- Published on
OpenAI's Altman on memory shortage: Global DRAM should be prioritized for data centers...
- Authors
- Name
- DeepHorus
On October 30, OpenAI CEO Sam Altman, appearing on the YouTube podcast Cheeky Pint, said that against the backdrop of tight global DRAM memory chip supply, memory resources should be concentrated in data centers to improve utilization rather than dispersed into personal computers. In the interview, he focused on arguing the necessity of reducing computing costs and viewed memory allocation as a key link in the AI industry's computing power supply issue.
This statement comes against the backdrop of the rise of localized AI inference. As open-source models continue to improve, more and more users want to run AI models directly on their own computers to reduce dependence on cloud services. In the view of some users, this can both lower usage costs and ease pressure on centralized computing power.
Altman gave the opposite judgment. He said the way the supply chain works is "very counterintuitive," and if everyone occupies a large amount of limited DRAM in personal computers, while that memory may be used only 5% of the time, from the perspective of global resource allocation this is "actually very bad." He believes most people's intuition is exactly the opposite of the correct approach to dealing with severe supply constraints.
He argued that under limited memory supply, cloud resources should be utilized to the maximum extent, and limited resources should be used more efficiently through centralized deployment and higher utilization. This view continues his long-standing emphasis on reducing AI computing costs.
However, there are obvious flaws in this logic. The vast majority of people buy computers not merely to run AI inference locally; running games, professional software, multitasking, and the operating system itself all require system memory. The purpose for which PC users buy memory and the purpose for which cloud AI service providers purchase memory are not exactly the same, and the DRAM demand of the two cannot simply be regarded as mutually substitutable.
Judgments from the memory industry confirm the supply pressure. Samsung Electronics expects that, driven by AI infrastructure construction, tight memory chip supply may last until 2028. On its Q3 2026 earnings call, the company said that intensifying competition in AI infrastructure construction and the spread of personal AI agents are expected to continue driving memory demand growth.
Because new capacity is unlikely to significantly increase memory supply in the short term, tight supply is expected to continue until 2028, and consumer electronics will face memory cost pressure that is difficult to ease in the short term. The gaming hardware market has already seen price increases and tight supply.
Against the backdrop of simultaneous expansion of AI computing demand and consumer electronics demand, how to allocate limited DRAM production capacity between data centers and end devices will become an issue of ongoing contention among all parties in the industry chain. Altman's latest statement also reflects that AI companies' sensitivity to computing resource costs is rising.
