Sam Altman has called it ‘quite terrible’ to distribute scarce DRAM across personal computers where, he says, it is used only occasionally. The OpenAI chief executive argued that keeping memory more fully occupied in the cloud would make better use of limited supplies.
Speaking to Stripe co-founder John Collison on the Cheeky Pint podcast, released on 29 October 2026, Altman challenged the idea that running AI locally would remove infrastructure constraints. His argument concerned how intensively hardware is used after it has been manufactured.
The remarks addressed resource allocation during a supply shortage. In that exchange, he did not propose taking memory from existing PCs or restricting personal computer ownership.
What Altman Meant by ‘Quite Terrible’
The comments came around 29 minutes into the interview. Altman was discussing local inference, which means running an AI model on a user’s own hardware.
‘A lot of people want local inference,’ he said. He questioned the assumption that running open-source models on personal computers would eliminate the need for infrastructure inside data centres.
‘But from a global allocation perspective, actually, it’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer,’ Altman said. He then described that memory as being used ‘5% of the time’.
‘You want maximum utilisation in the cloud if you’re trying to deal with this very supply-constrained environment,’ he added. His argument was that shared infrastructure could keep scarce memory working more continuously.
What the 5% Figure Does Not Establish
DRAM stands for dynamic random-access memory. It is working memory used by computers. Moving AI processing from a remote service to a personal machine changes where that processing happens, but it does not remove its memory requirements.
Altman’s reasoning was that allocating hardware to many occasional users could be less efficient than pooling resources. A shared service could handle requests from different users at different times.
However, he gave no source or methodology for the 5% figure in this exchange. He did not define which computers or workloads it covered. The figure should therefore be treated as his characterisation of underuse, rather than an established measurement of personal computers generally.
Nor did he present a technical comparison between specific PCs and AI servers. The discussion supplied no calculation of potential memory savings or benchmark showing which workloads would benefit most from cloud deployment.
Why Memory Makers May Hesitate to Expand
Collison had asked why high prices for memory and other computing components were not drawing more suppliers into the market. Altman attributed memory manufacturers’ caution to their experience of repeated booms and busts.
He described a business in which substantial investment in production facilities had not consistently produced attractive returns. In his account, manufacturers remembered approaching bankruptcy after building too many fabrication plants.
Collison suggested that this history made companies conservative about capital spending. Altman agreed. The exchange explained cautious investment, rather than a company-by-company examination of manufacturers’ finances.
That difference is significant because Altman was discussing two separate responses to constrained supply: producing more hardware and using existing capacity more intensively. His preference for cloud utilisation addressed the second.
OpenAI’s Plans to Expand Computing Capacity
Earlier in the interview, Altman said OpenAI was learning to build its own data centres. He also said the company intended to direct its robotics programme towards making more of them.
Collison questioned whether construction labour was the main obstacle, pointing instead to permitting and component constraints.
Altman responded that robots could eventually help produce more components. These were proposed routes to expansion, rather than evidence that the bottlenecks had been resolved.
Altman also acknowledged that computing markets could experience shortages, gluts and periods of excessive enthusiasm. Nevertheless, he expected demand for sufficiently capable AI models at sufficiently low prices to continue growing.
His remarks make his preference clear: expand supply while keeping scarce hardware busy.
What the interview does not establish is how much memory cloud deployment would save across real workloads, or when that advantage would outweigh the reasons users choose to run AI locally.