The memory shortage got deep enough that HP started buying the chips Washington tried to wall off. The tell is that the Chinese supplier isn't even discounting them.

Somewhere on a contract-manufacturing floor this summer, a budget laptop is coming down the line for a brand you have heard of, and inside it, soldered near the processor, is a small rectangle of memory carrying a name that was not there a year ago. The machine will ship to a shop in Jakarta or São Paulo or Warsaw — somewhere, in any case, that is not the United States. Nobody will announce it. The buyer will never know. But the presence of that chip is one of the more telling facts to come out of the memory market this year, because the name stamped on it is CXMT, a company the United States placed on its restricted-trade list, and the companies quietly designing it into their laptops are HP, Asus and Acer. Western brands are putting Chinese memory in their computers, and the reason they are doing it is the whole story.

Let me be careful with what is known, because the reporting here rests on people close to the supply chain rather than on any company's confirmation. According to reports out of Asia this week, several of the largest PC makers finished qualifying CXMT's DRAM — the everyday working memory in a computer, distinct from the storage — around the middle of this year, and have begun using small volumes of it in a limited number of notebook models built for markets outside the United States. The scale, by every account, is modest so far. But qualification is the hard part, the months of testing a component must pass before a brand will vouch for it, and qualification is not a thing you do idly. You do it when you can see you are going to need a door, and you want it open before the room fills with smoke.

The tell is the price

Here is the detail that tells you what is actually happening, and it is the one most likely to be skipped. CXMT's memory is reportedly not cheaper. It is priced roughly in line with Samsung's — not at the discount you would expect from a newer entrant trying to buy its way onto a bill of materials, but at parity with the incumbent it is supposedly challenging. Sit with that, because it inverts the usual logic of why anyone buys the Chinese part. Nobody is putting CXMT in these laptops to save a dollar. If price were the motive, parity would kill the deal. They are putting it in because when they go to buy the memory they have always bought, from the suppliers they have always used, the answer increasingly is that there is none to be had at any sensible price, or none to be had at all beyond the current quarter. A buyer does not qualify a sanctioned supplier and pay full freight for its output unless the alternative is an empty slot on the board. The price is the tell. The shelf is bare.

Nobody is putting CXMT in these laptops to save a dollar. They are putting it in because when they go to buy the memory they have always bought, the answer is that there is none to be had.

And it is bare for a reason that has nothing to do with laptops. The world's leading-edge DRAM is made, to a first approximation, by three companies — Samsung, SK Hynix and Micron — who between them hold something close to ninety percent of the market. That concentration is the kind of thing this column returns to, because it is the hidden architecture of almost every hardware story: a handful of firms, a few plants, one upstream dependency, and the whole world downstream of it. For thirty years those three made the memory the world runs on and priced it like the commodity it was. Then the AI build-out arrived and asked them for a different product — high-bandwidth memory, the exotic stacked kind that sits beside an AI accelerator — at a margin the commodity business never came close to. So they did the rational thing. They pointed their wafers at the richer customer.

The residual

This is the mechanism underneath the shortage, and it is worth stating precisely, because "shortage" makes it sound like a shortfall of effort when it is really a decision about allocation. A wafer of silicon can be turned into commodity memory for a laptop or, with far more processing, into the stacked memory an AI data center will pay almost anything for. It cannot be turned into both. By the estimates now circulating in the industry, artificial-intelligence infrastructure will consume something like seventy percent of the world's memory output this year. The everyday memory that goes into phones and PCs and cars — the stuff that was the whole business a few years ago — has become the residual, the portion left over after the data centers have taken theirs. DRAM prices have roughly tripled in eighteen months. Some manufacturers, unable to buy enough new memory, have reportedly begun harvesting older components to keep production lines moving. When the residual gets thin enough, buyers go looking for supply wherever it exists, and the largest pool of memory the big three are not already committing to data centers happens to sit in Beijing.

So the door that opened this year was not opened by policy or by any thaw in the trade fight. It was opened by scarcity, which is a more powerful force than either, because it works on the purchasing manager who has a build target and no parts. CXMT is the fourth-largest DRAM maker in the world now, with something on the order of seven or eight percent of the market against the big three's ninety. A year ago that share was a rounding error to a Western brand, a supplier you did not need and did not want the complications of using. This year it is the marginal ton of memory that keeps a product line alive, and the complications have become the price of doing business.

Follow it one link further

The comforting reading of all this is that China has simply solved commodity memory and is now flooding in to fill the gap, and if you stop at the brand on the laptop you might believe it. Follow the dependency one link further, though, and the picture gets more constrained on both sides. CXMT cannot supply as much as the West would now take, because it is reportedly steering a large share of its output to its own domestic customers first — Huawei chief among them — for whom it is not a convenience but a strategic necessity. The memory reaching HP and Asus is what is left after China's own champions are served, which is to say it is a residual of a residual. This relief valve is real, but it is narrow, and it is throttled from the far end by priorities that have nothing to do with a laptop in Warsaw.

And CXMT's ability to widen it runs straight into the reason its name is notable in the first place. This is a company under United States trade restrictions, which is precisely why the chips are kept out of American-market machines and why the Western brands using them do so quietly — both to stay clear of sanctions exposure and to avoid provoking the three incumbent suppliers they still depend on for the other ninety percent of their memory. Those same restrictions bear on the tools CXMT can buy to expand. The most advanced lithography and etch equipment, the machines that make it possible to build memory competitively at the leading edge, are exactly what the export rules are designed to keep out of a company like this. CXMT is reportedly weighing a second big fabrication plant in the Yizhuang district southeast of Beijing, a project that with its existing site could more than double its capacity to over six hundred thousand wafers a month. A leading-edge memory fab costs well upward of ten billion dollars to build. But the constraint on Chinese memory has never really been the will to spend or the wafers to start; it has been the tools to finish, and that wall is still standing while the door beside it swings open.

• What is reported: HP, Asus and Acer using small volumes of CXMT DRAM in limited notebook models for non-US markets, after qualifying the chips around mid-2026.
• The tell: CXMT is priced at parity with Samsung, not at a discount — so this is about availability, not cost.
• The cause: AI infrastructure is set to take ~70% of memory output this year; the big three (Samsung, SK Hynix, Micron, ~90% of DRAM) redirected wafers to high-bandwidth memory, leaving commodity memory as the residual; DRAM prices have roughly tripled in 18 months.
• The catch: CXMT prioritises Chinese customers such as Huawei, and US restrictions on advanced tools cap how fast it can expand — even as it eyes a second Beijing fab.

Geography, as ever, is destiny

Step back and the shape of it is almost too neat. For years the United States has worked to wall Chinese memory out of the global supply chain, and CXMT sits on the restricted list as part of that effort. This year the West's own AI ambitions created a memory shortage so acute that Western brands began, discreetly, buying from the very company the walls were built around — not because the politics changed, but because the physics of wafer allocation left them no other seat. The dependency the restrictions were meant to prevent is arriving through the side door that the restrictions' own second-order effects propped open. Nobody designed this. It accreted the way chokepoints always do, one rational purchasing decision at a time, until a laptop brand you trust is quietly shipping memory from a fab its own government sanctioned.

And notice who is standing in that doorway. It is not the flagship phone or the premium ultrabook, where memory is a manageable slice of the cost and the brand can afford to pay up for a name it prefers. It is the budget laptop for the market outside the United States, the machine bought by the family or the small business for whom a hundred dollars of memory inflation is the difference between a purchase and a wait. The AI build-out priced them out of the memory they used to take for granted, and the part filling the hole is a Chinese chip that came in at full price through a door scarcity forced. That door can close as fast as it opened — a shift in the shortage, a tightening of the rules, a decision in Beijing to keep more of the output at home — and if it does, the person who feels it first is not in a data center. It is the buyer at the far, thin end of the residual, holding the machine that got built out of whatever memory the world had left over.