October 25, 2026 — Two numbers are moving through Taiwan's supply chain this week. First: domestic DRAM, led by CXMT (ChangXin Memory), now holds roughly 10% of the market in the current price cycle. Second: at the current expansion pace, CXMT capacity may pass Micron in 2027 and take third place worldwide. The detail that should actually catch a buyer's eye is the pricing posture — CXMT is not discounting. Some server-grade modules quote above the incumbents. What it is competing on is the contract.
How the share was won
The gain is structural, and it was handed over by the incumbents. Samsung, SK hynix and Micron have shifted capacity hard toward HBM for AI servers, tightening standard DDR5 supply and sending prices vertical — we covered 32 GB kits going from RMB 800 to RMB 3,500 in September. Module makers and system builders who could no longer find — or afford — supply turned to CXMT. This was not a slogan about domestic substitution; it was an empty shelf. In a shortage the second source becomes the first call — and CXMT knows the window cuts both ways: a capacity ramp needs customers, and customers need certainty.
What 10% actually buys
Share at that scale changes the negotiating table in three ways. Pricing stops being a take-it-or-leave-it number, because a credible second source caps how far the incumbents can push during a shortage. Lead times get a second queue to stand in, which in 2026 has been worth more than any discount. And supply risk spreads across two vendors instead of one — the argument system integrators have been making to their own customers for years, now finally available on memory.
None of this requires CXMT to win on cost. It requires capacity that shows up when contracted, a compatibility list that keeps growing, and qualification work done on real boards. The 10% says the first two are happening; the third is on the buyers.
Pricing parity, terms that differ
Taiwan-based suppliers report CXMT quotes within a hair of the big three. The business terms are from a different era:
| Term | The big three | CXMT (supply-chain accounts) |
|---|---|---|
| Long-term agreements | Three to five years minimum, volume and price locked | No aggressive LTA required |
| Prepayment | High upfront ratios | More flexible |
| Penalty clauses | Strict | Lighter |
| Pricing level | The benchmark | Close; some server modules above |
| Capacity rank | Top three | May reach third in 2027 |
For buyers, the expensive thing in 2026 is not the die — it is certainty. A big-three LTA welds your price and volume for three years; bet the cycle wrong and you lose twice. Flexible terms hand that choice back to the buyer.
Why it refuses the price war
Everyone in Chinese manufacturing can recite the solar and battery script: capacity ramps, prices crash, everyone bleeds. CXMT is explicitly not taking that role. Two reasons. The market is badly undersupplied, so discounting buys nothing. And its production cost per line still sits above the incumbents' mature fabs, so it has no ammunition for a price fight anyway. The supply-chain read on the cycle: prices peak in 2027, balance or slight surplus in 2028, and a real price war only after 2029.
Three takeaways
- CXMT is winning on terms, not price. Update the stereotype of cheap domestic memory: it sells the opposite of a certainty premium — elasticity. Counter-intuitive, but that is what the terms say.
- 10% is a byproduct of the up-cycle; 2028 is the real exam. Once supply balances, contract flexibility stops being scarce and the cost curve takes over. That is the test still ahead.
- For system builders, the window is now. Dual sourcing — an incumbent for volume, a domestic line being qualified in parallel — has never been cheaper to set up than during a shortage.
A qualification checklist for Pearl River Delta system integrators
- Run compatibility first. Check the memory QVL for every target board and server; stress 2-DIMM-per-channel full population, ECC behavior and long memtest runs.
- Phase the rollout. Start with office and test fleets; only move to production after a full quarter in the field.
- Buy the supply chain, not the sticker. Skip the long lock-in; order in batches and keep the optionality — that is exactly the room CXMT's model creates.
- Watch two signals. The big three's HBM capacity share (which sets standard DRAM supply) and whether CXMT's 2027 capacity actually lands.
- Hold a buffer. Until the cycle peaks, four weeks of safety stock on critical parts beats another round of price negotiation.