CXMT's DRAM ramp and HBM timeline — is China's entry the exogenous shock that breaks oligopoly pricing discipline?
The question
What is CXMT's current DRAM production capacity, yield rate, and timeline to HBM production — and how likely is their market entry to break the three-player DRAM oligopoly's pricing discipline, threatening the memory capital-cycle thesis? Context: memory-cycle-v1(.1) is a long-only bet premised on Samsung/SK Hynix/Micron holding capacity discipline; the 2026-06-23 Stratechery piece flagged CXMT as a Phase-2 disruptor, and the vault has no dedicated coverage of Chinese-entrant progress — the single biggest un-researched risk to the thesis.
What we already know (from the vault)
- The thesis is HBM-anchored: HBM shortage is the binding constraint, and every HBM wafer crowds out ~3 commodity DRAM wafers, cascading tightness into DDR5/DDR4/NAND. Oligopoly pricing power rests on the big three deliberately restraining greenfield capacity while shifting to premium HBM [[2026-05-18-memory-cycle-v1-1]].
- The executable thesis already names a China disqualifier: "CXMT or YMTC achieves volume HBM3E equivalent (low-prob 2026, high-prob 2027)" — and its own bear case flags YMTC/CXMT as "18-24 months from being a meaningful DDR5 supplier," with non-HBM DRAM (~75% of maker revenue) "one capacity announcement away from a glut" [[2026-05-18-memory-cycle-v1-1]].
- The Phase 2→3 indicator brief concluded every primary signal still reads deep Phase 2 (prices accelerating, >40wk lead times, inventories at 2018 lows), and explicitly named Chinese entrants as "the most likely agent to force Phase 3 from outside the big three" — a supply source the oligopoly's discipline can't control [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]].
- The 2026-06-23 Stratechery reference already logged CXMT's Q1 2026 revenue +700% YoY and a $4B Shanghai IPO, plus HP in talks to use CXMT chips for Asia-market products (DRAM fungibility lets buyers mix CXMT-for-Asia / big-three-for-West) [[2026-06-23-stratechery-memory-chips-china-microsoft-deepseek]].
- TrendForce backfill notes flagged CXMT as under-covered relative to potential impact (only 5 mentions vs Samsung 120 / SK Hynix 87 / Micron 85) — a known blind spot in the vault's pricing anchor [[2026-05-17-trendforce-backfill-notes]].
What the web says
- DRAM wafer capacity is scaling fast. CXMT ran ~265k wafer starts/month (kwspm) end-2025, projected to ~300-350k by end-2026, with additions of ~85k (2026) / 70k (2027) / 80k (2028) — enough to become the industry's #3 supplier by wafer capacity (not by bit or by value) (SemiAnalysis; DigiTimes).
- But bit-share is still small and the oligopoly is intact. Big three held >90% of DRAM shipments in Q4 2025 (SK Hynix 33.2%, Samsung 32.6%, Micron 25.7%). CXMT bit-share ~9% (2025) rising to ~11% (2026) / 12% (2027) (SemiAnalysis).
- CXMT competes in commodity DDR, at a cost disadvantage. Mass-producing DDR5/LPDDR5X on a 16nm-class node; but 1Q26 DDR5 cost-per-bit runs >30% higher than the big three. In today's undersupplied market it prices only ~5-10% below competitors, and posted a 70% operating margin (vs SK Hynix 73% / Samsung 81% / Micron 84%). Analysts expect DDR5 yield parity ~late 2026, after which CXMT would use aggressive pricing to capture the low-to-mid end (PCs, phones, appliances, Asia channel) (SemiAnalysis).
- HBM is years away and slipping. CXMT is roughly two generations behind (industry at HBM3E/HBM4; CXMT at ~HBM2, targeting HBM3). HBM3 mass production has been postponed beyond 2026; HBM3 8-Hi yield ~25% overall (35% front-end × 70% back-end) and unlikely to break 40% until H2 2026; 12-layer HBM targeted only 2027 (DigiTimes — HBM3 slip, paywalled headline only; Tom's Hardware).
- The HBM barrier is structural, not just execution. No EUV access caps DRAM at ~1γ; the HBM4/4E base-die logic needs leading-edge nodes CXMT can't get (SMIC stuck ~7nm, no TSMC orders on geopolitics) — using a memory-process base die costs ~10% bandwidth. MR-MUF packaging materials and NAMICS-type supplier relationships are hard to replicate. HBM is a categorically higher wall than DDR5 (ChinaTalk).
- Discipline is holding right now. TrendForce reports the recent China DDR5 "price drop" was secondary-market modules only, impact confined to a niche channel; major-supplier contract prices stayed completely stable, and DRAM remains undersupplied by high-single-digits (2026) to low-to-mid-teens (2027) (TrendForce, Apr 2026).
Convergences and contradictions
- Convergence: Vault and web agree CXMT is a commodity-DDR story, not an HBM story, and that HBM — the thesis's load-bearing lever — is structurally protected for the current cycle. The thesis's own "high-prob 2027" HBM-volume disqualifier looks, if anything, too aggressive: CXMT's HBM3 slipped past 2026 and volume HBM3E parity is not a 2027 event.
- Tension: capacity (#3 by wafers) vastly overstates competitive weight (~9-11% bits, >30% cost gap, near-zero HBM). The single number that would matter — CXMT bit-share crossing into oversupply territory — isn't here yet; TrendForce still models undersupply through 2028.
- The real contradiction to watch: the thesis leans on commodity tightness (the crowd-out cascade), and commodity DDR is exactly where CXMT scales fastest and cheapest. A fast DDR5 yield ramp + ~155k kwspm of 2026-27 additions is the one channel that could relieve the ~75%-of-revenue commodity book ahead of schedule.
Synthesis for RDCO
Net threat to oligopoly pricing discipline over the next 4-8 quarters: MEDIUM — but decomposed, it's LOW on the thesis-critical HBM lever and MEDIUM (rising toward HIGH in the 2028+ window) on the commodity-DDR flank. The thesis is intact for the current Phase 2 window; CXMT reads as a Phase-3 accelerant, not a Phase-2 breaker.
On HBM — where the oligopoly actually extracts its 80%+ margins — CXMT is not a factor in the thesis window. Two generations behind, ~25% HBM3 yield, mass production slipped past 2026, and a hard base-die ceiling (no EUV, no TSMC, SMIC at 7nm) that no amount of state subsidy fixes on a 4-8 quarter clock. This is thesis-confirming: the vault's own disqualifier ("CXMT volume HBM3E, high-prob 2027") is outrun by reality. HBM premium pricing, and the crowd-out that starves commodity DRAM, stays structurally protected through at least 2027.
On commodity DDR4/DDR5, the picture is genuinely more uncertain and is the correct place to worry. CXMT is adding ~85k wspm/year, targeting DDR5 yield parity by late 2026, and openly intends to use price to take the low-to-mid end. What caps the near-term threat: (1) a >30% cost-per-bit disadvantage means CXMT can't sustainably undercut in a tight market without bleeding — today it prices near parity because the market is short; (2) TrendForce confirms contract prices are stable and the only observed weakness was a niche secondary-module channel; (3) the market is modeled undersupplied through 2028. The threat activates on a state-change, not a level: if/when the market flips from undersupply to balance, CXMT's low-margin volume becomes the marginal price-setter that deepens and accelerates the Phase 3 downturn — precisely the "China is the wildcard that forces Phase 3 from outside the big three" scenario the indicator brief named. It sets a lower ceiling on low-end/Asia-channel commodity ASP and shortens the runway on the far side of the peak; it does not break discipline while supply is still short.
Actionable read for the position: keep memory-cycle-v1.1 posture unchanged — this does not trip a disqualifier and the Phase 2 anchors still hold. But CXMT graduates from "narrative wildcard" to a tracked exogenous variable: instrument CXMT bit-share (the SemiAnalysis 9%→11%→12% path), DDR5 yield-parity confirmation (~late 2026), and any big-three contract-price crack that is not confined to the secondary channel. The moment two of those move together with the vault's inventory-days / ASP-rollover indicators, CXMT flips from accelerant to co-trigger and the Phase 3 timing pulls forward — most plausibly into the 2027-2028 window, not the next 4-8 quarters.
Open follow-ups
- Build a CXMT bit-share + wafer-capacity time series as a new memory-cycle anchor (SemiAnalysis / TrendForce quarterly) — the vault tracks big-three but not the exogenous entrant.
- Set a DDR5 yield-parity tripwire (~late 2026): confirm/deny CXMT reaching cost-per-bit within ~10% of the big three, the point at which aggressive pricing becomes sustainable rather than loss-making.
- Watch contract-price integrity, not spot: distinguish niche secondary-module softness (benign) from a genuine major-supplier contract crack (Phase-3 signal). TrendForce monthly bulletins already feed this.
- Track the YMTC/NAND parallel (2x capacity by end-2027) — same entrant dynamic on the NAND side of the cascade, not covered here.
- Re-key the thesis's HBM3E-volume disqualifier from "high-prob 2027" to reflect the confirmed HBM3 slip beyond 2026 — the current wording is more alarmist than the evidence supports.
- Monitor US policy loosening (Apple/HP lobbying to use CXMT/YMTC): demand-channel reopening could either extend the cycle (more buyers) or accelerate share migration to Chinese supply.
Related
- [[2026-05-18-memory-cycle-v1-1]] — executable thesis this brief stress-tests; contains the CXMT/YMTC disqualifier and commodity-glut bear case
- [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]] — Phase 2→3 indicator set; named Chinese entrants as the exogenous Phase-3 trigger
- [[2026-06-23-stratechery-memory-chips-china-microsoft-deepseek]] — Stratechery trigger doc; CXMT +700% YoY revenue, $4B IPO, HP talks
- [[2026-05-18-hbm3e-hbm4-capacity-timeline-phase-b-end]] — HBM capacity-online timeline the CXMT HBM gap should be read against
- [[2026-05-17-trendforce-backfill-notes]] — DRAM-spot pricing anchor; flagged CXMT as under-covered
Sources
Vault
- [[2026-05-18-memory-cycle-v1-1]] —
~/rdco-vault/01-projects/investing/theses/2026-05-18-memory-cycle-v1-1.md - [[2026-06-28-chip-memory-cycle-phase2-phase3-indicators]] —
~/rdco-vault/06-reference/research/2026-06-28-chip-memory-cycle-phase2-phase3-indicators.md - [[2026-06-23-stratechery-memory-chips-china-microsoft-deepseek]] —
~/rdco-vault/06-reference/2026-06-23-stratechery-memory-chips-china-microsoft-deepseek.md - [[2026-05-18-hbm3e-hbm4-capacity-timeline-phase-b-end]] —
~/rdco-vault/06-reference/research/2026-05-18-hbm3e-hbm4-capacity-timeline-phase-b-end.md - [[2026-05-17-trendforce-backfill-notes]] —
~/rdco-vault/01-projects/investing/anchors/dram-spot/2026-05-17-trendforce-backfill-notes.md
Web
- China's CXMT Is Set to Challenge DRAM Incumbents — SemiAnalysis
- China's CXMT muscles into DRAM's top tier — DigiTimes
- CXMT HBM3 timeline slips, mass production unlikely in 2026 — DigiTimes (paywalled, headline only)
- Chinese semiconductor industry gears up for domestic HBM3 by end 2026 — Tom's Hardware
- Will China Hit the HBM Wall? — ChinaTalk
- China's DDR5 Price Drop Driven by Secondary-Market Modules; Impact Confined to Niche Channel — TrendForce