produce because someone else's venture capital and someone else's bondholders were paying the difference. That was fantastic while it lasted, and it produced a generation of products whose unit economics only work at the subsidised price.
When the subsidy normalises, three things move at once:
Price. Inference gets more expensive per token, or at least stops falling, while capability keeps rising. Both directions are real, and they pull different line items.
Concentration. The provider you integrated is the same provider whose refinancing you now read about in the news. Uptime becomes a credit event, not an SLA.
Consolidation. Some of the capacity you depend on gets bought, repriced or switched off. The neocloud that hosted your fine-tune is exactly the kind of company Musk was talking about.
The other side of the ledger, honestly
It would be lazy to call this the end of AI. The day after OpenAI disclosed a model that appears to have settled a 27-year-old conjecture at the base of complexity theory, "bubble" started sounding like wishful thinking. The capability curve has not bent. The rails from the railway bubble were still there after the railway bubble, and they carried the next fifty years of traffic.
Two things are also pushing prices down at the same moment. Open-weight models are landing within a point of frontier quality at half the price, and the hardware that was financed at 15% is going to end up on the secondary market at a discount. Cheap capacity and cheap weights are real deflationary forces, and they will outlive this quarter's credit stress.
So the honest position is this: AI is not a bubble in the sense of a technology that does not work. It is a bubble in the sense of a financing structure that was always going to be unwound, and the unwinding is now in progress. Builders do not need to predict the bottom. They need to stop pricing their products as if the subsidy is permanent.
What to do before your 2027 budget
The aratech takeaway
The roughest three weeks in AI's short history did not damage the technology. They removed a subsidy and revealed a cost structure. That is not bad news for people who build things that work. It is bad news for decks that were priced on someone else's balance sheet.
The winners of 2027 will not be the teams with the biggest model budget. They will be the teams that know exactly what each AI feature costs to run, that can move a workload between providers in an afternoon, and that kept the parts of the system where their own judgement compounds. That is unglamorous engineering. It is also the only kind that survives a refinancing.
If you want to know what your AI stack actually costs, and what it would cost if the subsidy ended tomorrow, we run that audit in five days and hand you the numbers in priority order. Talk to aratech.