CXMT has shipped two versions of HBM3E to Huawei, Cambricon, MetaX, and Iluvatar CoreX: a standard design and a revised version carrying changes meant to lift manufacturing yield. Only 20 to 30 percent of the chips are usable. Testing places the part closer to SK Hynix's previous-generation HBM3 than to the HBM3E it is named after. And it will cost more than comparable foreign products.
No normal business sells that product. A memory maker with a third of its rival's yield, a generation behind on performance, and a higher price tag is not competing. It is being funded.
The comparison figure is the whole story in one number: SK Hynix runs around 70 percent yields on HBM. CXMT runs 20 to 30. Low yield means a large share of output is scrapped or reworked, and the depreciation and manufacturing cost allocated to each surviving die balloons. The resulting price is not a mistake. It is the invoice for doing this without the equipment that makes it cheap.
Low yield means a large share of output is scrapped or reworked, and the depreciation and manufacturing cost allocated to each surviving die balloons.
The equipment gap is visible in the silicon itself. CXMT is attempting HBM3E with only 3,000 through-silicon vias per layer. Samsung's HBM2 used about 5,000. SK Hynix's HBM3 uses more than 8,000. Front-end yields sit near 30 percent, and the bottleneck is TSV maturity — the vertical wiring that turns a stack of DRAM into one coherent part.
Fewer vias, lower density, lower bandwidth headroom. CXMT is targeting an 8-high stack and is unlikely to move to 12-high while the engineering problems persist. The part is late, thin, and expensive by design, because the alternative is not having one.
That alternative was foreclosed in December 2024, when Washington tightened export restrictions on advanced HBM bound for China. The ban did not just remove the chips. It removed the tools — EUV lithography and top-tier hybrid bonding equipment — that make stacked memory manufacturable at scale. Chinese firms turned to unofficial channels, paying several times normal cost for unstable supply.
The cost surfaced downstream fast. Huawei, Cambricon, MetaX, and Iluvatar CoreX raised prices on current and next-generation AI chips by 20 to 50 percent as memory procurement costs climbed. Cambricon's latest sales missed expectations as chip supply tightened. Memory shortage became compute shortage, and compute shortage became a ceiling on the entire domestic AI buildout.
This is where the balance sheet matters. CXMT holds 10 percent of global DRAM revenue — fourth-largest supplier of conventional memory, up from 4 percent a year earlier — and posted an operating margin near 83 percent in the second quarter. That profit is not a verdict on manufacturing prowess; Korean analysts note that Chinese accounting treatment and state support can inflate the figure, and depreciation from new capacity will pull it down. But it is real money, and it is the reason CXMT can sell HBM3E below cost while its customers absorb the difference.
The strategy has a name inside the industry: breakthrough first, optimize later. Deliver samples before yields justify delivery. Let customers adjust PCB layouts and underlying firmware to match the chip, a collaborative compromise that builds a working system out of immature parts. Fix the yield in production, not in the lab.
Whether that works depends on the supply chain behind the fab, not the fab itself. Naura, Maxwell, and U-Precision are developing domestic tools for HBM production, but they lack mass-production experience, and inspection and testing capability remains a bottleneck. Samsung's recovery from its own HBM yield crisis — HBM3 stuck at 10 to 20 percent in 2024, before HBM4 reached stable early yields — came from an ecosystem: its own foundry for base dies, advanced packaging, DTCO coordination between memory and logic. CXMT has the fab expansion. It does not yet have the ecosystem.
The expansion is coming anyway. CXMT is pushing a new Shanghai DRAM fab toward a 2027 production start, on top of existing capacity in Hefei and Beijing. When that depreciation lands, the 83 percent margin will compress. The bet is that by then, HBM yield has climbed and the domestic stack has a customer base that cannot buy anywhere else.
That customer base is the part export controls cannot reach. The Sugon 8000 — "Dengfeng" — is a 100,000-card supercluster at the National Supercomputing Center in Zhengzhou, built on Hygon x86-compatible CPUs and DCUs, wired with Sugon's own InfiniBand-like scaleFabric interconnect, running ParaStor storage and immersion phase-change liquid cooling for single-cabinet megawatt density. It was delayed for years after the 2019 sanctions on Sugon and THATIC forced the Chinese to continue chip development without AMD's assistance.
What Sugon has not disclosed is nearly everything that would let outsiders evaluate it. No processor count. No accelerator type. No memory type or capacity. No rack count. No power requirements. No Linpack score. The system is plugged into China's National Supercomputing Internet, already used for 300 applications from large models to drug discovery to quantum computing, and the world cannot benchmark it.
That opacity is the point, not an oversight. China now has at least four exascale-class machines — LineShine, which took the TOP500 top spot in June with 2.2 exaflops, plus Oceanlite, Tianhe-3, and now Dengfeng. Only LineShine has a published Linpack score. The rest exist whether or not anyone can measure them.
Which is why the framing of this contest keeps missing. PwC's South Korean member firm warned in October that China's growing capacity in conventional DRAM, NAND, mature-node foundry, and assembly and packaging could reshape competition — that Chinese production could influence global supply, pricing, and investment decisions, and that Korea needs to move toward higher-value products and defend cost competitiveness. That warning is not about HBM. It is about the floor underneath HBM.
CXMT already has 10 percent of DRAM revenue and the margin to lose money on advanced memory for years. If that conventional-memory capacity compresses Korean margins while CXMT's HBM climbs from 20 percent yields toward something usable, the question stops being whether CXMT matches SK Hynix.
It becomes whether Beijing can absorb losses long enough that the answer no longer matters. The Shanghai fab is scheduled for 2027. The samples are already at Huawei. The Sugon 8000 is running workloads nobody outside China can score.