Nomura has launched coverage on ChangXin Memory Technologies, or CXMT, with a Buy rating and a target price of RMB 116. The report’s front page also listed the IPO price at RMB 8.66 and implied upside of 1,239.5%. According to the ChainCatcher article, the note is the highest target price the author has seen so far from a foreign institution on CXMT, and one of the most aggressive valuation calls now in the market.
How Nomura arrived at RMB 116
The target price comes from applying a 20x multiple to Nomura’s projected 2028 earnings per share of about RMB 5.8.
The key variable is that 20x earnings anchor. Nomura builds it in two steps. First, it uses Micron as the closest comparable. In Nomura’s view, Micron is CXMT’s nearest peer, and Micron’s two-year forward price-to-earnings ratio over the past five years has generally traded in a 5x to 15x band, with a midpoint around 10x.
Second, Nomura adds what it sees as a China market valuation premium. The comparison it uses is ACM Research Shanghai and its U.S.-listed parent, ACM Research. For the same business, their A-share and U.S. market valuation relationship has, in Nomura’s telling, remained in a 1x to 3x range over time. On that basis, Nomura assumes CXMT could trade at 1x to 3x Micron’s valuation, which gives a 10x to 30x range and a midpoint of 20x.
In other words, the RMB 116 target is effectively the product of Micron’s historical valuation center, a China market premium factor, and Nomura’s own 2028 profit forecast.
Revenue and profit forecasts point to an aggressive earnings ramp
Nomura’s operating forecasts are unusually strong.
- Revenue: RMB 61.799 billion in 2025 actuals, RMB 290.666 billion in 2026 forecast, RMB 560.788 billion in 2027, and RMB 773.323 billion in 2028.
- Net profit attributable to the parent: RMB 1.875 billion in 2025 actuals, RMB 130.315 billion in 2026 forecast, RMB 277.248 billion in 2027, and RMB 393.070 billion in 2028.
From those figures, Nomura derives a 2026-2028 revenue CAGR of 63% and a net profit attributable to shareholders CAGR of 74%. Diluted earnings per share rise from RMB 2.05 in 2026 to RMB 5.79 in 2028.
Using the IPO price of RMB 8.66, the implied price-to-earnings ratios would be 4.2x for 2026, 2.1x for 2027, and 1.5x for 2028. The report also gives a 2028 price-to-book ratio of 0.6x. As the article puts it, if Nomura’s model holds, an investor buying at the IPO price would, three years later, be paying less than the company’s book value per share. Return on equity reaches 84% in 2026 in the model. By the end of 2028, cash is projected at RMB 1,085.5 billion and net cash at RMB 933.5 billion.
The article flags one detail that is easy to miss. Nomura projects 2026 net profit after tax at RMB 173.753 billion, but minority shareholders would receive RMB 43.438 billion of that amount, leaving only RMB 130.315 billion attributable to the parent. In Nomura’s model, minority interest stays near 25% of the capital structure. That means reading the earnings line without focusing on profit attributable to shareholders would overstate profit by roughly one quarter.
The boldest assumption sits in gross margin
The article argues that the most aggressive part of Nomura’s model is not simply the growth rate in revenue or earnings, but the gross margin path. Nomura puts CXMT’s gross margin at 41.0% in 2025, then 83.7% in 2026, 89.2% in 2027, and 90.5% in 2028.
The reason is the cost line barely moves in comparison. Operating cost is listed at RMB 36.465 billion in 2025, RMB 47.451 billion in 2026, and RMB 73.560 billion in 2028. Over the same period, revenue climbs from RMB 61.799 billion to RMB 773.323 billion. Revenue rises by 12.5x over three years, while operating cost only doubles.
That implies Nomura expects most of CXMT’s incremental revenue in the coming years to come from higher pricing rather than proportionally higher spending, with much of that price increase flowing almost directly to profit. The article notes that memory upcycles have, at times, shown much higher price elasticity than cost elasticity. Even so, extending that condition all the way through 2028 and taking gross margin to around 90% effectively treats the best stretch of the cycle as if it were normal.
Demand side: memory consumption could rise more than sevenfold in five years
The foundation for those forecasts is Nomura’s view of future memory demand, centered on agentic AI.
The report breaks a single agentic task into eight stages: user request arrival, model weight loading, prefilling, reasoning and planning, tool calls, context integration, multi-step iteration, and response generation.
Nomura says the crucial pressure points are the sixth and seventh steps. Each time an external tool is called, the returned result has to be folded back into the context window, and the model must process that now-expanded context again. The KV cache grows with it. Because an agent repeatedly runs a loop of reasoning, tool use, and integration, the context keeps getting longer. Nomura therefore sees peak memory pressure appearing in the seventh stage, multi-step iteration, where memory demand could approach the capacity limit of HBM.
From there, Nomura lays out a multiplicative framework: memory demand equals number of users times usage duration times task complexity times inference token consumption times agent penetration. Because the model is multiplicative, even moderate growth in each factor can produce a sharp expansion in total demand. Nomura estimates that by 2030, the number of concurrent agentic tasks run globally each year will be 50 times the 2026 level.
On the other side of the ledger, the report also looks at efficiency gains from memory-saving techniques, including KV quantization, grouped-query attention, or GQA, PagedAttention, prefix caching, and multi-head latent attention, or MLA. In theory, Nomura says, these tools can stack and compress memory usage by 4x to 40x. In practice, the report says actual savings are unlikely to exceed 5x. The reason given is that the single biggest source of compression comes from quantizing weights and KV states, such as moving from FP32 or FP16 down to INT8 or INT4, and that route has an accuracy floor. The other methods are additions on top rather than full substitutes.
After netting the two sides out, Nomura’s conclusion is that global memory usage will still expand by more than 7x from 2026 to 2030, with a CAGR above 60%, even after applying a 4x efficiency discount. The estimate also assumes zero growth in non-AI applications.
Nomura adds another data point from outside semiconductors. Cloudflare’s CEO has publicly said that, for the first time in internet history, bot traffic has overtaken human traffic. The report then raises a qualitative question: if AI no longer needs humans to issue every task one by one and can keep running work continuously on its own, where is the ceiling for demand? Nomura’s answer is that only three constraints remain: whether people are willing to grant that authority, especially given security concerns; whether infrastructure is sufficient, including power, chips, data centers, and, most importantly, talent; and the capex limits of cloud providers and enterprises.
Supply side: the industry still cannot catch up
Nomura estimates that global memory suppliers can grow bit output at only 30% to 40% on a compound basis, short of the 60%-plus growth it sees on the demand side. In that setup, the supply-demand gap persists for a long time.
The report says this gap cannot be closed simply by spending more money. It identifies four bottlenecks: cleanrooms, equipment, materials, and people. Among them, experienced engineers are the hardest to scale quickly.
Nomura also discusses mitigation routes. One is to offload to NAND, which offers capacity more than 100 times that of DRAM but at lower speed, while another is to push high-bandwidth flash, or HBF. Still, Nomura says HBF is not yet ready for mass production, and replacing DRAM with NAND at large scale would tighten NAND supply as well. Its conclusion is blunt: these measures may slow the trend, but they do not reverse it.
The report also pushes back on a common market concern that aggressive capacity expansion by producers could flip the cycle. Over the past 15 years, the DRAM industry’s capital intensity has mostly stayed in a 25% to 45% range. But because AI demand has rapidly expanded the market since September 2025, that ratio could fall to only 15% in 2026. Judged against history, that means current investment is not high relative to market size, and the market can absorb capex well above current levels.
Nomura adds, however, that as memory prices have risen sharply since September 2025, both producers and customers are gradually recognizing that price stability is healthier. Long-term agreements, or LTAs, are already being signed. From 2028 onward, Nomura expects storage makers to shift from pushing price higher toward more measured capacity additions.
CXMT itself: capacity, market share, yields and product roadmap
On company-specific fundamentals, Nomura lays out a clear capacity path for CXMT.
- By the end of 2025, Hefei and Beijing together are projected to reach about 280 kwpm, or thousand wafers per month.
- By the end of 2026, that rises to 350 kwpm.
- Another 100 kwpm is added in each of 2027 and 2028.
- By the end of 2028, total capacity reaches 550 kwpm.
Separately, CXMT is building HBM packaging capacity in Shanghai with a potential scale of 50 kwpm. That capacity is for packaging rather than DRAM fabrication.
On spending, Nomura uses a rule of thumb that each additional 1 kwpm of DRAM capacity requires about $100 million. CXMT raised RMB 57.9 billion in its July 27 STAR Market IPO, and the total could rise to RMB 66.6 billion if the greenshoe is exercised in full.
Assuming full utilization, Nomura estimates CXMT’s wafer shipment CAGR at 20% to 25% from 2026 to 2030. Add a process-node transition every two years, and bit output CAGR reaches 40% to 45%.
That compares with an industry bit growth rate of 30% to 40% and demand growth of more than 60%. So Nomura’s view is that CXMT can outgrow peers, continue taking share from global competitors, and still operate in a market that remains undersupplied overall.
The report’s market share target has CXMT’s global DRAM share rising from about 10% now to around 18% by the end of 2028. As a reference point, Nomura notes that Micron’s share is above 20%, and that its market capitalization stood at $960 billion as of July 20, 2026.
Process technology, HBM progress and channel pricing checks
On technology, Nomura estimates CXMT’s mainstream node in 2026 will be in the 1x-1y range, roughly 16 to 17 nanometers. DDR5 yield is put at about 80%, while DDR4 yield is estimated above 90%. Starting in 2027, CXMT is expected to move into the 1z-1a range, roughly 10 to 15 nanometers, and gain mass-production capability for HBM3. The report says the process migration is achieved with DUV rather than EUV.
For HBM specifically, Nomura says CXMT has already sent HBM3 samples to a leading domestic ICT company and others, though certification will take time. HBM3e remains under development.
The pricing section is one of the most detailed parts of the article. Based on channel checks with leading Chinese smartphone makers and PC/server vendors, Nomura says CXMT prices its products below overseas leaders, but not by much, at about 0% to 20% lower. The main reason cited is policy support for local procurement.
There is still a real cost gap, though. Nomura estimates that because CXMT’s mainstream process lags global leaders by about five years, it gets only several hundred good dies per wafer, versus more than 1,000 for overseas leaders. As a result, CXMT’s wafer average selling price in 2026 is estimated at about $14,000 to $15,000. As process technology moves from 16-17 nm to 14-15 nm in 2027 and 2028 and bits per wafer improve, that figure is projected to rise to $21,000 to $25,000.
Domestic substitution and customer mix
The report also cites a broader market statistic. According to WSTS, China accounts for roughly 25% of the global DRAM market in 2025. But Nomura estimates domestic manufacturers, including CXMT and others, represent only about 10% of global output value. That implies domestic DRAM self-sufficiency of only around 30%, which Nomura treats as direct evidence that CXMT still has substantial room to gain share in its home market.
On customers, more than 85% of CXMT’s revenue is generated through distributors. The top-five customer concentration ratio was 68.08% in 2025. End customers disclosed in the prospectus materials include Alibaba Cloud, ByteDance, Tencent, Lenovo, Xiaomi, Transsion, Honor, OPPO and vivo.
By product mix, LPDDR accounted for 66.43% of 2025 revenue, while DDR accounted for 31.87%. In 2023, DDR contributed 20.16%. The article says the rise is mainly tied to the ramp in DDR5 server products.
Nomura also notes that CXMT and GigaDevice are working together on DRAM-on-logic wafer-on-wafer, or WoW, stacking aimed at edge AI applications such as intelligent automotive cockpits, high-end smartphones, PCs and robots. In Nomura’s view, that technology could begin to show meaningful momentum from the end of 2027.
The article’s conclusion: better seen as an upside case than a steady-state baseline
The author’s closing judgment is that the 90%-level gross margin assumption is plainly not sustainable. Extending a cyclical peak into a long-term norm does not fit the way the memory industry operates.
On that reading, the report’s earnings assumptions and target price should not be treated as a long-term midpoint. They are better understood as an upper-bound scenario under conditions where supply is extremely tight and both pricing and profitability sit in a highly optimistic state.
The original piece ends with a short line: “Logic and technology, we are still on the way.”