Frontier Labs Pots are Committed

Anthropic told the world this spring it had crossed a $47 billion revenue run-rate. OpenAI cleared $25 billion. Two years ago those were rounding errors. Now they’re the two fastest revenue ramps in the history of software, and both companies are still setting money on fire to post them. OpenAI burns north of $27 billion this year. The plan, as reported, calls for that figure to roughly double to around $63 billion in 2027.

Sit with that for a second. The plan is not to lose less next year. The plan is to lose more, and to lose it faster, on purpose.

If you’ve never played serious poker, that reads as madness. If you have, it reads as a specific and well-understood spot. It’s called being pot committed. You’ve pushed so many chips into the middle that the math of the next call flips: even when you’re fairly sure you’re beaten, calling is correct, because folding forfeits everything you’ve already staked. So you call. Not because you think you’ll win the hand. Because the chips already in the pot tell you that you have to.

That is exactly where the frontier labs are sitting. Half a trillion dollars of committed capital is already in the middle. Walking away means writing all of it off. So they keep shoving — into a hand that the open-source models are quietly drawing out on, one cheap river card at a time.

The 99% already left the building

Here is the part the burn figures don’t capture on their own. The expensive thing the labs are racing to build is, for almost everyone, already free.

DeepSeek shipped DSpark this week and put it on Hugging Face for anybody who wants it — 60 to 85% faster inference, running across architectures it doesn’t even own. Zhipu’s GLM-5.2 sits in the top ten most-used models on OpenRouter, a platform that brokers more than four hundred of them. The Wall Street Journal reported this month that GLM-5.2 now matches Anthropic’s flagship Mythos at finding software vulnerabilities in some scenarios, at roughly fifteen percent of the cost. Open weights have become the daily driver. They clear the bar for the boring ninety-nine percent of what anyone actually asks a model to do.

The frontier still owns a real piece of ground. Tomasz Tunguz mapped it cleanly a few weeks back with his AI Problem Matrix: sort work along two axes, closed-loop versus open-loop, and infinite demand versus finite. The quadrant that prints money is closed-loop plus infinite demand — work where the machine can verify its own output and there’s always more of it to do. Software engineering lives there. Security scanning lives there. Ad optimization and cryptography live there. That’s the quadrant where being twelve months ahead of the open floor is worth paying for, because a test suite or an exploit count closes the loop without a human in the way.

That quadrant is genuine. It is also small. Most of the economy does not live there. Most companies run on lower-value or slower-turning work that either never touches the frontier or gets handled perfectly well by a model that costs a rounding error. And you cannot service a trillion dollars of capital expenditure on the thin slice of work that genuinely has to be bleeding edge. The math doesn’t bend that far. To return even ten percent on a trillion, the frontier needs something like a hundred billion dollars a year of durable profit. Today the two flagships, combined, run about seventy-two billion in revenue and lose money doing it. The gap between those two numbers is the whole story, and it is not closing on the timeline the capital needs.

Squeezed from both ends

It gets worse for the frontier, because the slice of work it can defend is being pinched from the other side too.

The open floor takes the commodity eighty percent. Fine. But the remaining twenty percent — the hard part, the edge cases, the operational reality that decides whether a thing survives contact with the world — turns out to still need a human. Jonathan Beard wrote a quietly devastating essay this week called “The 80% Problem,” arguing that AI relocated the old ninety-ninety rule rather than repealing it: the model gets you to a working draft with startling speed, and the last fifth, the part where the engineer actually lived, is still the part that breaks.

Ford just proved it with a checkbook. The company rehired more than three hundred veteran quality inspectors after its automated systems couldn’t match them. “Artificial intelligence is a fantastic tool,” its hardware engineering VP told reporters, “but it’s only as good as the information you use to train it.” Translation: we automated the easy eighty percent, the cars got worse, and we quietly went and found the graybeards we’d let go. OpenAI’s own data says the same thing from the inside — eighty percent of agent tasks now take more than thirty minutes, which is another way of saying the work that’s left is the work that’s actually hard.

So look at the box the frontier is in. The cheap floor eats the commodity majority. The expensive human still owns the high-value tail. What’s left for pure, sold-by-the-token frontier intelligence is a sliver in the middle — real, defensible, and nowhere near a trillion dollars wide.

The ceiling nobody priced

Now add the part that turns a bad business into a doomed one: politics.

For the frontier equity to pay off, AI has to do more than help. It has to eat the labor market — replace the employees, keep the companies, bank the margin. That’s the only version of the future where a trillion in capex earns its return. But run that forward and you get mass unemployment, and mass unemployment gets you political unrest, and political unrest does not let a handful of named American companies keep shipping the thing that’s killing taxpayers’ jobs.

You can already hear the first tremor. This week Apple’s Tim Cook said price increases were “unavoidable,” driven by a memory and storage shortage that AI demand is making worse. Apple raised prices on fourteen products by around twenty percent. Within a day, Alexandria Ocasio-Cortez was calling for Congress to break up the company — and, to her credit, she named the actual cause, pointing at the AI race and the energy strain of data centers. She diagnosed a physical-supply problem correctly and then reached for the antitrust hammer anyway, because that’s the only tool the populist kit carries. Breaking up Apple does not produce a single additional gram of high-bandwidth memory. The cure is aimed at a disease the patient doesn’t have.

Here’s the cruel asymmetry, though. When the political backlash to job loss finally lands, it can only grab what it can see and subpoena: OpenAI, Anthropic, named, domestic, regulatable. It cannot touch a GLM weight file sitting on a laptop in Ohio. The regulation will throttle the exact players who can’t do the mass damage, and it will sail right past the open and Chinese models that can. Anthropic and OpenAI will kill jobs. They just won’t be allowed to be the trillion-dollar mass executioner racking up monopoly margins while they do it. The floor does the politically explosive work and pays no political price. The frontier takes the heat for capability it can barely monetize.

Why none of this is a scandal

The temptation is to call this a con. It isn’t, and the honest version is more unsettling than a con would be. Everyone in the picture is behaving rationally, and rational actors produce predictable outcomes.

The labs are pot committed, so they spend — that’s correct play. The regulators want to look essential, so they gate — also rational. The bankers love a scarce product, so they cheer the gating. And the most striking confirmation came from the least excitable institution on earth. The Bank for International Settlements, the central bank of central banks, used its flagship annual report this week to line the trillion-dollar AI build-out up against canal mania in the 1830s, the British railway bubble of the 1840s, and the dot-com crash of 2000 — each a real breakthrough that drew more capital than returns could justify, each ending in recession. Using contest-theory modeling, the BIS found that as competitive pressure drives capital expenditure higher, the sector’s net surplus declines and can turn negative. That is the academic translation of a poker table where everyone is pot committed. Your gut and the world’s central bankers ran the same hand and got the same answer.

No villains required. As we wrote back on May 6 in No One Set Off My Evil Detector, the capital cycle doesn’t need bad actors; it just compresses principle into convenience until everyone’s self-interest crests on the same week.

The gate is the exit

Which finally explains the behavior that makes no sense any other way: why would a company beg to be regulated? Anthropic asks for it. OpenAI happily accepted a regime where the government has to approve who gets GPT-5.6. Companies do not normally lobby for a chaperone.

They do it when the chaperone is the product. “Our models are so powerful the government has to ration them through vetted experts” is the single greatest line ever written for an IPO roadshow, and you cannot buy it. A regulator has to grant it to you. The gate isn’t safety and it isn’t even a moat in the usual sense. It’s a credential — a state-stamped certificate of dangerous importance, worth more to a pre-IPO equity story than any benchmark score. The booming green head in the throne room tells the public the models are too strong to be sold freely. Pay no attention to the man behind the curtain timing the liquidity event.

And the reason it works is that it’s about sixty percent true. There really are sovereign reasons to gate Mythos for the people defending critical infrastructure. The best cover stories always carry a real payload. That’s what makes this one durable instead of laughable.

So watch what the smart money is actually doing. Anthropic is racing toward an $800 billion listing. OpenAI’s IPO chatter is everywhere. If the hand can’t be won — and the matrix, the floor, and the politics all say it can’t be won at the scale the capex demands — then the last correct move is to cash your chips before the river and let the index funds hold the cards. The old line from Rounders: if you can’t spot the sucker at the table in the first half hour, you are the sucker. At this table, the sucker is whoever’s holding frontier equity when the music stops.

We’ve seen this movie

None of this means the technology is fake. That’s the trap people fall into every cycle, and it’s wrong every time. The railways were real. The fiber was real. In 1999 a fortune got spent burying glass across the country, and the demand the builders promised did eventually show up — about a decade late. By the time it arrived, the companies that laid the cable had gone bankrupt, and a second wave bought their fiber for pennies on the dollar and built Google and YouTube and Netflix on top of it. The track survived. The shareholders didn’t.

That’s the shape here. The intelligence is real and it will diffuse into everything, mostly through the cheap floor, mostly to the benefit of the people using it rather than the people who paid to invent it. Arsalan Tavakoli of Databricks, who can see enterprise AI budgets as well as anyone, told a room this week that every software monopoly falls in the next twenty-four months. The capability compounds. The capital that chased the frontier gets impaired. Both things are true at once, which is exactly what a technological revolution financed as a bubble looks like from the inside.

So the move for anyone watching this with their own money on the line is the one that survives either ending: own the table, not the players. In the fiber bust the cable-layers went broke and the landlords got paid. The pick-and-shovel layer here — power, memory, the data-center operators renting out compute — collects in the boom and collects again selling the wreckage in the bust. Remember the tell from this week: even Google, the most vertically integrated company alive, with its own chips and its own data centers and its own power, is so short of compute it’s paying SpaceX something like $920 million a month to rent GPUs. The guy holding the megawatts gets paid no matter who wins the hand.

The frontier isn’t the safe seat. It’s the most expensive one, and the players who built it are already counting the exits.

---

This is the long-form Signal/Noise. The shorter email edition — the number, the three moves, and the five stories — went out this morning. As always: this is analysis, not investment advice. We’re describing how rational actors behave, not telling you which chips to push.

— Signal/Noise by CO/AI