The consensus memory trade rests on a single assumption: Chinese supply cannot arrive before the 2027 demand peak. That assumption is now carrying more weight than the evidence behind it. ChangXin Memory Technologies (CXMT) is reported to be running 8-high HBM3 at roughly 25% overall yield, around 30% through the front end, with only about 70% of those stacks surviving back-end packaging. Close to 80 of every 100 stacks fail final test. Read as a verdict, that is a failure. Read as a position on a learning curve, it is an ordinary early ramp. And while DRAM analysts argue about it, the NAND half of the same trade has already broken. Counterpoint Research placed China’s Yangtze Memory Technologies Co (YMTC) third in global NAND bit shipments in Q2 2026 with a 14% share, ahead of both Kioxia and Micron. China is behind in DRAM and HBM. It is already winning in NAND. The market is pricing both as one trade.
The Insight: yield slope, not yield level, is the variable that settles this
A 25% stacked-die yield is not evidence that CXMT cannot build high-bandwidth memory. It is evidence that CXMT sits where every HBM manufacturer has sat early in a new stack generation, before through-silicon via formation, thermocompression bonding and known-good-die screening are tuned. SK hynix now runs above 90% on HBM3. It did not start there, and the informal “golden yield” bar for volume production is around 80%. The question reduces to one derivative: how many quarters does CXMT need to travel from 25% to 60%, and from 60% to 80%? If that climb takes two years, the 2027 supply gap stays open and incumbent pricing power holds. If it takes four quarters, funded by a balance sheet that does not need the interim wafers to be profitable, Chinese HBM demand gets served locally from 2027 and the marginal buyer leaves the incumbents’ order book. Everybody is modelling the level. Nobody is modelling the slope.
Key facts
- CXMT 8-high HBM3 yield: about 30% front-end, roughly 70% of those stacks passing back-end test, near 25% overall, per a South Korean industry report relayed by Chosun Ilbo on 9 September 2026. SemiAnalysis modelled 35% and 70% in June 2026, also landing near 25%.
- Incumbent benchmark: SK hynix has run above 90% on HBM3 since mass production began in 2022. The industry “golden yield” bar is about 80%.
- NAND ranking: YMTC took 14% of global NAND bit shipments in Q2 2026, third behind Samsung at 25% and SK hynix at 22%, per Counterpoint Research. It ranked fifth on revenue.
- DRAM share: CXMT reached 10% of global DRAM revenue in Q2 2026, up from 4% a year earlier, cutting the Big Three’s combined share to about 87%.
- Capital: CXMT raised RMB 57.9 billion (about $8.6 billion) on the Shanghai STAR Market on 27 July 2026, a record for the board, closing its debut up 466% from an RMB 8.66 offer price.
- HBM concentration: Q2 2026 HBM revenue share was SK hynix 50%, Samsung 33%, Micron 18%. SK hynix held 64% a year earlier.
- Supply gap: TrendForce expects HBM to absorb 30% of DRAM wafer input but only 13% of bit supply by end-2027, against 18% and 8% at end-2025.
What CXMT and YMTC have actually achieved
Start with what is not in dispute. CXMT is now a top-four DRAM supplier by revenue and the fastest-growing on the board. Counterpoint measured its DRAM revenue up 716% year on year in Q2 2026, taking it to a 10% global share from 4%. That is the first time a Chinese producer has reached double digits, and it pushed the three-decade oligopoly of Samsung, SK hynix and Micron below 90%.
The July listing changed the funding picture as much as the share data changed the competitive one. CXMT raised RMB 57.9 billion on the STAR Market, beating SMIC’s RMB 53.2 billion record, and closed its debut at RMB 49 against an RMB 8.66 offer price. What matters is where the money goes. Of roughly RMB 29.5 billion in net proceeds, the prospectus allocates RMB 13 billion to DRAM technology upgrades, RMB 9 billion to next-generation DRAM research and RMB 7.5 billion to wafer line upgrades. There is no high-bandwidth memory line item.
The HBM programme is real but it is being funded off-balance-sheet
That omission is the most under-read fact in the file. CXMT is clearly working on HBM: it reached risk production of HBM3E in September 2026, with small volumes at Alibaba’s T-Head unit and AI chipmaker Cambricon for qualification. Yet the prospectus points investors at conventional DRAM. The inference is that the HBM effort is carried by state channels rather than public equity, the structure that lets a 25% yield persist without commercial consequence. A listed Western manufacturer scrapping three of every four stacks would halt the line. A national champion keeps running wafers, because the wafers are how the learning happens.
Counterpoint estimates CXMT at roughly 9% of DRAM bit shipments today, rising toward 11% by 2028, with capacity going from about 320,000 wafer starts per month to 420,000 by 2027. SemiAnalysis models roughly 350,000 by end-2026 against an estimated 385,000 at Micron. On raw wafer input the gap is close to closed. What remains is bit density, cost per bit and stacked-die assembly.
NAND is a different story entirely
YMTC does not have a yield problem worth discussing. It is mass-producing 267-layer 3D NAND on its Xtacking architecture, its two Wuhan fabs run a combined 200,000 wafers per month, and three further fabs designed at 100,000 each would take it toward 500,000. Counterpoint’s Q2 2026 data has YMTC at 14% of global NAND bit shipments, up 22% year on year, third ahead of Kioxia and Micron.
The nuance the headlines skipped is the revenue split. YMTC came third on bits and fifth on revenue, behind Kioxia and Micron, because its mix is still weighted to consumer applications. Enterprise SSDs took 48% of global NAND bits in Q2 2026, almost double the 26% a year earlier, and that is where the price per bit sits. YMTC has won volume without yet winning value. For anyone tracking the names levered to the NAND upcycle, bits set the floor and mix sets the margin.
Neil Shah, Vice President of Research at Counterpoint Research, framed the DRAM side bluntly: “The question is not about ‘how’ but ‘when’ CXMT will be able to break into the Big Three Memory Club.”
How the incumbents and their customers are responding
The incumbents are not behaving like companies that believe Chinese entry is imminent. They are behaving like companies that have sold out. Micron has committed effectively all of its HBM output through end-2026 and booked its 2027 capacity, with data-centre customers asking for roughly 50% more supply than it can commit. That is the posture of a supplier with no incentive to build ahead of demand, the condition that leaves an opening for a subsidised entrant.
Within HBM, concentration is loosening. Counterpoint’s Q2 2026 read has SK hynix at 50% of HBM revenue, down from 64% a year earlier, Samsung at 33% and Micron at 18%. SK hynix has completed HBM4 development, claiming 10 Gbps data rates and a 40% power efficiency gain; Micron has shipped HBM4 samples rated to 11 Gbps. Nvidia’s Rubin Ultra is expected to carry 384GB of HBM per GPU, the reason HBM keeps consuming more of every DRAM fab’s output.
There is a second-order effect that gets missed. Every wafer redirected into HBM leaves the conventional DDR5 market, and the vacancy that creates in China is the space CXMT has filled. Samsung’s own HBM pivot handed CXMT a domestic commodity DRAM opening it did not have to fight for. TrendForce measured DRAM contract prices rising roughly 90% to 95% quarter on quarter in early 2026, NAND up 55% to 60%. The rotation in DRAM share between Micron and SK hynix reflects the same reallocation.
Hyperscalers have started doing what buyers always do in a shortage: qualifying more suppliers. TrendForce notes that persistent tightness is pushing customers to diversify their supplier base, a shift it expects to benefit CXMT. Korean equities read the same tape, with Samsung and SK hynix trading on AI capex headlines.
Sanjay Mehrotra, Chief Executive Officer of Micron Technology, has been explicit about why demand is structurally different this cycle: “Today there is no AI without memory. AI systems need more memory. They need higher performance memory. They need lower power memory. So, the value of memory, that equation has totally changed.”
MS Hwang, Research Director at Counterpoint Research, put the incumbent share dynamic in context: “Although SK hynix registered record earnings sequentially and annually, it grew slower than the competition as its market share declined.”
Market impact and data analysis
The cleanest way to see why treating DRAM, HBM and NAND as one China trade is a mistake is to line them up side by side. China’s position is materially different in each.
| Segment | Chinese position (Q2 2026) | Incumbent position | Gap |
|---|---|---|---|
| NAND (bit shipments) | YMTC 14%, third | Samsung 25%, SK hynix 22% | Closed on volume |
| NAND (revenue) | YMTC fifth, behind Kioxia and Micron | eSSD mix drives value | Wide, mix-driven |
| Conventional DRAM | CXMT 10% of revenue, from 4% | Big Three about 87% | Narrowing quarterly |
| DRAM capacity | CXMT approx. 320k wafer starts/month | Micron approx. 385k | Near parity |
| HBM yield | CXMT approx. 25% on 8-high HBM3 | SK hynix above 90% | Wide, and decisive |
| HBM generation | HBM3E at risk production | HBM4 qualifying and sampling | One full generation |
On pricing, a caution. Figures circulating in market commentary that put HBM3e at roughly $1,500 to $3,000 per unit and HBM4 at $2,200 to $4,160 could not be traced to a primary research publication and do not reconcile with per-stack pricing. What is sourceable is narrower: reports place SK hynix HBM4 12-high in the mid-$500s per stack against HBM3E 12-high in the mid-$300s, a premium above 50%.
The supply arithmetic is where the CXMT yield curve enters the model. Morgan Stanley has modelled a DRAM shortfall of roughly 17% in 2026 and 15% in 2027. Add a scenario: if CXMT reaches 60% HBM yield during 2027 and serves a third of Chinese accelerator demand internally, the incumbents lose a buyer with nowhere else to go. That does not collapse pricing. It compresses the tail of the cycle, where most of the terminal value in memory equity valuations sits.
Yang Wang, Principal Analyst at Counterpoint Research, on the duration of the squeeze: “The impact is expected to continue through H2 2027, as it will take several quarters for memory supply expansion to materialize.”
Export controls and the regulatory tension
The policy layer makes the yield curve unusually hard to forecast, because it acts on both sides of the equation.
Since December 2024, US rules have applied country-wide controls on exports of advanced HBM to China, written around memory bandwidth density rather than raw bandwidth so they capture designs using more, smaller dies. The same package added controls on 24 categories of semiconductor manufacturing equipment and listed around 140 entities including toolmaker SMEE. Foreign-produced HBM is reached through the advanced computing Foreign Direct Product rule. In September 2025 the Bureau of Industry and Security adopted the 50 percent rule, extending restrictions to majority-owned affiliates.
CXMT’s own status is more ambiguous than commentary suggests. It is not on the BIS Entity List, despite repeated consideration since 2023. It is on the Department of Defense Section 1260H list, named in January 2025 and reaffirmed in June 2026. That distinction matters: 1260H carries procurement and reputational consequences, not the Entity List’s licensing burden.
The 2026 escalation is legislative, not regulatory
The most consequential 2026 development is the Multilateral Alignment of Technology Controls on Hardware Act, introduced in April 2026 by Senators Jim Risch, Pete Ricketts and Andy Kim as S.4281, with a House companion at H.R.8170. It cleared the House Foreign Affairs Committee 44-0. The bill would push allied suppliers to adopt country-wide controls on covered equipment with a policy of denial, and threaten expanded US jurisdiction over firms in allied nations, including ASML, Nikon and Canon, if their governments fail to align within 150 days.
The Dutch and Japanese dimension is no longer theoretical. Pressure connected to the affiliates rule contributed to the Netherlands moving against Nexperia in late 2025, triggering Chinese retaliation and real automotive supply chain disruption. That episode is the template for how equipment controls now propagate: the restriction lands on a tool or an owner, the counter-move on a downstream assembly bottleneck.
The tension is straightforward. Controls slow CXMT’s access to advanced lithography, bonders and metrology, flattening its yield curve. They also ensure Chinese accelerator makers cannot buy SK hynix or Micron HBM, handing CXMT a captive customer base willing to accept 25% yields. Denying the tools and creating the demand at once buys time rather than wins.
What happens next
Three things are worth watching, each with a testable trigger.
First, CXMT’s HBM3E qualification results at T-Head and Cambricon in the first half of 2027 are the highest-information datapoint in the sector. A pass at commercial volume, even at poor yield, means Chinese accelerators stop being HBM-constrained by 2028. A failure or an open-ended qualification window means the incumbents keep the 2027 and 2028 order book intact. Because the HBM effort sits outside the IPO proceeds, the signal will come from Chinese accelerator launch specifications, not from filings.
Second, YMTC’s mix shift into enterprise SSDs will reprice NAND before it reprices DRAM. YMTC already has the bits. What it lacks is qualification into data-centre eSSD programmes, where 48% of NAND bits now go. If it converts part of its consumer volume into enterprise product during 2027, its revenue rank moves from fifth toward third and Kioxia and Micron lose leverage in the segment carrying the margin.
Third, the MATCH Act’s 150-day alignment clock, if enacted, forces a visible Dutch and Japanese decision during 2027. Neither government wants to hand its toolmakers jurisdictional exposure to Washington, nor a repeat of Nexperia. Partial alignment with carve-outs is the likely outcome, which means CXMT’s equipment access degrades gradually rather than abruptly, and the yield curve flattens rather than breaks. That is the scenario the market is least prepared for: neither Chinese failure nor Chinese parity.
The practical implication is that DRAM and NAND exposure should be underwritten separately, not as one China risk. The divergence between DRAM-heavy and HBM-heavy earnings mixes will widen as HBM4 volumes ramp.
Chey Tae-won, Chairman of SK Group, has argued that AI memory supply constraints could persist into 2030 because of wafer shortages and the physical limits on scaling production. If he is right, the window in which a 25% yield can be improved without commercial penalty is longer than most models assume.
Frequently asked questions
Is the 25% CXMT HBM3 yield figure verified?
It is reported, not disclosed. A South Korean industry account relayed on 9 September 2026 put 8-high HBM3 front-end yield near 30%, with roughly 70% of stacks passing back-end test. SemiAnalysis independently modelled about 35% and 70% in June 2026. Both land near 25% overall. CXMT has published no yield data itself.
Which Chinese company overtook Micron and Kioxia in NAND?
Yangtze Memory Technologies Co, or YMTC. Counterpoint Research data reported on 13 August 2026 placed YMTC third in global NAND bit shipments for the second quarter of 2026 with a 14% share, behind Samsung at 25% and SK hynix at 22%. The metric is bit shipments, not units and not revenue.
Does YMTC now out-earn Micron and Kioxia?
No. YMTC ranked third on bit shipments but fifth on revenue in the same quarter, behind Kioxia and Micron. Its mix remains weighted toward consumer applications rather than enterprise SSDs, which took 48% of global NAND bits in Q2 2026 and command materially higher prices per bit.
Is a 25% yield unusually bad for early HBM production?
Not for a first-generation stacked product. The volume-manufacturing benchmark is around 80%, and SK hynix now exceeds 90% on HBM3, but it reached that over several years from a much lower base. The meaningful question is the rate of improvement, not the level.
Is CXMT on the US Entity List?
No. Despite repeated consideration since 2023, CXMT is not on the Commerce Department’s BIS Entity List. It is named on the Department of Defense Section 1260H list, first in January 2025 and again in June 2026. Advanced HBM exports to China are separately restricted country-wide under December 2024 rules.
How large is the projected 2027 memory supply gap?
Morgan Stanley has modelled DRAM shortfalls of roughly 17% in 2026 and 15% in 2027. TrendForce expects HBM to consume 30% of DRAM wafer input by end-2027 while contributing only 13% of bit supply, mechanically tightening conventional DRAM. Counterpoint expects the squeeze to run through the second half of 2027.