CXMT Starts HBM3E With 20-30% Yields: What Reuters Reports and What Is Verified

China's largest DRAM maker, CXMT, is reportedly about to start making HBM3E, the high-bandwidth memory that sits next to AI accelerators, even though most of what comes off the line does not meet spec. In an exclusive dated October 28, 2026, Reuters reports, citing sources, that yields are about 20% to 30%, against around 70% at SK Hynix, and that limited production could begin within about two months.

This is a "claimed versus verified" post. The number is specific and plausible, but it rests on unnamed sources. Below is what is attributed to whom, what earlier reporting adds, what nobody has confirmed, and why it matters if you build on or buy AI hardware. It extends our earlier coverage of China's STAR50 selloff and CXMT and of the wider memory shortage.

TL;DR: claimed versus verified

What Reuters actually reports

According to the Reuters exclusive as syndicated by Investing.com, CXMT expects to begin HBM3E production in about two months, with limited output. Reuters says only "a small proportion" of its chips currently meet specifications. It contrasts the 20% to 30% yield with roughly 70% at SK Hynix, the HBM leader, and says low yields raise cost because so many chips must be discarded or reworked.

The story adds that CXMT has sent samples to several unnamed domestic AI chip designers, and describes the report as adding yield and timing detail to an earlier piece from The Information. The Techmeme summary of the same Reuters article adds that the HBM3E's performance is closer to SK Hynix's previous-generation HBM3.

We could not open the original Reuters page directly (the syndicated copies we tried returned access errors), so details here come from the Investing.com copy and the Techmeme summary. Where we only have a second-hand summary, the table above says so.

The earlier trail: The Information and SemiAnalysis

This is not a surprise from nowhere. At the end of August, The Information reported that CXMT had begun producing HBM3E in small quantities, as relayed by Korea JoongAng Daily. That article adds several data points:

• CXMT plans to ramp up next year, and Chinese customers could use the chips commercially as early as next year if they pass final validation.
• Alibaba's chip unit T-Head and Cambricon are reportedly testing CXMT's HBM3E for AI accelerators. That is reported, not confirmed by those companies in the article.
• SemiAnalysis estimated in June that 8-layer HBM3 had roughly 35% front-end and 70% back-end yield, about 25% overall, and that 12-layer HBM3 and HBM3E would be lower.
• Samsung, SK Hynix and Micron are mass-producing HBM4, a generation ahead.
• CXMT raised about 12 trillion won through a Shanghai STAR Market listing in July, and reported first-half revenue up 874% year over year to 150.3 billion yuan.

The 25% SemiAnalysis figure is for 8-high HBM3, not HBM3E, so it is context rather than corroboration of the new 20% to 30% number. It does show that a low-20s-to-30% range is in line with outside estimates for CXMT's earlier generation. Treat the agreement as consistent, not confirming.

Why yield is the number that matters

HBM stacks several DRAM dies vertically and connects them with thousands of tiny vertical links. A flaw in any die or any connection can sink the whole stack, so stacked memory compounds yield losses. Front-end yield (making good dies) and back-end yield (stacking and packaging) multiply together. That is why SemiAnalysis's 35% and 70% combine to roughly 25%.

At a 25% overall yield you throw away three of every four stacks. The cost of good parts rises accordingly, and a company can only supply customers if it either has large capacity or accepts the loss. One of Reuters' reported reasons CXMT is pushing ahead anyway is demand: Chinese AI chipmakers have struggled to obtain enough advanced memory to raise output. Our look at Nvidia paying suppliers for packaging capacity, Amkor and the $3 billion total, shows the same pattern outside China: advanced packaging and stacking are the bottleneck.

How CXMT compares with the leaders

HBM3E is a fifth-generation high-bandwidth memory used in AI accelerators, per Korea JoongAng Daily. The three incumbents, SK Hynix, Samsung and Micron, have already moved their volume production to HBM4, so CXMT is shipping samples of a product the leaders consider mature. That gap matters for two reasons.

First, accelerator designers pick memory early in a chip's life. A design that targets HBM3E bandwidth cannot simply swap in HBM4 later, so Chinese designers who qualify CXMT parts are locking in a lower bandwidth ceiling. Second, the leaders keep improving yield as a product ages, while CXMT is starting from a low base. Closing a yield gap of 20% to 30% against roughly 70% usually takes quarters of process tuning, not weeks.

None of that means the effort is futile. A domestic supplier that reaches even half the leaders' yield changes negotiating leverage and reduces dependence on grey-market HBM. The open question, which only filings and customer announcements can answer, is how fast the curve rises.

Export controls make imperfect memory valuable

Reuters notes that after US export controls on certain advanced HBM tightened in December 2024, Chinese firms reportedly bought supplies through unofficial channels at several times the usual cost. Against that backdrop, even costly and low-yield domestic HBM has a buyer. Accelerator designers in China who cannot import Korean HBM may accept worse economics for guaranteed supply.

This ties to a thread we have followed: how companies route around limits, from leasing chips abroad to Samsung and SK Hynix upgrading their China memory plants under licenses. CXMT's HBM effort is the domestic-supply leg of the same story.

What is not verified

1. The yield figure. It is an anonymous-source number. CXMT has not publicly disclosed yield, stack configuration, bandwidth, packaging partner or completed customer qualification.
2. Volume. Neither Reuters nor the earlier reports give production quantities. "Limited" could mean engineering samples.
3. Customers. The customers are described only as unnamed domestic designers. T-Head and Cambricon testing comes from The Information via Korean press.
4. Performance. "Closer to HBM3" is a characterization by sources, not a benchmark.
5. Price. A secondary summary says one source expects CXMT's HBM3E to cost more than overseas rivals, but we could not confirm that line in the Reuters text we read.
6. Timing. Two months for production and "next year" for a ramp are plans, which slip in semiconductors.

How to check this yourself

• Read the original Reuters piece and note which claims carry "two people familiar" attribution.
• Watch CXMT's STAR Market filings and quarterly reports for language on HBM revenue or capacity.
• Look for named customers: a T-Head, Cambricon or Huawei announcement that mentions domestic HBM would be real confirmation.
• Compare with SK Hynix's own commentary; our SK Hynix Q3 2026 write-up covers HBM4 overtaking HBM3E and why the company sees shortages through 2030.

What this means for what you build or pay

For most builders outside China, nothing changes this quarter. CXMT's HBM3E is a generation behind, in small supply and aimed at Chinese accelerators. The indirect effects are more interesting.

• Memory prices. If China's AI chipmakers can partly self-supply, they may buy fewer HBM stacks and DRAM from Korean vendors, but at these yields the relief is small. Our RAM price and local inference cost analysis explains why AI demand keeps pulling consumer memory prices up.
• China's AI chip roadmap. The binding constraint on domestic accelerators is memory. Better domestic HBM, even at poor yield, lifts the ceiling on how many accelerators can ship.
• Planning. If you buy or rent compute tied to Chinese-made accelerators, watch memory capacity as a leading indicator of supply.

The bottom line

Reuters' report is credible and fits earlier estimates, but the headline 20% to 30% is a sourced estimate, not a disclosed number. The honest summary: CXMT is reportedly moving to HBM3E production with weak yields, one generation behind the leaders, because supply-starved Chinese chipmakers will take it. We will add an update if CXMT, a named customer or a filing confirms or contradicts the figures.

Related reading

• China's STAR50 falls 40%: what it means for AI builders
• SK Hynix Q3 2026: HBM4 overtaking HBM3E
• Michael Dell on the four-year memory shortage
• Samsung and SK Hynix upgrading China memory plants
• Nvidia pays suppliers for AI packaging
• Tencent leases 100,000 Oracle AI chips

Figures are accurate as of October 28, 2026 and rely on anonymous-source reporting that may be revised.