Altman vs your PC: running AI at home is a terrible waste

OpenAI CEO Sam Altman sat down with Stripe co-founder John Collison on the Cheeky Pint podcast, in an episode released on October 29. The talk was long and covered agents, math and data centers, but one short clip is what went viral.

In it Altman explains why he thinks local AI models are a bad deal for the world as a whole. Plenty of people assume that open models mean you can just run everything on your own computer and skip the giant data centers. He disagrees: DRAM is in short supply right now, and having bits of it sitting in every home PC, where it is busy about 5% of the time, is in his words "quite terrible". In the cloud the same memory runs close to full time, so with hardware this scarce, he argues, that is where it belongs.

The line the clip spread with, "a terrible allocation of scarce resources", is the poster's caption. Altman's own wording was "quite terrible", same point, softer phrasing. He said it while discussing why memory makers are cautious about building new fabs.

The local model crowd pushed back fast. The main counterpoints: the memory shortage was caused by AI accelerator buyers, since fabs are shifting capacity toward them, and idle hardware only costs its owner. On top of that, a model on your own machine can't get repriced overnight, hit a rate limit or be quietly swapped for another version.

So the man selling the cloud explained why owning is worse than renting from him. The logic holds up, it just happens to be convenient.