CXMT Deep Dive: The 550,000-Wafer/Month Capacity, 18% Share, and Breakthrough Path Behind Nomura’s RMB116 Price Target
CXMT Deep Dive: The 550,000-Wafer/Month Capacity, 18% Share, and Breakthrough Path Behind Nomura’s RMB116 Price Target
目录
TL;DR
I. Understanding RMB116: Pricing a Difficult 2028 End State in Advance
II. First Correct Four Errors in the Source Model: The Sell-Side Model Is Useful, but Should Not Be Copied Verbatim
What Exactly Has the New Information Changed?
First, Quantify the 2026 Earnings Step-Up
III. Why AI Genuinely Needs More Memory: The Bottleneck Is Shifting from “Insufficient Compute” to “Insufficient Data Delivery”
Bit Demand, Wafer Volume, and Revenue Are Not the Same Thing
IV. Why Supply Cannot Keep Up: Capacity Expansion Requires More Than Buying Equipment
V. How CXMT Reaches 550,000 Wafers of Capacity: Distinguishing Nominal Capacity from Effective Output
How Much Qualified Output Can Ultimately Be Retained from 550,000 Wafers?
VI. Process Nodes, Yields, and Wafer Pricing: The Real Profit Breakthrough
Breaking Revenue Down to the Wafer Level
VII. How to Break Through to an 18% Share: Adding More Production Lines Is Not Enough
VIII. HBM and WoW: Worth Assigning Option Value, but Not Yet Profit
IX. Nomura’s 2026—2028 Model: Why Profit Appears to Be on an Accelerator
Depreciation, Inventory, and Free Cash Flow Form Another Income Statement
How Dependent Is RMB116 on the “20x” Multiple?
10. Sanctions Scenario: Nomura’s So-Called “Worst Case” Actually Only Removes Shanghai Capacity
A Genuine Stress Test Must Examine How Four Risks Amplify One Another
11. The Domestic Supply Chain Will Benefit, but CXMT’s Capacity Expansion Should Not Be Equated with Linear Order Growth
12. How CXMT Can Break Through: Three Scorecards Matter More Than the Price Target
1. Manufacturing Scorecard: From Installed Capacity to Qualified Bits
2. Product Scorecard: From Domestically Available Products to High-Value Products
3. Supply-Chain Scorecard: From Substitution Capability to Stable Maintenance
The Shortest Validation Chain for Determining Whether RMB 116 Remains Valid
XIII. Conclusion: RMB 116 Is the Score After Completing Every Difficult Maneuver, Not the Starting Line
本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读
On CXMT’s first day of trading, Nomura assigned it an RMB116 price target, more than 12 times above its RMB8.66 issue price. Whether that valuation is ultimately justified depends on three things: 550,000 wafers of monthly capacity, an 18% global share, and how much of the near-ceiling margin can actually be realized.
TL;DR
The arithmetic behind RMB116 works; the difficulty lies in each multiplier. Nomura multiplies its 2028 forecast EPS of RMB5.79 by a 20x P/E multiple to derive RMB115.8, rounded to RMB116. Based on the RMB8.66 issue price, this implies 1239.5% upside. However, RMB5.79 in EPS itself requires CXMT to generate RMB393.07 billion in attributable net profit in 2028, while the 20x multiple requires the market to value the company at approximately twice Micron’s historical median P/E even at the profit-cycle peak. This is a dual extrapolation of “high earnings × high valuation.”
The price target values CXMT as a global memory giant. Nomura’s model assumes approximately 67.884 billion shares after full exercise of the greenshoe option. At RMB116, this implies a total market capitalization of approximately RMB7.87 trillion, or approximately US$1.16 trillion at the report’s exchange rate—higher even than the US$960 billion Micron market capitalization cited in the report. Investors are effectively betting that CXMT will approach the global top three in scale, earnings, and valuation status within three years.
Monthly capacity of 550,000 wafers is the nominal year-end 2028 endpoint. Nomura estimates that CXMT’s monthly capacity will increase from 280,000 wafers at end-2025 to 350,000 wafers at end-2026, with another 100,000 wafers added in each of 2027 and 2028, ultimately reaching 550,000 wafers. Accounting for intra-year commissioning, utilization, and yields of 75%–85%, the model implies effective wafer output of approximately 400,000 wafers/month in 2028. Earnings ultimately depend on stable output of qualified wafers, not the amount of equipment installed in fabs.
The 18% share is the optimistic upper bound for global DRAM bit-output share, not a company disclosure. The official figure is a 7.67% global share by revenue in the fourth quarter of 2025. Nomura’s report, by contrast, uses starting points of approximately 8%, approximately 9%, and approximately 10%, then compares CXMT’s 40%–45% bit growth with industry supply growth to derive an approximately 18% bit share by end-2028. The denominator, measurement basis, and timing differ, so 7.67% cannot be connected directly to 18%.
Pricing is the most aggressive assumption in Nomura’s earnings model. The model assumes that global DRAM prices rise from US$4.0/GB in 2025 to US$13.7/GB in 2026, an increase of 245% year over year. CXMT’s price per wafer is then assumed to rise from US$14,000–15,000 in 2026 to US$21,000–25,000 in 2027–2028. As a result, gross margin jumps from 41.0% in 2025 to 83.7% in 2026 and rises further to 90.5% in 2028. If pricing, yield, or process migration fails to materialize at any stage, earnings will not increase in a straight line.
The direction of AI demand is credible, but growth above 60% remains an aggressive assumption. Agentic AI will increase demand for model weights, KV cache, context integration, multi-round iteration, and cold-data storage, benefiting conventional DRAM, HBM, and SSDs. Nomura further assumes that the number of agentic tasks in 2030 will be 50 times the 2026 level. Even if efficiency technologies reduce memory use per task by 4 times, total memory usage would still grow by more than 7 times from 2026 to 2030. The direction is credible; the slope must be verified.
“How to break through” depends on passing three gates. The first is manufacturing: migrating from 16–17nm to 14–15nm, improving DDR5 yields, and increasing bits per wafer. The second is products: expanding from an LPDDR base into server DDR5, HBM, and DRAM-on-logic wafer stacking. The third is the supply chain: raising the sell-side estimate of approximately 40% domestic-equipment content while addressing constraints in advanced photoresists, spare parts, and equipment servicing. If any one of these three gates is not passed, the 550,000-wafer capacity and 18% share remain figures on paper.
The sanctions scenario shows that RMB116 does not offer a particularly wide margin of safety. Nomura delays only two Shanghai phases totaling 200,000 wafers/month of advanced DRAM capacity and estimates that 2028 revenue and attributable net profit would be 30% and 33% lower, respectively. This scenario does not yet fully incorporate process migration, yields, equipment maintenance, or valuation-multiple compression, so it cannot be considered a true worst-case scenario.
I. Understanding RMB116: Pricing a Difficult 2028 End State in Advance
Nomura initiated coverage of CXMT on July 27, 2026, with a Buy rating and an RMB116 price target. The valuation formula is very simple:
2028 forecast EPS of RMB5.79 × 20x P/E = RMB115.8.
The enormous gap between the RMB8.66 issue price and the RMB116 price target is not a mysterious premium generated by a complex model. It comes from two choices: first, setting the valuation year directly at 2028, bypassing the low 2025 earnings base and the 2026 earnings surge; second, assigning CXMT a 20x P/E multiple rather than applying Micron’s historical 5x–15x range and approximately 10x median.
Nomura’s explanation is that comparable assets in China generally command higher valuations than in the US market, while CXMT’s global share is also rising, justifying a 100% valuation premium to Micron. This reasoning is not entirely without basis, but it simultaneously pushes both “share gains” and the “valuation premium” close to their upper bounds.
This is the most important point for understanding RMB116. It is not a conservative answer to “what is domestic DRAM worth?” It is the answer to “what would CXMT theoretically be worth if, by 2028, its share approaches Micron’s, its earnings reach the sell-side’s high-cycle forecast, and it continues to enjoy a valuation premium for Chinese assets?”
Our previous comprehensive company deep dive based on the prospectus already explained CXMT’s business, financials, products, and fundraising foundation. This article addresses only the new questions arising after the listing: why Nomura is willing to assign an RMB116 price target and what hurdles must be cleared for that target to materialize. The previous report is available here: CXMT Deep Dive: DDR5 Volume Ramp, RMB50 Billion in First-Half Attributable Net Profit, and the Cycle Test for China’s DRAM Leader.
II. First Correct Four Errors in the Source Model: The Sell-Side Model Is Useful, but Should Not Be Copied Verbatim
Nomura’s report provides a complete 2026–2028 model, but there are several conflicts between the appendix and the main tables. These errors must be corrected before discussing the price target further.
First, 2025 attributable net profit was overstated by 10 times. Page 25 of the report states approximately RMB18.7 billion, while both the main financial table and the Shanghai Stock Exchange prospectus report RMB1.875 billion. Consolidated net profit in 2025 was RMB7.144 billion, while attributable net profit excluding non-recurring items was RMB5.316 billion. These three measures cannot be conflated.
Second, the report gives two different figures for 2026 capacity: 300,000 wafers/month and 350,000 wafers/month. The appendix states that the company’s target is total capacity of 300,000 wafers/month by end-2026, while the core earnings model uses 350,000 wafers/month. CXMT’s official materials do not disclose absolute monthly capacity, so 350,000 wafers can only be described as Nomura’s estimate, not as the company’s plan.
Third, the direction of LPDDR’s revenue share in 2025 was stated incorrectly. LPDDR’s share declined from 74.54% in 2023 to 66.43% in 2025; it did not “increase to 66.43%.” DDR’s share was what actually increased, rising from 20.16% to 31.87%, driven by the server DDR5 volume ramp.
Fourth, the report mixes shares calculated using different denominators. The company officially disclosed a 7.67% share of global DRAM revenue in the fourth quarter of 2025. Nomura’s main text states “currently approximately 10%,” its charts show close to 9%, and it ultimately forecasts an approximately 18% share of global DRAM bit output in 2028. Revenue share is affected by pricing and product mix, while bit share more closely reflects physical output. They are not the same metric.
In addition, the report’s EV/EBITDA, EV/EBIT, certain balance-sheet figures, and ROA cannot be reproduced using data on the same pages, so this article does not use those multiples. The governance structure also follows the prospectus: CXMT has no controlling shareholder or actual controller, and this article does not adopt the conflicting appendix description that identifies Zhu Yiming as the controlling shareholder and actual controller.
This does not mean the entire report is unusable. On the contrary, its capacity, pricing, demand, and stress-test analysis is highly valuable. The correct approach is to use historical data according to official filings, clearly label future figures as Nomura forecasts, and exclude internal inconsistencies from the investment conclusion.
What Exactly Has the New Information Changed?
Compared with the previous company deep dive based on the prospectus, Nomura’s report does not change CXMT’s business identity or the fundamental conclusions regarding DDR5 volume growth, domestic substitution, and DRAM’s strong cyclicality. What it genuinely adds are four testable quantitative assumptions: nominal capacity of 550,000 wafers/month by end-2028, an approximately 18% share of global DRAM bit output in 2028, attributable net profit of RMB393.07 billion in 2028, and an RMB116 price target based on a 20x P/E multiple.
These four figures move the question from “does CXMT have room to grow?” to “how much growth must materialize?” The prospectus can demonstrate that the company already has three 12-inch wafer fabs, held a 7.67% global revenue share in the fourth quarter of 2025, and achieved 95.73% capacity utilization in 2025. It can also demonstrate that DDR5 and LPDDR5/5X have entered mass production. However, it cannot prove that capacity will definitely reach 550,000 wafers/month in 2028, that global share will definitely reach 18%, or that attributable net profit will reach RMB393.07 billion.
Accordingly, this article makes three adjustments to the previous conclusions. First, confidence in the direction has increased: listing proceeds, DDR5 volume growth, and high utilization provide a more solid starting point for capacity expansion. Second, the confidence interval around earnings forecasts has widened: Nomura simultaneously pushes industry pricing, capacity, yields, process nodes, and the valuation multiple toward their upper bounds. The more specific the figures become, the more interconnected risks they expose. Third, the monitoring focus shifts from “whether China has domestic DRAM” to “how many qualified bits each wafer can produce, what price they can command, and whether advanced production lines can be maintained reliably.”
First, Quantify the 2026 Earnings Step-Up
2026 is the first observation year for determining whether Nomura’s model has a credible starting point. The officially reviewed data are: first-quarter 2026 revenue of RMB50.8 billion, consolidated net profit of RMB33.012 billion, and attributable net profit of RMB24.762 billion. The company expects first-half 2026 revenue of RMB110 billion–120 billion, consolidated net profit of RMB66 billion–75 billion, and attributable net profit of RMB50 billion–57 billion. The first-half figures have not been audited or reviewed and do not constitute an earnings commitment.
At the midpoint of the company’s first-half forecast range, revenue would be approximately RMB115 billion and attributable net profit approximately RMB53.5 billion. To reach Nomura’s full-year 2026 forecasts of RMB290.666 billion in revenue and RMB130.315 billion in attributable net profit, the second half would still need to generate approximately RMB175.666 billion in revenue and RMB76.815 billion in attributable net profit. This implies a second-half attributable net margin of approximately 43.7%, which would not represent a further increase from the first-half midpoint of approximately 46.5%. However, second-half revenue would still need to be approximately 52.8% higher than the first-half midpoint.
This breakdown is important. Nomura’s 2026 earnings forecast is not wholly detached from the first-quarter results and the company’s expected first-half range. The genuinely aggressive element is that it requires high pricing, high utilization, and a high net margin to be sustained throughout the year and then continue expanding in 2027–2028. The company has explicitly warned that the high growth recorded in the first half of 2026 may not be sustainable. The first-quarter attributable net margin of 48.7% therefore cannot be mechanically annualized, and a short-term pricing peak cannot be treated directly as a long-term earnings midpoint.
Nomura’s forecast implies that attributable net profit will jump from RMB1.875 billion in 2025 to RMB130.315 billion in 2026, approximately 69.5 times the previous year’s level. Such enormous operating leverage means the model is relatively insensitive to minor deviations but extremely sensitive to a cyclical reversal: if selling prices decline faster than costs, earnings will retreat more rapidly than revenue. The first question investors need to verify is not the 2028 outcome, but whether volume, pricing, and margins can hold their three respective steps in the second half of 2026.
III. Why AI Genuinely Needs More Memory: The Bottleneck Is Shifting from “Insufficient Compute” to “Insufficient Data Delivery”
During the early stage of large-model training, the market focused primarily on GPU counts. With the rise of inference and agentic AI, systems increasingly resemble pipelines that continuously read, retain, and append context. Compute chips may become faster, but if model weights, KV cache, tool outputs, and historical context cannot be delivered to processors in time, GPUs will still sit idle waiting for data.
An agentic task generally proceeds through eight stages: receiving the request, loading model weights, prefill, inference and planning, calling tools, integrating context, multi-round iteration, and generating results. A standard Q&A; task may require only one round, while an agent repeatedly invokes search, code, files, and APIs, feeding each new result back into the context. As the number of iterations increases, the KV cache expands, and memory pressure typically peaks at this stage.
This simultaneously drives demand across three layers:
SRAM and HBM, closest to the processor, address bandwidth and latency;
Host-side DDR provides more capacity than HBM and supports models, cache pools, and system memory;
SSDs and HDDs store RAG data, checkpoints, logs, and cold data.
AI memory demand therefore benefits more than just HBM. HBM is the most expensive and offers the highest bandwidth, but standard server DDR, enterprise SSDs, and high-capacity storage are also integral to a complete inference system. For CXMT, the more realistic near-term benefit still comes from DDR5 and LPDDR5/5X, rather than directly counting HBM, which remains under validation, as core-business profit.
Nomura’s most instructive point is that it does not ignore efficiency technologies. Quantization, GQA, PagedAttention, prefix caching, MLA, and other techniques can materially reduce memory consumption per task. The report estimates that, in theory, multiple technologies could cumulatively deliver savings of 4—40x, but the combined savings in real-world systems are more likely to remain below 5x. Its base case applies an efficiency adjustment of approximately 4x.
The issue is that Nomura also assumes the number of agentic tasks in 2030 will be 50x the 2026 level. Dividing 50x task growth by an approximately 4x efficiency improvement still implies more than 10x growth in agent-related memory consumption. Even if non-AI demand does not grow, global memory demand could still achieve a bit CAGR of more than 60% during 2026—2030, representing cumulative growth of more than 7x.
The direction of this reasoning is sound, but the magnitude is aggressive. It requires the number of users, task frequency, task duration, context length, and agent penetration all to increase simultaneously. If agents remain concentrated in simple Q&A;, enterprises impose strict cost limits, or model architectures compress KV cache more aggressively, the 60% estimate would be revised downward.
The most prudent way to assess AI demand is not to debate whether “7x is too aggressive,” but to continuously monitor three facts: whether cloud providers are increasing server memory configurations, whether long-term DDR5 and HBM agreements are being extended, and whether commodity DRAM prices can hold after new capacity comes online. As long as all three persist, the directional thesis remains valid. If only model-call volumes continue growing while memory value per server declines, the earnings trajectory will be weaker than Nomura’s model suggests.
Bit Demand, Wafer Volume, and Revenue Are Not the Same Thing
The easiest mistake when assessing this AI memory cycle is to equate bit-demand growth directly with revenue growth. A bit is a physical unit of storage capacity, while revenue equals shipment volume multiplied by price. Process migration allows each wafer to yield more bits, efficiency technologies reduce consumption per task, and product mix assigns very different prices to the same capacity. Even at identical capacity, HBM, server DDR5, and mobile LPDDR5X differ materially in bandwidth, packaging, qualification, and pricing.
Nomura’s upside scenario amplifies two variables simultaneously: memory bit demand growing at a CAGR of more than 60% during 2026—2030, and global DRAM prices rising from US$4.0/GB in 2025 to US$13.7/GB in 2026. If supply genuinely remains constrained over the long term, both variables can rise together. Once new capacity from Samsung Electronics, SK hynix, Micron, and CXMT comes online, however, supply growth will first pressure prices and then affect manufacturers’ returns on capacity expansion.
The report’s supply-demand table shows negative inventories and utilization rates above 100% after 2027. In the physical world, inventories cannot remain negative for an extended period, nor can production-line utilization remain consistently above 100%. These figures should be interpreted as “model-implied unmet demand,” indicating that prices or supply must adjust further, rather than as real-world conditions that can be entered directly into an income statement.
Demand analysis must therefore pass through three layers of conversion: whether agentic tasks increase; whether those tasks translate into actual memory configurations; and whether incremental bits ultimately become products that CXMT can deliver and collect payment for. The first layer concerns AI applications, the second concerns system architecture, and only the third concerns the company’s manufacturing and sales. Verifying model-call volumes alone cannot prove CXMT’s revenue, while verifying DRAM price increases alone cannot prove an 18% share.
IV. Why Supply Cannot Keep Up: Capacity Expansion Requires More Than Buying Equipment
DRAM may appear to be a standardized product, but expanding capacity involves more than moving machines into a fab. Cleanrooms, lithography, etching, deposition, inspection, materials, engineers, and yield must all be in place. If any one of these falls behind, nominal capacity cannot become qualified bit output.
Nomura estimates that global DRAM industry bit supply will grow by approximately 30%—40% during 2026—2030, below its demand-growth assumption of more than 60%. The report also notes that the DRAM industry’s capital expenditure over the past 15 years has generally equaled 25%—45% of market revenue, but this ratio could fall to approximately 15% in 2026 as prices surge and the market expands. Absolute capital expenditure is not declining; rather, the revenue denominator is rising faster, making supply expansion appear more restrained.
However, the report’s 30%—40% supply-growth figure in the main text is not fully consistent with its charts. Total DRAM production in the charts implies a CAGR of approximately 24% during 2026—2030. This further demonstrates why a single industry growth figure cannot be mechanically used to project market share: if industry supply grows by only 24%, CXMT’s 40%—45% bit growth would be sufficient to gain substantial share; if the industry also reaches 40%, CXMT’s relative advantage would be much smaller.
China’s self-sufficiency gap offers another avenue for growth. Nomura estimates that China accounted for approximately 25% of global DRAM consumption in 2025, while Chinese manufacturers generated only approximately 10% of global production revenue, corresponding to a domestic self-sufficiency rate of approximately 30%. This 30% is not “global market share”; it is the proportion of Chinese demand met by Chinese manufacturers. As long as server, smartphone, and PC customers are willing to increase local memory procurement, CXMT can grow through domestic substitution without capturing every overseas customer.
Domestic substitution does not, however, guarantee unconditional orders. Server memory must be qualified for stability, service life, compatibility, and long-term supply. Customers’ willingness to allocate share to local suppliers does not mean they will accept lower yields or higher total costs. CXMT’s ultimate breakthrough still depends on manufacturing capability, not a policy label.
V. How CXMT Reaches 550,000 Wafers of Capacity: Distinguishing Nominal Capacity from Effective Output
kwpm means thousands of wafers per month. 550 kwpm represents nominal capacity of 550,000 12-inch wafers per month. It is an installed-capacity target for year-end 2028 and does not mean production can run at 550,000 wafers per month from January through December 2028. New production lines require equipment installation, commissioning, qualification, and yield ramp-up, so average annual capacity is inherently lower than the year-end figure.
Nomura estimates that nominal capacity will rise from 280,000 wafers per month in 2025 to 550,000 wafers per month in 2028, representing a CAGR of approximately 25%. The model further assumes utilization close to 100% and yields of approximately 75%—85%, resulting in effective output of approximately 400,000 wafers per month in 2028. The 400,000-wafers-per-month figure, rather than 550,000, is closer to the wafer volume that could actually contribute to that year’s income statement.
The funding requirements for this capacity expansion are not inherently implausible. Nomura estimates that each additional 1 kwpm of DRAM capacity requires approximately US$100 million. Adding 270 kwpm during 2025—2028 would therefore theoretically require approximately US$27 billion. The report projects cumulative capital expenditure of RMB216 billion during 2026—2028, equivalent to approximately US$31.9 billion at the exchange rate used in the report, broadly sufficient to cover the additional capacity and some technology upgrades.
The model nevertheless remains overly smooth. The company’s three fundraising-backed investment projects are expected to undergo phased acceptance only through 2028, while new production lines will bring depreciation, engineering delays, and initially low yields. Nomura nonetheless holds annual capital expenditure constant at RMB72 billion during 2026—2028 while assuming utilization near full capacity and a rapid return to positive cash flow. In reality, the faster the capacity expansion and the newer the technology, the more volatile initial yields and expenses are likely to be.
The correct interpretation of 550,000 wafers is therefore not that “CXMT is certain to double,” but that capacity passes through a three-stage funnel:
Fabs and equipment create nominal capacity;
Nominal capacity is converted into operable wafer capacity;
Operable wafers are converted into qualified DDR5, LPDDR5X, or more advanced products.
Investors should monitor each stage of the funnel rather than focusing only on the final 550,000-wafers-per-month figure.
How Much Qualified Output Can Ultimately Be Retained from 550,000 Wafers?
A purely mechanical example illustrates the difference between nominal capacity and effective output. If nominal capacity of 550,000 wafers per month operates at the company’s 2025 utilization rate of 95.73% and is then multiplied by Nomura’s estimated blended yield of 75%—85%, it corresponds to approximately 395,000—448,000 qualified wafer equivalents per month. Nomura’s modeled effective output of approximately 400,000 wafers per month lies near the lower end of this range.
This example should not be treated as a 2028 forecast because the 2025 utilization rate cannot be directly extrapolated to new fabs, while blended yields will change with DDR5, HBM, and node migration. Its purpose is to reveal the model’s structure: there is not much margin for error between 550,000 and 400,000 wafers per month. If average utilization at the new lines is only 85% and blended yield is only 70%, effective output would fall to approximately 327,000 wafers per month, approximately 18.2% below Nomura’s assumption.
On a full-year basis, 550,000 wafers per month corresponds to nominal annual capacity of 6.6 million wafers, while 400,000 wafers per month corresponds to annualized effective output of 4.8 million wafers. However, 550,000 wafers per month is a year-end 2028 target, and equipment added during the year will not operate at full capacity from January. Actual full-year 2028 output will therefore still depend on the commissioning schedule. Multiplying year-end capacity by 12 months is one of the most common and dangerous sources of overestimation in valuation models.
Process migration can also change the economic meaning of “one wafer.” More advanced nodes can produce more dies and bits per wafer, but may initially reduce yields. If node migration increases bits per wafer by 30% while yield falls from 85% to 70%, the increase in qualified bits may be only approximately 7%. Technology upgrades translate into profit only when node-density gains exceed yield losses; otherwise, they translate into additional depreciation and scrap.
Official disclosures show that the three fundraising-backed investment projects have aggregate investment of RMB34.5 billion, of which RMB29.5 billion is expected to come from proceeds, with phased acceptance anticipated through 2028. This confirms that 2028 is an important engineering milestone. Completion of these projects, however, proves only that asset-construction and R&D; milestones have been met; it does not mean that capacity, yield, customer qualification, and cash collection will all be completed simultaneously. Before the 550,000-wafers-per-month target can be incorporated into valuation, it must be broken down quarterly into equipment move-in, trial production, capacity utilization, yield, and shipments of qualified bits.
VI. Process Nodes, Yields, and Wafer Pricing: The Real Profit Breakthrough
CXMT’s official materials disclose only its first- through fifth-generation process platforms, without specifying the nanometer nodes. The company’s product page shows that its DDR5 products cover 16Gb and 24Gb per die, with speeds of up to 8000Mbps; its LPDDR5X products cover 12Gb and 16Gb, with speeds of up to 10667Mbps. The company also disclosed that its 8533Mbps and 9600Mbps LPDDR5X products entered mass production in May 2025, while its 10667Mbps product began customer sampling.
Nomura believes CXMT’s mainstream process technology lags overseas leaders by approximately 5 years. Each wafer yields only several hundred dies, compared with more than 1,000 for leading manufacturers. This is the most important counterargument in the entire report. Capacity can be increased through capital expenditure, but output per wafer and yields require accumulated process expertise.
To break through, CXMT must accomplish three things simultaneously:
Migrate to more advanced nodes to increase bits per wafer;
Increase the share of high-value products through DDR5 and LPDDR5X;
Improve yields to reduce the cost per qualified chip.
Nomura’s increase in per-wafer pricing from US$14,000—15,000 to US$21,000—25,000 incorporates industry price increases, node migration, and product-mix improvement. If global DRAM prices decline and CXMT cannot offset this with higher-density products, per-wafer pricing will not rise as assumed in the model.
Breaking Revenue Down to the Wafer Level
Nomura’s model can also be reduced to a more straightforward formula: annual revenue approximately equals average monthly effective wafer output multiplied by 12 months and then by revenue per wafer. Based on the report’s charted effective output of approximately 250,000, 330,000, and 400,000 wafers per month, implied annual effective wafer output in 2026—2028 is approximately 3.00 million, 3.96 million, and 4.80 million wafers, respectively. The corresponding revenue per wafer is approximately RMB96,900, RMB141,600, and RMB161,100, while operating cost per wafer is approximately RMB15,800, RMB15,300, and RMB15,300.
This recalculation reveals the true source of the 90.5% gross margin. Nomura does not assume that unit manufacturing costs will continue to decline sharply. Instead, it raises revenue per wafer by approximately 66% from 2026 to 2028 while keeping unit costs at approximately RMB15,000. Node migration increases the number of bits per wafer, while industry price increases raise the price per bit; both must materialize simultaneously. Capacity expansion without an increase in value per wafer would not produce the modeled income statement.
Under the 2028 model, each additional 10kwpm of effective output corresponds to approximately RMB19.3 billion in annual revenue. If effective output is 10% below the base case, or revenue per wafer is 10% lower, revenue would mechanically decline by approximately RMB77.3 billion in either case; if both variables are 10% lower, revenue would be approximately 19% below the base case. Nominal capacity of 550,000 wafers is merely the starting input to the formula; revenue per wafer is the most sensitive multiplier behind the RMB116 price target.
VII. How to Break Through to an 18% Share: Adding More Production Lines Is Not Enough
Nomura expects CXMT’s share of global DRAM bit output to rise from approximately 10% currently to approximately 18% by the end of 2028. This trajectory consists mainly of two components: wafer shipments growing at a 20%—25% CAGR, with node migration lifting bit growth further to 40%—45%.
The logic is sound, but the numbers look closer to an upper-bound scenario. The report’s charts show CXMT’s bit growth at approximately 47%, 50%, and 40% in 2026—2028, versus approximately 29%, 19%, and 20% growth in global DRAM output. Starting from a 10% share, the relative growth rates would broadly raise the share to approximately 17%; starting from the official revenue share of 7.67% or the approximately 8% cited in the report’s appendix would produce a result closer to 13%—14%. Reaching 18% requires a higher starting point, stronger product pricing, or faster industry share migration.
The 18% figure should therefore not be treated as a static target, but broken down into four areas for verification:
Output share. Whether new capacity translates into effective output on schedule;
Product share. Whether DDR5, LPDDR5X, and server memory surpass lower-end mature products;
Customer share. Whether leading cloud, smartphone, PC, and server customers move from trial orders to long-term procurement;
Value share. Whether unit pricing approaches that of overseas leaders, rather than relying on low prices to gain volume.
LPDDR accounted for 66.43% of CXMT’s revenue in 2025, while DDR accounted for 31.87%. The rising DDR share indicates that server DDR5 is becoming a second growth engine. More than 85% of the company’s revenue is still generated through distributors, and its top five customers account for 68.08% of core-business revenue. The distribution model supports rapid product rollout, but also reduces transparency into end-market inventory and actual customer orders.
A share breakthrough therefore cannot be assessed solely through shipment announcements. More robust signals would include a sustained rise in the revenue contribution from server DDR5, higher revenue from direct sales or channels with visibility into end customers, no further deterioration in top-five-customer concentration, broader capacity coverage under long-term agreements, and a gradual narrowing of price discounts.
VIII. HBM and WoW: Worth Assigning Option Value, but Not Yet Profit
Nomura says CXMT has sent HBM3 samples to leading domestic information and communications technology companies, that HBM3E remains under development, and that the company is expected to have HBM3 mass-production capability from 2027. The company’s official materials have not confirmed this specific progress, so these points should be treated only as sell-side research indications.
HBM is not simply “faster memory” than conventional DDR. It requires more complex stacking, bonding, testing, packaging, and customer qualification, while the yield ramp is also more difficult. The approximately 50kwpm of HBM packaging capacity in Shanghai is likewise a Nomura estimate, not company capacity that has already been completed and entered mass production. For valuation purposes, HBM belongs in the upside scenario, not in base-case profit.
Another emerging technology is DRAM-on-logic wafer-on-wafer stacking. Nomura believes CXMT and GigaDevice may cooperate to advance this approach, bringing DRAM closer to logic chips to provide approximately 1TB/s of bandwidth, with potential applications including smart cockpits, high-end smartphones and PCs, and robotics. The report expects progress could accelerate from late 2027.
The value of WoW is that it is not identical to HBM. Edge AI requires higher bandwidth and lower power consumption, but may not require the cost of data-center-grade HBM. If the process and ecosystem mature, CXMT could extend its DRAM capabilities into physical AI devices.
However, this remains a technology option. Before it can be incorporated into the model, investors need to see at least a tape-out, customer qualification, a mass-production timeline, yields, and orders. Without such evidence, 1TB/s is only an architectural target, not revenue.
IX. Nomura’s 2026—2028 Model: Why Profit Appears to Be on an Accelerator
The profit figures are the most eye-catching part of this table, but costs are what investors should really monitor. From 2025 to 2026, revenue grows by 370%, while costs rise by only approximately 30%, causing gross margin to jump from 41.0% to 83.7%. Capacity expansion alone cannot explain this change; the real driver is DRAM pricing.
Earnings per share must also pass through the profit-attribution gate. In 2025, consolidated net profit was RMB7.144 billion, but net profit attributable to shareholders of the listed company was only RMB1.875 billion, representing approximately 26.2% of consolidated net profit. In its 2026 model, Nomura forecasts consolidated net profit of RMB173.753 billion and attributable net profit of RMB130.315 billion, raising the attributable share to 75%, where it remains in 2027—2028.
This means that the approximately 69.5x increase in attributable net profit from 2025 to 2026 is driven not only by the approximately 24.3x increase in consolidated net profit, but also by an increase in the attributable share from 26.2% to 75%. Why the new capacity and profit can be attributed to shareholders of the listed company at this ratio is another gate that must be cleared to support EPS of RMB5.79.
As a purely mechanical sensitivity analysis, if 2028 consolidated net profit still reaches Nomura’s RMB524.093 billion forecast but the attributable share remains at the 2025 level of 26.2%, attributable net profit would be approximately RMB137.55 billion, EPS approximately RMB2.03, and a 20x P/E would imply approximately RMB40.5. This is not an alternative forecast; it simply shows that the scope of consolidation, subsidiary ownership, and non-controlling interests directly affect the price target, and that group-level profit alone is insufficient.
Nomura’s global supply-and-demand table assumes that the average DRAM price rises from US$4.0/GB in 2025 to US$13.7/GB in 2026, an increase of 245% year on year, followed by a further increase to US$18.6/GB in 2027. Global DRAM market revenue rises from US$168.7 billion in 2025 to US$747.4 billion in 2026 and exceeds US$2 trillion in 2030.
This is not a conventional cyclical upturn, but an extreme scenario in which memory becomes globally scarce infrastructure. If prices rise along this curve, an explosion in CXMT’s profit would not be surprising; if prices rise by only half as much, or decline early in 2027, a gross margin above 90% would be difficult to sustain.
The model also assumes cumulative free cash flow of RMB981.116 billion in 2026—2028, cash exceeding RMB1 trillion and net cash of RMB933.488 billion by the end of 2028, alongside zero dividends and zero M&A; over the 3 years and fixed annual capital expenditure of RMB72.0 billion. It retains nearly all the high profits on the balance sheet. This cash accumulation demonstrates upside potential, but also serves as a reminder that any additional capacity expansion, M&A;, dividends, or price decline would change the terminal value.
The real foundation of the RMB116 price target is not 550,000 wafers, but the following system of linked assumptions:
Nominal capacity of 550,000 wafers × approximately 400,000 wafers of effective output × more advanced nodes × higher per-wafer pricing × 90.5% gross margin × 20x P/E.
Any deviation in any component requires the price target to be recalculated.
Depreciation, Inventory, and Free Cash Flow Form Another Income Statement
At the end of 2025, CXMT had RMB183.024 billion of fixed assets, RMB24.68 billion of depreciation for the year, and RMB29.39 billion of inventory. For a manufacturer, capacity expansion first enters the balance sheet and then flows through the income statement via depreciation; unsold products enter inventory and may subsequently require impairment charges if prices decline. A wafer fab does not stop depreciating when DRAM prices fall.
Nomura’s model forecasts annual capital expenditure of RMB72.0 billion in each year from 2026 to 2028, while free cash flow reaches RMB147.446 billion, RMB333.351 billion, and RMB500.319 billion, respectively. This means high selling prices and high margins not only cover incremental capital expenditure but also allow cash to accumulate rapidly. By the end of 2028, the model projects more than RMB1 trillion of cash and RMB933.488 billion of net cash, which in turn raises equity value.
At least 3 cash-flow variables need to be checked quarterly. First, whether operating cash flow moves in line with net profit, or whether profit is tied up in inventory and receivables; second, whether RMB72.0 billion of capital expenditure is sufficient to cover capacity of 550,000 wafers, node migration, and HBM/WoW investment; and third, whether depreciation on new assets has been fully reflected in costs. If actual capacity-expansion spending is higher, or high-priced inventory accumulates, free cash flow will expose the problem before net profit does.
Gross proceeds from the base IPO issuance were RMB57.919 billion, with net proceeds of RMB57.638 billion; if the over-allotment option is exercised in full, gross proceeds could reach up to RMB66.607 billion and net proceeds up to RMB66.31 billion. This is a substantial amount, but it still provides only part of Nomura’s projected RMB216.0 billion of aggregate capital expenditure in 2026—2028. In the long run, capacity expansion must ultimately be funded through internally generated operating cash flow rather than relying solely on a one-off public listing.
As of the end of 2025, the company still had accumulated uncovered losses of RMB36.65 billion, constraining its near-term dividend capacity. Nomura’s model also assumes zero dividends and zero M&A; from 2026 through 2028, with nearly all profits retained by the company. This treatment supports cash accumulation, but it is not the only possible capital-allocation path. If the company accelerates fab construction, acquires technology, or begins paying dividends, net cash and the target valuation would need to be adjusted.
How Dependent Is RMB116 on the “20x” Multiple?
Breaking the valuation into a matrix is more useful than focusing solely on RMB116. The table below is not a new price-target forecast; it presents only a mechanical sensitivity analysis using 2 variables: 2028 EPS of Nomura’s RMB5.79 base case, approximately RMB3.88 under the delayed Shanghai 200kwpm scenario, and approximately RMB2.9 if base-case profit is halved; the valuation multiples are 10x, 15x, and 20x.
This table shows that RMB116 requires both earnings and the valuation multiple to be at the upper end. Even if Nomura’s base-case profit is fully realized, a market valuation of only 15x P/E would mechanically imply approximately RMB86.9; even if the market still assigns a 20x P/E, a delay in Shanghai capacity would reduce the mechanical valuation to approximately RMB77.6. If both profit and the valuation multiple revert toward the middle of their ranges, the valuation contraction would be multiplicative rather than additive.
Nomura’s rationale for using 20x is CXMT’s rising market share and the Chinese market’s valuation premium relative to Micron. That premium is justified only if the quality of share gains is sufficiently high: share captured through advanced products, stable customers, and cost improvements can command a premium, whereas share gained through industry shortages and low pricing more closely resembles cyclical profit and should not receive the same multiple.
RMB116 is therefore not merely an earnings forecast, but also a forecast of market sentiment. If 2028 happens to coincide with a peak in DRAM pricing, investors may lower the P/E due to concerns about the next round of supply additions; a 20x multiple would be easier to sustain only if the industry enters a period of stable structural growth. Whether the profit peak and valuation peak can occur at the same time is the second layer of risk most easily overlooked in the price target.
10. Sanctions Scenario: Nomura’s So-Called “Worst Case” Actually Only Removes Shanghai Capacity
Nomura identifies the MATCH Act, Entity List designation, or similar restrictions as the principal risk. The report estimates that approximately 40% of CXMT’s current equipment is domestically sourced, while advanced photoresists still rely primarily on Japanese suppliers. Restrictions on overseas equipment, spare parts, services, or high-end materials would affect migration to 14–15nm and more advanced nodes.
The report’s stress test does only one thing: it delays the two Shanghai phases, each with 100kwpm of advanced DRAM capacity, or 200kwpm in total. The results are as follows:
The absolute figures in the delayed scenario are mechanically derived by applying the declines disclosed by Nomura to its base-case forecasts; they are not a new price target from Nomura. If the 20× valuation multiple remains unchanged, 2028 earnings per share would decline from RMB 5.79 to approximately RMB 3.88. In practice, however, sanctions would likely compress not only earnings but also the valuation multiple.
More importantly, this scenario still preserves most of the assumptions regarding industry price increases, node migration, and gross margin. It does not fully model restrictions on servicing the installed equipment base, photoresist supply disruptions, yield deterioration, customers shifting orders, or wasted capital expenditure. A more accurate label would therefore be the “Shanghai 200kwpm delay scenario,” rather than a genuine worst-case scenario.
For CXMT to overcome this barrier, equipment localization cannot be assessed solely by equipment count. It must also consider:
Whether domestically produced equipment can maintain yield and stability after entering production lines;
Whether etching, deposition, cleaning, inspection, and other processes can provide an integrated suite of capabilities;
Whether high-end photoresists, silicon wafers, gases, CMP materials, and critical components have secondary sources;
Whether existing production lines can remain operational if overseas equipment spare parts and services are interrupted;
Whether the introduction of domestically produced equipment increases depreciation or lengthens qualification cycles.
An increase in equipment localization from 40% to 60% would not constitute a genuine breakthrough if yields declined materially. The real standards are output, cost, and maintainability.
A Genuine Stress Test Must Examine How Four Risks Amplify One Another
The first is pricing risk. A decline in DRAM prices would directly reduce revenue per wafer while exposing high-cost inventory to impairment. Falling prices typically also change customer purchasing patterns: distributors may first destock, putting simultaneous pressure on the company’s shipment volumes and prices.
The second is manufacturing risk. If node migration and new-fab ramp-up fall short of expectations, bits per wafer, yield, and effective output would all decline. Because depreciation and R&D; expenses have already been incurred, every RMB 1 reduction in revenue could have a profit impact materially greater than RMB 1.
The third is supply restrictions. Constraints on overseas equipment, critical components, photoresists, or services would not only delay incremental Shanghai capacity but could also affect maintenance of existing production lines and validation of advanced nodes. Nomura’s Shanghai-delay scenario covers only incremental capacity volumes and does not fully capture efficiency at the installed base.
The fourth is customer and product risk. More than 85% of the company’s 2025 revenue came through distributors, while its five largest customers accounted for 68.08% of core-business revenue. If server DDR5 customer qualification proceeds more slowly than expected, capacity may be redirected toward lower-priced products. If channel inventory is difficult to see through, reported shipments may again run ahead of genuine end demand.
These four risks are clearly interconnected. Tighter restrictions would slow node migration; slower node migration would reduce bit output and the share of high-end products; insufficient high-end products would increase dependence on industry spot prices; and weaker cash flow following a price decline would constrain the next round of capital expenditure. Conversely, if yield and customer qualification exceed expectations, an improved product mix could offset part of the pressure even if industry pricing does not reach Nomura’s extreme assumptions.
A genuine downside scenario should therefore not simply remove a capacity figure. A more reasonable sequence is to first reduce effective output, then lower price per wafer, incorporate depreciation and inventory pressure, and finally reduce the valuation multiple. RMB 116 would contract rapidly if all four layers came under simultaneous pressure, reflecting the nonlinear nature of the “high earnings × high valuation” model.
11. The Domestic Supply Chain Will Benefit, but CXMT’s Capacity Expansion Should Not Be Equated with Linear Order Growth
The domestic partners identified by Nomura span etching, thin-film deposition, bonding, inspection, cleaning, wafers, chemicals, gases, CMP, and assembly and testing. AMEC, NAURA Technology, Leadmicro Nano, Piotech, Precision Measurement Electronics, Skyverse Technology, ACM Research Shanghai, National Silicon Industry Group, Anji Microelectronics, Dinglong, JCET, Tongfu Microelectronics, and others may all benefit from CXMT’s capacity expansion and higher localization rate.
However, the order cycles of equipment and materials companies differ from CXMT’s earnings cycle. CXMT first places equipment orders, allowing equipment suppliers to recognize revenue. Once the production line is installed, it enters validation, after which CXMT ramps yields. Stable wafer output only follows once yields improve. The products must then still undergo customer qualification and sales. The supply chain cannot directly equate “CXMT’s 550,000-wafer target” with proportionate revenue growth for every supplier.
CXMT will also balance localization against profitability. Nomura explicitly notes that the company will consider the impact of domestically produced equipment on yield and profit margins, rather than mechanically replacing equipment simply to increase the localization ratio. Better-performing domestic equipment is more likely to enter advanced-node production, while processes where performance remains insufficient may initially penetrate mature production steps.
Supply-chain investors should focus on two types of companies: those already supplying CXMT’s mass-production lines with products that are consumed consistently as wafer volumes increase, and those addressing critical processes subject to the strongest overseas restrictions and presenting the greatest substitution challenges. Companies that may only be “providing samples” but lack mass-production validation should not prematurely receive full capacity-expansion valuations.
12. How CXMT Can Break Through: Three Scorecards Matter More Than the Price Target
The question of “how to break through” in the title ultimately comes down to three scorecards.
1. Manufacturing Scorecard: From Installed Capacity to Qualified Bits
The most important metrics are not monthly capacity but effective wafer output, yield, and bits per wafer. If CXMT can steadily improve DDR5 yield from Nomura’s estimate of approximately 80% while completing node migration, it can increase bit output without a proportionate increase in fab space. Conversely, the greater the nameplate capacity, the heavier the depreciation and scrap burden arising from low yields.
2. Product Scorecard: From Domestically Available Products to High-Value Products
DDR5 and LPDDR5X are already commercial products, while server DDR5 is central to improving the profit mix. HBM3, HBM3E, and WoW represent the next phase. Product breakthroughs should be assessed through revenue contribution, customer qualification, and pricing—not launch-event specifications. A rising server-revenue mix, narrower unit-price discounts, and customers progressing from sample evaluation to mass production are what will convert the technology narrative into earnings.
3. Supply-Chain Scorecard: From Substitution Capability to Stable Maintenance
Equipment localization, secondary material sources, and spare-parts and service capabilities jointly determine production-line resilience. Replacing only new equipment without the ability to maintain the installed equipment base would still leave operations vulnerable to restrictions. Similarly, addressing only mature nodes without supporting migration to 14–15nm would not deliver Nomura’s projected price per wafer.
In each subsequent quarter, the following checklist will be most important to monitor:
The Shortest Validation Chain for Determining Whether RMB 116 Remains Valid
Despite the many variables, quarterly monitoring does not require rebuilding the entire model every time. The shortest path can be standardized into five steps. Step 1 is to assess revenue, gross margin, and operating cash flow to determine whether pricing gains have genuinely converted into cash. Step 2 is to examine the DDR5, LPDDR5X, and server product mix to determine whether earnings come from industry price increases or product upgrades. Step 3 is to assess capacity utilization, transfers from construction in progress to fixed assets, and depreciation to determine the conversion efficiency of new lines from investment to output. Step 4 is to use only global market-share data from the same institution and based on the same denominator, avoiding the conflation of revenue share, bit share, and China’s domestic self-sufficiency rate. Step 5 is to place the latest earnings per share into a 10×, 15×, and 20× valuation matrix.
These five steps follow a clear sequence. If revenue grows but cash flow deteriorates, inventory and receivables should be investigated first. If capacity increases but gross margin declines, yield and product mix should be examined first. If market share rises but unit-price discounts widen, volume growth may not be creating equivalent value. If earnings materialize but the P/E multiple declines, the market has begun pricing in the next round of supply additions. Only when volume, price, cash flow, and the valuation multiple improve simultaneously will RMB 116 increasingly reflect operating reality.
Source hierarchy must also be maintained. Absolute monthly capacity, specific nanometer nodes, HBM sampling, and an 18% bit share currently come primarily from Nomura’s forecasts. They should not be elevated to company-confirmed facts merely because multiple secondary articles have repeated them. Company announcements can confirm the listing, financing, financials, product mass-production status, capacity utilization, and revenue-based market share calculated on a consistent basis. Each new item of official incremental disclosure should replace the corresponding sell-side assumption; items without new disclosure should remain scenarios rather than being filled in with rumors.
If any 2 of the following occur for 2 consecutive quarters—price declines, no yield improvement, rising inventory, or delayed server-product qualification—the earnings scenario should be proactively revised downward. Only if high utilization, operating cash flow covering capital expenditure, a rising share of advanced products, and consistent-basis market-share gains are all observed for 2 consecutive quarters would there be grounds to increase the weighting of the upside scenario. With this tracking framework, the price target becomes a function of operating results rather than a slogan that fluctuates daily.
XIII. Conclusion: RMB 116 Is the Score After Completing Every Difficult Maneuver, Not the Starting Line
There is little dispute over CXMT’s strategic value. It is China’s largest and the world’s fourth-largest DRAM manufacturer, and has progressed from “whether it can produce DRAM” to “whether it can sustain profitability through global cycles.” DDR5 volume ramp-up, China’s self-sufficiency gap, and memory demand from AI inference all provide it with a rare window of opportunity.
The significance of Nomura’s RMB 116 price target is that it extrapolates this window to the most optimistic end-state: the number of agentic tasks increases substantially, DRAM supply remains structurally tight, and prices stay elevated; CXMT’s nameplate capacity reaches 550,000 wafers/month, with effective output of approximately 400,000 wafers/month; process nodes, yields, and product mix improve simultaneously; its global bit share approaches 18%; its advanced Shanghai capacity remains unrestricted; net profit attributable to shareholders reaches RMB 393.07 billion in 2028; and the market assigns a 20x P/E multiple.
These conditions are not independent of one another. A failed process migration would simultaneously pressure yields, revenue per wafer, market share, and gross margin; a decline in DRAM prices would simultaneously pressure revenue, profit, and valuation multiples; and an escalation in sanctions would affect both capacity expansion and maintenance of existing capacity. RMB 116 may look like a single number, but it is underpinned by a series of tests that must all be passed simultaneously.
Therefore, what matters most for CXMT is not whether the price target can be reached in one step, but whether uncertainty can be reduced each quarter: whether 350,000 wafers/month is delivered, whether DDR5 yields improve, whether server revenue increases, whether pricing discounts narrow, whether the use of domestic equipment and materials expands without impairing yields, and whether HBM and WoW progress from samples to orders.
If these indicators are delivered consistently, 550,000 wafers/month and an 18% share will shift from a sell-side upside scenario to a trackable operating path. If profits derive only from a short-term price surge while technology, yields, and customer adoption fail to improve in tandem, RMB 116 will merely be elegant arithmetic that stacks the cyclical peak, the upper bound of market share, and a valuation premium.