OpenAI chief executive Sam Altman has stirred a fresh debate about how the world’s limited supply of computer memory should be used, arguing that packing large amounts of RAM into personal computers is a poor use of a resource that artificial intelligence data centres could put to far better work. His remarks, made on the Cheeky Pint podcast, land as memory prices climb sharply, and they cut to a question the industry has been circling for months: when a vital component is in short supply, who should get it first?
Altman did not suggest that computers should do without memory. RAM is the working space in which a machine holds the programmes and data it is actively using, and without enough of it even routine tasks become sluggish. His point was narrower. Much of the memory installed in homes and offices, he argued, sits idle for most of the day, while the same chips inside a shared data centre could work almost constantly for many different users.
Sam Altman Says PCs Are Wasting Limited RAM
The conversation turned on a growing trend: running artificial intelligence models directly on a personal computer rather than sending every request to a distant server. This practice, known as local AI inference, has obvious attractions. It can keep personal data on the owner’s own machine, work without a constant internet connection, and free users from monthly cloud subscriptions. Hardware makers have responded with machines carrying ever larger amounts of memory, marketed as ready for on-device AI.
Altman questioned whether spreading memory thinly across millions of individual machines is the wisest response to a genuine shortage. He described the global picture as quite terrible from the standpoint of allocating resources well, a blunt phrase that captured his frustration with the mismatch between where memory sits and where it is actually used. According to his estimate, a typical personal computer may draw on its available memory for only around five per cent of the time, leaving expensive hardware largely dormant while data centres elsewhere cry out for capacity.
AI Data Centres Are Driving RAM Demand
That tightness is not theoretical. The rapid expansion of artificial intelligence infrastructure has placed extraordinary pressure on the global memory market, and the consequences are now visible in shops and on price lists. Large AI systems need vast quantities of memory to hold their models and process requests, and the operators building these systems have been buying DRAM and high-bandwidth memory on a scale the industry has never seen before.
High-bandwidth memory, a specialised form of DRAM designed to feed data quickly to powerful processors, has become one of the most prized components in technology. The three companies that dominate DRAM production, Samsung, SK Hynix and Micron, have increasingly directed their factories and investment towards the server and AI market, where contracts are larger and margins are higher. Long-term deals with cloud providers, including OpenAI’s own Stargate initiative, have locked in enormous future demand.
The knock-on effect for consumers has been severe. Analysts at IDC have described an unprecedented memory shortage in late 2025, warning that the imbalance could persist well into 2027, while market trackers have recorded steep rises in the price of ordinary memory kits. A computer that was affordable a year ago can now cost noticeably more for the same specification.
Local AI Could Reduce Dependence on Cloud Services
Yet Altman’s efficiency argument runs into a powerful counter-argument that many computer users feel instinctively. Local AI offers independence. A model running on your own machine does not need to phone home, does not hand your documents to a third party, and does not stop working when a subscription lapses or a service changes its terms. For people handling sensitive work, and for anyone wary of concentrating ever more power in a handful of cloud giants, that independence is worth a great deal.
There is also a practical wrinkle. Running larger models locally genuinely demands serious hardware, and that demand is precisely what has made high-memory machines so popular. Altman himself leads a company whose main business is cloud-based AI services, and critics will note that his preferred answer conveniently favours the centralised model on which his company depends.
What This Means for PC Users
For ordinary buyers, the immediate message is less about philosophy than about prices and choices. Memory remains essential for everyday computing, and nobody is suggesting that computers should be sold with too little of it. What Altman’s comments highlight is that the cost of that memory is now shaped by forces far beyond the PC market, above all by the appetite of AI data centres.
Buyers planning an upgrade may therefore find that paying for extra memory costs more than expected, and that machines marketed for local AI carry a premium reflecting the shortage as much as their performance. Those who mainly browse, stream and handle office work may find modest memory perfectly adequate, while creators who genuinely need large amounts must weigh the higher price against how much they will actually use. In the longer term, if supplies recover, the case for spreading memory widely becomes easier; if shortages persist into 2027 and beyond, as some analysts fear, pressure to reserve scarce chips for shared infrastructure will only grow.
