For years the pitch from Chinese memory makers was price. Buy from us, the argument went, and you pay less than the Koreans charge, and you are not on anyone's export list.
This summer that pitch changed direction. When Huawei asked for relief from rising DRAM prices, CXMT (长鑫科技) held firm, according to Reuters, and the standoff ended with engineers from a Huawei-linked equipment vendor being told to leave the memory maker's R&D cleanroom floor.
The quarterly report filed at the end of October shows what that kind of leverage looks like on a P&L.
Key takeaways
- CXMT reported third-quarter revenue of RMB 118.71 billion (≈ US$16.6 billion), up 613% year on year and 19.3% on the previous quarter.
- Gross margin reached 89.53%, up 1.94 points sequentially — a margin that would be startling in any manufacturing business, let alone memory.
- Attributable net profit was RMB 54.43 billion for the quarter and RMB 132.04 billion for the first nine months, against a loss a year earlier.
- TrendForce puts CXMT fourth in global DRAM revenue at 9.5% in the second quarter, behind Samsung, SK hynix and Micron.
- Contract prices are still climbing: TrendForce expects conventional DRAM up another 10–15% in the fourth quarter.
The quarter
Revenue for July to September was RMB 118.71 billion, roughly US$16.6 billion at 7.15 to the dollar. Net profit was RMB 88.98 billion; the figure attributable to shareholders, after minority interests and one-offs, was RMB 54.43 billion. Nine-month revenue reached RMB 269.02 billion, up 738.51%, and nine-month attributable profit RMB 132.04 billion, against a loss of RMB 5.28 billion in the same period of 2025.
Two lines explain the rest. Research spending for the nine months was RMB 11.07 billion, up 73% — large in absolute terms, but only 4.12% of revenue, down from about 20% a year earlier, because revenue grew faster than the budget. Operating cash flow for the nine months was RMB 254.75 billion, up 1,864%, and total assets nearly doubled to RMB 672.54 billion.
A memory maker earning an 89.5% gross margin is not a normal business. It is what a shortage looks like when demand so far exceeds supply that the seller sets price rather than accepts it.
Why the shortage landed there
The AI buildout did not mainly add demand for ordinary DRAM. It removed the supply. Samsung, SK hynix and Micron have poured advanced wafer capacity into HBM and high-capacity server modules, because that is where the margin is, and every wafer switched to HBM is a wafer not making LPDDR5X for a phone.
TrendForce's read of the fourth quarter is that the shortage persists: conventional DRAM contract prices up 10–15% sequentially, NAND up 15–20%, server DRAM still undersupplied, and constraints in component supply, packaging and test capacity limiting how fast anyone can rebalance. The same research house expects HBM's blended ASP to rise 121% in 2027.
That is the mechanism behind CXMT's margin. It is not that CXMT makes a better chip than Samsung. It is that the big three vacated the segment CXMT occupies, and Chinese buyers have nowhere else to go.
Who is paying
Reuters reported in July, citing people familiar with the matter, that CXMT had been raising prices on Huawei for months, that Huawei's pushback did not work, and that in some cases the Chinese maker charged more than Samsung did for comparable server modules.
The same reporting described a five-year ByteDance agreement worth more than US$7 billion in July, following a Tencent deal Reuters put at more than US$3 billion in June. Those are not spot purchases. They are companies locking in supply because they expect the shortage to outlast their patience.
The squeeze has a domestic cost too. Reuters reported that several Chinese electronics firms complained to the industry ministry this year that price rises were delaying product launches.
Global readers should internalise this. For a decade, Chinese system makers treated domestic memory as a cheaper substitute bought under duress. In 2026 it became the scarce input they compete for, and the price umbrella is now set at home rather than imported.
What the money is for
CXMT is spending. Reuters reported new fabs in Shanghai and Hefei, a third under discussion, and plans to more than double capacity to above 600,000 wafers a month. The company has said its sixth-generation process platform completed engineering-sample validation in the quarter, and it announced mass production of LPDDR6 in September: 12,800 Mbps peak, about 60% more peak bandwidth and 20% less power than LPDDR5X at 10,667 Mbps, according to the company's own comparison.
Note what is missing. Reuters reported that CXMT's HBM remains about two generations behind its rivals, and HBM is where the AI accelerator bill actually goes. CXMT's share of global DRAM revenue was 9.5% in the second quarter, per TrendForce — real, and still less than half of Micron's 23.3%.
Honest limitations
The financials come from CXMT's quarterly report filed with the Shanghai Stock Exchange and summarised by Chinese financial media. I have read the headline tables, not the full notes, and segment margins are not disclosed, so the 89.53% is a company-wide figure that mixes commodity DRAM, specialty products and whatever HBM-related work is already running.
The pricing anecdotes are Reuters reporting based on unnamed sources; the companies declined to comment. TrendForce's share and price figures are estimates built from supplier revenue, not audited data. I have not verified how much of the profit came from price, volume or cost reduction; the company attributes it to all three.
What readers can do now
- If you buy DRAM anywhere in the stack, assume 2027 allocation is a contract problem, not a spot problem. TrendForce expects PC DRAM supply to shrink in 2027 as capacity moves to servers.
- Watch CXMT's R&D line as a share of revenue, not its absolute size. At 4.12% it is thin for a company that needs to close an HBM gap, and that ratio is the clearest public signal of how hard it is trying.
- Treat domestic Chinese memory as a price setter from here. Procurement models that assume a 10–20% discount to Korean product will be wrong in the other direction.