Everyone can read SemiAnalysis on X. Almost nobody maintains a public, revisioned ledger of CXMT HBM monthly yield / stacks against the Huawei stockpile clock. That cell was blank in our Memory hub on purpose. Today we open it — with anchors we can defend, and months we still mark 待填.
This is not a claim that we have factory telemetry they lack. It is a claim that we own the tracking product: two clocks, conflicting public estimates kept in separate columns, and a standing invitation to overwrite a row when a better source lands.
Why this cell
From David's Stockpile: foreign HBM burns down; CXMT is the domestic ramp. Best-case framing we already published — ~2M stacks / year ≈ 250–300K Ascend 910C-class cards (÷8). From Semi's June CXMT thread: equipment bans, multi-gen DRAM lag, larger HBM gap, and still not a near-term cycle-killer for Big-3. Same company. Two clocks. Yield is the hinge.
Discipline. We do not invent a monthly yield %. If a row has no dated source, it stays empty. Survey rumors that conflict stay side by side — not averaged into false precision.
Anchor ledger (public sources only)
| As-of | Metric | Value | Source | Confidence |
|---|---|---|---|---|
| 2025-09 | CXMT HBM stacks (annual thesis) | ~2M / yr → ~250–300K 910C | Semi Ascend ramp → Aileena huawei-hbm | site ← Semi |
| 2025-09 | Foreign stockpile | ~13M stacks → ~1.6M 910C; burns by year-end framing | Semi Ascend ramp → Aileena huawei-hbm | site ← Semi |
| 2025 YE (est.) | CXMT HBM wafer starts | ~5,000 WPM (~<2% of ~265k total) | TechNews / market est. on IPO filing read | secondary |
| 2026 YE (est.) | CXMT HBM wafer starts | ~30,000 WPM | TechNews path (→ ~55k by 2027 YE) | secondary |
| 2026 (alternate) | CXMT HBM wafer allocation | ~60,000 WPM / ~20% of ~300k total | DigitalToday / foreign-media chain | secondary — conflicts with TechNews path |
| ~2026 mid | HBM3 8Hi package yield (est.) | ~25% comprehensive | TechNews market-person estimate | rumor / secondary |
| ~2026 | HBM3 initial yield (alt. band) | <60% vs leaders >80% | LavX / industry write-up | secondary — different band |
| 2026 | Mass-production target | HBM3 by end-2026; samples to Huawei et al. | Tom's Hardware / CNBC / Counterpoint via CNBC | quoted-secondary |
| 2026-07 | IPO proceeds use | Filing read: funds → commodity DRAM, not HBM line | TechNews on STAR filing | secondary |
| 2026-06 | Cycle stance | CXMT expansion ≠ near-term supercycle killer | Semi X 2071767487662768547 | quoted |
| 2026-08 | Monthly yield % | 待填 | — | open |
| 2026-08 | Monthly stacks shipped | 待填 | — | open |
Conflicts we refuse to flatten
- 5k → 30k → 55k WPM (TechNews path) vs ~60k WPM / 20% HBM allocation rumors — both secondary; we keep both rows.
- ~25% 8Hi yield vs looser “<60% / near Samsung” headlines — do not average.
- WPM ≠ stacks. Wafer starts are not finished HBM stacks. Converting without die-per-wafer + package yield is fanfic. We will not ship that conversion until we have an explicit source for the bridge.
- Semi's ~2M stacks / yr remains the Huawei-path working thesis on this site until a better primary replaces it — labeled site ← Semi, not “measured August output.”
Open months (copy forward)
| YYYY-MM | HBM gen | Yield % | Stacks shipped (est.) | Source | Notes |
|---|---|---|---|---|---|
| 2026-08 | HBM3? | 待填 | 待填 | — | Watch opened |
| 2026-09 | 待填 | 待填 | |||
| 2026-10 | 待填 | 待填 | |||
| 2026-11 | 待填 | 待填 | |||
| 2026-12 | 待填 | 待填 | End-2026 MP target check |
How this is “ours”
Semi knows its own model. What they do not ship as a public product is this: a dated, conflict-preserving yield watch wired to the Ascend stockpile essay, the memory-tax naming firewall (SPHBM4 ≠ Google SP-HBM), and the rule that YMTC does not fill HBM. The edge is the ledger + discipline, updated when sources move — not a pretended private yield number.
Next overwrite wins. If you have a primary (filing footnote, teardown, Semi paywall paste), we replace a row and bump the date.
References
- SemiAnalysis — Huawei Ascend Production Ramp (HBM bottleneck, 2025-09)
- SemiAnalysis X — CXMT deep + supercycle (2026-06-30)
- TechNews — CXMT HBM push may be weaker than rumored (2026-07-09)
- CNBC — CXMT debut vs memory giants (2026-07-31)
- Tom's Hardware — CXMT HBM3 samples / end-2026 mass-production target
- Companion — The Memory Tax Moves
Internal twin: data/research/2026-08-semianalysis-dram-hbm-memory.md §4.3