Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
The story the market told itself on 27 July was simple enough. China’s largest memory manufacturer came to market, raised a fortune, and immediately became a threat to the three companies that have run the DRAM industry for two decades. Micron fell. Samsung fell. SK Hynix fell. The narrative wrote itself before anyone had finished reading the prospectus.
That reading was wrong, and the way it was wrong is more interesting than the event itself. What listed in Shanghai that Monday was not primarily a competitor. It was the most leveraged available position on the price of conventional memory, brought to market on a fraction of its own share count, at the precise moment the market was starting to reprice the cycle it depends on. Investors treated a price event as a supply event, sold the wrong securities, and then spent the rest of the week selling for an entirely different reason without appearing to notice the substitution.
A listing that broke three records in one session
The mechanics deserve stating plainly because they do most of the analytical work. CXMT priced at 8.66 yuan per share and raised 57.92 billion yuan, roughly 8.6 billion dollars, in what Reuters recorded as the largest mainland Chinese semiconductor offering ever completed. Proceeds could reach 66.61 billion yuan if the over-allotment option is fully exercised. It was Asia’s biggest initial public offering of the year and mainland China’s second largest of any kind, behind only Agricultural Bank of China in 2010.
Then trading opened. The stock reached 55.03 yuan intraday before closing at 49 yuan, a gain of 466 percent, which lifted market capitalisation to about 3.3 trillion yuan, near 488 billion dollars. That displaced Industrial and Commercial Bank of China, valued at 2.6 trillion yuan, as the most valuable company listed on a mainland exchange. Turnover reached roughly 141 billion yuan, a single-session record no A-share had previously touched.
Here is the number that should have anchored every subsequent piece of analysis and did not. Only 6.73 percent of the enlarged share count was freely tradable. The remainder sat under lock-up. Against that sliver of supply, the retail tranche drew 9.4 million individual orders worth 7.07 trillion yuan, a subscription rate of 212 times.
A valuation set under those conditions is not a market verdict. It is a queue.
The revenue is a price, not a position
The financial disclosures are genuinely extraordinary, and they are also very widely misread. According to the prospectus figures released through the Shanghai Stock Exchange, CXMT reported 2025 revenue of 61.799 billion yuan, up 155.6 percent. First-quarter 2026 revenue then reached 50.80 billion yuan, a rise of 719.13 percent year on year. The company guided first-half revenue to between 110 and 120 billion yuan, meaning six months of 2026 should deliver close to twice the whole of 2025.
Read those numbers as evidence of market share capture and you arrive at the consensus conclusion. Read them correctly and you arrive somewhere else entirely.
The work published by SemiAnalysis, summarised in its deep dive on the company, is explicit on the point. Revenue moved from roughly 3.3 billion dollars in 2024 to around 8.6 billion in 2025 to approximately 7.3 billion in the first quarter of 2026 alone, with full-year 2026 potentially exceeding 50 billion dollars. The driver is average selling prices, not share gains. The product mix remains concentrated in commodity DDR and LPDDR.
That distinction changes everything. TrendForce recorded conventional DRAM contract prices rising by something in the region of 93 to 98 percent quarter on quarter in the first quarter of 2026, which lifted total memory industry revenue by 81 percent in a single quarter. CXMT did not take that revenue from Micron. It received it from the same cycle Micron is riding, with roughly the same sensitivity in the same direction, and with considerably less to fall back on if the cycle turns.
The company said as much itself. The prospectus cautions against extrapolating from recent performance, noting that DRAM prices swung from 7.89 dollars per gigabyte to 1.78 dollars per gigabyte inside the first half of 2023. Very few companies warn their own investors this directly. Fewer still are rewarded with a 466 percent debut for doing so.
Whose profit is it
There is a second disclosure that almost no coverage engaged with. Reported first-quarter net profit was 33.012 billion yuan on a total basis and 24.76 billion yuan attributable to shareholders, and the gap between those figures is not rounding. SemiAnalysis puts approximately 74 percent of net profit as attributable to minority interests. Hefei state-owned entities hold over 30 percent of the shares. Alibaba appears as a shareholder of nearly 4 percent. The company controls its fabs through a concert party arrangement and declares no de facto controlling shareholder.
So the listed vehicle captures materially less of those headline earnings than the headline implies, and the residual is claimed by a structure whose incentives are not obviously those of a public minority holder. Any valuation multiple built off reported group profit is measuring something the buyer does not own.
What a 6.7 percent float actually prices
The float question generalises well beyond this listing, which is why it belongs in the argument rather than in a footnote.
When under 7 percent of a company trades and demand exceeds it by two orders of magnitude, the clearing price is determined by the shape of the queue rather than by any assessment of the asset. The valuation that results is real in the sense that transactions occurred at it. It is not informative in the sense that it tells you what the business is worth. Treating it as a signal, as much of the commentary did when it observed that CXMT had reached a meaningful fraction of Micron’s market capitalisation, is a category error.
It is also a mechanically unstable number. Lock-up expiry introduces supply on a known schedule. Index inclusion introduces demand on a known schedule. Neither has anything to do with DRAM. The price will move substantially on both, and those moves will be narrated afterwards as verdicts on Chinese memory competitiveness, which they will not be.
The lithography report that moved the wrong stock
Hours after the debut, The Information reported that a state-backed Shanghai manufacturer had begun producing domestic immersion deep ultraviolet lithography machines, with first units bound for SMIC, Hua Hong and CXMT. The two stories fused in the market’s mind into a single thesis: money plus tools equals supply.
The volumes are the reality check. As Tom’s Hardware set out in its account of the report, output targets are roughly five machines in 2026 and about 20 in 2027. Most components are domestic but critical parts still arrive from Japan. The manufacturer was not named, though sources described teams assembled from several Chinese companies including the state-backed startup Shanghai Yuliangsheng, whose immersion tool SMIC has reportedly been testing since September 2025.
Set that against ASML. The Dutch company recognised revenue on 279 DUV systems in 2025, of which 47 percent were immersion, generating 12 billion euros in DUV system sales. Chief financial officer Roger Dassen told analysts in July that ASML expects to ship about 130 immersion systems in 2026, plans to raise immersion capacity by 30 percent in 2027, and is examining a further 30 percent for 2028. Bank of America, JPMorgan and BNP Paribas each estimated that 20 domestic tools would touch roughly 2.4 percent of ASML revenue.
ASML nonetheless fell close to 6 percent on the session, with intraday losses nearer 8 percent. CNBC’s account of the caveats carried the more useful observation, from SemiAnalysis, that scaling production of the machine itself is the most consistently underestimated obstacle, alongside tool performance, fleet reliability, the surrounding ecosystem, and the poor economics of competing against fully depreciated ASML equipment already installed and running.
None of which makes the report unimportant. It changes optionality. It does not change 2027 supply.
Three of the five named companies
The policy backdrop supplies an irony that has gone almost entirely unremarked.
The MATCH Act, filed as H.R. 8170 with a Senate companion, cleared the House Foreign Affairs Committee in April by 44 votes to nil. Its stated design is to designate as covered facilities every chipmaking site operated by CXMT, Hua Hong, Huawei, SMIC and YMTC, including subsidiaries and affiliates, and to apply restrictions equivalent to the Entity List across exports, servicing and technical support. It would give the Netherlands and Japan 150 days to align their own controls or face unilateral enforcement.
Three of the five entities named in that bill are the first customers for the domestic scanner.
The servicing provision is the part that matters. Restricting new sales slows a fab. Restricting maintenance on installed equipment eventually stops one. A domestic tool does not need to match ASML on throughput or yield to be worth building under that threat. It needs only to exist, which reframes what the export control regime has actually purchased. It bought time, and it simultaneously guaranteed a captive, price-insensitive, politically protected domestic buyer for whatever emerged. That is the demand condition every infant industry requires and almost never obtains.
The bill is not law. It cleared a committee. Anyone modelling it as settled is modelling a hope.
The list that was published for an hour
The instability of the policy input is best illustrated by an episode from February that has largely dropped out of the coverage.
The Pentagon published an updated Section 1260H list of Chinese military companies, then asked the Federal Register to withdraw it from public inspection roughly an hour later, without explanation. The withdrawn version appeared to remove CXMT and YMTC. As WilmerHale documented in its client alert, the version eventually published on 8 June added 65 entities in total and kept both memory manufacturers on the active list. That publication followed a summit between the American and Chinese presidents in Beijing by a matter of weeks.
Within six months, then, the designation status of the company that just became China’s most valuable listed enterprise was reversed, un-reversed, and separately targeted by draft legislation moving through Congress. Investors are being asked to underwrite a valuation whose central variable has moved three times in half a year and could move again on a diplomatic schedule nobody outside two governments can observe.
What the market actually sold
The most useful evidence against the consensus reading arrived within 48 hours.
If the selling had been about Chinese supply, it would have concentrated in the names Chinese supply threatens and then stabilised. Instead it broadened and accelerated for a completely unrelated reason. SK Hynix guided 2026 capital expenditure up by 50 percent to at least 31 billion dollars, and that guidance, delivered alongside a record 76 percent operating margin, triggered a considerably larger rout than the CXMT listing had. The Philadelphia Semiconductor Index fell 19 percent across July, its worst month since 2008, with every constituent below its 50-day moving average. Chip stocks shed more than a trillion dollars across the episode.
Micron lost over 40 percent in the month. SK Hynix fell more than 30 percent, and its New York listing, completed on 10 July as the largest ever American IPO by a foreign company, was already down 23 percent from its debut. Federal Reserve rate concerns and geopolitical risk added to the pressure. Alphabet raised its own capital expenditure guidance into the same window.
That is not a market pricing a new entrant. That is a market pricing whether the existing entrants are spending too much, which is the opposite anxiety. Investors sold Micron on Monday because of China and then kept selling all week because of Korea, and the fact that the two motives are close to contradictory attracted very little comment.
The divergence nobody priced
The final piece of evidence lands on the same side. TrendForce research published on 30 July projects that DRAM and NAND will diverge sharply in 2027. DRAM supply stays constrained, with the sufficiency ratio around negative 1 to negative 2 percent in 2026 and the gap widening thereafter, because new capacity will not ramp meaningfully until the second half of 2027 and substantial output is not expected until 2028. High bandwidth memory also consumes far more wafer input per bit, so rising wafer starts do not convert proportionally into supply. NAND, by contrast, is expected to loosen in the second half of 2027 as fresh capacity arrives against soft consumer demand.
That single forecast explains something the contagion narrative could not. NAND-exposed names fell hardest during the episode, and the standard explanation was indiscriminate selling. But NAND has a genuine 2027 supply story, and the names concerned had risen enormously into it. Nothing indiscriminate is required.
It also sharpens the position on CXMT. The company’s economics sit almost entirely in conventional DRAM. Its high bandwidth memory capacity is estimated at around 5,000 wafers per month against total capacity approaching 350,000 by the end of 2026, and mass production of competitive HBM3 remains a problem it has not solved. It is therefore the purest available expression of the conventional DRAM price, with no meaningful HBM cushion, listed at 488 billion dollars off a 6.73 percent float.
Which produces the inversion. If your view is that Chinese capacity will eventually break the memory cycle, the security that expresses that view most efficiently is not a short in Boise or Suwon. Those companies hold the high-margin AI memory franchises, the customer relationships and the balance sheets to survive a downturn. The instrument with the highest sensitivity to the conventional DRAM price and the least protection against its reversal is the one that just listed in Shanghai.
The market spent a week treating that instrument as the threat rather than as the trade. It is a familiar sort of error. The most leveraged position in a cycle usually arrives dressed as the thing that will end it.
