Sam Altman: RAM Belongs in Data Centers, Not PCs

Sam Altman: 'It's a shame that RAM is inside the user's computer' captures OpenAI's argument against running AI models locally. The company's CEO says scarce memory should serve data centers rather than sit unused in personal computers.

Altman made the comments during an interview with the YouTube channel Cheeky Pint, published on October 30. He was responding to people who see open-source models running on personal computers as a way to reduce reliance on cloud services.

"A lot of people want local inference. And somehow, they assume that the problem is solved now because we are going to have open-source models, we’re going to be able to run them on our computers, and so we’re not going to need all the stuff in the data center.

But 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 and being used five percent of the time.

You want maximum utilization in the cloud if you’re trying to deal with this very supply constrained environment. And somehow many people’s intuition is the opposite of that"

Why Altman opposes local inference

Local inference runs a trained AI model directly on a device. Processing can use hardware such as a computer or phone's CPU, GPU or NPU, rather than sending data to third-party cloud services such as ChatGPT and Claude.

Altman argued that local AI leaves much of the available memory idle. He said limited RAM is distributed across users' computers even though it is used only 5% of the time. In his view, concentrating memory in data centers would make better use of scarce global resources.

His position assumes data center construction and expansion will continue. It also assumes RAM operating continuously in those facilities will provide more value than memory installed in users' devices.

OpenAI's business model sits behind the debate

OpenAI has released several open-source models, but its core business remains based on closed models running in data centers. That makes local inference more than a technical choice: it could affect how users access and pay for AI models.

Tom's Hardware described Altman's comments as a defense of OpenAI's business model. Its view is that personal computers do not currently threaten OpenAI, but local inference could reduce what users pay for models such as GPT-6 Luna.

The dispute therefore covers two competing approaches. Local models keep processing on users' hardware. Cloud models place computing and memory in provider-owned data centers, where companies can manage shared capacity and charge for access.

Memory supply adds pressure

AI infrastructure companies are signing long-term contracts to secure memory supplies. Micron has signed 26 strategic agreements covering more than 35% of its projected revenue by 2030, with $32 billion in committed customer deposits.

Samsung is reportedly finalizing contracts covering 75–80% of its projected memory chip production by 2027. SK Hynix is also pursuing similar agreements.

This cycle differs from earlier periods of high chip demand. Manufacturers previously expanded factories, then faced oversupply as demand cooled. This time, they have raised prices multiple times instead of significantly increasing capacity, keeping supply below demand and extending price pressure.

Analysts expect the memory chip shortage to last at least until 2028. That gives manufacturers greater negotiating power while companies compete for future production.

Why computer prices matter

Steve Bonifiled of The Verge said the outcome is similar whether Altman meant users' personal computers or devices used by AI application developers. In his view, billions of dollars spent by companies on RAM orders are helping make computers, phones and other technologies more expensive.

The argument makes this News relevant beyond OpenAI. It connects local AI, cloud computing, memory allocation and hardware prices. Readers weighing local models against cloud services can see the trade-off clearly: local inference uses hardware already owned by users, while cloud inference concentrates demand in data centers that are now competing aggressively for scarce memory.