Sam Altman says local AI wastes scarce DRAM resources when PCs use it only 5% of the time - Big Tech and the digital economy

OpenAI CEO Sam Altman argues that running AI models locally is an inefficient use of limited DRAM compared with concentrating inference in data centers. His view contrasts with users who choose local models for privacy, offline access or specific needs.

Sam Altman discussed the allocation of computing resources during an appearance on John Collison’s Cheeky Pint podcast. The OpenAI CEO focused particular attention on local AI, where users run open models on their own computers instead of relying on cloud data centers.

Altman described local inference as a poor use of scarce resources. He argued that many people have computers equipped with DRAM that is used only around 5% of the time, while cloud infrastructure can concentrate that memory and keep it working more intensively.

According to Altman, the current shortage of memory makes that contrast especially important. He said that, while supply remains restricted, the priority should be maximizing utilization in data centers rather than distributing memory across individual home computers.

The comments came in a broader discussion about the shortage of computing capacity and the difficulty of expanding memory production quickly. Micron, Samsung and SK Hynix control global DRAM production, and the source text says TrendForce reported price increases of between 90% and 95% during the first quarter of 2026.

Memory manufacturers have announced new production facilities and have also shifted capacity toward HBM for data centers, reducing some consumer-focused output. However, expanding advanced memory production cannot be done immediately, which has contributed to the pressure on supply.

Altman also pointed to the need for new participants in the supply chain. OpenAI is developing custom chips with Broadcom, while companies including SpaceX, Amazon, Meta, Microsoft and Google are also working on hardware or infrastructure intended to address their computing needs. Even custom chips would still require DRAM, although they could allow systems to be optimized more closely for their workloads.

OpenAI’s position on local models adds another element to the debate. In August 2025, the company launched gpt-oss, its first open-weight models since GPT-2, designed to run locally. Altman had also said earlier that year that OpenAI had been wrong about open source.

Users who run AI locally may nevertheless have reasons unrelated to resource efficiency. The source text highlights privacy, offline operation and specific requirements as possible motivations, particularly for people who do not want to depend on a company such as OpenAI. From that perspective, the value of local inference is not measured only by how frequently the computer’s memory is used.

Sources: Xataka.