China Memory Chip Production 2026
By Axis Intelligence Research
Co-authors: Sophie Winslow | Last updated: 1 September 2026 | License: CC BY 4.0
China’s largest DRAM maker earned RMB 61.799 billion in 2025 and posted its first attributable net profit — RMB 1.875 billion — after two years of heavy losses. Axis Intelligence Research finds its structural cost load fell from 182.33% of revenue to 58.17% in two years, the single clearest measure of Chinese memory reaching commercial scale.
Quick Answer
CXMT Corporation, China’s only volume DRAM manufacturer, reported audited 2025 revenue of RMB 61.799 billion, up from RMB 9.087 billion in 2023 — a compound annual growth rate of 160.78%. Axis Intelligence Research calculates its Memory Fixed-Cost Burden at 58.17% of revenue for 2025, down from 182.33% in 2023. Gross margin swung from −2.19% to 41.02% across the same three years.
Key Findings
- According to Axis Intelligence Research, CXMT Corporation’s Memory Fixed-Cost Burden — depreciation, amortisation and R&D combined as a share of revenue — fell from 182.33% in 2023 to 93.67% in 2024 to 58.17% in 2025.
- CXMT Corporation reported audited revenue of RMB 61.799 billion for 2025 and attributable net profit of RMB 1.875 billion, its first profitable year, according to its Shanghai Stock Exchange prospectus signed 9 July 2026.
- CXMT Corporation carried a consolidated accumulated deficit of RMB 36.650 billion as of 31 December 2025, meaning a single profitable year has not erased its losses.
- Axis Intelligence Research finds that none of the three fundraising projects named in CXMT Corporation’s prospectus is allocated to high-bandwidth memory, despite RMB 34.5 billion of planned project investment.
- The US Bureau of Industry and Security added Yangtze Memory Technologies to the Entity List effective 16 December 2022 under a presumption-of-denial licence policy, published at 87 FR 77505.
What Is China’s Actual Memory Chip Output in 2026?
There is a gap between what can be counted and what gets repeated, and on this topic the gap is unusually wide.
Almost every widely circulated figure for Chinese DRAM and NAND capacity — wafer starts per month, bit-output share, fab-by-fab ramp schedules — originates with a small number of subscription research firms running bottom-up supply models. Those models may well be good. They are also unpublished, unaudited, and impossible for a reader to check. Axis Intelligence Research does not republish them as fact.
What changed in July 2026 is that one of the two companies became legible. CXMT Corporation (长鑫科技集团股份有限公司) filed a prospectus with the Shanghai Stock Exchange to list on the STAR Market, audited by Deloitte, covering the three financial years 2023 through 2025. That document is the first time China’s DRAM industry has had to show its accounts to a securities regulator. This analysis is built on it.
YMTC remains private and publishes nothing comparable. We treat that asymmetry as a finding rather than a hole to fill, and we say plainly below what is not knowable.
The Two Companies, and Why They Are Not Symmetrical
CXMT makes DRAM — the volatile working memory in servers, phones and PCs. Yangtze Memory Technologies makes 3D NAND, the non-volatile storage in SSDs and phones. They are the two halves of China’s memory ambition and they sit in completely different information environments.
CXMT was founded on 13 June 2016 and operates three twelve-inch DRAM wafer fabs across Hefei and Beijing. It now files audited accounts. YMTC was placed on the US Entity List in December 2022, has no listed securities, and discloses no revenue, capacity or shipment data. Any number you read about YMTC’s output is an estimate by a third party.
Sophie Winslow, Industrial Technology: The instinct in this sector is to reach straight for wafer starts per month, because it is the number fab operators themselves use. But wafer capacity is an input, not an output. Two fabs running identical wafer volumes can ship wildly different quantities of sellable memory depending on die size, node and yield — and yield is exactly the number nobody discloses. A filing that tells me depreciation, inventory provisions and unit price movement tells me more about whether a fab is actually working than a capacity estimate ever will.
How Fast Did CXMT Revenue Actually Grow?
The audited series is unambiguous.
| Financial year | Revenue (RMB bn) | Attributable net profit (RMB bn) | Gross margin | Source |
|---|---|---|---|---|
| 2023 | 9.087 | −16.340 | −2.19% | CXMT prospectus, SSE |
| 2024 | 24.178 | −7.145 | 5.00% | CXMT prospectus, SSE |
| 2025 | 61.799 | 1.875 | 41.02% | CXMT prospectus, SSE |
Figures audited by Deloitte Touche Tohmatsu CPA LLP. Retrieved 1 September 2026.
Revenue rose 6.8-fold over two years. The prospectus states a revenue CAGR of 160.78% for the period, and we recomputed that independently from the audited revenue series and reached the same figure to two decimal places.
Two things deserve more attention than the growth rate.
First, the 2025 attributable profit of RMB 1.875 billion is smaller than the profit excluding non-recurring items, which was RMB 5.316 billion. When a company’s underlying operating result exceeds its headline result, something one-off pushed the headline down. That is the opposite of the pattern that usually needs scrutiny, and it makes the 2025 turn look more solid rather than less.
Second, the accumulated deficit stood at RMB 36.650 billion at the end of 2025. One good year against nine years of building does not clear the board.
The Price Cycle Behind the Numbers
The prospectus reproduces an Omdia series putting DRAM at a high of USD 7.89 per gigabyte and a low of USD 1.78 per gigabyte across 2015 to 2025, with the trough in the first half of 2023. Axis Intelligence Research calculates the peak-to-trough ratio at 4.4.
CXMT’s own average selling price for its main DRAM products rose 55.08% in 2024 and a further 33.69% in 2025. Roughly speaking, the company’s revenue explosion is a volume story and a price story at once, and the filing is candid that the price half is not durable: it states directly that continued sharp price increases are not sustainable.
The Memory Fixed-Cost Burden: An Axis Intelligence Research Metric
Memory fabs are among the most capital-intensive industrial assets on earth. The question that determines whether a memory company survives is not what it earns at the top of a cycle but what it must pay regardless of the cycle. That number is not published anywhere, so we built it.
MFB — Memory Fixed-Cost Burden measures the annual cost load a memory manufacturer carries irrespective of what its product sells for, expressed as a share of revenue.
Formula:
MFB = (fixed-asset depreciation + intangible amortisation + R&D investment) ÷ revenue × 100
All three inputs are disclosed line items in the audited filing. Nothing is estimated.
| Year | Depreciation | Amortisation | R&D | Revenue | MFB |
|---|---|---|---|---|---|
| 2023 | 10.555 | 1.343 | 4.670 | 9.087 | 182.33% |
| 2024 | 14.875 | 1.432 | 6.341 | 24.178 | 93.67% |
| 2025 | 24.680 | 1.675 | 9.593 | 61.799 | 58.17% |
All values RMB bn. Source: CXMT Corporation prospectus, Shanghai Stock Exchange. Computed by Axis Intelligence Research.
Worked example for 2025: (24.680 + 1.675 + 9.593) ÷ 61.799 × 100 = 58.17%.
Every reading was computed twice — once in floating-point and once in exact rational arithmetic — and both agreed to four decimal places.
What MFB is measuring. Depreciation sits mostly inside cost of sales; R&D sits in operating expense. MFB deliberately adds across that boundary because it is not a margin — it is the total structural load per unit of revenue, the thing that does not shrink when prices fall. A reader wanting margin should read the gross margin column above instead.
What it shows. In 2023, CXMT spent RMB 1.82 on fixed structural costs for every RMB 1.00 of revenue. By 2025 that was RMB 0.58. The fall is driven far more by the revenue denominator than by cost restraint: absolute depreciation more than doubled over the period and R&D more than doubled too. This is what scale arriving looks like in a capital-intensive industry — you do not spend less, you finally sell enough.
Scope. MFB is computed from consolidated figures and reflects both price and volume movement in the denominator. Because 2025 revenue benefited from a sharp DRAM upcycle, the 2025 reading flatters the underlying position; a price reversal would push MFB back up without any change in the company’s cost base. That sensitivity is the point of the metric, not a defect in it.
Sophie Winslow: Depreciation was 116.16% of revenue in 2023 and 39.94% in 2025. Run that backwards and you can see what the last downcycle did to this company — it was paying for fabs at full freight while selling memory below cost. The instructive thing is that the depreciation line never fell. It went from RMB 10.6 billion to RMB 24.7 billion while the burden ratio collapsed. Nothing got cheaper. The denominator arrived.
Which CXMT Fab Entities Actually Generate the Revenue?
This breakdown does not appear in English-language coverage, because it requires reading the subsidiary disclosures in the filing rather than the summary tables.
CXMT operates through three principal manufacturing subsidiaries, each separately audited.
| Subsidiary | Location | 2025 revenue (RMB bn) | 2025 net profit (RMB bn) | Control | Source |
|---|---|---|---|---|---|
| ChangXin Memory Technologies | Hefei | 29.025 | 7.556 | 100% owned | CXMT prospectus, SSE |
| ChangXin Xinqiao | Hefei | 17.931 | 3.328 | 73.01% of voting rights | CXMT prospectus, SSE |
| ChangXin Jidian | Beijing | 17.702 | 4.267 | 75.32% of voting rights | CXMT prospectus, SSE |
Entity-level audited figures before consolidation adjustments.
Here is the part that matters, and the reason this table needs a warning attached. Those three revenue figures sum to RMB 64.658 billion — which Axis Intelligence Research calculates as 104.63% of the consolidated 2025 revenue of RMB 61.799 billion.
They cannot be presented as a clean split, because intercompany transactions are eliminated on consolidation. Anyone dividing these three numbers to produce “share of CXMT output by fab” would publish a chart that is arithmetically impossible. We are publishing the entity-level figures because they are real and useful, and publishing the elimination alongside them because the subtraction is not ours to guess.
What the figures do support: the Beijing entity, ChangXin Jidian, converted revenue to profit more efficiently than the larger Hefei Xinqiao entity in 2025 — RMB 4.267 billion on RMB 17.702 billion versus RMB 3.328 billion on RMB 17.931 billion. Two entities of near-identical size, materially different results.
Note also that CXMT does not fully own two of its three main production entities. It controls ChangXin Xinqiao and ChangXin Jidian through concert-party agreements rather than outright majority equity, holding 30.68% and 31.72% of the shares respectively. The filing lists the possible unwinding of those agreements as a risk to consolidation itself.
Is CXMT Building HBM? What the Prospectus Actually Allocates
High-bandwidth memory is the product that matters most for AI accelerators and the one where China is furthest behind. The prospectus answers the question by omission.
The three named use-of-proceeds projects are:
| Project | Total investment (RMB bn) | Funded from offering (RMB bn) |
|---|---|---|
| Memory wafer manufacturing line technology upgrade | 7.50 | 7.50 |
| DRAM memory technology upgrade | 18.00 | 13.00 |
| Forward-looking DRAM research and development | 9.00 | 9.00 |
| Total | 34.50 | 29.50 |
Source: CXMT Corporation prospectus, Shanghai Stock Exchange.
Axis Intelligence Research finds no high-bandwidth memory project among them. Every named allocation goes to wafer lines, process upgrades and DRAM research for products the company already makes.
That is an observation about a document, not a claim about company strategy. A company can pursue HBM using operating cash flow, subsidiary balance sheets or state funding without naming it in a prospectus, and CXMT generated RMB 36.520 billion of operating cash flow in 2025. But the use-of-proceeds table is the most legally scrutinised statement of spending intent a Chinese issuer makes, and HBM is not in it.
Sophie Winslow: Ask what the omission costs to make. Naming an HBM project in a prospectus invites questions about equipment dependencies the company would rather not itemise in a public filing. Leaving it out costs nothing and forecloses nothing. I would not read the absence as evidence that no HBM work exists — I would read it as evidence about what this company is willing to write down in front of a regulator.
What Is Known and Not Known About YMTC’s NAND Production?
Yangtze Memory Technologies is the harder half of this question, and honesty about that is more useful than filling the space.
What is documented. The Bureau of Industry and Security added YMTC to the Entity List through a final rule effective 16 December 2022 and published 19 December 2022 at 87 FR 77505. Thirty-six entities were added in that rule. YMTC was listed alongside Hefei Core Storage Electronic Ltd under the destination of China and Yangtze Memory Technologies (Japan) Inc under the destination of Japan, with a licence requirement for all items subject to the EAR and a presumption of denial.
One detail from that rule is worth extracting because it is routinely missed. YMTC was added without the footnote 4 designation that brings the Entity List foreign direct product rule into play — the designation applied in the same rule to the Cambricon entities. The stated rationale for YMTC was risk of diversion to already-listed parties rather than support for military modernisation. Different rationale, different licence architecture, and a materially different practical restriction.
In a separate proceeding, the US Patent and Trademark Office Director vacated institution and denied two YMTC petitions challenging Micron NAND patents on 15 January 2026, in a decision designated informative (IPR2025-00098 and IPR2025-00099). The patents at issue were US 8,945,996 and US 10,872,903.
What is not known. YMTC’s revenue, bit output, wafer capacity, layer count in volume production, yield and global NAND share are not disclosed by the company and are not obtainable from any regulator. Figures circulating for these are third-party estimates. Axis Intelligence Research does not publish them as facts and does not average them into a range, because averaging unsourced estimates produces a number with false authority and no provenance.
This is the honest state of the record. If YMTC files or discloses, we will update this section and say what changed.
Who Buys Chinese DRAM?
The prospectus names end customers, which is unusual and useful.
CXMT’s top five customers accounted for 68.08% of main business revenue in 2025, down from 74.12% in 2023 but still high. The direct buyers are largely semiconductor distributors; the named end customers include Alibaba Cloud, ByteDance, Tencent, Lenovo, Xiaomi, Transsion, Honor, OPPO and vivo.
Two observations follow. The customer list is almost entirely domestic, which means demand is exposed to Chinese cloud capex and Chinese handset cycles rather than to global memory demand. And Alibaba is both customer and shareholder: Zhejiang Alibaba Cloud Computing subscribed RMB 6.1 billion in the June 2025 funding round and holds 3.85%.
Ownership Structure
| Shareholder | Stake | Notes |
|---|---|---|
| Hefei Qinghui Jidian | 21.67% | Largest holder; itself has no actual controller |
| Hefei ChangXin Integrated Circuit | 11.71% | Ultimately Hefei SASAC |
| National IC Industry Investment Fund II | 8.73% | The “Big Fund” second phase |
| Hefei Jixin | 8.37% | Employee holding platform |
| Anhui Provincial Investment Group | 7.91% | Anhui SASAC |
| Zhejiang Alibaba Cloud Computing | 3.85% | Also a customer |
Source: CXMT Corporation prospectus, Shanghai Stock Exchange.
CXMT reports no controlling shareholder and no actual controller. No single holder exceeds 50%, and the filing states that no shareholder can determine the appointment of more than half the board. State capital is present at multiple levels without any one entity holding formal control — a structure worth noting given that the company’s subsidiary was added to the US Department of Defense Section 1260H list of Chinese military companies on 8 June 2026, a listing CXMT discloses and states will not materially affect operations.
How Exposed Is CXMT to the Next Downcycle?
This is the question the audited numbers are best equipped to answer, and the answer is more nuanced than either the bull or bear case.
Arguing for resilience: operating cash flow reached RMB 36.520 billion in 2025, up from negative RMB 7.272 billion in 2023. The gearing ratio fell from 66.22% to 54.24%. Inventory write-down provisions fell from RMB 3.444 billion in 2023 to RMB 0.822 billion in 2025, indicating inventory is currently saleable at cost. And the RMB 29.5 billion raised in the offering adds a cushion the company did not have in the last trough.
Arguing for fragility: net fixed assets reached RMB 183.024 billion at end-2025, 54.34% of total assets, and the filing warns depreciation will keep rising as construction completes. Inventory grew to RMB 29.390 billion. The accumulated deficit of RMB 36.650 billion remains. Customer concentration is 68.08%. And the entire 2025 result rests on an ASP that rose 33.69%.
Where we disagree with the prevailing framing. Coverage of CXMT has converged on wafer capacity as the measure of threat to the incumbents — the recurring claim that CXMT will approach Micron’s wafer volume. Wafer capacity is the wrong yardstick for that specific conclusion, and CXMT’s own filing makes the case better than we can: it reports 2025 gross margin of 41.02% against Samsung, SK hynix and Micron holding a combined 91.85% revenue share by Omdia’s reckoning, and states plainly that its process technology still lags the top three. A company can match a competitor’s wafer starts and remain well behind on bits shipped and further behind on revenue per wafer. The capacity framing consistently overstates near-term commercial impact because it silently assumes equivalent die size and yield.
Methodology
Collection. Every figure in this analysis was retrieved from a source document opened during production on 1 September 2026. The primary source is the CXMT Corporation prospectus (招股意向书) signed 9 July 2026 and filed with the Shanghai Stock Exchange for the STAR Market listing, with financial statements for 2023–2025 audited by Deloitte Touche Tohmatsu CPA LLP and Q1 2026 figures reviewed but not audited. Regulatory facts were taken from the Federal Register final rule at 87 FR 77505 and from the USPTO Director’s decision in IPR2025-00098.
Currency. All figures are stated in Renminbi as filed. We have not converted to US dollars, because doing so requires selecting an exchange rate and date that the source does not specify, and an unsourced conversion would introduce a number the filing does not support. The prospectus reports in units of RMB 10,000 (万元); we have converted to billions for readability, and the CSV preserves the original filed unit.
Formulas. MFB is defined above with its inputs shown. Other Axis calculations — the big-three combined share of 91.85%, the peak-to-trough price ratio of 4.4, the subsidiary-sum ratio of 104.63% — are each disclosed inline where they appear and carry their full formula in the method_note column of the dataset.
Arithmetic verification. All calculations were run twice in Python, once using floating-point and once using exact rational arithmetic via fractions.Fraction, and were required to agree before publication. One discrepancy surfaced and is disclosed here: the prospectus states cumulative 2023–2025 R&D investment of RMB 20,605,016,700 while the three annual figures it discloses sum to RMB 20,605,016,600 — a rounding difference of RMB 100. The filing itself notes that rounding may cause such tail-end differences. We report the annual figures as filed and have not silently reconciled them.
Sources we declined to use. Wafer capacity, bit output and market-share forecasts from subscription research firms were excluded, because they cannot be verified by a reader and their underlying models are not published. Where a commercial research figure appears inside the regulated filing — the Omdia market shares and price series — we report it, attribute it to Omdia in plain text, and flag it as non-primary in the dataset rather than presenting it with the authority of the audited accounts.
Scope. This analysis covers what regulatory filings disclose. It does not estimate Chinese national DRAM or NAND production volume, and it does not model YMTC. Those figures do not exist in verifiable form, and the absence is stated rather than filled.
About This Dataset
The complete fact table backing this analysis is published as china-memory-chip-production.csv — 112 rows covering every figure that appears above, each carrying its value, as-of date, source organisation, source document, source URL, retrieval date, a primary-source flag, an Axis-calculated flag, and for every calculated row the full formula.
Licence: CC BY 4.0 (plain text; attribution required).
Citation: Axis Intelligence Research, China Memory Chip Production 2026: CXMT and YMTC in Audited Numbers, 2026.
Citation Formats
APA: Axis Intelligence Research. (2026). China memory chip production 2026: CXMT and YMTC in audited numbers. https://axis-intelligence.com/china-memory-chip-production/
MLA: Axis Intelligence Research. “China Memory Chip Production 2026: CXMT and YMTC in Audited Numbers.” Axis Intelligence Research, 2026, axis-intelligence.com/china-memory-chip-production/.
Chicago: Axis Intelligence Research. “China Memory Chip Production 2026: CXMT and YMTC in Audited Numbers.” Axis Intelligence Research, 2026. https://axis-intelligence.com/china-memory-chip-production/.
Frequently Asked Questions
Does CXMT’s wafer capacity matter more than its bit output?
Bit output matters more, and the two diverge. Wafer capacity counts silicon discs entering the line; bit output counts sellable memory leaving it. The conversion depends on die size, process node and yield, and CXMT’s filing states its process technology still trails Samsung, SK hynix and Micron. A capacity figure that matches an incumbent therefore does not imply matching supply. Bit output and yield are not disclosed by any Chinese memory maker.
Why does CXMT’s depreciation keep rising even as it becomes profitable?
Because fabs under construction are not yet being depreciated. Net fixed assets reached RMB 183.024 billion at end-2025, and the prospectus states the carrying value will increase further as capacity building continues. Depreciation rose from RMB 10.555 billion in 2023 to RMB 24.680 billion in 2025 and will keep climbing as new lines transfer from construction-in-progress to fixed assets. Profitability arrived through revenue growth outpacing that increase, not through it stopping.
Did CXMT capitalise its R&D to flatter its first profitable year?
No — it did the opposite. CXMT capitalised RMB 1.734 billion of R&D in 2024, a capitalisation ratio of 27.34%, and capitalised nothing at all in 2025. Expensing all RMB 9.593 billion of 2025 R&D depressed the reported profit rather than inflating it. For a company crossing into profitability for the first time under regulatory scrutiny, that is a conservative choice.
What does the Entity List actually prevent YMTC from buying?
It imposes a licence requirement on all items subject to the Export Administration Regulations, reviewed under a presumption of denial. In practice that covers US-origin semiconductor manufacturing equipment, components and software. YMTC was listed without the footnote 4 designation that extends the foreign direct product rule to certain foreign-made items, which the same December 2022 rule applied to other listed entities — so the restriction, while severe, is narrower in reach than the toughest designation available.
Is China’s DRAM industry profitable at current prices, and what happens if prices fall?
CXMT reached 41.02% gross margin in 2025 on an average selling price that rose 33.69%. Its own prospectus states that continued sharp price increases are not sustainable and warns that industry oversupply could return as Samsung, SK hynix and Micron bring planned capacity online. Because the cost base measured by MFB is fixed and the revenue denominator is not, a price reversal would push the burden ratio back up without any change in spending.
Who actually controls CXMT?
Formally, nobody. The company reports no controlling shareholder and no actual controller. The largest holder, Hefei Qinghui Jidian at 21.67%, itself has no actual controller. State capital appears through Hefei SASAC, Anhui SASAC and the National IC Industry Investment Fund II, but no single shareholder can appoint a majority of the eleven-member board. The company flags this dispersed structure as a governance risk in its own filing.
Can I reuse the numbers on this page?
Yes. The dataset is published under CC BY 4.0 with attribution to Axis Intelligence Research. Every figure carries its source URL and retrieval date so you can verify it against the filing directly rather than taking ours on trust.