The CEO who warned rivals about reckless spending signed half a trillion in eleven months. How to read the number without getting played.
In early 2026, Dario Amodei was on record warning the industry about investing too fast — the kind of measured caution you’d expect from the safety-first lab. Eleven months later, per The Decoder, Anthropic has reportedly signed up to $517 billion in compute deals. OpenAI’s line runs to roughly $750 billion through 2030. Sam Altman, who leads on spend, now frets aloud about “unsustainable silliness.”
Everyone is warning about the fire while pouring the accelerant. I’ve watched this movie four times — dot-com fiber, mobile, crypto, cloud — and the numbers always arrive pre-inflated for headlines. Before you repeat $517 billion in a planning meeting, run the checklist.
1. Subtract the tense.
“Signed” is not “spent.” These are committed capacity reservations — multi-year contracts for GPU clusters, data-center power, and cloud allocation — not cash out the door this quarter. The $517 billion is a ceiling stretched across years, and ceilings have clauses. Treat it as a spending intention, because that’s what it is.
2. Find the denominator nobody prints.
The headline gives you the commitment and hides the revenue. Neither Anthropic nor OpenAI generates income within shouting distance of these figures. That gap isn’t a rounding error — it’s the whole story. This is AGI-race positioning dressed as capex. A factory commits capital against orders it can forecast; these labs are committing against a capability they hope to reach.
3. Ask what a “compute contract” actually locks in.
In practice it’s some blend of reserved accelerator clusters, guaranteed power and floor space in facilities that may not be built yet, and preferential cloud capacity. The value depends entirely on delivery dates, unit economics, and whether the hardware generation you reserved is still competitive when it ships. A dollar committed to 2029 silicon is not a dollar of anything today.
4. Read the warning as a tell, not a principle.
Amodei’s caution and Altman’s “silliness” line are the same move from opposite chairs. When a frontier CEO publicly worries about overspending, check whether they’re spending anyway. Both are. The public caution is competitive theatre — a way to sound responsible while the contracts get signed. Don’t call it dishonesty. Call it what it is: nobody in this race can afford to be the lab that under-provisioned compute and got left behind.
5. Weight the pressure, not the personality.
The hypocrisy framing is satisfying and mostly a distraction. Amodei didn’t change his mind about risk; he changed his read of the game. If your competitor reserves the world’s GPU supply and you don’t, their model trains and yours waits in a queue. Capacity is the constraint that decides who ships next-gen first. That structural fact, not any executive’s flip-flop, is driving both numbers.
6. Watch the exits, not the entrances.
The signing is the easy and loud part. The number that will actually matter is the first quietly renegotiated or unwound commitment — the tell that the capacity got booked faster than demand materialised. Every prior buildout cycle ended not with a bang but with a footnote in a filing. Set a reminder for six months out and watch for the walk-back.
If you do one thing: stop quoting $517 billion as money and start quoting it as a bet. It’s a wager that whoever holds the most reserved compute gets to AGI-adjacent capability first, and that the revenue shows up before the invoices do. Maybe it does. But a commitment is a promise to pay, and the frontier labs are now making promises an order of magnitude larger than anything their income statements can cover. The warnings were the honest part. The contracts are the answer.
