06-reference/research

cxmt q2 2026 gross margin anomaly

2026-11-04·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
memory-cyclecxmtdramgross-marginaccountingcost-per-bit

All three margin figures are real, but the comparison is not a cost test — and CXMT's own prospectus corroborates the >30% cost gap rather than contradicting it

The question

Verbatim: "Why did CXMT's Q2 2026 gross margin (87.6%) exceed SK hynix (83.2%) and Micron (84.9%)? Product mix, quarter timing, depreciation/subsidy accounting, or a real cost-gap narrowing? Does the CXMT prospectus disclose per-unit cost, utilization, or inventory days that would allow a direct cost-per-bit estimate?"

Context: this closes the first and fourth open follow-ups of [[2026-09-12-cxmt-ddr5-cost-per-bit-yield-parity-tripwire]], which flagged ChangXin Memory Technologies' (CXMT's) Q2 2026 margin as "the most interesting number in the brief" and could not explain it. This brief sizes a framework parameter and audits a disclosure. It contains no position recommendation, price target, or buy/sell call.

What we already know (from the vault)

What the web says

Premise check, figure by figure.

Convergences and contradictions

Synthesis for RDCO

Verdict on the four candidates: product mix is supported and is sufficient on its own; quarter timing is supported and small; two specific Chinese-accounting channels are supported and sized; a real cost-gap narrowing is not supported and the new evidence argues against it. Confidence: medium-high on mix and timing, medium on the accounting channels, medium-high on the rejection of a cost-gap closure.

Why mix alone closes the gap, by two independent routes. Route one uses Micron's business units. The four unit operating margins weighted by unit revenue reconcile to 81.3%, against a consolidated gross margin of 84.6%, so Micron allocates essentially all operating expense to units and the unit-level gross margin is roughly the unit operating margin plus about 3.3pp. That puts Micron's Mobile and Client unit at a gross margin near 89% and its HBM-heavy Cloud Memory unit near 81% (Ray derivation; the uniform operating-expense allocation is an assumption, since research-and-development intensity differs by unit). Route two solves the blend. Writing Micron's cost-of-goods-sold ratio as a revenue-weighted average of its DRAM and NAND ratios, with NAND at 15-20% of revenue and a NAND gross margin of 60-70%, the DRAM-only gross margin solves to about 89% (Ray derivation; the NAND weight and margin are assumptions this brief could not verify, because Micron publishes no DRAM-versus-NAND split). Two methods converge on Micron DRAM-only at roughly 88-90% for a quarter whose consolidated print was 84.9%, which is above CXMT's 87.59%. The ordering in the premise reverses once scope is matched. The mechanism is the one the vault already documents: 2026 was a conventional-DRAM price spike, HBM was pre-sold under multi-year agreements inside floor-and-ceiling bands, and NAND margins trailed, so a producer's margin rank in 2026 is mostly a ranking of exposure to spot-repriced conventional DRAM and only faintly a ranking of cost. CXMT, at 65.9% LPDDR with no HBM and no NAND, had the purest exposure available. SK hynix, carrying the largest HBM book and a NAND business, had the least, which is exactly the observed order.

Quarter timing is real but second-order, and the accounting channels are larger than expected. Micron's calendar-Q2 equivalent, month-weighting fiscal Q3's 84.9% over April and May with fiscal Q4's guided 86% over June, is roughly 85.3-86.5%, so timing accounts for about 0.4-1.6pp of the 2.7pp headline gap. On accounting, the subsidy question could not be answered: the filing draft discloses no grant policy and no grant amounts, so whether CXMT uses the net method, which reduces an asset's carrying value and therefore its depreciation and its cost of goods sold, or the gross method, which routes grants to other income below gross profit, remains open and is the largest unresolved confounder in this brief. Two adjacent channels are sizeable and quantifiable. First, written-down inventory: CXMT carried an impairment provision of RMB 1,986M at 2025-06-30, the residue of RMB 15.6bn of 2022-2023 write-downs, and it is selling that inventory into 2026 boom prices at a written-down carrying value. If the full balance was consumed in H1 2026, it is worth roughly 1.3pp of H1 2026 gross margin and up to about 2.3pp of Q2 2026 (Ray derivation; full-consumption-in-period is an assumption, and the Q2 figure uses a Q2 revenue of roughly RMB 85-90bn backed out of the disclosed H1 total, since CXMT's Q1-versus-Q2 revenue split was not sourced). That single channel is the same order of magnitude as the entire anomaly. Second, overhead absorption: CXMT's H1 2025 depreciation of RMB 11,349M is roughly 76% of its H1 2025 revenue, which cannot all sit in cost of revenue alongside a positive 12.72% gross margin, so a large share is being capitalised into an inventory balance that tripled from RMB 9.4bn to RMB 28.1bn. At a turnover ratio of 1.2-1.6 against Micron's typical two-to-three turns, those low-cost layers reach the income statement slowly and land in high-price quarters. The implied depreciation rate on net fixed assets of 12.5-14.0% (Ray derivation from the disclosed series) hints at useful lives longer than Micron's five-to-seven-year equipment lives, but with no useful-lives table disclosed this stays a hint, not a finding.

The load-bearing result: the margin-residual method, run where it has discriminating power, corroborates the >30% cost gap. Boom-quarter margins are a terrible cost instrument because the metric is violently leveraged there. At an 87.6% margin, one percentage point of gross margin corresponds to roughly an 8% move in unit cost or a 9% move in price, so the whole 2.7pp headline gap is equally well explained by CXMT's unit cost being about 18% lower than Micron's or by its realised revenue per unit of output being about 22% higher, and nothing in public data separates the two. Pre-boom periods are the opposite. Taking CXMT's own H1 2025 main-business gross margin of 12.72% against Micron's hard-disclosed overlapping fiscal Q3 2025 GAAP gross margin of 37.7%, the cost-of-goods-sold ratios are 0.873 and 0.623. Their ratio is 1.40, and applying the lineage's 5-10% CXMT price discount gives a cost-per-unit-of-output gap of 26-33% (Ray derivation). That brackets SemiAnalysis's >30% almost exactly, from two sources neither of which is SemiAnalysis. Three corrections push it wider, not narrower: Micron's 37.7% is blended and its DRAM-only figure was higher, so a scope-matched comparison implies a gap above 40%; CXMT's 12.72% excludes RMB 628M of write-downs that Micron-style accounting would have put in cost of goods sold, worth roughly 4pp of margin; and CXMT's periods carry ramp underloading. The honest read of the direction is that the >30% figure stopped being single-sourced tonight, and the appropriate posture is that the gap was intact as of mid-2025 with no public evidence of closure since.

What to change in the instrumentation. Three reversible edits, no posture change to [[2026-05-18-memory-cycle-v1.1]]. First, retire Indicator 1 of the parent brief as written. Its crossing condition, Q2's margin gap persisting through Q3, has now technically fired at 89.53%, and this brief shows that firing carries no information about cost, because the condition tests exposure rather than cost. Replace it with a mix-and-period-matched comparison, CXMT total gross margin against Micron's Mobile and Client unit margin plus about 3.3pp, on the nearest fiscal quarter, with a note that the residual is a price-versus-cost composite and not a cost reading. Second, re-key the cost-gap refresh to down-quarters. The parent brief's Indicator 2, margin beta in the first quarter of falling contract prices, is the right instrument and this brief explains why: margin differences regain discriminating power as margins fall, and the pre-boom comparison above is the proof of concept. Add the mechanical correction that CXMT's Chinese-standards margin excludes write-downs that Micron's includes, so any down-cycle comparison must add CXMT's asset-impairment loss back into its cost of goods sold before comparing. Third, mark Indicator 4 partly satisfied. The >30% estimate now has a second, independent, filing-based corroboration for mid-2025, so the open item narrows to whether the gap moved during 2026, which no public source addresses.

Answering the second half separately: no, the prospectus does not permit a direct cost-per-bit estimate, and the gap is specific. The binding absence is volume. The filing draft discloses no bit shipments, no wafer output, no wafer starts, and no capacity utilisation rate, and it gives no absolute per-unit cost or per-unit price, only a year-on-year percentage change in the unit selling price of major DRAM products. Without a bit or wafer denominator, no cost-per-bit level is recoverable by any route. Three weaker things are recoverable. A relative unit-cost index follows from combining gross margin with the disclosed price change: unit cost fell about 44% in 2023 and rose about 44% in 2024 (Ray derivation), but "per unit" means per piece and CXMT was migrating from DDR4 to denser DDR5 and LPDDR5 over exactly that window, so mix migration contaminates it and it is emphatically not per-bit. Inventory days are recoverable from the disclosed turnover ratio: roughly 248, 223 and 257 days for 2022, 2023 and 2024, with H1 2025's 1.19 ratio reading as about 153 days if it is a half-year figure and about 307 if annualised, an ambiguity the document does not resolve (Ray derivation; compare the Micron days-inventory-outstanding series in [[2026-09-07-memory-supplier-inventory-days-dio-anchor]]). And the margin-residual cost gap above, which is the most useful thing the document supports. One caveat on the whole answer: this is the 2025-12-29 filing draft read through an unofficial translation, covering only through H1 2025. A final registration prospectus or a listed-company annual report would carry updated periods and may add a cost-composition table; neither was reached in this run.

Why this is in the vault

It closes two of the four open follow-ups in [[2026-09-12-cxmt-ddr5-cost-per-bit-yield-parity-tripwire]] and changes what the CXMT cost leg is instrumented with. Concretely: Indicator 1 of that brief's monitoring table should be retired rather than marked as fired, the ">30% cost gap is single-sourced" caveat carried through four briefs in the CXMT lineage can be downgraded, and any future CXMT-versus-peer margin comparison in the memory-cycle thesis needs the write-down add-back correction documented here before it is used.

Open follow-ups

Related

Sources

Vault

Web — hard disclosure

Web — secondary

Flagged, not retried