CXMT HBM3E: What Low Yields Mean for AI Chips
CXMT plans to produce HBM3E despite low reported yields. See what that means for costs, Chinese AI chip supply and competition with memory leaders.
CXMT's HBM3E Bet: Why China Is Producing Memory at Low Yields
China's biggest memory chipmaker, CXMT, is reportedly preparing to produce HBM3E - the high-bandwidth memory used alongside advanced AI processors. The eye-catching detail is the yield: only 20% to 30% of the chips are reportedly usable, versus around 70% at SK Hynix. Reuters reported the figures, citing people familiar with the matter. The Next Web said it hadn't independently verified the report.
That yield gap matters because it doesn't just affect factory efficiency. It changes the cost per usable chip, how much inventory a supplier needs to meet demand, and whether AI chipmakers can actually get enough memory to ship their systems. Still, CXMT's choice to press ahead may not be about undercutting prices or matching performance head-on right away. Chinese AI firms have limited access to advanced memory, and even an expensive domestic option can make strategic sense if the alternative is simply not having enough supply.
This article breaks down what the reported yield means, why CXMT may start with limited production before it "catches up," and what this could change for China's AI chip market - and the global HBM race.
What does a 20% to 30% HBM yield mean?
Yield is the share of manufactured chips that meet the required specifications. If a production run makes 100 chips and 25 pass the tests, the yield is 25%. The rest may be defective, fail performance targets, or need additional work before they can be used. With a product as complex as high-bandwidth memory, yield is a straightforward way to estimate how much usable output comes out of a manufacturing effort.
Reuters' sources put CXMT's HBM3E yield at about 20% to 30%. They put SK Hynix's yield at around 70%. On those numbers, CXMT would likely have to make more units to end up with the same count of usable chips. That doesn't automatically mean costs are exactly three times higher - cost depends on materials, equipment, process steps, testing, rework, and other factors. But a big yield gap does increase the cost per "good" chip.
Why yield is especially important for HBM
HBM isn't just a memory chip that you slot into a system. It's built by stacking memory dies so data can move quickly between the memory and an AI processor. That stacking and packaging work adds complexity, and customers still need parts that meet strict performance and reliability requirements. A device that passes a basic test can still fall short once it's integrated into a finished system.
So yield becomes a system-level problem, not just a factory metric. The supplier needs enough acceptable memory, and the customer needs to be able to integrate it into their AI accelerators and validate the full package. A promising sample is a starting point, not the same thing as a steady stream of production-ready components.
What the yield gap does and does not tell us
A single yield number is a snapshot. It's not a permanent label for a company's skill. Yields can improve as engineers adjust processes, spot recurring failure modes, and refine designs. CXMT has reportedly sent customers two versions of its HBM3E: a standard version and an updated one aimed at improving yields. That suggests the company is still working through production issues rather than treating the first version as fully settled.
At the same time, you can't infer everything from one figure. The report doesn't spell out the test conditions, how much was produced, or which steps are driving the losses. It also doesn't say how fast yields might move. The cautious takeaway is simpler: if the numbers are accurate, CXMT likely has a meaningful efficiency gap versus SK Hynix, which makes reliable, economical volume harder to achieve.
Why is CXMT starting production before yields improve?
Starting limited production with low yields can sound pointless - especially if the goal were purely the cheapest memory. It makes more sense if you treat early manufacturing as part of a longer supplier-and-customer development cycle. CXMT can put samples and early products into Chinese AI chip designers' hands, get feedback from real integration work, and learn how its process behaves when you scale it. Waiting for perfect yields before involving customers would slow down that learning.
According to Reuters, CXMT has sent HBM3E samples to several Chinese AI chip designers and expects to begin limited production within two months of the report. The word "limited" matters. It doesn't suggest CXMT is ready to replace established suppliers across the market. Instead, it points to a constrained start - enough output to support testing, early deployments, or specific customer needs.
The value of learning by producing
Manufacturing improves through a feedback loop: build, test, diagnose, adjust, and build again. With memory, customer testing adds an extra layer - because it shows how a component behaves in the kind of system where it will actually be used. That includes compatibility, performance under real workloads, reliability, and the practical friction of integration with an AI processor.
Early production can be useful beyond the profit on each chip. It can help CXMT identify which process changes improve usable output, and it can show how customers react to the product's performance and cost. The catch is that the loop only works if the supplier can ship enough units for real testing, and if customers are willing to spend time on integration.
The trade-off: progress versus cost
Low yield makes early output expensive. Reuters' sources said CXMT's chips would likely cost more than comparable foreign products, even though the product is reportedly closer to SK Hynix's older HBM3 than to its HBM3E. That's a tough commercial proposition if customers have dependable alternatives at competitive prices.
But buying decisions aren't based only on performance and unit price. Customers also consider availability, whether supply will keep coming, policy constraints, and what it costs to delay an AI chip launch. If advanced memory is scarce or hard to access, even a pricier domestic option may still be worth testing. That's a business calculation - not proof that the product is already cost-competitive.
There's also a hard limit. If yields stay low, production costs remain high and customers see a meaningful performance gap, early demand may not justify large-scale expansion. CXMT needs manufacturing learning to translate into better yields, stronger product performance, and dependable supply. Production is one step in that process, not evidence that the economics already work.
What could CXMT's HBM3E mean for China's AI chip market?
Right now, the immediate significance of CXMT's move is less about knocking out global HBM leaders and more about whether Chinese AI chipmakers can secure another source of memory. AI processors need fast memory to keep data moving during training and inference. If processor makers can't obtain enough compatible HBM, they can end up stuck - even if their own chips are ready.
Reuters' sources said Chinese AI chipmakers were struggling to get enough advanced memory to increase output. The report also said US export controls tightened restrictions on some advanced HBM products bound for China in December 2024. Chinese firms have reportedly tried to source memory through unofficial channels, where it can cost several times more. Together, those conditions help explain why a domestic supplier might attract attention even before it matches established products on cost or performance.
Memory shortages can change chip economics
When a key component becomes scarce, its price can reshape the economics of the final product. Reuters previously reported that Huawei, Cambricon, MetaX, and Iluvatar CoreX raised prices for their AI chips by 20% to 50% as memory costs rose. It also reported that Cambricon's latest sales missed expectations as chip supply tightened. The point is that memory constraints don't stay neatly inside the memory market - they spill over into pricing and sales.
A domestic HBM option could give customers another route to test and, later, source memory. But "another supplier" isn't automatically "enough supply." CXMT's reported low yield limits how much usable output it can produce, and early production is expected to be limited. A domestic source could relieve specific shortages without solving the wider supply problem.
The practical test is customer deployment
The strongest evidence of progress won't be a headline about a product, or even a batch of samples. It's whether customers can qualify the memory, integrate it into working AI systems, and receive enough consistent units to actually deploy those systems. In other words: quality of supply matters as much as the existence of a domestic option.
For AI chip designers, the evaluation looks practical:
- Availability: Can CXMT deliver enough usable memory on a predictable schedule?
- Compatibility: Can the memory work with the customer's processor and system design?
- Performance: Does it meet workload requirements, including data throughput?
- Reliability: Does it hold up consistently in testing and real use?
- Economics: Does the total system cost make sense versus available alternatives?
If CXMT improves yields, those questions may start to look more favorable. If it doesn't, customers may still use its memory selectively - but limited output and high costs will likely cap how widely adoption spreads. Market impact will depend on production results and customer qualification, not on how ambitious the goal sounds.
How far behind are CXMT and the HBM leaders?
CXMT is a proven memory manufacturer, but conventional DRAM and HBM aren't the same yardstick. Reuters' report said CXMT holds about 10% of global DRAM revenue, making it the fourth-largest supplier of conventional memory, according to Counterpoint Research. CXMT also said in September that it had caught up with Samsung and Micron in conventional memory. Those claims relate to DRAM, though, and they don't automatically show that CXMT is at the same level in HBM.
The HBM market is still concentrated around suppliers whose products are already qualified and shipping at scale. Counterpoint Research put SK Hynix at 50% of global HBM revenue in the second quarter, Samsung at 33%, and Micron at 18%. Those numbers describe the position of established competitors, not how much capacity any one customer can immediately rely on.
Revenue share and manufacturing yield are connected to competitiveness, but they're not the same metric.
A performance gap as well as a production gap
Reuters' report points to two issues. First is manufacturing efficiency: the reported 20% to 30% yield for CXMT versus around 70% for SK Hynix. Second is product capability: one source said CXMT's HBM3E was closer to SK Hynix's older HBM3 than to its HBM3E. Even if yields improve, the product still has to meet customer requirements in real systems.
Meanwhile, competitors keep moving. SK Hynix said on Wednesday that HBM4 exceeded its target yield and is expected to surpass HBM3E in shipments and revenue in the fourth quarter. Samsung said in July that it expanded HBM4 sales and shipped HBM4E samples. Those statements don't mean every next-generation product is available in unlimited volume everywhere, but they do show the benchmark moving while CXMT works to establish HBM3E production.
What to watch next
A more useful way to track CXMT's progress is to follow three milestones instead of fixating on one launch date:
- Yield improvement: Does the usable share rise from the early range?
- Customer qualification: Do AI chip designers move from samples to validated products and place repeat orders?
- Consistent volume: Can CXMT deliver enough units over time at a cost customers can accept?
Each milestone depends on the one before it, but none guarantees the next. Better yield can reduce cost without solving a performance gap. Qualification can prove the product works without proving it can ship at scale. And volume is only meaningful if reliability and economics hold up.
The balanced view is that CXMT's plan is a serious attempt to build domestic HBM capability, but the reported numbers don't yet justify treating it like a near-term peer to SK Hynix, Samsung, or Micron in advanced memory. Its impact will come down to whether early manufacturing experience becomes repeatable quality and usable supply for customers.
Conclusion
CXMT's reported plan to start limited HBM3E production despite yields of 20% to 30% is best read as an early manufacturing and supply-chain move - not proof that it has caught up with established HBM suppliers. Low yields make each usable chip harder and more expensive to produce, and the reported performance gap adds another challenge.
For China's AI chipmakers, even a flawed domestic option could be worth testing when advanced memory is hard to get. But the real test is measurable progress: higher yields, successful customer qualification, and consistent production at a cost the market can tolerate. Those are the signals to watch - more than the fact that production is starting.