The Man Buying the Memory Says Yours Is Wasted

The man who buys the world's memory has decided you're wasting yours.

On a podcast with Stripe co-founder John Collison, OpenAI's CEO called it "quite terrible" for everyone to hold a share of the world's limited DRAM in a personal computer, used 5% of the time. Scarce memory, he says, belongs in cloud servers running at full use. Yahoo Finance and PC Gamer both reported the remarks.

Notice the shape of the argument. It is not about what your computer can do. It is about what he would rather do with the parts inside it. "Terrible" is doing a lot of work there. Terrible for whom?

OpenAI is one of the biggest buyers of the memory he calls short. In October 2025 Altman signed letters of intent with Samsung and SK Hynix for Stargate, and SK Group said that could put projected demand at up to 900,000 DRAM wafers a month, as much as 40% of global output. Projected, not bought. But the direction is clear.

Then the prices. TrendForce reports DRAM contract prices rose roughly 93% to 98% in the first quarter of 2026. A 32 GB DDR5 kit now runs about $500, up from about $400 in early April, per Tom's Guide. Analysts expect the shortage to last at least through 2027.

And here is the part worth sitting with. Altman asked why high prices have not drawn new suppliers. "The invisible hand is failing right now," he said, per Yahoo Finance. The man whose company is signing for up to 40% of the world's memory, watching the market he is helping to corner, wondering aloud why the market will not fix itself.

Some people online turned their anger into threats against Altman. That is wrong, and it is not the argument. The argument is simpler: he wants the memory, and he has a story about why you should not.

Here is the part he will not say. A model that runs on a machine you own has no per-call bill. It keeps your data on the device. It works offline. It cannot change its terms overnight. We wrote about this before, when a model that fits on a 128 GB laptop went from a score of 3 to 50 on the Artificial Analysis index in two years. AI you can hold, on hardware you already own.

The honest limits are real. A quantized local run typically scores a point or three lower. Speed drops on very long inputs. A laptop trails the best cloud models on the hardest work. And memory prices are up, so a well-equipped laptop costs more this year. But it is bought once. A monthly bill returns every month.

For a business, the move is not all or nothing. A rule of thumb, not a measurement: put repeated, private, high-volume work on machines you own. Drafts, summaries, sorting, searching your own files. Keep the hardest work in the cloud. First step: add up what you spend on AI each month and sort it by task.

The man selling the subscription has a reason to call your memory a waste. You have a reason to keep it.

Use the tools →

Sources: