Sam Altman has called the compute boom an "unsustainable absurdity," and the mechanism he describes for it unwinding is OpenAI's own progress. If the company cuts costs and improves efficiency fast enough, the expensive infrastructure projects being launched right now turn into bad investments. It is a warning about the industry's capital spending, delivered by a buyer who has already ordered capacity of his own.
Sam Altman on OpenAI's next model and the backlash against AI
Source: the-decoder.com
The second risk he names is more conventional and arrives sooner. A broad market downturn would make it harder for OpenAI to pay for the capacity it has already committed to. Altman considers that risk manageable.
On selling compute to other companies, he stopped short rather than saying no: OpenAI has no plans to do it, but he declined to rule it out. That is the sentence worth marking. A company that has ordered heavily and expects its own unit costs to fall has one graceful exit available, which is to become a supplier to everyone else. Leaving the door open is what a hedge looks like before anyone calls it one.
What these comments do not settle is whether his view of infrastructure has actually changed. Nothing in them says OpenAI is ordering less. The shift is in tone, and it is audible: he is more cautious now than he was before Anthropic chief executive Dario Amodei accused him of expanding compute on a "you only live once" basis.
Read plainly, this is a man positioning in public against both branches of his own bet. If efficiency arrives on schedule, the projects being financed today stop making sense, and he has now said so out loud. If the market turns first, OpenAI is paying for capacity it ordered under better conditions, and he has said that is manageable. Manageable is doing the heaviest lifting of any word in the set of comments, and it is the one piece of arithmetic he did not show.