They want the memory

Sam Altman says it's "quite terrible" for the world's scarce memory to sit in your computer. This week the price of a graphics card, an Intel forecast, a firmware hack and a satellite over Svalbard all told the same story about who gets the world's chips.

There is a moment about twenty-nine minutes into the latest Cheeky Pint, the podcast where Stripe's John Collison drinks a beer with someone important and asks them how the world works, where the conversation drifts from superintelligence to something much smaller and much closer to home. Collison has been asking why the memory shortage hasn't fixed itself yet, since high prices are supposed to be their own cure, and Sam Altman answers him in one word. "Bad." Then he says that the invisible hand "is failing right now," and that DRAM makers, who spent decades investing twenty billion dollars to make nineteen, are suddenly about to make something like a hundred billion dollars of free cash flow a year each, and that they remember every previous time somebody told them this boom was different and they built too many fabs and nearly went bust.

That part is fair, and honestly it's one of the clearer explanations of the memory market anyone has given on a podcast this year. It's the next part that people clipped.

"A lot of people want local inference," Altman says. "And somehow, they assume that the problem is solved now. Because, we are going to have open-source models. We're going to be able to run them on our computers." And then: "From a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer. And, being used 5% of the time. You want maximum utilization in the cloud if you're trying to deal with this very supply constrained environment."

You can guess how that landed. Engadget went with "Sam Altman Is Upset At You For Hoarding RAM In Your Own Computer." Tom's Hardware pointed out, a bit drily, that Altman has data centres full of memory and seems to think the problem is the 32GB in your desktop. On X the replies turned into a long gallery of people photographing their gaming rigs and home servers, some angry, some joking about hiding DIMMs under the floorboards, and a smaller and more thoughtful group saying that privacy, latency and simply owning the thing you use are worth more than a utilisation number.

I want to be careful here, because the quote is easy to make worse than it is. Altman wasn't calling for anyone to confiscate gaming PCs. He was talking specifically about local AI inference, and his argument is a real one: a stick of memory in a server rack is busy almost all the time, a stick of memory in a laptop mostly isn't, and in a shortage the busy one does more good per gigabyte. The 5% figure is his and comes with no source, which MetaPatchAI's write-up of the transcript flags plainly. But the logic holds together on its own terms.

The trouble is whose terms. Because when you step back from the podcast and look at what else happened in the last week, the quote stops sounding like an economics lecture and starts sounding like a description of a decision that has already been made for us.

What a graphics card costs now

Start with the most ordinary thing a normal person might want to buy with a chip in it.

Nvidia launched the GeForce RTX 5080 at $999. On 28 October, VideoCardz surveyed Micro Center, Best Buy, Newegg and other US retailers, counting only new cards actually in stock, and Notebookcheck's write-up of that survey found the cheapest verified offers at $1,979.99, for local pickup only, and $1,999.99 at Best Buy online. That is roughly double the launch price, about $800 more than two months ago, and almost exactly what the flagship RTX 5090 cost when it launched. One analyst, Josh Walrath, noted the going price had climbed from around $1,200 to as much as $2,200 for some models. Nvidia hasn't explained it publicly, though The Elec reports the company has said consumer GPU shortages come from memory supply it can't control.

Memory is the root of it, and the numbers there are almost funny in how relentless they are. TrendForce expects PC DRAM contract prices to rise another 13 to 18 percent this quarter, after 18 to 23 percent last quarter. A 16GB DDR5 module that cost $247 in September was $275 in October. An 8GB DDR5 module jumped 24 percent in a single month. And the bit that should worry anyone who plans to buy a laptop next year is TrendForce's estimate that Samsung and SK hynix may cut PC DRAM supply by more than 40 percent in 2027, once you count the volumes being set aside for strategic clients. Their late-October bulletin describes suppliers keeping inventory "extremely low" with "strong bargaining power," PC makers sitting on piles of stock they can't sell at the new prices, and memory being steered toward server modules.

None of that is a conspiracy. It's exactly what Altman described: the memory makers can sell the same wafer as high-bandwidth memory for an AI accelerator or as server DRAM for a hyperscaler, at far better margins than anything a PC maker will pay, and they would be strange businesses if they didn't. The consumer part of the market isn't being banned. It's being outbid.

The chipmaker that couldn't make enough chips

Then there was Intel, which on Thursday night put out a forecast that knocked about 29 percent off its share price on Friday, one of its worst single days in decades.

What makes the Intel story strange is that the quarter itself was good. Revenue was $16.5 billion, up 21 percent on a year ago, and the data centre business grew 78 percent, which CFO David Zinsner called the strongest growth there in fifteen years. The problem was everything around it. Intel guided the next quarter to $15 billion to $16 billion, well short of what analysts expected, because PC makers are ordering fewer chips. They're ordering fewer chips because memory has made their PCs too expensive to sell. According to IBTimes, Intel now expects the PC market to shrink about 12 percent this year and another 10 percent next year.

And on the server side, where the demand is enormous, Intel can't make enough Xeons to fill the hole, partly because of a shortage of the advanced packaging substrates its newest chips need and partly because of an explosion earlier in October at its fab in Israel that killed two workers and shut the plant.

Lip-Bu Tan's line on the call is the one worth keeping. "The industry continues to navigate one of the most significant supply shortages in history," he said, per Business Insider, "driven by the unprecedented demand for super intelligence." It's a remarkable sentence for the head of the company that built the personal computer industry to say out loud. The thing that used to be the core of his business has become the part that gets squeezed so the other part can grow. Dell fell 12 percent the same day, HP 10, AMD 9. Micron, which sells the scarce memory and just posted a record $54.2 billion of quarterly revenue against $11.3 billion a year earlier, fell 6 percent too, which tells you the market isn't quite sure who wins when everyone downstream is struggling.

The half-speed switch

Here is where the week gets properly interesting, because while Altman was explaining why the memory in your machine is wasted, a few people on GitHub were proving that the chip in your machine had been holding back on purpose.

For three generations of GeForce cards, the RTX 20, 30 and 40 series, Nvidia's own published specs have shown a particular kind of AI arithmetic running at half speed. When the tensor cores multiply FP16 or BF16 numbers and add up the results in higher-precision FP32, which is the mode most training and a lot of inference actually rely on, consumer cards issue those instructions at half the rate of the workstation cards Nvidia builds on the very same silicon. It's product segmentation, the same reason a cheaper car might have a limited top speed in software, and it's been sitting in the spec sheets for years.

In the past few days, as TechSpot and Tom's Hardware both reported, a GitHub user called Zagitta published gpu-unlock, a small Linux kernel module that lifts the limit on RTX 30 cards. Two forks followed, one for the 40 series and one for the 20 series. The way it works is a little breathtaking. The speed settings live in two registers the normal driver isn't allowed to write. But Nvidia's signed boot firmware turned out to have a stack buffer overflow, found after Nvidia published a debug version of that firmware encrypted with a known public test key. The module uses that flaw to open the lock, sets the registers to full rate, resets the card, and then lets Nvidia's ordinary driver load as if nothing happened. Nothing gets flashed, and a full power-off puts everything back.

The gains are real but smaller than the headline. The Ada fork's developer saw an RTX 4090's raw FP16 issue rate double from 165.6 to 331.6 teraflops, but an independent owner's real matrix-multiply tests showed 12 to 34 percent. On an RTX 3090 Ti capped at 450 watts, BF16 throughput went from 87.5 to 157.4 teraflops. Prompt processing and image generation get faster. Chatbot token speed mostly doesn't, because that depends on memory bandwidth rather than arithmetic, which, if you've been following along, is a little bit of a punchline in itself.

It's an exploit, it's unsupported, the Turing version warns that running tensor cores flat out might damage hardware, and the project's own safety page warns about fake Windows installers floating around. I wouldn't run it on a card I couldn't afford to lose. But I keep thinking about the shape of it. In the same week that the most powerful person in AI said home hardware is a poor use of scarce resources, a handful of hobbyists showed that a lot of that home hardware had been deliberately running at half its ability, and fixed it with a four-byte write.

The landlord's rules

The other way to get compute, if you can't buy it, is to rent it, and that door got narrower this week too, at least for one very large group of customers.

The Financial Times reported on Monday, citing five people familiar with the matter, that Nvidia has told cloud and data centre customers worldwide that at least 51 percent of their services must be sold to non-Chinese companies. It wasn't written down. The instruction went out verbally through Nvidia's sales and regional teams over recent months, it applies to new contracts rather than existing leases, and it isn't limited to Asia. The point is to close the gap in export controls where a chip stays physically outside China while its computing power is rented over the internet to firms inside it.

The Seoul Economic Daily's account of the same reporting has the detail that makes the scarcity concrete: a rack of Nvidia B300s that sells for about $600,000 internationally reportedly traded inside China last month for 18.5 million yuan, roughly $2.8 million. A Singapore provider told the FT it's hard to find enough non-Chinese customers outside the US and that strict enforcement would clearly hurt its business. On X, Chris McGuire argued the 49 percent cap is still far too loose.

You can think the policy is right or wrong. What I notice is that the chip company is now doing something that looks a lot like allocation, deciding by phone call which landlords may rent to which tenants, and that it's doing it because a chip in someone else's building has become a strategic asset. That's the same logic as Altman's quote, just pointed at a country instead of a living room.

The satellite over Svalbard

And then, because this industry never does anything by halves, Google sent a chip into orbit to get away from the shortage altogether.

On 30 October, Travis Beals, who leads the effort, wrote on Google's blog that the Project Suncatcher prototype satellite, launched on 1 October on a SpaceX Falcon 9, had reached its "Hello World" moment. During a ground-station pass, engineers uplinked a prompt to a Gemma 4 model running on a data-centre-class TPU in orbit, asking for a haiku about being in space. As the satellite passed over Svalbard, it sent this back:

Blue marble below
Data streams in silent grace
Starlight feeds my core.

It's not a good haiku, and I mean that fondly. But it's the first answer from a data-centre-class AI chip in space, and the reasoning behind Suncatcher is the same pressure we've been circling all week: power, cooling and land on Earth are running out faster than the chips can be made, so maybe you go where the sunlight is constant and nobody objects to the planning application. The next tests are about how the TPUs cope with heat and radiation, and the next milestone is two linked satellites in 2027. Rohan Anil, meanwhile, has been writing about training models across "compute islands," separate clusters with patchy links between them, which is a fairly elegant way of saying that the people building this stuff are now designing for a world where compute is scattered wherever it can be found.

It's a moonshot, and a lot of the commentary has rightly pointed out how far a fridge-sized satellite is from a data centre. But it tells you how the people with the most money are thinking. When the ground gets crowded, they don't wait. They look up.

Who it's actually for

So here's the week in one line: the memory is going to the servers, the servers are going to the labs, the labs are going to orbit, and the rest of us are being told, politely, that the cloud will look after it.

I don't think that's the whole story, though, and the bit that gives me the most hope came from a fintech company of all places. At AI Engineer in New York, Revolut's David Abitbol described moving about 2,000 engineers onto GLM-5.3, an open-weight model, self-hosted on around 56 Nvidia B200s. The blended cost came out at roughly 30 cents per million tokens, about a tenth of what they'd been paying the big labs, and because it was cheap enough to use without thinking about it, engineers started using it for everything from code to tickets and docs. The premium models are still there for the hard cases.

That isn't local inference on a laptop. It's a bank deciding that renting intelligence by the token from someone else is a dependency worth getting out of, and owning the model and the machines instead. Which is, on a bigger scale, exactly the instinct Altman was describing as a misunderstanding of supply chains, and exactly the instinct all those people posting photos of their rigs were defending.

Altman is right about utilisation. A busy chip does more work than an idle one, and in a real shortage that matters. But utilisation isn't the only thing people are buying when they buy a computer. They're buying the ability to run something without asking, without a usage pool, without someone else's rules about which tenants are allowed, and without the price changing because a hyperscaler signed a bigger contract. The fact that a single podcast line could make so many people angry tells you how many of them already feel that slipping away.

The invisible hand may well be failing, as he says. What we're watching this week is what replaces it. For now it looks like the highest bidder gets the memory, the chip company decides who rents, and the rest of us find out what that means at the checkout.