RAM belongs to people

OpenAI CEO Sam Altman said on a podcast that it is actually terrible how much scarce memory sits in home computers. In his view it belongs in data centers. That surprised me. But I understand where that line of thinking is heading.

What he said

On October 29, 2026, Sam Altman, the CEO of OpenAI, was a guest on the Cheeky Pint podcast, hosted by Stripe cofounder John Collison. The talk turned to the shortage of DRAM, the working memory in computers. According to Yahoo Finance, Altman said that from the point of view of global allocation it is “quite terrible” for everyone to keep a piece of the world's limited DRAM supply in their own computer. At home, he said, memory is busy only about 5% of the time. In a data center it would be used far more (Tom's Hardware).

That 5% is his estimate. He cited no measurement. And he did not say that people should not own memory or that it should be taken from them. Critics read that into his words later (TechTimes). Elon Musk answered him with one line: “the people deserve to have their RAM too.”

Why memory is scarce

Large language models run on graphics cards. Those cards have their own memory, called HBM. But HBM is DRAM too, and it is made by the same three companies that make the memory in your computer: Samsung, SK Hynix and Micron. Production is shifting to HBM for data centers, where the margins are higher, and less is left for ordinary memory (CDR.cz, in Czech). In December 2025 Micron even shut down its consumer brand Crucial (Gigacomputer, in Czech).

Why does this matter to people who run local models? A model that does not fit into the graphics card's memory runs from the computer's main memory. If you want a model at home, you need RAM. A lot of it.

What it costs

Prices went up brutally. The same 64 GB DDR5 kit I have in my computer cost around CZK 5,400 (about $250) in November 2024 (Živě.cz, in Czech). Today, on the Czech price comparison site Heureka, I find it from roughly CZK 38,000 (about $1,760), and typically around CZK 40,000 (about $1,850), as of November 4, 2026. That is roughly seven times more. A smaller kit tells the same story. In September 2025, 32 GB of DDR5 cost CZK 2,390 (about $110) in the Czech Republic (Lupa.cz, in Czech). By January 2026 it was CZK 10,890 (about $500) (Živě.cz, in Czech). More than four times as much in four months. (Dollar amounts use the exchange rate of November 4, 2026.)

There is a claim going around the internet that RAM is more expensive than gold by the gram. It is not. If you weigh a module and divide the price by its weight, memory still comes out roughly 5 to 16 times cheaper than gold. On November 4, 2026, Czech dealers were buying gold at CZK 2,899 (about $134) a gram (Kurzy.cz, in Czech). But memory is getting more expensive faster than gold. South Korean export prices for DRAM chips rose 12.5 times in about three and a half years. Gold rose 2.4 times over the same period (TrendForce).

Competition, not waste

For the big players, hardware at home is competition. Every model running on your computer is a model you are not paying them for. Some argue against local models on safety grounds, others on resources. Altman has now said the second argument out loud.

But they are partly to blame themselves. In my view, the move to local models stands on four legs, like a table.

The first is refusals. When users get refused even on safe, completely normal tasks, they have to solve it somewhere else, and that means locally. Safety layers also often miss the mark. Out of 392 crisis hotline offers in my archive, two were justified, as I write in the article on false alarms. And a model can refuse a scene with nothing objectionable in it at all (Refusal is a distribution). From what I see, the roleplay community moved to local models mainly because of this.

The second is privacy. Plenty of people do not want to send their diary or questions about their own health to the cloud.

The third is cost. A subscription you pay every month. Hardware you buy once.

The fourth is control. Nobody swaps out a local model under your hands overnight. Every version change in the cloud throws people's established work off. And if you want to fine tune a model on your own data, you need to have it with you.

A double bottom

Altman is right about one thing. Memory at home does nothing most of the time. But there is irony in his sentence. What he is really saying is: your RAM is lying idle, give it to us. And yet it is precisely the hunger of data centers that is driving memory prices so high that an ordinary person will soon be unable to build a computer that can run a local model at all.

So the efficiency argument and the reality of the market push in the same direction. Toward renting instead of owning. A car also sits in a parking lot most of the day, and still nobody wants to trade it for a taxi that can refuse to stop for them at any moment. Utilization is not the only value things have. Some things are valuable simply because they are yours.

Model outputs are translated; the content is unchanged.