The True Future of CXMT: Beyond 'Domestic Substitution'

Recently, the storage industry has seen a continuous rise in heat.

Global DRAM prices have surged significantly, with Samsung, SK Hynix, and Micron accelerating capacity layouts around HBM. CXMT went public on the STAR Market with a fundraising scale of nearly 57.9 billion yuan, and its market value once exceeded 3 trillion yuan after listing. Following the company's IPO, the topic of 'Hefei state-owned capital betting on CXMT and gaining a trillion-yuan paper profit' has become a hot market discussion.

CXMT, a storage company still in a phase of rapid capacity expansion, why was it able to raise nearly 57.9 billion yuan at once?

The easiest answer the market gives is domestic substitution. But if one only understands CXMT as a company filling the domestic DRAM gap, it is easy to underestimate the opportunities it faces and the risks it will encounter in the future. This is because the changes brought about by AI are no longer just about increasingly powerful chip computing.

From large model training and inference to databases, intelligent agents, and edge devices, the entire computing system is becoming increasingly dependent on memory capacity, bandwidth, and data transfer efficiency. In a sense, storage is becoming the 'second type of computing power' in the AI era.

Therefore, to judge the future of CXMT, three questions truly need to be answered:

• First, can it grow from a follower in the global DRAM market into a long-term player with scale advantages?
• Second, can it move from ordinary DDR and LPDDR into higher-value markets such as server memory and HBM?
• Third, does its current profit growth come from long-term competitiveness improvement, or did it just happen to catch the storage price upcycle?

01 CXMT has crossed the 'can it be made' hurdle, Next, it must solve the 'can it be continuously profitable' problem.

From 2023 to 2025, CXMT's operating revenue was 9.087 billion yuan, 24.178 billion yuan, and 61.799 billion yuan, respectively, with a two-year compound growth rate of 160.02%. During the same period, the company's net profit attributable to the parent company was a loss of 16.34 billion yuan, a loss of 7.145 billion yuan, and a profit of 1.875 billion yuan. Looking at this set of data, CXMT seems to have completed the leap from huge losses to profitability. But this does not mean the company has escaped the cyclical nature of the storage industry.

From 2023 to 2025, CXMT's gross margin for its main business was -2.19%, 5.00%, and 41.02%, respectively. If calculated based on the caliber before the reversal of inventory write-downs, the gross margins for the three years were -113.27%, -4.72%, and 37.80%.

(Source: CXMT Prospectus)

Such drastic fluctuations in gross margin indicate that DRAM remains a typical capital-intensive, highly cyclical industry. When prices fall, companies must not only bear product price cuts but also shoulder equipment depreciation, inventory impairment, and high R&D expenditures; when prices rise, fixed costs are rapidly diluted, and profits can explode. This cyclical elasticity was even more evident in 2026.

In the first quarter of 2026, CXMT achieved an operating revenue of 50.8 billion yuan, a year-on-year increase of 719.13%; the net profit attributable to the parent company reached 24.762 billion yuan. The company expects its operating revenue for the first half of 2026 to reach 110 billion to 120 billion yuan, and the net profit attributable to the parent company to reach 50 billion to 57 billion yuan.

But at the same time, TrendForce expects that the contract price of traditional DRAM will rise by 55% to 60% month-on-month in the first quarter of 2026, and may still rise by 58% to 63% in the second quarter.

Entering the third quarter, the DRAM market is still in short supply, but the rate of price increases has begun to slow down. TrendForce expects that the contract price of traditional DRAM will rise by 13% to 18% month-on-month in the third quarter. Among them, AI servers and general-purpose servers are still supporting server DRAM demand, but consumer electronics customers such as PCs and smartphones have gradually approached their cost tolerance limits.

The latest disclosed performance of SK Hynix also reflects the intensity of this industry boom.

In the second quarter of 2026, SK Hynix's operating revenue reached 79.32 trillion won, and its operating profit reached 60.54 trillion won, both hitting record highs for a single quarter. The company attributed its performance growth mainly to the expansion of sales of high-value-added products such as HBM, AI server DRAM, and enterprise SSDs.

The significant profit growth of both CXMT and SK Hynix during the same period shows that this is not a business turning point for a single company, but an industry cycle driven by rising storage prices, AI server demand, and the volume release of high-value products.

(Source: TrendForce)

Therefore, CXMT's current performance explosion contains at least two factors: one is the improvement in capacity, shipment volume, and product structure; the other is that DRAM prices have entered a very strong upward cycle.

This is also where one needs to remain clear-headed when analyzing CXMT. The 2026 profit cannot be simply linearly extrapolated. To judge the long-term value of CXMT, it is more important to observe how much gross margin, market share, and customer stickiness the company can retain after storage prices return to normal.

02 Moving from 9% to 15% may be the scale threshold CXMT must cross

CXMT's prospectus cites Omdia data stating that in the fourth quarter of 2025, the company's sales revenue share in the global DRAM market reached 7.67%, ranking fourth in the world. Counterpoint, in an analysis released in July 2026, estimated that CXMT's current DRAM bit shipment share is about 9%, and it is expected to rise to about 11% by 2028.

It should be noted here that 7.67% is based on sales revenue, and 9% is based on bit shipments; the two cannot be directly compared.

But they together illustrate a fact: CXMT has entered the ranks of major global DRAM manufacturers, but it has not yet entered the first tier composed of Samsung, SK Hynix, and Micron. Counterpoint also provided an empirical judgment: for a DRAM company to establish relatively stable economies of scale, its global market share may need to be close to 15%.

Why is 15% important? Because the DRAM industry has extremely strong economies of scale. From 2023 to 2025, the total cash paid by CXMT for the construction of fixed assets, intangible assets, and other long-term assets reached 164.627 billion yuan, while the cumulative operating revenue for the same period was about 95.064 billion yuan.

In other words, the capital expenditure over the past three years is equivalent to 1.73 times the operating revenue for the same period. At the same time, the company's cumulative R&D investment over the three years reached 20.605 billion yuan. As of the end of 2025, there were 6,259 R&D personnel, accounting for 32.43% of the total number of employees. In such an industry, a share that is too small means that R&D, depreciation, and production line upgrade costs cannot be fully diluted. Only when yields, capacity utilization, and shipment scale continue to increase can a company form true cost competitiveness. Therefore, for CXMT, moving from about 9% to 15% is not just about selling more memory. It determines whether the company can move from 'relying on continuous investment to catch up' to a new stage of 'using operating cash flow to support the next round of R&D and capacity expansion'.

03 What AI is really changing is the position of storage in the computing system

In the past, when talking about AI chips, the market often only focused on GPUs. But as AI moves from training to large-scale inference, computing power is no longer the only bottleneck. How models read, save, and move data is becoming equally important. The 'Database Development Research Report (2026)' points out that large model inference can usually be broken down into two stages: Prefill and Decode.

Prefill mainly handles input context; the more input tokens, the greater the matrix calculation that needs to be performed. Decode requires generating tokens one by one and continuously reading the historical KV Cache.

In the Decode stage, the system bottleneck comes not only from computing power but also from video memory bandwidth, cache access efficiency, and data transfer efficiency.

And KV Cache itself is one of the most important dynamic video memory resources in the process of large model inference. The longer the input content, the longer the context, and the more users processed simultaneously, the larger the video memory occupied by the KV Cache. When video memory capacity is insufficient, even if the GPU computing unit is idle, the number of concurrent requests the system can carry will still decrease.

(Source: Zhiwang Technology)

This means that after the scale of AI inference expands, the system bottleneck will gradually shift from 'whether there are enough computing units' to 'whether data can be sent to the computing unit fast enough'.

Databases are further amplifying this demand. The 'Database Development Research Report (2026)' points out that traditional vector indexes have a high reliance on memory, and billion-level vectors often require hundreds of GB of memory. Technologies like DiskANN move some indexes and raw vectors to SSDs, essentially to alleviate memory capacity pressure.

The report also points out that GPU databases have long been constrained by video memory capacity limits and CPU-GPU data migration costs. New GPU-native databases are combining HBM with high-speed flash memory to build a new multi-level storage system.

(Source: CCSA TC601)

Even more worthy of attention are AI agents. The 'Database Development Research Report (2026)' cites data disclosed by Supabase in June 2026, stating that more than 60% of new databases on its platform have been automatically created and deployed by AI agents or AI coding tools.

The data generated by AI in the future will no longer be just one-time data from the model training stage, but will also include the agent's long-term memory, conversation context, vector indexes, multi-modal data, and continuously generated token-level records. The report points out that these workloads will bring about high-frequency token-level writes, persistent context, long-term memory management, and dynamic resource scheduling needs.

(Source: CCSA TC601)

These data will eventually enter different levels of storage systems:

GPUs and AI accelerators need high-bandwidth HBM; cloud servers need large-capacity DDR5; database and agent platforms need DDR5 and CXL memory pools; mobile phones and edge devices need low-power LPDDR. This shows that AI storage does not just mean HBM. HBM solves high-speed data supply near the GPU, server DDR solves system-level memory capacity, and LPDDR undertakes the low-power inference needs of mobile phones, AI PCs, and edge devices.

This just explains why CXMT does not have only one growth path.

04 CXMT has captured server DRAM, but HBM is the real value leap

In 2023, CXMT's revenue from the server field was only 395 million yuan, accounting for 4.60% of its main business revenue. By 2025, the company's server revenue had reached 15.97 billion yuan, and its proportion rose to 26.51%. In two years, CXMT's server revenue grew about 40 times.

At the same time, mobile devices remain CXMT's largest source of revenue. In 2025, the company's mobile device-related revenue reached 36.381 billion yuan, accounting for 60.40% of its main business revenue. This shows that CXMT's current foundation is still LPDDR and mobile devices, but the company's revenue structure has begun to migrate rapidly to server DDR.

This change is basically consistent with the global DRAM demand structure. CXMT's prospectus cites Omdia's forecast that servers will account for about 50% of global DRAM demand in 2025 and may rise to 71% by 2030. From 2025 to 2030, server DRAM capacity demand is expected to grow from 19,953 MGB to 69,340 MGB, with a compound growth rate of 28.29%.

At the same time, edge AI may also form a second demand curve. CIC Consulting expects that under the specific statistical caliber of AI boxes and all-in-one machines, China's edge AI inference chip revenue will grow from 626.6 million yuan in 2024 to 5.7 billion yuan in 2030, with a compound growth rate of 44.4%.

The report expects that the shipment volume of related chips will grow from about 800,000 units in 2024 to 12.1 million units in 2030, with a compound growth rate of 56.6%. These edge AI devices may not necessarily need expensive HBM, but they generally need DDR or LPDDR as system memory.

(Source: CIC Consulting)

Therefore, CXMT's relatively certain short-term increments may come from server DDR5 and LPDDR used in edge devices; while what truly determines the upper limit of the company's long-term value is HBM.

05 Why say HBM determines the valuation ceiling of CXMT

Jiashi Consulting estimates that the global HBM market size will grow from 16.5 billion USD in 2024 to 60 billion USD in 2026, and is expected to reach 130 billion USD by 2030. The rapid growth of the HBM market is behind the continuous increase in AI chip requirements for storage capacity and bandwidth.

The data listed in the report shows that NVIDIA's H100 is equipped with 80GB HBM3, with a bandwidth of about 3.35TB/s; H200 increases to 141GB and 4.8TB/s; B200 further reaches 192GB to 288GB, with a bandwidth of about 8TB/s. From a technical generation perspective, the single-stack bandwidth of HBM3 is about 640-819GB/s, reaching a maximum of 1.33TB/s; HBM3E reaches a maximum of about 1.2TB/s; HBM4 increases the interface bit width from 1024 bits to 2048 bits.

(Source: Jiashi Consulting)

Every upgrade of AI chips is simultaneously increasing the memory capacity, bandwidth, and storage value carried by a single card.

TrendForce data shows that in 2024, HBM only accounted for about 5% of DRAM bit shipments, but contributed about 20% of DRAM industry revenue.

This is the special feature of HBM. It consumes a limited number of bits, but can contribute revenue and value far higher than ordinary DRAM.

Currently, CXMT's publicly disclosed commercial products are still mainly DDR4, DDR5, LPDDR4X, and LPDDR5/5X. Therefore, at this stage, HBM cannot be regarded as a mature revenue source that the company has already realized.

At the same time, global HBM competition is still rapidly upgrading.

SK Hynix achieved mass production and shipment of HBM4 in the second quarter of 2026 and plans to continue to expand production in the second half of the year; HBM4E samples have also been delivered. The company also disclosed that it has completed discussions on long-term supply agreements with more than ten customers, including core customers, and continues to advance multi-year contracts with other major customers.

This dynamic further illustrates that what CXMT is chasing is not a static HBM market.

The competition among top manufacturers is no longer limited to bandwidth and stacking layers, but has extended to mass production yield, advanced packaging, joint customer development, and long-term capacity binding. For latecomers, even if they complete HBM product R&D, they need to enter the GPU manufacturer's product development cycle and strive for orders in a situation where customers have already signed long-term agreements with existing suppliers.

Counterpoint provided an optimistic scenario: if CXMT's 12-layer HBM3 can be successfully mass-produced and complete customer verification, the company's HBM revenue could reach about 2 billion USD by 2028.

Therefore, to judge the progress of CXMT's HBM, one cannot just look at 'whether there is R&D', but must look at three more specific questions:

• Can technology be converted into stable yield?
• Can GPU manufacturer and cloud customer verification be completed?
• Can it secure continuous orders after mass production?

Only when these three questions receive affirmative answers can CXMT truly transition from an ordinary DRAM follower to a player in the high-value AI memory market.

06 HBM is not just a chip, but an entire industrial chain capability.

The difficulty of HBM lies in more than just DRAM wafer manufacturing. The "2026 HBM Chip Industry Report" points out that compared to standard DRAM, HBM also requires TSV (Through-Silicon Via), wafer thinning, temporary bonding and debonding, stacked packaging, GMC materials, and high-complexity testing. The report estimates that for the same storage capacity, HBM may consume more than three times the DRAM wafer capacity of standard DRAM.

From production line construction and equipment procurement to TSV yield debugging, new HBM capacity typically takes 12 to 18 months to be released. This means that even if CXMT masters DRAM wafer manufacturing, it does not mean it can naturally replicate standard DRAM production lines for HBM.

HBM competition involves a complete system composed of manufacturing, packaging, materials, equipment, testing, and customer collaboration. The report shows that the global HBM market is highly concentrated, with SK Hynix, Samsung, and Micron accounting for nearly 100% of total shipments.

In the upstream segment, the report states that Japanese manufacturers hold over 90% of the market share for GMC materials used in HBM, and the customer verification cycle for some materials can reach two years; temporary bonding equipment, high-end die bonders, and testing equipment are also concentrated in the hands of a few manufacturers. This is why CXMT's capacity expansion may have a deeper impact on the domestic semiconductor materials industry.

07 CXMT's capacity expansion, may drive a supply chain longer than that of DRAM.

Data from Handing Think Tank shows that the market size of key integrated circuit materials in China grew from 66.47 billion yuan in 2019 to 113.93 billion yuan in 2023, and is expected to reach 258.96 billion yuan by 2028. Among this, the manufacturing materials market is expected to reach 185.38 billion yuan, accounting for over 70% of the key materials market. However, there are significant differences in local supply capabilities for different materials.

(Source: Handing Think Tank)

The report shows that the top five global silicon wafer manufacturers hold over 85% of the market share; Japan's JSR, Tokyo Ohka Kogyo, and Shin-Etsu Chemical together account for about 52% of the global photoresist market. The localization rate of domestic ArF photoresist is less than 5%, the localization rate of CMP polishing pads is less than 10%, and the overall localization rate of key materials is about 15% to 20%. In the silicon wafer sector, the domestic localization rate for 8-inch wafers has reached about 60%, but the localization rate for 12-inch wafers is still only 10% to 15%.

CXMT owns three 12-inch DRAM wafer fabs located in Hefei and Beijing. In the future, whether CXMT expands its DDR5 capacity or attempts to enter HBM, it will increase demand for 12-inch silicon wafers, electronic gases, wet electronic chemicals, lithography materials, CMP materials, and advanced packaging materials.

This supply chain transmission has already begun to manifest in Hefei.

Currently, the second phase of CXMT's 12-inch memory wafer manufacturing base is still under construction. The Xinqiao Integrated Circuit Science and Technology Park surrounding the main plant area has already gathered semiconductor equipment companies such as NAURA Technology, and Hefei Payton Storage, located just across the street from CXMT, mainly undertakes high-end memory chip packaging and testing. Its Hefei packaging and testing plant is already at full capacity and continues to expand based on customer demand.

From a broader industrial perspective, CXMT has already driven more than 400 upstream and downstream enterprises to gather in Hefei, covering segments such as memory chip design, EDA, equipment, materials, and packaging and testing. The output value of the Hefei integrated circuit industrial chain has also grown from approximately 18 billion yuan in 2016 to 151.4 billion yuan in 2025.

However, an increase in the number of enterprises and geographical clustering does not mean that the key supply chain has been fully localized.

For material and equipment companies, being close to CXMT only provides more convenient conditions for sample submission, verification, and technical services. Whether they can truly enter the bulk procurement system still depends on product performance, yield stability, and continuous delivery capabilities.

Therefore, the industrial value of CXMT cannot be measured solely by how much memory it can sell.

If its capacity expansion can drive domestic material and equipment companies to complete product verification, CXMT could become an important gateway for local semiconductor material companies to enter the advanced memory supply chain. Conversely, if key materials, advanced packaging, and testing equipment remain concentrated in the hands of a few suppliers, CXMT's expansion speed and HBM yields will also be constrained by supply chain capabilities.

08 There are three possible outcomes for CXMT's future.

The first is the baseline scenario.

CXMT continues to increase its market share by relying on DDR5 and LPDDR5/5X, with the proportion of server revenue continuing to rise, reaching a bit shipment share close to the 11% predicted by Counterpoint by around 2028. In this case, CXMT will become the fourth major player in the global DRAM market that cannot be ignored, but its main competitiveness will still come from the scale, cost, and supply assurance capabilities of standard DRAM.

The second is the optimistic scenario.

CXMT's global market share gradually approaches 15%, server memory becomes the main source of growth, and it simultaneously completes HBM mass production and customer verification. Once it enters HBM, the company's revenue growth will no longer rely solely on shipment volume, but also on the increase in product value. Only at this stage will CXMT have the opportunity to grow from a "domestic DRAM leader" into an important participant in the global AI memory industry.

The third is the stress scenario.

If DRAM prices re-enter a downward cycle while HBM mass production, yields, and customer certification progress are lower than expected, the company will simultaneously face falling prices, increased depreciation, and high R&D expenditures. As of the end of 2025, CXMT's accumulated undistributed profit was still negative 36.65 billion yuan. From 2023 to 2025, the company's total fixed asset depreciation reached 50.111 billion yuan.

This means that as long as capacity utilization drops, the massive fixed costs could once again compress the company's profit margins.

Therefore, when observing CXMT in the future, what is truly important is not how much quarterly profit has grown, but five indicators:

• Can the global bit share continue to approach 15%?
• Can the proportion of server revenue continue to exceed the current 26.51%?
• Can HBM generate commercial revenue and complete verification by core customers?
• Can gross profit margins remain stable after memory prices return to normal?
• Can massive capital expenditures be converted into advantages in yield, cost, and cash flow?

Conclusion

What is most worth noting about CXMT is not whether it can continue to expand production.

The nearly 57.9 billion yuan in IPO fundraising, three 12-inch wafer fabs, and more than 160 billion yuan in capital investment over the past three years have already proven that it has the ability to continue expanding.

What truly determines the company's future is whether it can complete two leaps:

The first is moving from a bit share of about 9% to a higher share with economies of scale.

The second is moving from "quantity growth" in DDR and LPDDR to "value growth" in server memory and HBM.

AI will not automatically benefit all memory companies. It will only further amplify the importance of memory bandwidth, capacity, yield, and supply chain collaboration.

Therefore, the true future of CXMT is not just domestic substitution. It is whether, in the process of AI turning memory into a "second type of computing power," it can grow from the fourth-ranked player in the global DRAM market into a memory giant that truly possesses technical influence and product pricing power.