Sam Altman: “It’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer”

Local inference and scarce memory

OpenAI CEO Sam Altman answered questions from Stripe co-founder John Collison in the Cheeky Pint podcast episode “Why Sam Altman is against a single superintelligence,” published on Oct. 29. In the episode’s section on compute, Altman said supply chains work in ways that are “super non-obvious” and offered an example. Many people, he said, want local inference, meaning models that run on users’ own computers rather than in data centers. They assume that because open-source models can run on their computers, the infrastructure in data centers will not be needed.

According to the episode transcript, Altman said: “From a global allocation perspective, actually, it’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer.” He added that this memory is used 5% of the time. In such a supply-constrained environment, he argued, the goal should be maximum utilization in the cloud, though many people’s intuition is the opposite. The 5% figure is not based on a measurement or a cited source; it is Altman’s own estimate, given during the conversation.

John Collison@collision

OpenAI’s @sama joined me at the pub (and had his first Guinness) to discuss the PHP exploit, math weekend at OpenAI, agent swarms, and how he thinks AI leads to deflation.

The DRAM makers’ investment cycle

In the same episode, Altman also described how memory makers used to invest. For a very long time, he said, a DRAM producer would invest $20 billion of capital spending to make $19 billion; in some years it made $21 billion, in others $18 billion. The topic the episode is named after is many AIs rather than a single superintelligence. “I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of 1 trillion IQ join society,” Altman said.

Memory prices

Altman’s remarks come as memory prices climb quickly. According to a TrendForce News report dated Oct. 20, DRAM contract prices rose about 80% quarter over quarter in the second quarter of this year and about 40% in the third. The report expects a rise of about 20% in the fourth quarter and attributes the slowdown to a higher comparison base, weaker end-market purchasing power and price ceilings in long-term agreements. A TrendForce press release dated Sept. 30 projects a 10% to 15% rise in conventional DRAM contract prices for the fourth quarter; the two figures come from different publications.

According to the same release, server DRAM will remain undersupplied in the fourth quarter. PC DRAM supply could decline in 2027 as more capacity shifts to servers, and PC makers are buying aggressively on expectations that supply will stay tight in 2027. SK Group Chairman Chey Tae-won said at Computex in Taipei on June 2 that the memory bottleneck will continue through 2030 and that SK hynix will double its wafer production capacity within five years.