CXMT H1 2026 — The Plant Predates the Windfall. The Next Build Has Already Started.

The only tension that matters

Three months ago this newsletter covered a company that had lost RMB 31.8B (~US$4.74B) across 2022–2024. Two months later it covered the largest listing on record: priced at RMB 8.66 (~US$1.29) on July 27, 2026, opened at RMB 49.50 (~US$7.38), up as much as 531% intraday to roughly RMB 3.3 trillion (~US$492B) on a float of 6.73%. On Friday August 28, for the first time, CXMT filed its own numbers with its own risk factors attached, and they are the largest in the complex: revenue of RMB 150.31B (~US$22.4B at ~6.71 CNY/USD, approximate) against RMB 15.44B (~US$2.30B), up 873.64%; net profit attributable of RMB 77.61B (~US$11.57B) against a loss of RMB 2.33B (~US$347.5M); operating cash flow of RMB 131.16B (~US$19.55B), up 2,985.64%.

Every one of those numbers is clean. Net profit excluding non-recurring items was RMB 78.79B (~US$11.74B) — higher than the headline, because non-recurring items netted to −RMB 1.19B (~US$177.0M), mostly RMB 1.34B (~US$199.2M) of fair-value losses. The one-offs were a drag, not a prop. Cash exceeded accounting profit by a wide margin. Basic EPS was RMB 1.2893 (~US$0.19), weighted ROE 81.06% from −5.75%, total assets RMB 468.08B (~US$69.76B, +38.98%), net assets attributable RMB 134.72B (~US$20.08B, +137.38%). Against pre-IPO guidance of RMB 110–120B (~US$16.4–17.9B) of revenue and RMB 50–57B (~US$7.45–8.49B) of profit, the print cleared the top of both bands — revenue by 25.3%, profit by 36.1%.

And the stock closed the reaction session at RMB 58.01 (~US$8.65), down 1.01%. That is the tension. Not whether the print was real — it was, in cash — but what it is evidence of, and this is exactly the seam where the house edge lives: the gap between the narrative and what the data supports. The narrative is that an AI memory champion has arrived at scale. The filing describes something narrower and more reversible.

Read the revenue and the cost side by side

One comparison carries the whole argument, and stringing it through a sentence is exactly the thing a reader cannot hold. The chart does it in a glance.

Now decode it. The gap between the first two bars is the 84.84% main-business gross margin, and the company explains it without euphemism: rapid growth in global compute demand and capacity reallocation by the major manufacturers left DRAM in shortage, prices rose sharply, and that drove the increase in gross profit on its main products. Incremental revenue arrived at close to no incremental cost, which is what happens when the same wafers clear at a much higher price. That is a price margin, not a process margin — and a margin explained by price is direction-agnostic in exactly the way a margin explained by mix is. It reverses when the spread reverses. What that margin does not tell you is anything about the plant underneath it — for that you have to leave the half and read the build.

The issuer marked its own top

Companies do not usually write the bear case into their first filing. This one did. The risk factors state that between 2015 and 2025 DRAM ranged from a high of US$7.89/GB to a low of US$1.78/GB in the first half of 2023, that prices sit at a high level now, and — verbatim in substance — that the sustained sharp increase in prices is not sustainable. Then it names the mechanism: international DRAM makers are raising capital expenditure on improved profitability, the planned new capacity is considerable, and if demand growth lags the pace of supply release, the risk of the industry's supply-demand structure reversing will keep rising.

Decode that and it is a scheduling statement, not a disclaimer. Management is telling you that the margin which produced RMB 77.61B has a shelf life set by other people's fabs. Which makes the question of what CXMT is doing with the windfall the only question that matters — and the answer is in the fixed-asset note, not the income statement.

The capex is a build cycle, not a decision

In the most profitable half in the company's history, cash paid to acquire fixed assets, intangibles and other long-term assets was RMB 21.59B (~US$3.22B), down from RMB 24.11B (~US$3.59B). Fixed assets at period end were RMB 181.56B (~US$27.06B), slightly below the RMB 183.02B (~US$27.28B) carried at the end of 2025, because depreciation of RMB 14.51B (~US$2.16B, +27.85%) outran what was transferred in. Construction in progress rose to RMB 36.78B (~US$5.48B) from RMB 31.79B (~US$4.74B).

Read on its own, the capex line looks like restraint inside a windfall. It is not, and the prospectus says so. Capital expenditure ran RMB 35.43B (~US$5.28B) in FY2022, RMB 43.66B (~US$6.51B) in FY2023 and RMB 71.23B (~US$10.62B) in FY2024 before stepping down to RMB 24.11B across the first half of 2025. The peak was two years ago. This half's figure is the tail of a decline that began well before the price move, on a schedule set when the lines were commissioned — so the 10.45% fall is a phase of a build cycle, not a decision taken inside the best half the company has ever had. Anyone reading the single year-on-year number as evidence of harvesting is reading a construction timetable as a capital-allocation choice.

Read the series and the conclusion changes shape. The plant that earned this half was substantially standing before the cycle turned — fixed assets went from RMB 68.02B (~US$10.14B) at the end of 2022 to RMB 84.45B (~US$12.59B), then jumped to RMB 153.13B (~US$22.82B) through FY2024 as the equipment bought at the peak was installed and transferred into service, reaching RMB 183.02B by the end of 2025. That is the whole build, and all of it predates the windfall. What the half adds is the turn. Construction in progress ran RMB 26.56B (~US$3.96B) at the end of 2022, peaked at RMB 51.50B (~US$7.68B) in 2023 as the lines went up, then drained to RMB 41.21B (~US$6.14B) and RMB 31.79B as that work was capitalised — and in this half it rose again, to RMB 36.78B. Alongside it, contracted long-term asset purchases not yet recognised in the financial statements stood at RMB 52.44B (~US$7.82B) at June 30, against RMB 17.13B (~US$2.55B) six months earlier, +206.16%, about 2.43x the half's capex. Proceeds are earmarked for a wafer manufacturing line technical upgrade, a DRAM technology upgrade, and a forward-looking DRAM research project.

So the sequence is legible, and it is the ordinary sequence of a fab: build the shell, buy the tools, capitalise them, digest, then start again. Capex cash is the lagging record of that cycle; work-in-progress and the commitment note are the leading ones. Both leading indicators turned up in this half. Which means the interesting question is not whether CXMT is spending — it has ordered the next wave — but whether the wave converts at anything like the FY2024 cadence, into a market its own risk factors say is about to be supplied by everyone at once.

The orders are signed. The equipment is not installed, the process is not tuned, and the bits are not shipping. This is the same shape as the NVIDIA finding that a commitment is an allocation claim before it is a cost — supply commitments stepping from $119B to $279B "primarily related to the procurement of memory" — pointed the other way: on NVIDIA's page a commitment removed supply from everyone else; on CXMT's page a commitment is capacity that does not exist yet. Both are claims about a schedule. Neither is a fact about this half.

The word this report never says

Every headline attributed the print to artificial intelligence. Search the 192-page report for HBM and the count is zero. The product family that decides whether a DRAM maker sits in an accelerator bill of materials is simply absent. What the report does describe is a jump-generation roadmap — the first through fourth process platforms in mass production, a fifth in development — across DDR4 and LPDDR4X through DDR5, LPDDR5/5X and LPDDR6.

The risk factors then say it plainly: AI is one of the important drivers of this cycle, the industry is in a period of concentrated infrastructure investment, commercial application is still expanding, and the company's revenue share from that area during the reporting period was relatively low. The filing goes further and flags that compute-in-memory and on-chip-cache architectures, if industrialised at scale, could change the cloud GPU-plus-DRAM deployment model and weaken AI's pull on DRAM outright.

Decode the two together. CXMT was paid by the AI cycle and not for AI products. It was paid by the price of commodity DRAM that AI demand made scarce. That is a materially different exposure from the one the coverage describes, because a scarcity rent transfers to whoever adds supply next, while a product position does not.

Where the revenue actually went

The contract-disaggregation table settles the attribution, and it is the second place a chart beats a sentence.

Mobile-class memory outsold server-and-PC-class memory by RMB 8.72B (~US$1.30B). DDR was 46.30% of main business, up from 31.9% in 2025 — real mix improvement, and still the smaller half. Two distribution facts sit next to it and both cut the same way. By channel, RMB 133.48B (~US$19.89B), or 88.81% of revenue, moved through distributors rather than direct sales of RMB 16.83B (~US$2.51B); a company selling 88.81% of its output through intermediaries is renting the price, not owning the customer. And customers individually above 10% of revenue contributed RMB 80.15B (~US$11.94B), 53.32% of the total, against 59.17% a year earlier — concentration eased, but the named end customers are Alibaba Cloud, ByteDance, Tencent, Lenovo, Xiaomi, Transsion, Honor, OPPO and vivo, which is a list of Chinese handset and cloud buyers, not accelerator programmes.

The geography is the strangest disclosure in the document. Main-business revenue sourced from outside mainland China was RMB 95.79B (~US$14.28B) against RMB 54.25B (~US$8.08B) from inside it — 63.84% booked offshore, up from 36.88% a year earlier. Non-current assets run the other direction entirely: RMB 231.18B (~US$34.45B) in mainland China against RMB 97.0M (~US$14.5M) everywhere else, 0.04% of the non-current base, with total overseas assets of RMB 680.1M (~US$101.4M) or 0.15% of assets. Decode it: the revenue is going offshore fast while the fabs, the people and the political exposure stay put. On June 8, 2026 the US Department of Defense added subsidiary ChangXin Memory to its Chinese Military Companies list under Section 1260H of the FY2021 NDAA; the company states it will not materially affect daily operations or going-concern ability. That may well be right for operations. It is a strange fact to hold next to a revenue base that just became two-thirds foreign.

The strongest case against me

Here is the argument I have to beat, and it is good. I am reading a timing gap as a weakness when it is the textbook definition of operating leverage, and CXMT has just demonstrated the single most valuable thing a memory manufacturer can demonstrate — that it converts a price cycle into cash rather than into inventory. Operating cash flow of RMB 131.16B against net profit of RMB 77.61B is the opposite of the module layer, where in the same shortage Longsys's profit went into the warehouse and the warehouse is the position. Cash reached RMB 143.43B (~US$21.38B, 30.64% of assets, +175.87%) with a further RMB 38.83B (~US$5.79B) in trading financial assets — roughly RMB 182.26B (~US$27.16B) of liquidity before the RMB 66.61B (~US$9.93B) the July listing added. Receivables of RMB 6.65B (~US$991.5M) are 4.4% of revenue. Inventory write-down provisions were reversed to RMB 676.4M (~US$100.8M) from RMB 822.2M (~US$122.5M) as prices recovered above carrying value. R&D input rose 87.38% to RMB 6.86B (~US$1.02B), researchers went from 4,655 to 7,491 with 67.75% holding a master's degree or higher, and 943 invention patents were granted in the half against a cumulative 7,907 IP items. Signing RMB 52.44B of expansion into a shortage you expect to end is not harvesting a top; it is what a challenger with a cost gap and a fresh balance sheet is supposed to do, and buying equipment at the moment every competitor is also buying is the one thing a disciplined operator would deliberately sequence rather than rush.

Two things survive it, and they are both about the quality of the second-derivative disclosures rather than the headline. The R&D ratio fell from 23.71% of revenue to 4.56%, a 19.15-point decline, and RMB 1.32B (~US$197.0M) of the spend — 19.27% of the total — was capitalised for the first time rather than expensed, so what actually hit the P&L, RMB 5.54B (~US$825.3M), grew 51.27% and not 87.38%. Capitalising development cost in the one half where the P&L needed no help is a choice worth watching, not an accusation. And inventory net of provisions stood at RMB 35.13B (~US$5.24B) against RMB 22.93B (~US$3.42B) of first-half cost of sales — roughly 1.53x a half-year of COGS, carried at a basis struck before the spike. Neither breaks the steelman. Both say the same thing the capex line says: the reported half is better than the underlying half by an amount that is disclosed, small, and worth tracking.

How the memory complex reads it

The filing's own market data frames the read. Citing Omdia, it puts global DRAM at US$150.5B in 2025 growing to US$763.8B by 2030, a 38.39% CAGR, and gives 2025 share by sales as SK hynix 34.48%, Samsung 33.96% and Micron 23.41% — above 90% between them — with CXMT first in China and fourth globally by both shipments and sales.

Against this newsletter's standing prior that DRAM pricing plus hyperscaler capex set the cycle, the CXMT print is the supply-side print of the same trade the module layer took from the other end. Same shortage, opposite balance sheets: Longsys bought the position and waited for the cash; CXMT already has the cash and has not yet bought the position. Which is why the two names resolve on opposite lines — for a module house the settling number is cash conversion, and for a fab it is capital expenditure. The read-through for the incumbents is narrower than the headline implies: a fourth vendor whose AI-related revenue share is its own words "relatively low" and whose report never says HBM is not, on this evidence, competing for accelerator sockets. It is competing for the commodity bits that set everyone's price.

Valuation and what decides it

At Monday's close of RMB 58.01 (−1.01%, day range RMB 57.01–59.90 against a 52-week range of RMB 38.11–61.80), market capitalisation was approximately RMB 3.88 trillion (~US$578B) unverified — roughly RMB 3.94T (~US$587B) on the 67.88 billion shares outstanding, and on the order of 25x first-half profit annualised to RMB 155.21B (~US$23.13B). Published consensus carried a 12-month target near RMB 78.47, a high of RMB 129 and a low of RMB 16.1 unverified. An eight-fold spread between the most and least optimistic estimate is not analyst disagreement about a company; it is analyst disagreement about a price deck.

The reaction is the tell worth keeping. A company cleared the top of its own profit guidance by 36.1%, printed an 81.06% ROE, and the stock fell 1.01% the next session after running 6.70x from issue in a month. That is a record selling off on expectations, not flaws — the business is extraordinary and the entry already spent it.

The thesis settles on capital expenditure, and the checkpoint is dated: the FY2026 annual report, with a partial read in the Q3 disclosure before it. The comparison that matters is not the year-on-year half — it is the FY2024 peak of RMB 71.23B. I am wrong if full-year 2026 capex lands at or above roughly RMB 50B, construction in progress keeps climbing past the RMB 36.78B mark as the RMB 52.44B of commitments draws down, and main-business gross margin still holds above roughly 70% — that combination says the trough is over, the next build is funded out of the windfall, and the price margin is being converted into a cost-per-bit position. I am right if full-year capex stays near the RMB 43.19B the half annualises to, the commitment balance sits or is quietly revised, and gross margin compresses toward the cost base as new industry supply lands. Second checkpoint, and the better one: any disclosure that names HBM as a product rather than an absence.

Bottom line

CXMT's first filing as a public company is an extraordinary document, and nothing in this piece disputes it: RMB 150.31B of revenue, RMB 77.61B of profit, RMB 131.16B of operating cash flow, an 84.84% gross margin and an 81.06% ROE from a company that was losing money two years ago, with the ex-items number above the headline and the cash above the profit. There is no accounting story here to catch.

There is a composition story, and it only shows up once you stop reading the half in isolation. Read as proof that CXMT has become a scale DRAM manufacturer this year, the half overstates its case — mobile memory outsold server memory, HBM appears zero times, 88.81% of revenue moved through distributors, and the issuer itself wrote down that the price move is not sustainable. Read as proof that a fourth global vendor caught a shortage with a plant it had already paid for, the half is exactly what it appears to be: the build ran from 2022 to 2024, the peak spend was RMB 71.23B in FY2024, and the windfall arrived after the bill. The trap is the middle reading — treating a 10.45% dip in a post-peak half as evidence of restraint. It is not a decision; it is a construction timetable. What the half genuinely adds is that the trough looks finished: work-in-progress and commitments both turned up. The next report resolves it on one line, and that line is the cash the company actually pays for equipment — measured against FY2024, not against last June.