Pot Committed

The Last Rational Bet

Anthropic disclosed a $47 billion revenue run-rate this spring. OpenAI cleared $25 billion. Both are still lighting money on fire to get there. OpenAI burns north of $27 billion this year, and the plan calls for that number to roughly double in 2027. Read that again. The plan is to lose more.

That sounds insane until you've sat at a poker table. There's a moment called pot committed: you've pushed so many chips in that calling the next bet is correct even when you're pretty sure you're beat, because folding forfeits everything already in the middle. So you call. Not because you'll win. Because the chips already in the pot say you have to.

That's the frontier. Half a trillion dollars in, the rational move is to keep shoving, even into a hand the open-source models are quietly drawing out on. DeepSeek and GLM run at a fraction of frontier cost and clear the bar for the boring 99% of what anyone actually does with a model. The frontier still owns a real quadrant: the closed-loop, verifiable, fast-turn work where being twelve months ahead pays off, like code and security and ad tech. That 1% is genuine. It just isn't a trillion-dollar market. You don't service a trillion in capex on the thin slice of work that has to be bleeding edge.

So who pays? Nobody, at least not the people writing the checks. We've seen this movie. In 1999 a fortune got spent burying fiber. Demand was real, it just showed up a decade late, and by then the companies that laid the cable were bankrupt and the next guys bought their glass for pennies. The track survived. The shareholders didn't.

Here's what makes it a tragedy instead of a con: everyone is acting rationally. The labs are pot committed, so they spend. The regulators look essential, so they gate. And the BIS, the central banks' central bank, just ran the math and got the same answer your gut does: when everyone competes this hard, the whole sector's net return goes negative. Rational players, predictable wreck.

Which is why the smart ones are sprinting to go public. If the hand can't be won, 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, you're the sucker. At this table, the sucker is whoever's holding frontier equity when the music stops.

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

At COAI today: the full Signal/Noise — the poker math, the 1999 fiber-bust rerun, and who's left holding the cards when the frontier goes public — is live at getcoai.com.

Outsider Labs Advisory →

$63 billion. That's what OpenAI is on track to burn in 2027, up from roughly $27 billion this year, against revenue near $25 billion. The company is scheduled to lose more than it earns, and to lose it faster. That's not a death spiral. It's a rational one. When you're pot committed to staying at the frontier, the correct play is to keep raising, right up until the day you cash out and hand the table to someone else.

Split your stack at the 1% line. Most teams route everything to one frontier vendor out of habit. Map your workloads: the verifiable, high-value 1% that truly needs the bleeding edge, and the 99% an open-weight model clears for a fraction. Today, send the bulk to open weights and reserve frontier tokens for the closed-loop work that earns them.

Don't underwrite the frontier's burn in your contracts. The labs are losing money on purpose to stay ahead, so their pricing is a land grab, not a steady-state cost. A multi-year commit at today's frontier rate funds their bet, not yours. Today, shorten the term, add a price step-down clause, and keep an open-weight fallback wired in.

If you invest, own the table, not the players. In the fiber bust the cable-layers went broke and the second owners got rich. The pick-and-shovel layer — power, memory, the landlords renting out compute — gets paid in the boom and the bust. Today, weight toward the infrastructure that collects either way and treat frontier-lab equity as the romance, not the position.

Monday, June 29

The labs' own bankers say the bet doesn't pencil. The BIS compared the $1 trillion AI capex boom to canal mania and the railway bubble, warning sector returns could turn negative. When Basel reaches for the bubble word, the grown-ups have already left the room. (Full analysis above.)

Ford rehired the humans it automated away. The carmaker brought back more than 300 veteran quality inspectors after AI couldn't match them. The model nailed the easy 80%; the 20% that actually ships a car still has a pulse and a paycheck. The last mile is the whole job.

Google told Meta no. Google rationed Meta's Gemini capacity harder than any other client, and is itself renting GPUs from SpaceX at roughly $920 million a month. When the most vertically integrated company alive is short compute, "just build more" is dead on arrival.

China closed the cybersecurity gap, for free. The Wall Street Journal reports GLM-5.2 now matches Anthropic's Mythos at finding software bugs, open and at a sliver of the cost. You can gate a model. You can't gate a number that already escaped onto a million hard drives.

One company burned $500 million in tokens in a month and couldn't say on what. A Zapier survey found 37% of individual workers don't know AI costs anything, even as half of managers spend $100k-plus a month. The meter's running and most of the org can't see the dial.

— Harry and Anthony

Sources: