CXMT HBM3E: Production Planned at 20-30% Yield

China’s biggest memory chipmaker is going to start building a part that most of its output can’t yet pass as good, for China’s push to build alternatives to Nvidia’s AI chips. That’s the core of an exclusive Reuters report from Wednesday, October 28, and the number inside it is small enough to quote in full: 20% to 30%.

CXMT (ChangXin Memory Technologies) is China’s largest maker of DRAM, the ordinary memory in PCs and servers. Reuters, citing two people familiar with the matter, says it plans to begin limited production of HBM3E, the stacked high-bandwidth memory that sits next to AI processors. One source expects that within the next two months. CXMT didn’t respond to Reuters’ request for comment. The two sources weren’t named because the plans aren’t public.

What 20 to 30 percent means

Yield is the share of chips coming off a line that meet spec. Reuters says CXMT’s is about 20% to 30%, against around 70% at SK Hynix, the market leader. Reuters’ lead sentence says “as little as 20%”, so treat the range as an estimate from two people, not a measured figure.

Low yield is mostly a cost problem. Every bad chip is money gone. A large share of chips has to be thrown away or reworked, so each good one costs more. One source told Reuters CXMT’s HBM3E is expected to be pricier than comparable foreign products.

Performance is the second gap. In testing, one source said, CXMT’s chip came out closer to SK Hynix’s older HBM3 than to its HBM3E. CXMT has sent customers two versions, a standard design and an updated one with changes meant to lift yield, according to the other source. The sources add that CXMT started developing the technology internally as early as 2024.

Why ship it anyway

Because, per Reuters’ sources, Chinese AI chip designers can’t get enough of it. Washington tightened export controls on certain advanced HBM products headed to China in December 2024, and the designers haven’t been able to buy enough to raise output. Some have turned to unofficial channels, where the memory can cost several times the normal price.

That squeeze already showed up in chip prices. Reuters reported in September that Huawei, Cambricon, MetaX and Iluvatar CoreX had raised prices on current and next-generation AI chips by 20% to 50%, citing memory costs. CXMT has also sent HBM3E samples to several domestic AI chip designers, per Reuters. The Information reported CXMT’s HBM3E plans first; the yield and timeline details are Reuters’ own.

Detail	What Reuters’ sources say
CXMT HBM3E yield	About 20% to 30%
SK Hynix yield	Around 70%
Start of limited production	Within two months, per one source
Tested performance	Nearer SK Hynix’s HBM3 than its HBM3E
Price	Expected above comparable foreign parts

My read

I’d resist the easy line that 20% yield means this goes nowhere. Early yields at every memory maker start low, and Reuters doesn’t say how fast CXMT could improve. What the report does show is a buyer’s market turned upside down: when supply is blocked, a costly, weaker part is still worth making, because the alternative is idle chip lines.

Volume is the gap. I can’t tell it from the report. “Limited production” at 25% yield might be a few thousand good stacks or a lot more, and no source gave a number. Until there’s one, this is a signal about China’s memory shortage more than a threat to SK Hynix. For the policy side, see our earlier post on China’s draft licence rule for exporting its biggest AI models.

Watch for CXMT or its customers saying anything on the record, and for any word on shipped volume or on a yield improvement from the updated design.

Sources

Note on sources

I read the Reuters text as it appeared in search results for the Investing.com reprint, and the other three only as search summaries. I did not open the Reuters original, and every figure is an unverified claim from two unnamed sources.