CXMT: Three Years of Losses, One Half-Year Guided to Erase Them — and China's Only Seat at the Memory Table

Panoramic research report · longitudinal history + cross-sectional rivalry + synthesis Subject: ChangXin Technology Group Co., Ltd. — CXMT Corporation (Shanghai STAR Market: 688825; book-building July 16, 2026) Report date: 2026-07-13 · Data cutoff: prospectus registration draft (updated 2026-05-17), CSRC registration (2026-06-12) and the offering / inquiry announcements through 2026-07-13 Sources: SSE prospectus and CICC sponsor materials, CSRC filings, Omdia, Securities Times (STCN), Sina Finance, National Business Daily, Caixin, EastMoney, TrendForce/industry press; comparables from Yahoo Finance / Gurufocus / Seeking Alpha For information and research purposes only. Not investment advice. Several figures below are guidance ranges or single-source; each is flagged in the text and must be verified against the final official disclosure.

Before we start

Three days from now, on July 16, retail and institutional investors in China will place orders for the largest A-share IPO of 2026: about 6.69 billion new shares of a company that, eighteen months ago, the United States government placed on its Entity List. The company is ChangXin Technology — the holding entity above ChangXin Memory Technologies, the operating company the market simply calls "长鑫存储." It intends to raise roughly ¥29.5 billion (about $4.1 billion), the second-largest listing in the history of the STAR board, behind only SMIC.

Here is the fact that organizes everything else. In the three years 2022–2024, CXMT lost roughly ¥31.8 billion. In the single half-year ending June 2026, management has guided to a net profit of ¥50–57 billion — a number that, at the top of the range, would erase the entire accumulated loss of the prior three years and then some, earned in six months. If that guidance holds, this is one of the most violent swings from loss to profit any semiconductor company has ever printed. It is also, for exactly that reason, the number a careful reader should trust least until the audited half-year lands. We treat it as a guidance range throughout, and we build the case on the numbers that are already booked.

Those booked numbers are startling enough. Revenue went from ¥9.09 billion in 2023 to ¥61.80 billion in 2025 — a 155.6% jump in the last year alone. Net income attributable to shareholders went −16.34 → −7.15 → +1.88 billion, the first annual profit in the company's history arriving in 2025. Then the first quarter of 2026 booked ¥50.8 billion of revenue and ¥24.76 billion of net profit — a 48.7% net margin, in a business that ran a 3.0% margin for the full prior year. Something changed, fast, and the report is about what.

This is the fifth report in a series that has circled the memory industry from the outside — NVIDIA (the demand), Micron and SK hynix (the American and Korean incumbents), Sandisk (the NAND canary). Every one of those reports had a recurring background character it never quite put on stage: the China clock. CXMT is the China clock. This report brings it forward. It is organized in three parts. Part One is longitudinal: how a 2016 startup with no product became a sanctioned national champion with no controlling shareholder. Part Two is cross-sectional: the memory board as it stands — an oligopoly holding more than 90% of the market, and precisely where a 7.67% newcomer does and does not threaten it. Part Three is synthesis: the financial arc, what the IPO itself tells you as an object, scenarios, the valuation question this listing forces, and the falsifiable claims on which the whole thing rests.

One housekeeping note. All figures are in Chinese renminbi unless marked; we give dollar translations at roughly ¥7.2 per dollar where useful. Where a number is a guidance range, a single-source figure, or an industry estimate rather than an audited disclosure, the text says so — because on a company this new, pricing this week, the distance between a reported number and a hoped-for one is the whole analysis.

Part One · Longitudinal: from an empty fab to a sanctioned champion (2016–2026)

The founding bet: buy the market you cannot build

ChangXin Memory was established in 2016 in Hefei, the capital of Anhui province and, by design, the most aggressive semiconductor-subsidy city in China. The technical founder was Zhu Yiming (朱一明), who had already built GigaDevice — a NOR-flash and microcontroller company listed in Shanghai — and who understood the one thing that makes DRAM different from almost every other chip: you cannot enter it incrementally. DRAM is a scale-or-die commodity with three incumbents, decades of patent thickets, and a cost curve so steep that a sub-scale entrant loses money on every wafer it ships. China had tried before and failed. The ChangXin bet was that with enough patient state capital, a company could absorb years of deliberate losses to buy its way onto the cost curve — and that the Chinese domestic market, the largest DRAM-consuming market on earth, was a captive customer worth any price to serve.

For its first several years CXMT did exactly that: it lost money on purpose. The prospectus records net losses attributable to shareholders of ¥8.33 billion in 2022, ¥16.34 billion in 2023, and ¥7.15 billion in 2024 — deeper red as the ramp accelerated, then narrowing as yields improved. Cumulative losses across 2022–2024 ran to roughly ¥31.8 billion. This was not distress. It was the plan. Every one of those yuan bought wafer starts, process learning, and — most importantly — the yield curve that is the only real moat in memory.

The ownership inversion: from a founder's company to no one's

The most unusual fact in CXMT's prospectus is not financial. It is that the company declares itself to have no controlling shareholder and no actual controller (无控股股东、无实际控制人). No single holder owns more than 50%, and no shareholder can, by itself, decide the appointment of more than half the board. The history explains it: the company passed through three phases — a founding private team in control, then Hefei's state-owned assets commission taking the controlling position, and finally a diffuse structure in which no one party controls.

State capital from Hefei collectively holds about 45.16% ahead of the offering — the gravitational center, even if not a legal controller. The largest single holder is Hefei Qinghui Jidian (清辉集电) at 21.67%, itself a state-linked, no-controller vehicle; then ChangXin Integrated Circuit (长鑫集成电路) at 11.71%, the National IC Fund Phase II ("Big Fund II") at 8.73%, an employee platform (集鑫) at 8.37%, and Anhui Provincial Investment at 7.91%. Alibaba Cloud sits on the register at 3.85%, with other top institutions alongside. Zhu Yiming chairs the board and has pledged not to reduce his stake for ten years — but he is explicitly not the actual controller.

This matters two ways. For the bull, "no actual controller" is the tell that this is a state project wearing a corporate structure: the balance sheet has the patience of a government and the discipline of a listed company. For the bear, the prospectus itself lists it as a governance risk — decision-making efficiency, and the risk of a control contest after listing, are named in the filing. A company that answers to everyone answers to no one.

December 2024: the Entity List

In December 2024 the U.S. Department of Commerce added CMT / ChangXin Memory and a set of affiliates to its Entity List, folding the leading Chinese DRAM maker into the same export-control regime that already constrained SMIC and YMTC. This is public fact; the specific risk-factor language in the prospectus we did not obtain verbatim, and a reader underwriting this stock should read that section in the registration draft directly. The strategic meaning, though, is not ambiguous. It converted CXMT from a company that wanted to localize its equipment and materials supply chain into a company that must. Every advanced lithography, etch, and deposition tool it cannot buy from Applied Materials, Lam, Tokyo Electron or ASML is now a tool it must source domestically or do without — which is simultaneously the single largest execution risk in the story and the single largest reason the Chinese state will fund it without limit. The sanction is the threat and the moat at once.

The turn: 2025 into 2026

Then the cycle arrived. Beginning in the second half of 2025, DRAM prices rose — and kept rising, as the AI buildout pulled the entire memory complex into shortage and the incumbents steered capacity toward high-bandwidth memory, tightening the mainstream DDR/LPDDR market where CXMT actually sells. On the same wafers, at higher prices, a company built for scale flipped. 2025 closed with the first annual profit: ¥1.88 billion of net income on ¥61.80 billion of revenue. Then 2026 opened with a first quarter that looked like a different company entirely — ¥50.8 billion of revenue (up 719% year over year) and ¥24.76 billion of net profit. The half-year guidance that follows (¥50–57 billion of attributable profit on ¥110–120 billion of revenue) is the extrapolation of that quarter, and it is why this IPO is happening now, at this price, into this tape.

Part Two · Cross-sectional: the memory board and the 7.67% newcomer

The oligopoly, and where CXMT fits

DRAM is the most concentrated large market in technology. By 2025 revenue, per Omdia, SK hynix held 34.48%, Samsung 33.96%, and Micron 23.41% — three companies, more than 90% of the world's DRAM. Into that board walks CXMT, which Omdia placed at 7.67% for the fourth quarter of 2025, ranking it fourth in the world and first in China. That is the entire competitive fact of this company: it is the only DRAM maker on earth outside the three-way cartel with a share worth naming, and it got there in under a decade.

But share alone flatters the position. The three incumbents earn the bulk of their profit at the technology frontier — DDR5 at leading nodes, and above all HBM, the stacked high-bandwidth memory that sits beside every AI accelerator and carries the industry's fattest margins. CXMT's disclosed product line is mainstream: it has moved into volume production of DDR4 and LPDDR4X and up through DDR5 and LPDDR5/5X, running process platforms it describes as first- through fourth-generation (G1–G4). Industry commentary associates its fourth-generation node with roughly the 17nm class, and its HBM effort with early HBM2/HBM3 development rather than the HBM3E/HBM4 that SK hynix ships to NVIDIA — but the prospectus node figures and any HBM timeline we did not verify verbatim, and they should be read from the filing, not from us. The honest summary: CXMT is a genuine force in commodity DDR/LPDDR and a developing, not yet competitive, presence in the HBM tier where the incumbents make their richest money.

The import-substitution logic — and the China clock

The reason a 7.67% newcomer terrifies the incumbents is not its current share. It is arithmetic. China consumes a large fraction of the world's DRAM and, until CXMT, produced almost none of it. Every point of share CXMT takes inside China is a point the incumbents lose in their single largest market, and every wafer of Chinese domestic capacity that comes online is capacity that does not need to earn a Western cost of capital. This is the "China clock" that the Micron and SK hynix reports in this series kept glancing at: the risk that state-funded, price-insensitive supply floods the commodity tier of memory and structurally lowers its margin, exactly as it did in solar, in LED, and in LCD. CXMT is the clock's minute hand.

The bull case for China and the bear case for Micron are the same sentence. If DRAM behaves like every other industry China has entered at scale, the mainstream DDR tier deflates and commoditizes, the incumbents retreat upmarket into HBM, and CXMT inherits a growing slice of a lower-margin commodity — a good outcome for a company whose cost of capital is a provincial government, and a bad one for companies that answer to public shareholders. The bear case for CXMT is that memory has beaten this script before: it is more patent-fortified, more capital-intense, and more yield-dependent than solar or LEDs ever were, and the Entity List denies CXMT precisely the leading-edge tools that made the incumbents' cost curve possible. Both cases are live. The report does not pretend to resolve them; it insists you hold them at once.

The chokepoint: equipment

Everything upstream of the fab is the vulnerability. DRAM at competitive nodes needs EUV or extremely aggressive multi-patterning DUV lithography, advanced etch and deposition, and metrology — a tool chain dominated by five Western and Japanese firms, all now constrained from selling their most advanced equipment into a sanctioned Chinese memory maker. CXMT's answer, and China's, is domestic substitution: build the tools at home, accept a slower node cadence and a yield penalty, and let scale and time close the gap. Whether that works is the single most important unknowable in this report, and it is unknowable by design — the data that would answer it (real yields, real node, real domestic-tool share) is exactly the data neither the company nor the state will disclose. A reader should size the position to that uncertainty, not around it.

Part Three · Synthesis: the arc, the offering, and the price

The financial arc, in one table

The whole thesis is a margin story, and the margins are computable from disclosed figures. Here is the arc; the 2026 rows are, respectively, a reported quarter and a guidance range.

*2024 revenue of ¥24.18bn appears in a single secondary source and should be confirmed against the prospectus; 2023 and 2025 are multi-source. The 2026 H1 line is management guidance, not an audited result.

Read down the net-margin column and you have the entire bull case and the entire bear case in five numbers. The bull: a company that ran a 3.0% margin for all of 2025 printed 48.7% in a single quarter, because memory is a cyclical where price runs straight to the bottom line once the fab is paid for — and CXMT's fabs are largely paid for by the state. The bear: a 48.7% net margin in DRAM is a cycle-peak number, not a through-cycle number; the incumbents have earned margins like this before and given every point of it back. The difference between a 46% first half and a 6% one is not execution — it is the DRAM spot price, which no one at CXMT controls.

The offering as an object

Strip the narrative and look at what is actually being sold. CXMT is issuing about 6.69 billion new shares (up to roughly 7.69 billion if the greenshoe is fully exercised), representing about 10% of post-offering share capital — implying roughly 66.9 billion shares outstanding after listing. Half the initial offering is reserved for strategic placement. The raise is targeted near ¥29.5 billion, allocated across three named projects: ¥13.0 billion for DRAM technology upgrades, ¥7.5 billion for production-line upgrades and volume-manufacturing, and ¥9.0 billion for forward-looking DRAM R&D. Note what that allocation says: this is not a capacity-first raise but a technology-and-node raise — money pointed at closing the process gap the Entity List widened.

The timeline is nearly complete. The SSE accepted the filing on 2025-12-30; the STAR listing committee approved it on 2026-05-27; the CSRC assented to registration on 2026-06-12 (valid twelve months); the preliminary inquiry runs today, 2026-07-13; subscription is 2026-07-16; payment settles by 2026-07-20; and the listing date itself has not been officially announced — several outlets that wrote "lists July 16" conflated the subscription date with the debut, which by STAR convention falls a week or so after payment. The offer price and therefore the offering market capitalization are not yet set — they are being discovered in this week's inquiry. A back-of-envelope floor from the ¥29.5bn raise over ~6.69bn shares implies something around ¥4.4 per share, but that is our arithmetic, not a quoted price, and any "first-day +500%" figure circulating this week is pure speculation with no disclosed anchor.

The valuation question this listing forces

Here is the paradox the offering creates. In June 2026, Micron and SK hynix each crossed a $1 trillion market capitalization — and both trade at strikingly low forward multiples, Micron around 9–10× and SK hynix around 6–8× on secondary estimates, precisely because the market treats today's super-cycle earnings as partly transient (the trailing/forward P/E scissors that the Micron and SK hynix reports in this series dwelt on). Those are the comps. CXMT will price against a business earning a cycle-peak margin, into a market that is already discounting cycle-peak earnings elsewhere in the group at single-digit multiples.

That sets up the central tension. If CXMT prices at a multiple of its guided 2026 earnings anywhere near the incumbents' single-digit forward multiples, it will look cheap on the peak — and expensive the instant the cycle turns, because unlike Micron and SK hynix, CXMT has almost no HBM franchise to cushion a mainstream-DDR downdraft. If it prices at the growth multiple its revenue trajectory invites, it will be capitalizing a peak margin as if it were permanent. The domestic A-share market, with its appetite for scarce national-champion semiconductors, may well grant the second; the logic of the memory cycle argues for the first. Which one the inquiry produces this week is the single most important number this report cannot yet give you — because it does not exist until Friday.

Scenarios

Bull. The super-cycle runs into 2027; domestic-tool substitution proves good enough to hold the node cadence; CXMT compounds share inside China past 10%, uses the IPO proceeds to bootstrap a real DDR5/HBM position, and the state's patient capital lets it out-invest through the next downturn. The margin normalizes off a peak but from a structurally higher base, and China finally owns a memory champion. In this world the peak margin is a down payment.

Base. The cycle rolls over sometime in 2026–2027, as memory cycles always do. CXMT's margin falls hard — from the 40s back toward single digits — but the company stays profitable through the trough on state support and a paid-for asset base, keeps taking domestic share in commodity DDR/LPDDR, and remains a generation behind at the HBM frontier. A national champion, structurally, in the lower-margin half of the market.

Bear. The Entity List bites where it is designed to: the domestic tool chain cannot hold the node cadence, yields stall, and the leading edge pulls away faster than CXMT can follow. The super-cycle breaks, the 46% margin proves a one-off, and a company that capitalized a peak as if it were a trend re-rates violently. The stock becomes what the incumbents already are in the market's eyes — a peak-cycle asset — but without the HBM franchise or the balance-sheet independence to defend the trough.

The falsifiable claims

A report is only worth reading if it can be proven wrong. Here are the claims on which this one stands or falls:

• The H1-2026 guidance holds. If the audited first half comes in materially below the ¥50–57bn attributable-profit range, the entire "erased three years of losses in six months" framing collapses — and the pricing built on it with it.
• The margin is the cycle, not the company. If DRAM spot prices roll over and CXMT's net margin does not compress sharply within a few quarters, the "peak-margin" bear case is wrong and the business is more durable than the incumbents' own multiples imply.
• The Entity List binds the node. If, over the next two years, CXMT ships a competitive DDR5-class node and a credible HBM product despite the tool restrictions, the single largest bear risk is falsified — and so is much of the West's export-control theory of the case.
• The China clock ticks on the incumbents. If Micron's and SK hynix's mainstream-DDR margins visibly erode as CXMT scales, the import-substitution thesis is confirmed in the one place it can be measured — the incumbents' income statements, not CXMT's.

Every one of these is checkable against public disclosure over the next eight quarters. That is the point. This is a company priced this week on a half-year of guided numbers and a decade of state patience; the honest posture toward it is not conviction in either direction but a short list of things to watch, and the discipline to let the disclosures — not the debut-day tape — settle them.

Sources

• SSE STAR Market prospectus (registration draft, updated 2026-05-17, file 002170_20260517_MGLN.pdf) and CICC sponsor / underwriting materials; SSE project review page (auditId 2170); first- and second-round inquiry replies
• CSRC registration assent (2026-06-12); offering arrangement and preliminary-inquiry announcement (2026-07-09/13)
• Omdia: 2025 DRAM revenue shares (SK hynix / Samsung / Micron) and CXMT 2025 Q4 share (7.67%)
• Securities Times (STCN): ¥29.5bn raise and second-largest-STAR-IPO reporting; National Business Daily: CSRC registration and no-actual-controller governance
• Sina Finance: prospectus-update coverage, financials and H1-2026 guidance range; Caixin / EastMoney: timeline; Caijingu: ownership structure
• Comparables (Micron, SK hynix forward multiples; June-2026 $1T milestones): Yahoo Finance, Gurufocus, Seeking Alpha, 24/7 Wall St. — secondary/estimate sources, directional only

Report generated by the Aya Invest research pipeline. Every claim above traces to a public source; figures that are guidance ranges (H1-2026), single-source (2024 revenue), computed (net margins), or not yet set (offer price) are marked as such in the text. For information and research purposes only. Not investment advice.

FAQ

What is ChangXin Technology (CXMT) and what is the 688825 IPO?

ChangXin Technology Group (CXMT Corporation) is the holding company above ChangXin Memory Technologies, China's largest and most advanced integrated DRAM designer-manufacturer, founded in Hefei in 2016. It is listing on the Shanghai STAR Market under code 688825, book-building on July 16, 2026, and raising about ¥29.5 billion — the largest A-share IPO of 2026 and the second-largest in STAR history after SMIC. The offer price is set through this week's inquiry and was not public as of July 13.

How did CXMT swing from years of losses to a huge profit?

CXMT lost roughly ¥31.8 billion across 2022–2024 by design, buying its way onto the DRAM cost curve with state capital. When DRAM prices rose from the second half of 2025 — pulled up by the AI-driven memory shortage and the incumbents' shift toward HBM — the same fabs turned sharply profitable: a first-ever annual profit of ¥1.88bn in 2025, then ¥24.76bn of net profit in 2026 Q1 (a 48.7% net margin), with H1-2026 net income guided to ¥50–57bn. That H1 figure is a guidance range and should be confirmed against the audited disclosure.

Who controls CXMT, and why does the US Entity List matter?

CXMT reports no controlling shareholder and no actual controller; Hefei state-linked entities collectively hold about 45.16%, with the National IC Fund Phase II and others alongside, and chairman Zhu Yiming pledged a ten-year lock-up but is not the controller. In December 2024 the US added CXMT to its Entity List, restricting its access to the most advanced foreign chipmaking tools — simultaneously the company's largest execution risk and the reason Chinese state capital will fund its domestic-equipment substitution without limit.

Where does CXMT stand against Samsung, SK hynix and Micron?

Omdia placed CXMT at 7.67% of the DRAM market in 2025 Q4 — fourth globally and first in China — against SK hynix (34.48%), Samsung (33.96%) and Micron (23.41%) by 2025 revenue, the three of whom hold over 90% of the market. CXMT competes in mainstream DDR4/DDR5 and LPDDR; it is a generation behind in HBM, where the incumbents earn their richest margins. Its threat is the "China clock": state-funded domestic supply that could deflate the commodity DDR tier and erode the incumbents' margins in their largest market.