OpenAI chief executive Sam Altman said it is "quite terrible" for consumers to keep large amounts of memory in home computers that sit mostly idle, arguing the world's limited supply of DRAM should flow to cloud data centres instead. Speaking on the Cheeky Pint YouTube channel, Altman said that running AI locally on a home machine wastes a scarce resource: "You want maximum utilisation 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."
The comment is a defence of OpenAI's business. The company sells access to closed models that run in its data centres, and widespread local inference on hardware people already own would undercut subscriptions to products like its GPT-6 models. But Altman's framing also points at a real constraint. Memory and the chips around it have become a bottleneck for the AI buildout, and the companies racing to expand are competing for the same limited manufacturing capacity.
The demand side keeps growing. Amazon plans to deploy about two million more Nvidia GPUs across its data centres in 2027 and 2028, including Blackwell Ultra, Rubin and Rubin Ultra parts, on top of roughly one million Nvidia chips it began installing this year. Each of those systems needs high-bandwidth memory, and the pressure reaches ordinary DRAM too. On 31 October, Altman met executives from Samsung and SK Group in South Korea as tightening memory supplies and OpenAI's own commitments raised the stakes for its chip and data-centre partnerships, according to industry trade press.
The physical footprint is expanding to match. Foxconn's data-centre subsidiary Visionbay.ai has brought online what it calls Taiwan's largest Nvidia HGX B300 cluster, a 5-megawatt facility it says is already running above 90 per cent utilisation under long-term leases. Nvidia chief executive Jensen Huang has described the modern data centre as "one computer" a thousand acres across, drawing hundreds of megawatts to gigawatts. That scale is what turns a routine component like DRAM into a contested resource, and what puts a chief executive in the position of telling people what their own computers should contain.
Not everyone accepts the premise that the capacity is well used. In comments under one report, readers pointed to claims that AI data centres sit idle much of the time and that some buildout is defensive, meant to deny hardware to rivals rather than to serve demand. Those figures are unverified and sit uneasily against Visionbay's reported 90 per cent utilisation, so the true picture of how fully this hardware is used is not settled. What is clear is the direction of the argument: Altman is making the case that compute and memory belong in the cloud, and the buildout numbers from Amazon and Foxconn show the industry spending as though that case has already been won.