Altman Argues RAM Belongs in Data Centers, Not PCs, Exposing Global Memory Crisis

OpenAI CEO Sam Altman said in an interview on the Cheeky Pint podcast that keeping DRAM memory in personal computers rather than concentrating it in data centers is a "terrible" allocation of the resource on a global scale. The statement came during a conversation in which he also discussed AI's impact on mathematics and the RAM shortage affecting the tech sector.

According to Altman, many local-inference enthusiasts believe that open-source models running on home machines would solve the problem of infrastructure demand. He contests that logic: with the world's DRAM supply limited, distributing memory across millions of underutilized PCs wastes capacity that should be in the cloud.

The logic of maximum utilization

Altman's central argument is economic and logistical. A typical personal computer uses its RAM intermittently — on average, only 5% of the time, according to his estimate. In a data center, the same amount of DRAM can be dynamically allocated across thousands of simultaneous workloads, achieving near-total utilization.

"You want maximum utilization in the cloud if you're trying to deal with this really constrained supply environment," the CEO said. "And somehow, the intuition of many people is the opposite of that."

Who has already reserved production

The shortage is not hypothetical. Memory manufacturers such as Samsung and SK Hynix have signed long-term agreements with cloud service operators and AI infrastructure projects, including OpenAI's own Stargate. These contracts have reserved much of the world's RAM production for the coming years.

The direct result is already being felt by consumers: laptops, phones and other electronic devices are getting more expensive as manufacturers prioritize high-volume contracts with the corporate sector. The billions that cloud companies are willing to pay for memory chips are redefining the semiconductor industry's chain of priorities.

The local inference dilemma

Altman's remarks reignite a structural tension in the AI ecosystem: the promise of open models running locally versus the physical reality of memory supply. It is not just about gaming PCs or developer workstations, but any device that depends on DRAM — including professional laptops and machines geared toward AI development.

For the industry, the message is clear: in a scenario of constrained supply, the aggregate efficiency of the cloud trumps the comfort of local ownership. The question that remains is how to reconcile growing demand for privacy and computational autonomy with the reality of a finite and increasingly contested physical resource.