Altman calls personal DRAM a poor use of scarce memory in local-AI debate

Sam Altman thinks the DRAM in your laptop is a poor use of a scarce resource. Speaking in an interview on the YouTube channel Cheeky Pint, the OpenAI CEO argued that, from a global allocation standpoint, it is “quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer and being used five percent of the time,” according to Tom’s Hardware, as relayed by Breitbart on November 1.

The remark came about 29 minutes into the conversation, when the discussion turned to local inference, meaning running open-source models on a personal computer instead of calling a cloud service. Altman framed the issue as supply and demand: data centers can keep memory busy, while personal machines cannot.

“You want maximum utilization in the cloud if you’re trying to deal with this very supply-constrained environment,” he said. “And somehow many people’s intuition is the opposite of that.” That reframes the local-AI argument from whether small models are good enough to where scarce memory should physically sit.

The context is a real memory crunch. Breitbart notes that RAM prices have been surging as a global shortage worsens, with hardware makers including Apple and HP facing higher component costs. Against that backdrop, leaving DRAM in machines that use it a fraction of the time looks like idle capacity.

The position also tracks OpenAI’s business. Its main models run in data centers, even though the company has released some open models that users can run locally; Breitbart notes that Altman is, in effect, arguing for the infrastructure his company depends on most. That is why a passing interview comment travels: it marks the dividing line between local open models and cloud-hosted closed ones.

If memory is prioritized for the cloud, the cost advantage of running models locally could shrink further, leaving open-model inference on personal hardware closer to a hobbyist experiment than a mainstream option. Centralization would also keep inference capacity in the hands of a small set of infrastructure providers.

It is worth being precise about what this is: an off-the-cuff answer inside a long interview, not a product, pricing or policy announcement, and it arrived with no timetable. The things to watch are OpenAI’s actual moves on open weights and local deployment, and whether the memory shortage turns centralized inference from one executive’s preference into an industry default.