Sam Altman Wants the World's RAM in Data Centers
The statement was straightforward and triggered immediate backlash. Sam Altman, CEO of OpenAI, said on the Cheeky Pint podcast, hosted by Stripe co-founder John Collison, that it is "terrible" to have the world's DRAM memory concentrated in personal computers that use it only 5% of the time. The response was swift: the hardware community, gamers, and local computing enthusiasts did not like this idea at all — and it is not just personal pride from someone who built a machine with 64 GB of RAM.
What Altman Said
The exchange happened around the 29-minute mark of episode 39 of the Cheeky Pint podcast, published on October 29, 2026. Altman was asked about supply shortages, particularly the RAM crisis affecting laptops, consoles, and graphics cards. His answer boiled down to three sentences:
> "From a global allocation perspective, it's really terrible for everybody to have a piece of the scarce global supply of DRAM in their computer, used 5% 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."
The core point is not new in theory: shared servers have much higher utilization rates than individual desktops. The controversy lies in the context. In October 2026, OpenAI and other AI companies are buying massive quantities of DRAM and HBM (High Bandwidth Memory) from manufacturers. Samsung and SK Hynix have already signed long-term contracts with cloud providers and projects like Stargate (the joint venture between OpenAI, Saudi Arabia's PIF, and Oracle), securing much of the world's memory production for the coming years. DRAM prices have jumped from 20-30 cents per Gb to $2 per Gb — an increase of more than 600%. SK Hynix announced that the shortage will continue through 2030.
The irony did not go unnoticed: the CEO of one of the world's biggest RAM consumers telling people to stop using RAM in their own computers.
The Math Behind "5%"
Altman estimated, without presenting supporting data, that personal computers use their memory only 5% of the time. In hours, that equals roughly 72 minutes per day. A datacenter server operating continuously would use the same chip 20 times more hours. The analogy he suggested is straightforward: just as airlines want full planes and hotels want occupied rooms, cloud providers want servers running at maximum capacity.
But there are two hidden assumptions the argument ignores. The first is that datacenter construction has no limits. In practice, datacenters face power bottlenecks (each installation consumes megawatts), environmental permitting, and network infrastructure — problems that RAM in your PC does not have. The second is that the utility of a resource should not be measured only by time of use, but by the value it generates. A spreadsheet left open for 10 minutes may contain the tax declaration that prevents a fine; a game running for an hour may be a family's only form of entertainment. Measuring value solely by "utilization time" is a valid metric for datacenters, but insufficient for evaluating human wellbeing.
The Memory Market in Numbers
October 2026 numbers paint a picture of structural shortage, not cyclical. SK Hynix reported a record operating profit of 83.33 billion won (about $62 billion) in the third quarter of 2026, with an operating margin of 80% — the highest ever recorded. Standard DRAM reached $2.60 per Gb by the end of the quarter, more than four times the value recorded at the same period last year ($0.63). NAND also surged: from $3.82 to $30.61 per Gb in a year.
Samsung reported a 15-fold increase in its chip segment's quarterly profit, but its shares fell 6.2% because its outlook was less bullish than its competitor's — while SK Hynix predicts shortage through 2030, Samsung sees a tight market only until 2028. Both agree on one point: AI infrastructure demand will exceed industry supply in 2027.
Additional factors driving prices: trade tariffs, insufficient stockpiles, and the mass migration to DDR5 and HBM4. Micron discontinued its consumer memory brand, focusing exclusively on servers and datacenters. For anyone buying personal computers, the message is clear: prices must remain high at least until the end of 2027.
What This Means for Personal Computing
The tension between centralized (cloud) and decentralized (local) computing is at an inflection point. Since 2023, advocates of local computing have argued that AI models running on personal hardware would eliminate the need for huge datacenters. Altman refutes this idea with a resource allocation argument: if everyone keeps their RAM at home, the world's total supply is underutilized.
But this argument ignores issues that go beyond pure efficiency. Privacy is one: running AI locally means your data never leaves your machine. Latency too: local models respond instantly, without relying on a connection. And there is the aspect of sovereignty: who decides which services you can access when you depend on third-party servers?
Altman's implicit response is that economic efficiency justifies centralization. In a market with limited supply, allocating resources for maximum use in the most productive location (datacenters) seems rational. The problem is that the definition of "productivity" here includes only what can be monetized — and excludes benefits that have no price, such as privacy, autonomy, and control.
What We Observe
Datacenter RAM prices will continue climbing. The structural shortage predicted through 2030 by SK Hynix will not resolve itself — new factories take 2 to 3 years to build, and the investment needed is billions. Meanwhile, the cost of building a PC with a good amount of memory will rise, and manufacturers will prioritize datacenter contracts over consumers.
The question that remains: if AI needs all the RAM in the planet in datacenters, how far is it willing to go from people who want to use it in their own computers?
Sources: The Verge, Tom's Hardware, Mac Observer
✓ Independent sources cross-checked and verified before publishing
