Altman reportedly argued on the Cheeky Pint podcast that local AI inference creates a poor allocation of scarce DRAM. Local inference means running AI models directly on personal devices rather than through cloud servers. DRAM, or dynamic random-access memory, is the short-term working memory inside computers, phones, and servers.
According to that reported account, Altman called it quite terrible from a global-allocation standpoint for individuals to hold memory sitting unused roughly 95% of the time. He argued that pooled data-center memory could theoretically run at far higher utilization rates than consumer devices allow.
He also reportedly referenced an OpenAI-designed AI processor called “Jalapeno” and argued that the world shares an interest in using limited silicon efficiently. No primary source confirming the chip’s specifications or production status was available at the time of writing. Both the chip details and the podcast characterizations should be treated as reported rather than verified until a transcript or recording is confirmed.
Micron’s Exit Makes the Shortage Tangible
A major memory manufacturer is leaving the consumer market, and AI demand is the stated reason.
Micron announced in December 2025 that it would exit its Crucial consumer memory business. The company cited surging AI-driven demand, and said redirecting resources would allow it to support larger strategic customers in faster-growing segments, according to its investor press release.
HBM, or high-bandwidth memory, is a specialized chip built for AI accelerators. Reuters reported simultaneous pressure across DRAM, NAND flash, and HBM markets, affecting products from PCs to smartphones.
Micron’s decision reflects a manufacturer prioritizing enterprise and AI customers over consumers. It does not establish that consumer devices are the primary cause of the supply crunch. Capacity allocation, investment decisions, and corporate pricing strategy are also significant factors.
Who Benefits From This Argument
Altman’s utilization logic points toward centralized cloud services, which OpenAI sells.
OpenAI operates large-scale AI infrastructure and generates revenue when users access its models through cloud subscriptions. Users running open-source models locally on their own hardware represent a competing path that does not flow through OpenAI’s systems.
Altman’s utilization argument, whatever its technical merits, aligns with his company’s commercial interests. That does not make the argument wrong, but it is a relevant consideration alongside the engineering rationale.
Availability Has Value Too
A car parked in a garage is not a failed resource allocation.
Running models locally can improve privacy, preserve offline availability, and give users resilience when centralized services go down. Personal hardware that sits idle most of the day still delivers value by being ready when needed, much like a home appliance.
Concentrating memory in centralized data centers also concentrates decision-making power among a small number of companies. Users can become dependent on subscription terms and remote access policies they did not set and cannot change. These are ownership and policy questions, not purely engineering ones.
Who Gets to Decide How Resources Are Allocated
Utilization efficiency is one principle. It is not the only one.
Memory shortages are real, and manufacturer priorities are shifting toward AI infrastructure. Whether those pressures should also determine who owns and controls personal computing hardware is a separate question.
Accepting utilization as the governing principle for hardware ownership would shift significant computing power away from individuals and toward a small number of AI data centers. That shift is worth examining before it becomes the assumed default.
