Sam Altman says it’s better for RAM to be in a data center than on your desk

On the Cheeky Pint podcast, OpenAI's CEO argued that it is, in his word, terrible for the world's limited supply of DRAM to sit in personal computers that use it a small fraction of the time. Better, he said, to concentrate it in data centers where utilization stays high.

That is not an abstract point right now. As The Verge notes, Samsung and SK Hynix have long-term deals with cloud operators and projects like OpenAI's Stargate that reserve much of the world's memory output for the next few years, and that demand is part of why laptops and phones are getting more expensive.

Let's give the argument its due first. Utilization is a real thing. A gigabyte of memory serving thousands of requests in a well-run cluster does more work than the same gigabyte idling in a laptop. In a supply-constrained market, pooling scarce resources is a defensible engineering instinct, and cloud inference has made powerful models available to people who could never buy the hardware to run them.

But efficiency is not the only value on the table, and the person making the case is not a neutral party. OpenAI is one of the buyers whose demand is helping squeeze supply. An argument that concludes the scarce thing should live in your data centers rather than in customers' machines is, conveniently, also a business model.

Local inference matters for reasons a utilization chart doesn't capture: privacy, working offline, lower latency, freedom from a provider changing prices or terms, and resilience when a cloud service goes down. Open models running on personal hardware are one of the healthiest checks on concentration in this industry. If memory scarcity quietly prices ordinary people and small developers out of running their own models, that is a loss even if aggregate utilization goes up.

I'm also wary of how this framing could age. If the industry starts treating personal compute as wasteful by default, the end state is a world where nearly all serious AI runs on a handful of companies' servers, and users rent access instead of owning capability. That may be efficient. It is also a lot of power in very few hands.

Our read: Altman is right that the shortage is real and that cloud pooling helps in the near term. He's wrong if the takeaway is that local AI is a mistake. The better answer is more memory supply, not a smaller share of it for the people buying computers.

Memory makers have committed large shares of future production to hyperscalers and AI infrastructure projects, and consumer devices are feeling the effect in pricing. The comments came in a wide-ranging podcast interview in which Altman also discussed AI's impact on math.