OpenAI CEO Sam Altman has argued that keeping large amounts of RAM in personal computers is an inefficient use of the world’s limited DRAM supply, particularly when that memory is used only occasionally. Speaking on the Cheeky Pint podcast on October 29, 2026, Altman suggested that centralized cloud data centers can make better use of scarce memory resources than individual PCs running AI models locally.
The comments came during a discussion with Stripe co-founder John Collison about rising memory prices, AI infrastructure demand and the semiconductor industry’s ability to expand production. Altman specifically challenged the assumption that running open-source AI models on personal computers would reduce the need for large data centers.
“From a global allocation perspective, actually, it’s quite terrible,” Altman said while discussing the idea of people keeping significant amounts of the world’s limited DRAM supply in their own computers.
He suggested that memory in personal machines might be used just 5% of the time, arguing that cloud infrastructure could achieve higher utilization by serving multiple users and workloads.
“You want maximum utilization in the cloud if you’re trying to deal with this very supply constrained environment,” Altman added.
His 5% example was part of an argument about resource allocation, not a documented measurement of average PC memory utilization. The interview did not establish how much consumer RAM typically sits unused or provide comparative data demonstrating that cloud deployments are more efficient across all workloads.
Altman’s argument centered specifically on local AI inference, where users run models directly on their own hardware rather than sending requests to remote servers. He questioned the assumption that the availability of open-source models would automatically eliminate the need for centralized computing infrastructure.
That distinction matters because personal computers use RAM for considerably more than AI inference. Gaming, multitasking, content creation and other applications depend on locally installed memory, regardless of whether their owners use cloud-based AI services.
For gaming PC enthusiasts, the discussion also comes amid substantial pressure on the memory supply chain. SK hynix has warned that memory supply shortages could persist through 2030 as AI infrastructure demand continues to strain production capacity.
The company has also indicated that DRAM and HBM supplies are heavily committed through 2029, highlighting the limited availability of additional production capacity.
Although high-bandwidth memory used in AI accelerators differs from conventional DDR5 memory installed in gaming PCs, both are part of the broader DRAM manufacturing industry. Decisions about production capacity and investment can therefore influence the availability of memory products across different markets.
During the podcast, Collison pointed to high memory and GPU prices and questioned why manufacturers had not responded more quickly by expanding production.
Altman acknowledged the problem, arguing that memory manufacturers remain cautious because of the industry’s history of expensive factory investments followed by periods of oversupply and weak returns. He suggested that this experience helps explain why producers might hesitate to invest aggressively despite currently favorable market conditions.
The supply pressures have also prompted major technology companies to pursue longer-term procurement arrangements. Alibaba, Tencent and ByteDance have reportedly sought long-term DRAM supply agreements as competition for memory capacity intensifies.
Altman’s comments highlight a fundamental difference between local and cloud computing. Running AI models locally gives users greater control over their hardware and can allow supported workloads to operate without an internet connection or recurring cloud inference charges. Cloud systems, meanwhile, can distribute shared computing resources across customers rather than requiring each user to maintain dedicated hardware for occasional AI tasks.
However, greater utilization does not automatically establish that cloud computing is the better option for every user. Cost, privacy, network availability, performance requirements and the need to run other PC applications also influence that decision.
Altman did not announce any OpenAI policy concerning consumer RAM ownership, propose restrictions on personal computer memory, or provide evidence that shifting local AI workloads to cloud services would directly reduce retail RAM prices.
His remarks instead present a resource-efficiency argument at a time when AI infrastructure investment and consumer hardware manufacturers are competing for constrained semiconductor production capacity.
For PC builders and gamers, the immediate issue remains the availability and affordability of system memory. Altman’s proposed preference for greater cloud utilization does not resolve that supply challenge or reduce the amount of RAM required for demanding local applications.
