China’s ChangXin Memory Technologies plans to start production of HBM3E, an advanced high-bandwidth memory chip used with AI processors, even though only 20% to 30% of the chips it makes currently meet specifications, Reuters reported on October 28, 2026. The company, known as CXMT, has sent samples to several domestic AI chip designers and expects to begin limited production within the next two months, according to two people familiar with the matter who asked not to be named. CXMT did not respond to a request for comment, and SK Hynix did not immediately respond to the yield estimate.
A Shortage That Makes Waste Affordable
High-bandwidth memory, or HBM, is made by vertically stacking memory chips so they can exchange data rapidly with AI processors. It is a crucial component in the specialised chips used to train and run large language models, and it has become one of the tightest bottlenecks in the global AI hardware supply chain. When yields are low, production is expensive, because a large share of chips must be discarded or reworked. CXMT is pressing ahead anyway, the sources said, because Chinese AI chipmakers have struggled to obtain enough advanced memory to increase output.
That logic turns the usual manufacturing calculus on its head. In a normal market, a 20% to 30% yield would be a reason to keep a process in the laboratory. In a market defined by shortage, it can be a reason to ship: some usable output, even at a high unit cost, is worth more than no output at all. One of the sources told Reuters that CXMT’s HBM3E is expected to cost more than comparable products from overseas rivals, a premium that domestic buyers appear willing to absorb for now.
How Far Behind The Leaders CXMT Sits
The yield gap is only part of the distance CXMT must close. In testing, its HBM3E performance was closer to SK Hynix’s previous-generation HBM3 than to the South Korean company’s HBM3E, one source said. CXMT has sent customers two versions: a standard design and an updated version intended to improve yields. The company began developing the technology internally as early as 2024.
The established leaders, SK Hynix, Samsung Electronics and Micron Technology, dominate global HBM supply. SK Hynix held 50% of global HBM revenue in the second quarter, according to Counterpoint Research, followed by Samsung with 33% and Micron with 18%, with most revenue coming from HBM3E. Samsung said in July that it had expanded HBM4 sales and shipped HBM4E samples to major customers, putting CXMT at least one generation behind in commercial HBM production. CXMT, by contrast, held a 10% share of global DRAM revenue in the quarter, making it the world’s fourth-largest supplier of conventional memory chips, but it is not yet a major force in HBM.
Export Controls And The Domestic Push
The backdrop is Washington’s tightening of export controls on certain advanced HBM products bound for China in December 2024. Chinese chipmakers have increasingly sought the memory through unofficial channels, where it can cost several times more, Reuters reported. The controls, combined with a government-backed push to develop domestic alternatives to Nvidia’s AI processors at scale, have made local memory supply a strategic priority rather than a purely commercial question.
Reuters reported last month that Huawei, Cambricon, MetaX and Iluvatar CoreX had raised prices for current and next-generation AI chips by 20% to 50% as memory costs complicated efforts to ramp up production. That price pressure helps explain why a low-yield domestic product can still find buyers: when the alternative is constrained supply or grey-market pricing, an expensive local chip may still be the cheaper option once delivery certainty is factored in.
US export controls also limit CXMT’s access to some of the most advanced chipmaking equipment, making yield improvement harder. The updated HBM3E version sent to customers suggests that work is under way, but Reuters’ sources give no timeline for reaching SK Hynix’s roughly 70% yields.
Why It Matters
This matters because memory, not just processors, now sets the pace of AI expansion. The industry’s attention has focused on who can buy or build the fastest accelerators, as in our report on the $63 billion financing behind Anthropic’s chip bet and our report on OpenAI’s reported plan to buy Google AI chips. CXMT’s decision is a reminder that those accelerators are only as useful as the memory stacked beside them, and that the memory market’s geography is now a geopolitical question as much as a technical one.
For China’s AI chip designers, limited domestic HBM3E production could ease the most acute shortages even if it does not solve them. For SK Hynix and its rivals, it is an early warning that a low-yield entrant today can become a volume competitor tomorrow, particularly if domestic demand guarantees it a market while it climbs the yield curve. And for the wider AI industry, it underlines how much of the buildout depends on a handful of specialised components whose supply cannot be expanded overnight, as also seen in our report on SK Hynix’s $75 billion AI memory windfall.
Background
CXMT is China’s largest maker of dynamic random-access memory, the conventional memory used in computers and servers. Its move into HBM3E follows years of investment in advanced memory technology, and it comes as Beijing pushes for greater self-reliance across the semiconductor supply chain in response to US restrictions.
Source: Reuters report
0 Comments