Sam Altman says scarce computer memory is better used in data centers than in your PC

OpenAI CEO Sam Altman said it is "quite terrible" for scarce computer memory to sit in people's own computers instead of AI data centers, The Verge reported.  He spoke on Cheeky Pint, a podcast hosted by Stripe co-founder John Collison, in an episode posted Oct. 29.

The memory he meant is DRAM, short for dynamic random-access memory.  It is a computer's short-term workspace, the chips that hold whatever programs are running right now.  AI models need huge amounts of it.

Some people want to run AI models on their own laptops, called local inference, instead of sending requests to a company's servers.  Altman argued that home computers use their memory only about 5% of the time, while cloud servers stay busy.  "You want maximum utilization in the cloud," he said.

Altman is describing what economists call capacity utilization: how much of the time an expensive machine is actually doing work.  When something is scarce, the argument goes, it is better shared and kept busy than owned by many people and left idle.  It is the same logic behind car-sharing, or renting a tool instead of buying one that sits in the garage.

Memory is in short supply.  Samsung and SK Hynix, the South Korean chipmakers that make much of the world's DRAM, have long-term deals that reserve much of their output for cloud companies and projects like OpenAI's Stargate data centers, The Verge reported.  That is a big reason laptops and phones are getting more expensive, it said.

Prices are still climbing.  Market researcher TrendForce said Sept. 30 that it expects contract prices for standard DRAM to rise another 10% to 15% in the October-to-December quarter, because chipmakers keep giving priority to memory for AI servers.  It said memory supply for PCs could shrink in 2027 as more factory capacity shifts to servers.

Even critics grant Altman part of the point.  A server shared by many users can keep its memory busy far more of the time than a home PC, which spends much of the day idle.  Gizmodo's Tom McKay, who mocked the comments, wrote that there is "an unobjectionable point buried somewhere in there."  Altman did not say where his 5% figure came from.

But critics say the argument leaves out a lot.  Tom's Hardware noted that home memory is not just for AI; it also runs spreadsheets, web browsers and games.  It added that if people can run AI on their own machines, they may pay less for subscriptions to OpenAI's models.  McKay argued that Altman was describing "the distribution of resources optimal to him," since OpenAI sells AI from the cloud and the industry's buying is what is driving up memory prices.

Running AI on your own device has other benefits, too.  Your data can stay on the device, and it can work without an internet connection.  Apple, for one, says many of the AI models behind its Apple Intelligence features run on the iPhone or Mac itself for privacy, sending only harder requests to its own servers.  For most readers, the practical effects are price and control: memory demand is raising what new laptops and phones cost, and cloud AI usually means a monthly fee and handing your requests to a company.

Source:  The Verge, Oct. 30, 2026, Cheeky Pint, Oct. 29, 2026, TrendForce, Sept. 30, 2026, Gizmodo, Oct. 30, 2026, Tom's Hardware, Oct. 30, 2026 and Apple, Oct. 28, 2024.