The yield is the story
A production start is usually reported as a date. Here the date is the least useful number in the story. Yield decides what a start is worth, because every failed stack still consumed the wafer, the equipment time and the materials. As a simplified illustration, a 25% yield means roughly four attempted stacks for each passing stack, compared with about 1.4 attempts at a 70% yield. Actual production costs also depend on process complexity, materials, capacity and other factors not disclosed in the report. That gap does not disappear inside the factory. It lands in the unit cost, which is consistent with what the sources told Reuters: CXMT's HBM3E is expected to cost more than comparable products from overseas.
It also caps what this launch can do for supply. A 20% to 30% yield would limit qualified output relative to the number of stacks attempted. Combined with Reuters' description of a limited production start, that suggests an early-stage manufacturing effort rather than evidence of substantial near-term supply. Read the report as evidence of CXMT's progress toward HBM3E production: Reuters says samples in two designs have reached domestic AI chip developers, while limited production is planned within two months. Neither customer qualification nor meaningful commercial shipment volumes have been established. The second quarter's HBM revenue sat with three established makers, split 50% to SK hynix, 33% to Samsung and 18% to Micron.
Behind on the label as well as the line
The second discount is generational. One of Reuters' sources said CXMT's HBM3E performed in testing closer to SK hynix's HBM3, the previous generation, than to the HBM3E it is named for. And the leaders have already moved past the generation CXMT is entering: Samsung said in July that it had expanded HBM4 sales and shipped HBM4E samples to major customers. Reuters' own summary puts CXMT at least one generation behind in commercial HBM production, with US export controls limiting its access to some of the most advanced chipmaking equipment it would need to close the gap faster.
None of this makes the milestone trivial. CXMT is the world's fourth-largest conventional memory maker, with a tenth of global DRAM revenue, and HBM is the hardest product in its industry to manufacture. Getting a stack to pass specification at any yield is the step every established maker also had to take. The question the headline cannot answer is the slope: whether the updated design already sampling with customers lifts yields toward the range where the economics change, or whether the line stays a low-yield demonstration for another generation.
Why ship a line that mostly makes scrap
The reason to start anyway is in the demand side of the same report. Washington tightened export controls on certain advanced HBM products bound for China in December 2024. Since then, according to the sources, Chinese AI chipmakers have struggled to obtain enough advanced memory to increase output, and buyers routed to unofficial channels have paid several times the normal price. Reuters reported last month that Huawei, Cambricon, MetaX and Iluvatar CoreX had raised prices on current and next-generation AI chips by 20% to 50%, with memory costs complicating their production ramps.
Against that backdrop, an expensive domestic stack with a known yield problem can still be the rational purchase. If CXMT can qualify its HBM3E with domestic AI chipmakers and sustain production, buyers facing export restrictions and expensive alternative supply channels may have reasons to accept a higher-cost domestic option. But the report does not establish completed commercial purchases, dependable delivery volumes or the price customers would actually pay.
Sources
- [1] Reuters, "China's CXMT to start advanced AI memory production despite poor yields, sources say," October 28, 2026 (production plan, yields, samples and two designs, performance comparison, pricing expectation, Counterpoint market shares, export-control context; read in full via syndication)Read source
- [2] AI Weekly, "China's CXMT to Start HBM3E Memory Output at 20-30% Yields," October 28, 2026 (summary of the Reuters reporting, including the unofficial-channel pricing detail)Read source
- [3] Finimize, "China's CXMT Pushes Into HBM3E With Yields Near 20%," October 28, 2026 (summary of the Reuters reporting)Read source