The bet against itself

Part one of three on what happens to the market as intelligence commoditises. Next: cheap intelligence makes the incumbents richer, then the internet goes headless.

Two things are true in June 2026, and they should not be.

The first: the price of intelligence is collapsing. A fixed unit of capability gets about fifty times cheaper every year, and roughly two hundred times cheaper since the start of 2024. Gartner expects inference on a trillion-parameter model to cost providers 90% less by 2030 than it does today. The commodity is in freefall, on a schedule nobody controls.

The second: the companies that make the commodity are going public at the largest valuations in the history of private enterprise. Anthropic filed its S-1 on the first of June, with secondary shares trading north of nine hundred billion dollars. OpenAI restructured into a public-benefit corporation that values it around five hundred billion. SpaceX absorbed xAI into a single entity worth $1.25 trillion, now listing at closer to $1.75.

Put those side by side and the obvious question writes itself. CNBC asked it in May: if the product deflates fifty times a year, why is the market pricing the producers like monopolies? Cheap intelligence should be the thing that derails these IPOs rather than the thing that funds them.

The answer is that the market is not pricing the model. It is pricing what each lab does with the model once the model itself is worth nothing. Every one of these valuations is a bet against the thing the company is famous for.

Say the commoditisation claim out loud

It is worth being precise, because “intelligence is becoming free” is the kind of line that sounds rigorous and falls apart on contact.

What is actually falling is the price of yesterday’s capability. The cost to serve GPT-4-level reasoning has dropped about forty times a year since it shipped. Distillation keeps the open-weight models roughly three to six months behind the closed frontier, and that gap has held release after release. So the floor drops on a predictable clock: whatever was state of the art half a year ago is now cheap, and trending toward the cost of the electricity to run it.

The frontier itself is a different story. The best models are not getting cheaper to reach; they are getting more expensive to use well. An agentic task now spends five to thirty times the tokens a single answer used to, because the model plans, calls tools, checks its work, and tries again. Per-token deflation and per-task cost are pulling in opposite directions. The token is a commodity. The capability sitting at the very front is still scarce, and stays scarce for about six months at a time.

So the honest version of the thesis is narrow and still devastating: the moat on any given level of intelligence has a half-life of roughly six months. You cannot build a trillion-dollar company on an asset that depreciates that fast. Which means none of these companies is trying to.

Three companies, three bets, no models

Read the three IPOs as three different answers to the same question: what do you own once the weights are worthless?

Anthropic is betting on the enterprise. Its revenue went from about nine billion at the end of 2025 to a thirty-billion run-rate by April, and roughly four-fifths of that is enterprise. The single fastest line is Claude Code, an Anthropic product, at a two-and-a-half-billion run-rate of its own. That is the tell. Anthropic is not being valued as a model vendor. It is being valued as the company that wrapped the model in the reliability, the integration, and the accountability a business will actually pay for, and then sold that wrapper to a thousand enterprises paying over a million a year each. The model is the cheapest part of the thing they sell. The valuation rides everything bolted around it. (Worth noting the number is contested: OpenAI argues the thirty billion is closer to twenty-two once you net out cloud-resale accounting. It does not change the shape of the bet.)

OpenAI is betting on the habit. Five hundred billion dollars is not a claim about model quality, because model quality is the one thing OpenAI cannot keep. It is a claim about attention: that ChatGPT is now a verb, a daily default for hundreds of millions of people, and that defaults are sticky even when the thing behind them is replaceable. It is the same moat Google has owned for twenty years, except OpenAI is trying to manufacture in three years the distribution that Google was handed by inertia. Whether that holds is the open question in the whole sector.

xAI is betting on the physical layer. It has neither Anthropic’s enterprise book nor OpenAI’s consumer habit. It lost $6.4 billion in 2025 and is spending at a thirty-billion-a-year capex pace. So the bet is not on product at all. The merger with SpaceX is the strategy stated plainly: when you cannot win on what the model does, you win on owning the compute and the balance sheet to outspend everyone who can. The valuation is a bet on energy, silicon, and a captive trillion-dollar parent that never has to ask the public markets for permission. It is the most honest bet of the three, because it skips the model entirely and buys the ground it runs on.

The inversion

None of the three is priced as a model company. The model is the loss leader in all of them. The market has already agreed with the uncomfortable conclusion: the weights are not the moat, and everyone setting these prices knows it. That is why the valuations sit on top of enterprise lock-in, consumer habit, and raw compute instead of on benchmark scores.

There is a twist here for anyone who, like me, has spent two years arguing that the value migrates off the model and into the layer above it. It does. But it is not leaking out to a thousand application-layer startups picking up the commodity and reselling it. The labs watched that happen in slow motion and moved up the stack themselves. Claude Code is the proof: the most valuable application running on Anthropic’s models is Anthropic’s. The frontier labs are not the commodity producers about to be disrupted by their customers. They are the ones racing to stop being commodity producers first.

Each of these companies is worth a trillion dollars precisely to the degree that it can stop being the thing it is famous for, before its models are free. That is the bet. The IPO is just the market deciding which of them is most likely to win it.