Priced on a year of data. The labs have enough model. What they don’t have yet is time to prove the customer.

Start with what powerful institutions did, not with what they said. The White House asked OpenAI to hold back GPT-5.6, releasing it only to vetted partners on a customer-by-customer basis. Two weeks ago the administration pulled Anthropic’s strongest models offline; by the weekend it had authorised Mythos 5 back for roughly a hundred vetted infrastructure operators — but not Fable, the public-facing variant, and not general release. What looked then like a quarrel with one company now looks like a rule: the state inspects the US frontier before the public gets it, while cheaper Chinese open-weight models remain broadly available.

Then look at what Alibaba allegedly did to get the same capability the cheaper way. Anthropic claims the Chinese tech giant ran some twenty-five thousand fake accounts to pull Claude’s reasoning out through the front door. Put the two events together and the logic is plain. You do not inspect a toy before release. You do not run a year-long heist to copy a dud. The clearest evidence the frontier moved this week is not a benchmark score. It is what serious people did to control who gets to stand on it.

And yet the labs themselves stopped claiming the model is the constraint: the capability overhang. The field, it says, is past the experimentation phase. The model, in other words, is the part that is largely finished. Which moves the whole contest somewhere harder to win — and much harder to prove.

This is the heart of it. The labs grew on a bottoms-up motion — a consumer falls for the product, brings it to work, and adoption percolates up into the enterprise. That motion is real, and the spending caps prove it is also expensive: Amazon, Walmart, Cisco, Uber, and Meta all reined in employee AI this week as token costs bit, one executive saying flatly they had created a monster. But a capped pilot is not a committed enterprise. The question that decides the valuations is whether bottoms-up enthusiasm converts into the thing underneath them: long-term licences, signed at a known price, by a CFO who has rebuilt the company’s stack around the bet.

Picture that CFO. A year ago the price per million tokens was one number; today it is a fraction of that, and the pricing model has changed three times in eighteen months. She cannot model even a two- or three-year commit against a cost that still won’t sit still. And she faces a genuine fork. She can rebuild the whole stack around AI, the enormous-addressable-market future the valuations assume, or she can deploy it in the pockets where the efficiency is already obvious and wait for the rest to settle. Those two choices size the market an order of magnitude apart, and right now nobody knows which one is real. Faced with that, the conservative move is to run a pilot, cap the spend, and not sign a fixed-price renewal she may regret at the next pricing change. Multiply that hesitation across every large buyer, and the “signed enterprise revenue” beneath the valuations is softer than the headline numbers suggest.

The price routing is no longer hypothetical. Coinbase’s chief executive said the exchange halved AI spend while token use kept climbing, by defaulting its internal gateway to Chinese open-weight models — Kimi and GLM — alongside smarter caching and routing. Engineers can still pick any model; most never hit their old caps anyway. For a regulated financial institution, that is a deliberate trade: token economics over the compliance and security questions that come with Chinese-origin models under Beijing’s intelligence rules. The same week Snowflake’s chief executive benchmarked GLM against Opus on more than a hundred enterprise coding tasks, found them neck and neck at three attempts, and said the team wanted to make the Chinese model available to customers; Microsoft was still weighing DeepSeek for Copilot. Price has become the buyer’s primary lever — even as Washington tightens the gate on the US frontier.

That is also why the profitability claims deserve a careful read. Anthropic pointed to a profitable quarter; it also declined to promise another. Both labs are reportedly weighing token price cuts to fight each other for enterprise share, and rational firms do not start a price war from strength. A price war is what you run when growth isn’t arriving fast enough to cover the commitments.

Power is the constraint nobody disputes. Not chips. Compute and silicon are scaling, with new entrants and capacity arriving constantly. Power is the one that isn’t keeping pace, which is why every build-out now ships with its own generation attached: Microsoft and Chevron’s 2.67 gigawatts off-grid in West Texas, Washington’s US$17.5 billion in guarantees for reactors that won’t switch on until 2030, the FERC fight we tracked last week. And it is the one constraint where the West is structurally behind, because China is scaling generation far faster. That is the agreed number-one problem. But notice what solving it does not solve. You can pour the concrete in Texas and switch on the reactors and still not know whether the customer converts. Power is necessary. It is nowhere near sufficient.

Now sit the spend against the proof. The spend is staggering and visible: roughly US$160 billion of this year’s capex is borrowed, while AI revenue across the big three has only just edged past depreciation. Oracle alone cut twenty-one thousand jobs this week while planning to raise up to US$50 billion, half of it debt, to build capacity for the labs. That is the scale of the bet.

So how big does the payoff have to be to justify it? We won’t pretend to model that precisely — the inputs are too soft, and the labs’ own projections are the optimistic self-report this whole issue says to question. But the order of magnitude is enough. To clear spending on this trajectory, the three leading labs would together need to earn, within five years, revenue on the scale of an entire top-tier software category — and to do it while cutting prices to win the share.

Is that impossible? No. Cloud computing reached that scale, and looked just as unprofitable on the way. But it is an enormous claim, and it rests entirely on the one thing the layer above showed has not happened yet: the customer committing.

Which is where the clock comes in. OpenAI leaned toward 2027; Anthropic had filed but had not confirmed a matching delay by window close. The reason they point to is volatility, SpaceX floated and slid, but the comparison gives the excuse away. SpaceX is a full-stack, revenue-proven business: launch, Starlink, a decade of milestones, several lines that have already paid off for investors. A wobble in its debut says nothing about whether its model works. The labs are something else: magnificent, but single-product, and thin on the customer data that would prove the conversion. The same volatility cannot mean the same thing for both. The simpler read is that the labs need time the IPO would spend for them: time to turn pilots into contracts, to let pricing settle, to show the books once the books tell the story they are betting on. Power they can buy. That time, they cannot.