Sam Altman: RAM sitting idle in home PCs is 'pretty bad'

OpenAI CEO Sam Altman shared a notable view on local AI inference, and on people keeping DRAM in their home computers, in an interview with the YouTube channel Cheeky Pint. Looked at on a global scale, he said, the situation is 'pretty bad'.

The remarks drew attention because the head of a company with vast amounts of memory in its data centers was describing a desktop PC's 32 GB of RAM as a problem.

Altman's argument: the memory goes unused

In the interview, Altman addressed the assumption that people can run open-source models locally on their own machines. Weighing the idea of running local AI on systems people already own or want to build, he explained why he considers it a bad idea.

In his words, from a global deployment standpoint, it is pretty bad for everyone to hold a piece of the world's limited DRAM supply in their own computer and use it only 5 percent of the time. With supply this constrained, he said, the goal should be maximum utilization in the cloud. He added that many people's intuition runs in exactly the opposite direction.

The link to OpenAI's business model

Although OpenAI has some open-source models, the company's core business rests on closed-source models running in data centers. Altman's comments can therefore also be read as a defense of the company's own business model.

Laptops and gaming systems do not yet pose a serious threat to OpenAI. But if local execution becomes more widespread, users might one day pay less in subscription fees for models such as GPT-6 Luna. Altman's remarks suggest the company would prefer RAM to go into OpenAI's data centers rather than into home computers.

Hardware ownership worries and underlying assumptions

That stance also feeds a concern among enthusiast users: that they will be unable to own their hardware and software and will instead have to rent an agent and a computer in the cloud.

Altman's argument also rests on a few assumptions: that data center investment will not slow down, and that RAM running around the clock in AI data centers is worth more than memory used for spreadsheets, web browsing, gaming and the occasional local AI task.

The interview also touched on other topics, including this week's PHP security vulnerabilities, sandboxing models during training, and OpenAI's research in mathematics.