Chinese manufacturer CXMT announced on 20 September that its fifth-generation DRAM platform is in mass production, with two products already integrated into high-end Chinese smartphones. Wafer yield is said to be more than 50% higher than the previous generation. Chinese manufacturers now have a second source, beyond the three players that dominate this market.
In August we described the memory shortage as the most immediate constraint weighing on AI, the one that even affected the price of your next phone. Here is the link that was missing from the Chinese chain.
Why memory was the lock
We explained it at the time and it remains the central point. A mixture-of-experts model activates only a fraction of its parameters at any given moment, but the whole thing has to stay in memory.
Memory is therefore not a secondary component: it is what determines which model can run on which machine. And it is the line item where the tension was sharpest, to the point that data centres were competing directly with consumer electronics manufacturers.
A country with domestic accelerators but dependent on abroad for memory does not have an autonomous chain. It has a missing link, and that is where a restriction bites.
What this announcement closes
Put end to end with what we have covered in recent weeks, the sequence becomes legible.
A $532 billion five-year plan to quadruple computing capacity. A cluster of 100,000 domestic accelerators. A leading model served on it at a stated comparable cost. And now memory.
These are the four elements of a chain: capital, compute, software and components. None is at the level of the best in the world. Together, they form a combination that works.
We need to be precise about what this announcement does not say. This is memory for mobile devices, not the very high-bandwidth memory used in AI accelerators, which is a different product and markedly harder to manufacture. The yield figures are the company's own. And mass-producing for domestic smartphones is not the same thing as supplying a global market. The step taken is real, but it is not the one a headline might suggest.
The effect on the rest of the world
One point deserves to be noted, and it concerns your portfolio directly.
We wrote that the memory shortage was driving up consumer electronics prices, because component makers were prioritising their most profitable customers. The arrival of an additional producer mechanically eases that tension, at least for the products concerned.
This is the usual effect of a substitution policy: it aims at a country's self-sufficiency and, along the way, brings down prices for everyone. The same dynamic played out with solar panels and batteries.
What to take away
On Saturday, in our article on the racing metaphor, we wrote that what endures is not the technical lead but the infrastructure: the chips, the electricity, the factories.
This announcement illustrates exactly that point. A model can be caught up with in a few months; a memory industry takes a decade to build. The fact that this one is emerging now is the consequence of decisions taken years ago, long before the AI question was framed in these terms.
For Europe, this is the most useful piece of information of the week. It holds a critical link upstream — the equipment without which no advanced chip can be made anywhere. It manufactures neither accelerators nor memory at scale. Watching a supply chain close ranks elsewhere should at the very least raise the question of which one we want to hold on to.