CXMT: RMB131B nine-month profit, but Q3 barely tops Q2

CXMT, China's largest DRAM maker, now expects RMB131 billion to RMB132.5 billion of nine-month profit, but its third quarter earned almost exactly what its second did on 15% to 20% more sales, which says the memory shortage is still paying for volume while the price lift behind it is fading.

What happened?

• The filing: On the evening of 16 October 2026, CXMT (listed in Shanghai as Changxin Technology, code 688825) published a voluntary nine-month forecast, notice 2026-014. It expects revenue of about RMB265.0 billion to RMB270.0 billion for January to September, against RMB32.08 billion a year earlier, and net profit attributable to shareholders of about RMB131.0 billion to RMB132.5 billion, against a loss of RMB5.28 billion. The figures are preliminary, unaudited and under Chinese accounting standards.
• The reason given: CXMT said AI computing demand has grown quickly since the second half of 2025, global DRAM is in short supply, and prices have risen sharply. It also credited more output per wafer and capacity utilisation that stayed high.
• The third quarter, by subtraction: The half-year report gave first-half revenue of RMB150.31 billion and profit of RMB77.61 billion. Taking those away from the nine-month range, our arithmetic puts third-quarter revenue at RMB114.7 billion to RMB119.7 billion and profit at RMB53.4 billion to RMB54.9 billion.
• Against the second quarter: Yicai, working from CXMT's first-quarter and half-year reports, puts second-quarter revenue at RMB99.51 billion and profit at RMB52.84 billion. On those numbers, third-quarter revenue rose 15.3% to 20.3% and profit rose only 1.0% to 3.9%.
• The same evening: CXMT said chairman Zhu Yiming proposed a buyback of RMB1 billion to RMB2 billion of its A shares for employee incentive plans or to cut registered capital. The proposal still needs board or shareholder approval.

Why does it matter?

The direct signal. CXMT ranks itself first in China and fourth in the world in DRAM by shipments and sales, behind Samsung, SK hynix and Micron, which held more than 90% of the market in 2025 on Omdia figures cited in its half-year report. It sells DDR5 and LPDDR chips and modules to Alibaba Cloud, ByteDance, Tencent, Lenovo and Xiaomi, among others. A nine-month profit of more than RMB130 billion from a company that lost money a year ago is a direct reading of how much the DRAM shortage is worth to whoever owns wafer capacity.

The second-order signal: volume is still rising, the margin on it is not. The third quarter is where the shape changed.

• Our arithmetic: net margin was about 53% in the second quarter (RMB52.84 billion on RMB99.51 billion) and about 46% in the third quarter at either end of the range.
• Our arithmetic: the extra RMB15.2 billion to RMB20.2 billion of third-quarter sales kept only RMB0.6 billion to RMB2.1 billion of extra profit, roughly 4 to 10 fen of profit per extra yuan, against an average above 50 fen in the second quarter.
• CXMT did not explain the gap in its notice. Possible causes include depreciation from new capacity, higher costs, product mix, or smaller price increases. Those are hypotheses until the full third-quarter report shows gross margin and costs.

A second filing points at price. On 15 October Longsys, one of China's largest memory module makers and a DRAM buyer, said in its own forecast that the global memory market stayed tight in the third quarter but that DRAM and NAND price increases had "narrowed markedly" from earlier periods because of the high price base, and that the boost from rising memory prices to its earnings had slowed significantly. Two Chinese filings in two days, one from a seller and one from a buyer, say the same thing: the shortage is intact, but each quarter it lifts prices by less.

Why AI servers care. CXMT does not ship HBM, the stacked memory next to AI GPUs. Its weight lands on the conventional DRAM that AI servers also need in large amounts: DDR5 server modules for the CPUs, plus LPDDR. That market has been short partly because Samsung, SK hynix and Micron moved wafers to HBM; we traced that wafer shift in our HBM wafer-share analysis, and the HBM vs DDR5 explainer covers why the two compete for wafers. A fourth supplier adding 15% to 20% more sales in a quarter is extra conventional supply arriving exactly where that shift left the gap, and inside China first.

What does it change?

Near term (to year-end 2026)

• The price step is flattening before the volume step. If the third quarter's slower price lift holds, memory makers' quarterly profit grows more slowly than their shipments. For AI server builders in China, that is the first sign that memory is getting less expensive to add, not that it is getting easy to buy.
• CXMT's server push lands in a tight market. Yicai notes CXMT said this month it will launch a DDR5 RDIMM server module on its new 4F² DRAM architecture by year-end, after starting volume LPDDR6 output in late August. We followed that server test in our CXMT server DRAM Pulse. Qualification at Chinese cloud buyers is the step to watch, not the launch date.
• The buyback is small against the profit. Our arithmetic: RMB1 billion to RMB2 billion is under 2% of the nine-month profit forecast. It does not change how much capacity CXMT can fund.

Longer term (2027 and after)

• More of China's AI server memory comes from home. CXMT's half-year report says it aims to keep expanding capacity and global share. Each wafer it adds is a wafer of conventional DRAM that Chinese buyers do not have to source from the Korean and US makers, whose wafers then stay on HBM.
• The cycle is changing its driver. Longsys described the industry as moving from a consumer price cycle to an AI-driven technology cycle. If that is right, the margin leaders become the makers with HBM and high-end server parts, and a commodity-heavy maker's margin depends more on its own cost per bit.
• What we cannot see yet: CXMT's mix of server versus phone and PC memory, its contract versus spot exposure, and any capacity numbers. The full report should answer the first; the others may stay undisclosed.

What should we watch next?

• SK hynix third-quarter results and call, 28 October 2026, 9:00 Seoul time (08:00 Singapore). SK hynix set the date in a filing on 14 October. Its conventional DRAM pricing and its HBM share of sales show whether the slower price lift is a Chinese channel effect or global.
• Samsung Electronics full third-quarter results and call, 29 October 2026, 10:00 Korea time (09:00 Singapore). Samsung set the date on 8 October. The first split of its memory profit for the quarter, and a second read on DRAM contract prices.
• CXMT's formal third-quarter report, due by 31 October 2026. The notice says final figures come in that report. The test is gross margin: if it fell from the second quarter, the shortage is paying less per chip; if it held, the gap came from costs below gross profit.
• What would weaken the catalyst: SK hynix or Samsung reporting conventional DRAM prices still rising at earlier rates, which would make CXMT's flat quarter a company-specific cost story; or CXMT's report showing a one-off charge behind the third-quarter profit.

FAQ

Did CXMT report its third quarter?
Not yet. It published a nine-month forecast. The third-quarter figures here come from subtracting its reported first half, which is our arithmetic. The formal third-quarter report is due by 31 October.

If profit was flat, is the DRAM shortage over?
No. CXMT and Longsys both said supply was still short or tight in the third quarter. What changed is the size of each price increase. Longsys said those increases narrowed markedly; CXMT's flat profit on higher sales is consistent with that, though CXMT has not given its own reason.

Does CXMT make HBM for AI GPUs?
Not in volume today, based on its own product list of DDR, LPDDR and modules. Its effect on AI servers runs through DDR5 server memory and through Chinese cloud buyers such as Alibaba Cloud, ByteDance and Tencent, which it names as customers.

Why does a Chinese memory maker matter outside China?
Chinese buyers that take more domestic DRAM need less from Samsung, SK hynix and Micron. Those three can keep more wafers on HBM, which is the memory step AI GPU supply depends on.

Takeaway: CXMT's RMB131 billion-plus nine months say DRAM is still short, but a third quarter that barely beat the second says each quarter of shortage now adds less profit per chip; SK hynix on 28 October and CXMT's own report by 31 October are the next dated tests.

Figures come from the dated sources listed under References below. Our own arithmetic is shown and labelled as ours. Where a news report and the company filing differ, we use the filing and say so.

Last reviewed: 17 Oct 2026 · Next review: after CXMT's third-quarter report (by 31 Oct 2026)
Spot an error? Tell us and we will correct it and note the change here.
Our editorial standards →

References

1. CXMT (Changxin Technology), "2026 nine-month results forecast, voluntary disclosure", notice 2026-014, 16 Oct 2026
2. CXMT (Changxin Technology), "2026 Half-Year Report", 28 Aug 2026
3. Yicai, "CXMT expects more than RMB130 billion nine-month profit; chairman proposes buyback", 16 Oct 2026
4. 21st Century Business Herald (via Eastmoney), "CXMT's Zhu Yiming proposes RMB1–2 billion buyback", 16 Oct 2026
5. Longsys, "2026 nine-month results forecast", notice 2026-117, 15 Oct 2026
6. SK hynix, Form 6-K, third-quarter 2026 earnings conference call notice, 14 Oct 2026
7. Samsung Electronics, "Notice of 3Q 2026 Earnings Release Schedule", 8 Oct 2026
8. Economic Daily News, "Micron opens second Taichung advanced packaging and test plant", 15 Oct 2026
9. GlobalFoundries, "GlobalFoundries selected to operate two new semiconductor fabs in New York under landmark U.S.-Japan initiative", 14 Oct 2026
10. Second Order Wire, "NiSource signs Microsoft to 1,100 MW of Indiana data center power", 17 Oct 2026