Sam Altman Calls Local AI Inference ‘Quite Terrible’ Amid Global DRAM Shortage

OpenAI CEO Sam Altman used a new podcast interview to call running AI models locally “quite terrible,” and the reason had little to do with model quality. From a global allocation standpoint, he argued, putting the world's limited DRAM supply into personal computers that sit idle most of the day is a poor use of scarce memory.

His framing was about utilization. In a supply-constrained environment, he said, you want maximum utilization in the cloud, yet many people's intuition runs the opposite way. He pushed back on the idea that open-source models and local inference mean data centers will no longer be needed.

The remark landed against a worsening global memory shortage. Asked how the search for solutions was going, Altman answered with a single word: “bad.” He also said the invisible hand of the market was not working properly this time, a striking admission given how much AI capacity OpenAI buys.

He was, however, sympathetic to memory makers. Chip manufacturers remember past cycles in which overbuilding fabs and believing “this time is different” nearly drove them into bankruptcy, so they are adding capacity cautiously. That caution helps explain why the shortage is expected to persist rather than clear quickly.

Altman's position is also self-interested. Local inference offers better privacy and lower marginal costs, and it lets users avoid paying per-token cloud fees, which cuts directly at OpenAI's business model. Nvidia and Microsoft, meanwhile, are promoting high-memory PCs and local AI hardware, setting up a genuine fight over where inference happens.

For now, the CEO's message is that centralized cloud inference is the efficient answer while memory remains scarce, and that DRAM should be treated as a shared global resource rather than a consumer upgrade. Watch the memory price cycle, on-device hardware shipments, and OpenAI's own infrastructure commitments.