CXMT's Filing Says HBM Zero Times. Forecasts Say +207%.
Across all 192 pages of CXMT's H1 2026 interim report, the string "HBM" appears zero times. The Chinese equivalent, 高带宽内存, appears zero times too.
Meanwhile the sell-side numbers in circulation have CXMT shipping 125 million GB of HBM in 2026 and 1,179 million by 2028 — a 207% CAGR. This post measures the gap rather than picking a side.
A follow-on to the three-part series on CXMT's H1 2026 interim report.
KEY TAKEAWAYS
1. The filing mentions HBM zero times, and the July IPO prospectus contains no dedicated HBM project. The forecast has HBM growing at +207% a year through 2028.
2. Delivering that volume means allocating 23.5% of total DRAM wafer capacity to HBM by 2027. That is not a change a company makes quietly.
3. So the conclusion is not "the forecast is wrong." It is that the forecast dictates what has to appear in the next few filings for it to hold.
Two documents of different kinds
On one side is the interim report ChangXin Technology filed in August 2026. Unaudited, but a document the company puts its name to. The product section lists DDR5 and LPDDR4X/5/5X in volume production and LPDDR6 at customer sampling; the process roadmap runs to a fourth-generation platform in production and a fifth in customer qualification. HBM appears in none of it — not the product section, not the roadmap, not the risk factors.
On the other side is a Goldman Sachs Global Investment Research table. The report itself is paywalled; only the numbers from table images publicly cited on X are used here, and the basis for the assumptions is unknown. Worth stating up front.
The distinction matters. A filing records what has happened. A forecast records what will. The absence of HBM in one does not refute the other. What can be calculated is the size of the preparation the forecast implies.
What 125 actually asks for
HBM consumes far more wafer than commodity DRAM at the same capacity. Stacking dies and drilling TSVs cuts usable area, and stacking adds its own losses. The industry estimate quoted in the press is 3 to 4 times; the conservative 3x is used here.
The same table puts CXMT total DRAM supply at 5,536 / 7,382 / 9,837 million GB across 2026–2028. Converting the HBM volumes to wafer equivalent:
| Year | HBM (m GB) | Bit share | Wafer-equivalent share |
|---|---|---|---|
| 2026E | 125 | 2.3% | 6.5% |
| 2027E | 687 | 9.3% | 23.5% |
| 2028E | 1,179 | 12.0% | 29.0% |
On bits, 2027 looks modest at 9.3%. On wafers it is 23.5% — a quarter of the line turned over to HBM, and a quarter less commodity DRAM going out the door.
To do that in 2027, what has to exist now
Turning over a quarter of a line leaves traces in filings.
HBM is not just DRAM die. It needs TSV process, stack bonding equipment, a base die and a test line. Given lead times from order to production, 687 in 2027 means tools installed during 2026.
The July 2026 prospectus names three projects: 13.0B yuan for DRAM process advancement, 9.0B for next-generation research, 7.5B for wafer line upgrades. No HBM line item. The roughly 37.1B yuan of oversubscription has no disclosed allocation, so the room exists — it just has not been claimed.
Stated precisely: this is not "CXMT cannot do HBM." It is that no money or plan for HBM is visible in the public documents, which rules out neither undisclosed work nor a later filing.
The shape of the curve the forecast drew
The same table puts China HBM demand at 547 / 1,218 / 3,086 million GB. Dividing CXMT supply into that gives a self-supply rate of 23% → 56% → 38%.
It peaks in 2027 and falls back. Not because CXMT supply shrinks, but because domestic demand grows faster: 2028 demand is 2.5x 2027, while CXMT supply is 1.7x.
That tells you how the forecaster sees it. Not China self-supplying HBM — China closing in, then falling behind again.
Against the world
Global HBM demand runs 4,724 in 2026 to 9,713 million GB in 2028. CXMT's share goes 2.6% → 10.0% → 12.1%.
By 2028 that is an eighth of world HBM volume. In the same table Samsung HBM reaches 3,494 and SK hynix 3,968 million GB, so CXMT's 1,179 is about a third of either. Not small. Not board-flipping either.
What I actually watch
Rather than deciding whether to believe the forecast, it is more useful to fix the order in which its signals would appear.
| Order | What should appear | If it doesn't |
|---|---|---|
| 1 | An HBM line in the allocation filing for the 37.1B yuan oversubscription | 2027's 687 has no funding basis |
| 2 | HBM in the product section of a quarterly report, even at sampling stage | The lead time to production is already short |
| 3 | TSV and stack-bonding tool orders — a jump in construction in progress and capex | The line is not being prepared |
| 4 | A base die path — own logic process or an outside foundry | HBM4-generation parts are hard to reach |
Risks to this view
▶ The forecast source. Only part of a paywalled report was publicly cited, so the assumptions behind it could not be checked. Internal consistency was verified; plausibility was not.
▶ The wafer multiple. 3x is the low end of the quoted 3–4x range. At 4x, the 2027 wafer share rises to 29.1%. Either way the direction holds.
▶ What a filing is. An interim report describes the state at June 30. What started after that shows up in the next one.
The short version
When a forecast and a filing disagree, you do not have to pick one. Measure the gap and watch which way it closes. Here the closing has only two shapes — HBM appears in the disclosure, or the forecast comes down.
Next: the LPDDR6 reporting from late August, and where the phrase "world's first" was actually attached. Same class of gap, smaller scale.
Sources: ChangXin Technology Group H1 2026 interim report (unaudited, 688825); Goldman Sachs Global Investment Research Exhibits 37 and 50, from table images publicly cited on X; Tom's Hardware, July 2026. Everything here is from public sources.
Disclaimer: This post is for informational and educational purposes only. It does not constitute investment advice or a recommendation to buy or sell any security. All investment decisions are your own responsibility.
Comments
Post a Comment