A case for centralised AI computing

OpenAI chief executive Sam Altman has argued that scarce computer-memory supplies would be used more efficiently in cloud data centres than spread across personal machines running artificial-intelligence models locally. His comments frame a growing contest over where AI computation should happen as demand from data-centre operators puts pressure on the hardware supply chain.

Speaking on Stripe co-founder John Collison’s Cheeky Pint podcast, Altman said many people expect open-source models to make large data centres less necessary because inference can run on their own computers. He disputed the efficiency of that approach when dynamic random-access memory, or DRAM, is supply-constrained. Memory installed in individual devices may sit idle much of the time, he said, while shared cloud infrastructure can maintain higher utilisation.

The argument concerns allocation rather than whether local models work. A personal computer gives its owner direct access to hardware and can offer privacy, offline operation and independence from a service provider. Centralised systems instead pool expensive components among many customers and can shift capacity as demand changes. Altman’s position is that the second model makes better use of a limited global supply.

That view also aligns with OpenAI’s commercial interests. The company delivers closed cloud services and is pursuing a large infrastructure expansion. Sending more inference to data centres would increase reliance on providers such as OpenAI, while local execution can reduce recurring service use and place more control in users’ hands. The tension makes Altman’s efficiency claim both an engineering argument and a statement with clear business implications.

Gizmodo reported the remarks in a sharply critical article, arguing that concentrating memory in data centres would favour OpenAI while consumers face higher component prices. The outlet connected AI infrastructure demand with expensive DRAM used across computers and smartphones. Its broader characterisation of Altman’s motives is commentary, however, rather than a fact established by the podcast remarks.

The underlying trade-off is unlikely to disappear. Large cloud systems can keep specialised equipment busy and make powerful models available without requiring every user to buy capable hardware. Local AI can reduce latency and limit the information sent to remote servers, while allowing people and organisations to retain operational control.

Altman did not propose banning consumers from purchasing memory or preventing local inference. His narrower claim was that widespread, lightly used DRAM in personal computers is a poor global allocation during a shortage. The reaction illustrates why hardware policy around AI is contentious: decisions presented as efficient resource sharing can also determine who owns the infrastructure, controls access and collects payment for computation.