Sam Altman argues AI belongs in the cloud, not on home PCs

The case for the cloud

Altman said people want to run models locally rather than rely on cloud services, believing open models could reduce the need for data-center hardware. But, he argued, it is poor resource allocation for every household to hold part of the world’s limited supply of dynamic random-access memory.

His alternative is to use hyperscalers’ cloud services, including OpenAI’s. With resources constrained, he said, the cloud should be loaded as efficiently as possible—even if that runs against many people’s intuition.

The remarks landed badly with PC and gaming enthusiasts, who are already seeing prices for some components rise sharply. Earlier this month, Altman said the world should accept some bad consequences in exchange for the benefits of AI. That comment also drew a sharp reaction.

Who gets to run AI?

The backlash on Reddit was about more than memory allocation. One user called Altman selfish and said his comments reflected the mindset of a cult leader. Another said he seemed unable to understand why people want to own their own computers. A third argued that private companies driven by profit should not make consequential decisions affecting most people, while allowing that they might be wrong.

That dispute meets an awkward business reality for AI companies: their costs keep rising, and investors are concerned. Local inference on personal devices challenges the assumption that users must depend on expensive cloud services such as OpenAI’s. I think that is the more revealing tension here: a claim about efficient use of scarce hardware also protects the business model that depends on centralized access to it.

The supply pressure is real. PC makers say printed-circuit-board capacity is fully booked through the end of 2030. Altman’s remarks also reflect advanced AI labs’ desire to secure as many chips as possible as supplies tighten.

What the interview leaves unclear is how the costs and benefits of that allocation should be shared. Altman says everyone has a common interest in using limited global chip-manufacturing capacity well. For people who want to run models on hardware they own, efficiency alone does not settle who should get access.