Author: Dreams of the Fourth Dimension

Today,Nomura released its first coverage report on Changxin Memory, rating it as a buy, with a target price of 116 yuan. The front page of the report also noted an issuance price of 8.66 yuan, implying an upside potential of +1239.5%. This is also, to date, the highest target price given to Changxin Memory by foreign institutions that I have seen, and can be deemed the most aggressive valuation report in the current market.

1. How is the 116 yuan calculated?

Valuation of 20 times the forecasted earnings per share of 5.8 yuan for 2028 results in a target price of 116 yuan.

The key is the 20 times anchor. Nomura's derivation follows two steps. The first step is to reference Micron. Nomura believes Micron is the closest competitor to Changxin, and over the past five years, Micron's two-year forward P/E ratio has fluctuated roughly between 5 and 15 times, with a median of 10 times. The second step adds a valuation premium for the Chinese market. Nomura used a comparative sample: ACM Research in Shanghai compared to its U.S. parent company ACM Research, where the P/E ratios of the same business in A-shares and U.S. stocks have long maintained a relationship of 1 to 3 times. Based on this, Nomura assumes that Changxin will trade at a Micron valuation of 1 to 3 times, corresponding to a range of 10 to 30 times, taking the median of 20 times.

Thus, the figure of 116 yuan is essentially Micron's historical valuation center × the premium coefficient for the Chinese market × Nomura's own profit forecast for 2028.

2. Nomura's profit forecast is aggressively optimistic

Operating revenue: 617.99 billion yuan in 2025, 2906.66 billion yuan forecast for 2026, 5607.88 billion yuan for 2027, and 7733.23 billion yuan for 2028.

Net profit attributable to the parent company: 18.75 billion yuan in 2025, 1303.15 billion yuan forecast for 2026, 2772.48 billion yuan for 2027, 3930.70 billion yuan for 2028.

Nomura therefore gives a compound annual growth rate of 63% for revenue and 74% for net profit from 2026 to 2028. The diluted earnings per share rise from 2.05 yuan in 2026 to 5.79 yuan in 2028.

At an issuance price of 8.66 yuan, the corresponding P/E ratio would be 4.2 times in 2026, 2.1 times in 2027, and 1.5 times in 2028. The price-to-book ratio for 2028 is 0.6 times. This means that if Nomura's model holds, those who buy at the issuance price would have a book net asset that is more than double what they paid in three years. The return on equity would be 84% in 2026. By the end of 2028, cash on hand would be 10855 billion yuan, with net cash at 9335 billion yuan.

A detail that must be noted here is easily overlooked. Nomura predicts a post-tax net profit of 1737.53 billion yuan for 2026, but minority shareholders will take away 434.38 billion yuan, leaving only 1303.15 billion yuan for the parent. Changxin's equity structure contains a large amount of minority shareholder rights, which stabilize around 25% in Nomura's model. When considering Changxin's profits, it is important to look at the line attributable to the parent, or else one may overestimate by a quarter.

Then there is a point I want to remind: the most aggressive assumption in Nomura's model hides in the gross margin. For 2025, Changxin's gross margin is 41.0%, while Nomura forecasts 83.7% for 2026, 89.2% for 2027, and 90.5% for 2028.

How can it reach 90%? Because in this table, operating costs remain nearly unchanged. Operating costs are 364.65 billion yuan in 2025, predicted to be 474.51 billion yuan in 2026, and 735.60 billion yuan in 2028. During the same period, revenue rises from 617.99 billion yuan to 7733.23 billion yuan. In three years, revenue grows 12.5 times while operating costs only double.

In other words, Nomura assumes that the increased revenue for Changxin in these years comes largely from price increases rather than spending more, and that the increased prices almost entirely flow into profit. This is not without precedent in the upcycle of the memory industry; price elasticity indeed far exceeds cost elasticity, but projecting this continuously to 2028 and extending it to a 90% gross margin equates to treating the best segment of the cycle as the norm.

3. Demand Side: Nomura says memory usage will increase more than sevenfold in the next five years.

Supporting the above chart is Nomura's judgment on memory demand. Nomura's starting point is agentic AI. The report breaks down a single intelligent agent task into eight phases: user request arrives, model weights load, pre-fill, inference planning, tool invocation, context integration, multi-step iteration, and generating replies.

The key lies in the sixth and seventh steps. Every time an external tool is invoked, the returned results must be reassembled into the context, and the model has to process the lengthening context again, resulting in an increase in KV cache. The agent will repeatedly run the “inference - invoke tools - integrate” loop, with each loop extending the context further. Nomura judges that the peak memory pressure occurs at the seventh step, multi-step iteration, at which point it may approach the capacity limit of HBM.

Thus, Nomura proposed a multiplicative structure: memory demand equals the product of user numbers × usage duration × task complexity × inference token consumption × AI penetration rate. The multiplicative nature means that moderate growth in any single factor will lead to exponential increases in the total amount. Nomura estimates that by 2030, the number of concurrently operable intelligent agent tasks globally will increase by 50 times compared to 2026.

However, there are undeniable improvements in technological efficiency. Nomura specifically evaluated memory efficiency technologies, including KV quantization, grouped query attention (GQA), PagedAttention, prefix caching, and multi-head potential attention (MLA). These technologies can be stacked, theoretically compressing memory usage by 4 to 40 times. However, Nomura believes that the actual savings are unlikely to exceed 5 times. This is because the largest single source of compression effects comes from the quantization of weights and KV, which reduces FP32/FP16 to INT8/INT4, and this aspect has a precision lower limit; the remaining methods are supplementary to it.

After taking both sides into account, Nomura concludes that even with a 4 times efficiency discount applied, from 2026 to 2030, global memory usage will still increase more than sevenfold, with an annual compound growth rate exceeding 60%, while this still assumes zero growth for non-AI applications.

Nomura also cited a side evidence. Cloudflare's CEO openly stated that for the first time in internet history, bot traffic has exceeded human traffic. Nomura then raised a qualitative question: if AI no longer requires humans to give instructions one by one and can carry out tasks continuously on its own, what is the ceiling for demand? Nomura's answer is that at that point, only three constraints will remain: whether humans are willing to grant authorization (safety concerns), if the infrastructure is sufficient (power, chips, data centers, and most importantly, talent), and the capital expenditure limits of cloud providers and enterprises.

4. Supply Side: Trying to catch up is futile

What is the other side of the demand's compound growth rate of over 60%? Nomura estimates that the compound growth rate of bit output from global memory manufacturers will only be 30% to 40%, and the supply-demand gap will persist for a long time.

More troubling is that this gap cannot simply be filled by spending more. Nomura pointed out that the bottlenecks are clean rooms, equipment, materials, and personnel; semiconductor capacity expansion is hitting these walls, and among them, the hardest to develop quickly is skilled engineers.

Nomura also discussed mitigation measures, using NAND for offloading, which has a capacity over 100 times that of DRAM, but the drawback is slowness, pushing for high bandwidth flash memory (HBF). However, Nomura also stated that HBF is not yet prepared for mass production, and large-scale use of NAND as a substitute will only tighten NAND supplies too, offering limited relief for DRAM tightness. Nomura's original words were, these solutions only slow the pace of the trend, and do not reverse the trend.

In response to the market's greatest concern about manufacturers’ excessive capacity expansion leading to a cycle reversal, Nomura offered a counterpoint. Over the past 15 years, the capital intensity of the DRAM industry has generally been in the range of 25% to 45%, but due to the surge in market scale driven by AI demand since September 2025, the ratio might drop to only 15% in 2026. In other words, based on historical standards, current industry investments relative to market size are on the low side, and the market can absorb capital expenditures that are far above current levels.

However, Nomura also mentioned that as storage prices have soared since September 2025, manufacturers and customers will gradually recognize that stable pricing is healthier, and long-term supply agreements (LTAs) are being signed consecutively. Starting in 2028, Nomura believes that storage manufacturers will shift their focus from pushing prices to reasonable capacity expansion.

5. Changxin itself: capacity, share, yield, unit price

Returning to the company itself. Nomura provides this capacity pathway: by the end of 2025, Hefei and Beijing together will have about 280 kwpm (thousand wafers per month). By the end of 2026, this will increase to 350 kwpm. In 2027 and 2028, an additional 100 kwpm each will be added, reaching 550 kwpm by the end of 2028. Additionally, HBM packaging capacity is being built in Shanghai, with a potential scale of 50 kwpm, which is packaging rather than DRAM manufacturing.

How much will the capacity expansion cost? Nomura estimates that each 1 kwpm of DRAM capacity requires about 100 million US dollars. On July 27, Changxin raised 57.9 billion yuan in its IPO on the Sci-Tech Innovation Board, which, if the overallotment option is fully exercised, could reach up to 66.6 billion yuan.

Assuming full production capacity, Nomura estimates that Changxin's wafer shipment volume will have a compound annual growth rate of 20% to 25% from 2026 to 2030. Coupled with a process node advancement every two years, the bit output compound growth rate will reach 40% to 45%.

The industry's bit growth rate is 30% to 40%, Changxin's is 40% to 45%, and the demand growth rate is over 60%. Changxin is growing faster than its peers but still cannot catch up to demand. Therefore, Nomura's conclusion is that Changxin will continue to capture market share from global peers, while the whole industry remains in shortage.

The target for market share is to increase Changxin's global DRAM share from the current roughly 10% to about 18% by the end of 2028. As a reference, Nomura mentions that Micron's share is above 20%, with a market value of 960 billion dollars as of July 20, 2026.

Technically, Nomura estimates that Changxin's mainstream process in 2026 is at the 1x-1y node (around 16 to 17 nanometers), with a DDR5 yield of about 80%, and DDR4 yield over 90%; starting in 2027, it will advance to the 1z-1a node (around 10 to 15 nanometers), with capabilities for mass production of HBM3. The entire process does not use EUV and relies on DUV for advancement.

Regarding HBM, Nomura mentioned that Changxin has sent HBM3 samples to a leading domestic ICT company and others, but certification takes time. HBM3e is still in development.

The pricing section is the most informative part, in my opinion. Nomura's channel survey to leading domestic mobile phone and PC/server manufacturers revealed that Changxin's pricing is lower than that of overseas leaders, but not significantly, about 0% to 20% lower, mainly due to factors supporting local procurement policies.

However, the cost disparities are real. Nomura estimates that due to the mainstream process lagging behind overseas leaders by about five years, Changxin's good die per wafer is only in the hundreds, while overseas leaders exceed a thousand. Therefore, Changxin's average wafer ASP in 2026 is about 14,000 to 15,000 US dollars. As the processes transition from 16-17 nanometers to 14-15 nanometers in 2027 to 2028, with increased unit output, Nomura expects this number to rise to 21,000 to 25,000 US dollars.

Another background number is that according to WSTS, in 2025, China will account for about 25% of the global DRAM market, but Nomura estimates that domestic manufacturers (including Changxin and others) will only hold about 10% of the global share by value, meaning that the self-sufficiency rate of domestic DRAM is only around 30%. This is the direct basis for Nomura's belief that Changxin still has a large share to capture in the domestic market.

In terms of customers, over 85% of Changxin's revenue comes through distributors, with the concentration among the top five customers being 68.08% in 2025. The disclosed end customers in the prospectus include Alibaba Cloud, ByteDance, Tencent, Lenovo, Xiaomi, Transsion, Honor, OPPO, and vivo. In terms of product structure, LPDDR series accounted for 66.43% of revenue in 2025, while the DDR series accounted for 31.87%, up from 20.16% in 2023, with the ramp-up of DDR5 server products being the main reason for this increase.

Additionally, Nomura mentioned that Changxin is working with GigaDevice on DRAM-on-logic wafer-on-wafer (WoW) stacking, targeting automotive smart cockpits, high-end mobile phones/PCs, and robotics in edge AI scenarios, and Nomura believes this technology will start to gain significant momentum from the end of 2027.

I believe that the earlier assumption of a 90% gross margin is clearly unsustainable. Directly projecting the cycle peak as a long-term norm does not align with the operational laws of the memory industry. Therefore, the correct approach should be to view the profit assumptions and target prices not as a long-term central tendency but as upper limit estimates under conditions of extreme supply-demand tension and optimistically high pricing and profitability.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。