Altman: pool scarce DRAM in data centers, not personal PCs

Altman's argument rests on a history lesson he sketched about the memory business: DRAM makers have spent heavily on capex over the years without securing adequate profit, in his example investing $20 billion and recovering only $19 billion. That boom-and-bust record, he said, is why manufacturers stay cautious about adding capacity even as AI demand hands them large profits.

The local-inference question followed from there. Altman conceded that many people want to run open-source models on their own PCs rather than depend on data-center equipment, then countered that from a global resource-allocation view, scarce DRAM scattered across individual machines is wasteful. He did not, however, offer the quantitative comparison that would settle the point, and local AI carries benefits he did not weigh: it works without an internet connection and keeps data off external servers.

The remarks sit alongside OpenAI's own position as a provider of data-center-hosted models, and its disclosed work on custom chips and automated data-center construction. Whether his efficiency claim holds beyond the 5% illustration remains unverified in the interview.