What happened. On Friday, OpenAI cut the target on its funding round from $30 billion to about $20 billion, and the valuation it is asking investors for from $1.4 trillion down to $1.25 trillion. Some investors reduced their commitments. The first close is expected to come in under $20 billion. A consortium of UAE investors including the state-backed fund MGX is expected to be the largest participant, and as part of the deal OpenAI would gain the option to lease newly built compute capacity in the region. Wealthy individuals get a subsequent retail offering at the new, lower price.
The same day, the Nasdaq fell 2.9% — leaving it 7.8% below its all-time closing high from 6 October. More than $2 trillion of market value came off the Nasdaq 100 over the month of October alone. Google, Amazon and Meta shed roughly $900 billion between them across two sessions, according to the Financial Times.
Why it matters. Yesterday on this blog I argued that the market had changed its question — from is the demand real? to who is paying for the buildout? Friday was that question coming back as a number. The repricing did not stop at the corporate balance sheets where it started. It moved down one layer, into the private equity of the labs themselves.
That detail is worth dwelling on, because the order in which things repriced tells you more than the size of the move.
The stack repriced from the top down
AI is financed by a stack, and the stack has a fixed order of how quickly it can change price:
| Layer | The instrument | How fast it reprices |
|---|---|---|
| Frontier lab equity | OpenAI's private round | Days |
| Hyperscaler corporate debt | Alphabet's $28.7bn of notes; Microsoft's first US investment-grade bond sale in over nine years | Quarters |
| Neocloud credit | CoreWeave, Nscale, Firmus — loan books and credit default swaps | Continuously |
| Contractual compute obligations | Anthropic's $111.1bn with Google Cloud through July 2033 | Never, until someone renegotiates |
Equity is the softest layer. It reprices over a weekend: you simply say the number is smaller. That is precisely what happened on Friday.
But notice what that ordering means. The layer that repriced fastest is the layer carrying the most optionality, and the layer that reprices slowest is the layer carrying the most obligation. OpenAI can cut a round from $30 billion to $20 billion on a Friday. What it cannot do on a Friday is cancel the compute it has already contracted.
The duration mismatch is the real story
This is why the round size is not just a headline. Anthropic's S-1, filed on 26 October, lays the arithmetic out more honestly than any lab has before.
The revenue is short-cycle. Around 80% of it is consumption-based and billed per token. Contracts let customers cut usage or walk away with little or no notice. Only $10.8 billion of contracted remaining performance obligations was recognised within twelve months.
The cost is long-cycle and fixed. $111.1 billion with Google Cloud through July 2033. $110 billion with AWS through April 2036. $161.2 billion of Broadcom equipment leases. Up to $120.5 billion with SpaceX.AI at roughly $3.1 billion a month once fully delivered.
And the gap is already widening. Anthropic generated $1.70 of revenue for every dollar spent on computing and infrastructure in Q3, down from $2.08 in Q2 — an 18% deterioration in a single quarter, in the same quarter that revenue growth decelerated from 144% to 45%.
That is a duration mismatch. Short-cycle, repriceable, deflationary revenue funding decade-long, non-cancellable obligations. The arrangement only works if volume growth outruns price decline, indefinitely.
And here is the part that should worry you more than any market cap: these obligations do not reprice at all when sentiment changes. They reprice when a counterparty fails to deliver, renegotiates, or goes bankrupt. Anthropic's own filing notes that Amazon, Google and Microsoft are simultaneously its investors, its suppliers and its customers — and that some of them sell competing models. If one decides that selling you GPUs is no longer the best use of that capacity, "your contractual commitment" is a legal concept, not a guarantee of capacity.
The deflation is arriving from the open-weight side
Friday was not only about markets. On Thursday, Mistral released the weights of Large 4 — "Le Chonk" — a 1.05 trillion-parameter mixture-of-experts model with 52 billion active per token, a million-token context window, priced on Mistral's own API at $0.70 per million input tokens and $2.10 per million output. Artificial Analysis found its benchmark cost per task fell by 63%.
Now, the interesting part of that licence is not the "open" in open weights. It is the carve-out. The Mistral Ecosystem Open-Weight License is a modified MIT that permits commercial use, but requires a separate commercial licence from Mistral if you or your affiliates operate a Model as a Service — defined broadly enough to cover anyone providing inference or fine-tuning via API or hosted endpoint. Internal use is exempt. Reselling is not.
Read that for what it is: the labs are not conceding the API business to open weights. They are conceding the usage business and keeping the hosting business, which is where the margin sits. And in a market where the marginal customer's first question is now "what does this cost me per task?", the price of the frontier is being set at the bottom of the stack by weights you can download and run yourself — a 1.06 terabyte checkpoint on Hugging Face, around 610 GB in the NVFP4 build targeted at Blackwell.
This is the mechanism by which "raise prices" fails as a fix. If a frontier-class model is available at Mistral's API price, or as a self-hosted download, then the labs' pricing power is capped by somebody else's willingness to publish weights. And Mistral is not the only one. Cheaper Chinese models taking share is one of the reasons Anthropic's revenue-per-compute-dollar is falling.
There is a sharp irony here. Mistral also said Large 4 could give real attackers a boost, and that safeguard updates will reach its hosted API and future releases but not the checkpoints people have already downloaded. The lab is shipping the deflation at the same moment it declines responsibility for the deflation's consequences.
The counter-case, which is stronger than it looks
I want to give the bull case its full weight, because the lazy version of this post is a doom post and the data does not support one.
Demand is not the problem. Alphabet's cloud backlog reached $561 billion, Microsoft's commercial remaining performance obligation rose 73% to $677 billion, and AWS backlog was reported at $578 billion. That is over $2 trillion of contracted future revenue at three companies growing revenue between 39% and 103%. AWS grew 39.3% to $45.98 billion at a 38.7% operating margin. Google Cloud grew 103% to $30.7 billion with operating income up more than triple to $11.1 billion. The Gemini app reached a billion users. Bedrock token volume grew nearly seventeen-fold. This is not a demand story that has stopped.
Volume can outrun price. Anthropic's own evidence: when it cut pricing for a model launched last November, customer cost per task fell by about 40%, and revenue growth in the following quarter went from 48% to 132%. That is the Jevons paradox working in the labs' favour. If cheaper tokens create more demand than the price cut destroys, then "revenue per compute dollar" is the wrong statistic — the right one is revenue per task, and Anthropic's is going up.
Microsoft proves the buildout is financeable. Same week, comparable spend: $45.2 billion of capex, $18.6 billion of free cash flow — roughly $10 billion above consensus — Azure up 48%, and the stock up about 5%. The market is discriminating between balance sheets, not between AI and not-AI.
So the honest formulation of the bear case is not "the demand is fake." It is the one I keep landing on: the returns arrive later than the debt does. That is a financing problem, and financing problems have a specific shape. They resolve when the cost of capital crosses the cost of the obligation, when a covenant trips, or when the credit market stops rolling.
What actually breaks it
The market has been repricing the softest layer first, which is exactly why this still looks like noise rather than a crisis. It is not noise. It is a countdown with a visible ordering.
Here is the sequence I would watch, and it is specific:
- Anthropic's IPO prices. Marketing to investors from the week of 9 November, listing the week of 16 November, with Goldman Sachs and Morgan Stanley leading a 22-bank syndicate. If the book builds at $1.5 trillion and clears, the equity layer has a clearing price and the pressure releases. If it prices soft, the pressure stays on OpenAI and every private lab, because Anthropic's list becomes the only public reference point for what a frontier lab is worth.
- The October run-rate number. Anthropic left it blank in the S-1. Henry Blodget's calculation — month-to-month revenue growth falling from 88% in February to 4% in September — is the kind of series that either turns around or confirms the deceleration. A blank space in a prospectus is an invitation to imagine the worst number.
- The neocloud credit market. These companies need billions in outside capital and have no investment-grade ratings. Gartner's estimate is that 80% of neoclouds will be absorbed or cease to exist as independent companies by 2030. Firmus withdrew its IPO plans this month despite usage commitments from Meta and OpenAI. Nvidia is reportedly arranging billions in credit guarantees for a 460MW Nscale data centre Anthropic plans to use. That last one is the tell: a credit market that needs a chip vendor to guarantee the debt is not really a credit market.
- Hyperscaler capex guidance for 2027. Alphabet's finance chief has already said capex will increase significantly in 2027, from this year's $195–205 billion. Morgan Stanley forecasts the group runs roughly $550 billion of negative free cash flow next year. The question has already changed from "how much" to "what does that do to cash flow — before I ask what it buys."
None of those is a crisis yet. All of them are mechanical.
Where I land
I think the most useful thing to say about Friday is that it tells you where the pressure is, not that the boom is over.
The AI buildout is a financing arrangement with a specific shape: enormous front-loaded capital expenditure, revenue that arrives in arrears and reprices downward, and obligations to suppliers that get locked in years before the demand they are meant to serve has been proven. That shape has survived a hundred years of technology booms, because demand growth has historically outrun the cost curve. The question for this one is whether AI demand growth outruns financing costs, which is a much tighter constraint.
What changed this week is that the constraint stopped being theoretical. On Friday, the equity layer of a frontier lab repriced downwards for the first time in the cycle — and it did so because the market started asking where the money for the next tranche comes from, rather than whether the technology is any good.
There is a real possibility this resolves well. Backlogs of $2 trillion say demand is real. Microsoft's quarter says the capex can be financed out of operating cash flow by an operator with the balance sheet. Anthropic's price cut shows the volume response can be dramatic. If revenue per task keeps climbing and the cloud margins hold, the obligations get absorbed and the whole thing looks like an expensive infrastructure build that worked.
But notice what has to be true for that to happen. The labs have to keep selling tokens into a market where the price of a token is falling, in order to service commitments fixed years ago, with money that is getting more expensive. That is three things going right at once. It is not a crazy bet. It is a bet — and until Friday, nobody was pricing it.
Equity went first, because equity always goes first.
Sources: Financial Times via Reuters and Bloomberg Law on OpenAI's round; Reuters on the Nasdaq; Financial Times Markets on the $900bn Big Tech move; CNBC on capex guidance; CNBC and Fortune on Anthropic's S-1; The Irish Times on the October run-rate; Mistral on Large 4's weights and licence; VentureBeat via Artificial Analysis on cost per task.
Top comments (0)