Altman says running AI locally wastes scarce DRAM, advocates maximizing cloud utilization

On November 1, news broke that OpenAI CEO Sam Altman recently appeared on the podcast "Cheeky Pint" hosted by Stripe co-founder John Collison. In a setting of drinking beer and chatting, he offered a rather controversial judgment on the AI industry's most sensitive current issue—the shortage of compute and memory: putting scarce DRAM memory in personal computers to run local AI is a "pretty bad" global misallocation of resources.

Is running AI locally a waste of DRAM?

While discussing the tight memory supply chain on the program, Altman pointed the finger at "local inference." He said: "A lot of people want local inference, and somehow they think the problem is solved—because we'll have open-source models, we can run them on our own computers, so we won't need the stuff in data centers."

But he immediately went on to point out, "From a global resource allocation perspective, having a bunch of globally scarce DRAM memory sitting in everyone's home, while it's only being used 5% of the time, is actually pretty bad." Altman stressed that in an environment of extreme supply constraints, "what you need is maximum utilization in the cloud. But many people's intuition is exactly the opposite."

Altman's remarks about local inference "wasting memory" inevitably carry a vested interest. After all, OpenAI's business model is built on paid access to closed-source models in the cloud; if open-source models can run well on local devices, subscription demand could be weakened. Critics can fully point out that this CEO is defending his own data centers' memory needs.

But setting motives aside, the resource allocation problem Altman points out is not fictitious. At a time when DRAM supply is tight and prices keep climbing, whether the world's scarce memory should flow first to cloud inference clusters running 24/7, or be scattered across millions of personal computers that sit idle most of the time, is indeed an efficiency question worth taking seriously. The real problem is that this choice should not be made on everyone's behalf by the CEO of a cloud AI company, and also that not all enterprises and individual users are willing to send private information to the cloud.

The "invisible hand" is failing

Altman does not hide the severity of the memory shortage. When asked how progress on solutions was going, his answer was just one word: "Bad."

He pointed to memory manufacturers' historical trauma. Altman explained that DRAM makers had been through a brutal cycle of "investing $20 billion to build capacity and ultimately recovering only $19 billion," and had repeatedly come close to bankruptcy due to overexpansion. Therefore, even with the current AI demand explosion, these manufacturers remain unusually conservative in capital expenditure. He used a classic economics concept to describe the situation: "I think the 'invisible hand' is malfunctioning right now."

This judgment has some real data support. SK Hynix this week announced record quarterly operating profit and warned that the memory shortage could last until 2030; Samsung's quarterly profit also hit a new high for the fourth consecutive quarter. Memory manufacturers are enjoying the benefits of a seller's market, while downstream PC, phone, and game console makers are bearing soaring costs.

Superintelligence is inferior to "agent teams"

Another noteworthy view Altman raised in the interview is his clear opposition to a "single superintelligence." He dislikes the term "Agent Swarms" and prefers to call them "Agent Teams."

His reasoning: "Intuitively, I would rather have a trillion AIs slightly smarter than humans join society than one AI with an IQ of a trillion join society." Altman believes that a large number of AIs each slightly more capable than humans, competing and balancing one another, is closer to the way human society stays vibrant, rather than concentrating all capability in a single system.

This position also extends to the safety discussion. When asked whether models far beyond human level could be "put in a sandbox," Altman responded: "Sandboxes work to some extent, monitoring works to a further extent, but ultimately, you need enormous confidence in the model's alignment." He admitted that if facing an AI a million times smarter than the smartest human safety researcher, "would you really be willing to trust that sandbox?"

Mathematical breakthroughs and "the closing of an era"

Regarding OpenAI's recent breakthroughs in mathematics, Altman revealed that the company had deliberately lifted compute limits for a week to let models focus all-out on difficult problems, among which solving the Navier-Stokes equations cost about $10 million to $20 million. He expects that a year from now the same proof could be redone at 1% of the cost.

But amid the excitement, Altman described an unexpected emotion: "I was awake most of the night for two nights in a row because I was too excited to sleep. I was extremely excited and happy, but I also felt more sadness, melancholy, and a sense of something coming to an end than I expected." He compared the feeling to witnessing the closing of a chapter of an era—problems that human scholars had not solved for decades falling one after another before the model.

He also accidentally revealed the internal model's codename: "Model Name-Hold My Beer." He said that watching this model knock down, one after another, the hardest problems in researchers' academic careers across various fields was "an incredible sight."

Token pricing: a "fake" business unit

On the business side, Altman gave a candid assessment of the token billing mechanism commonly used in the AI industry. On the podcast he called tokens "a fake unit": "They're real in some sense, but they're not a good unit. From a business value perspective, they're fake."

But he also admitted that consumers' demand for "clear, transparent spending" is so strong that the industry has to keep this mechanism. He once asked users whether they would be willing to pay a higher price for the same output but priced in a more consistent number of tokens: "Even if we inflate the tokens, they would say: oh, I'd rather you inflate the tokens and make me pay more, because I know there's something real there."