The Seven Pieces of China's Chip Puzzle
The profit realization of China's domestic computing power industrial chain is still accelerating.
The seesaw battle between China and the United States on advanced manufacturing processes and equipment has not stopped, but a number of Chinese chip companies are no longer satisfied with "breakthroughs in single-point technology". Instead, they are building up the systematic capabilities of the entire chain from design, manufacturing, packaging and testing to storage step by step. They may not be able to confront TSMC and NVIDIA head-on at every node, but they have effectively moved the foundation of AI computing power a large step closer to local production.
This is not an easy story. The semiconductor industry is a tightly interwoven network of design, manufacturing, equipment, materials and software, which was expanded to the extreme through global division of labor in the past three decades. The special feature of 2026 is that this division of labor chain was forcibly interrupted by political forces, forcing Chinese companies to re-prove their capabilities in every link.
An AI chip involves at least seven key links from blueprint to mass production: IP/architecture licensing, chip design, wafer manufacturing, advanced packaging and testing, equipment and materials, EDA software, and storage. Looking at the full map of this industry, China has formed considerable competitiveness in five links: chip design, mature and partial advanced process foundry services, advanced packaging, NAND flash, and DRAM memory; IP licensing, lithography machines, high-end EDA, and HBM are still obvious weak points.
Today we focus on seven companies: HiSilicon, SMIC, Cambricon, Hygon Information, Yangtze Memory, ChangXin Memory Technologies, and JCET. They are exactly located in the five "competitive" links, forming the seven most noteworthy coordinates on the main chain of domestic computing power.
HiSilicon: The Anchor of Confidence on the Design Side
HiSilicon does not conduct independent financing and does not disclose financial reports, but it is the most core symbol in the narrative of China's semiconductor independence. After the export control in 2019 cut off the access to advanced process foundry services, the Kirin series chips once disappeared from flagship phones; with the breakthrough of local manufacturing capacity at the end of 2023, Kirin returned in a new form, and the closed loop of "design + local manufacturing" began to operate smoothly.
On the AI side, the Ascend series supports half of the domestic AI training chip market, forming a computing power base together with the Kunpeng CPU. The CANN software stack and MindSpore framework undertake the ecological task of competing with CUDA. Third-party estimates show that its revenue is expected to reach the level of 100 billion yuan. The real significance of HiSilicon does not lie in a specific number, but in the fact that it has proved that "China can design world-class AI chips" — which is the source of confidence for the entire domestic computing power narrative.
Market data also verifies this point. According to IDC statistics, in the first half of 2026, NVIDIA's share in China's data center AI chip market dropped to 34%, with shipments of 678,000 units, down 44% year-on-year; Huawei's shipments in the same period increased to 572,000 units, and its share rose to 28%, narrowing the gap between the two sides to only 6 percentage points. Alibaba's T-Head follows closely with 7%, while Cambricon, Kunlunxin, and Hygon account for 5%, 4%, and 3% respectively. Among the 11 Chinese chip companies tracked by IDC, 9 saw year-on-year growth in shipments. Xu Zhijun, Huawei's rotating chairman, stated more directly that according to the statistically available data, Ascend's share in the Chinese market has exceeded that of NVIDIA.
The product launch pace is equally intensive. On October 13, Huawei held a Kirin media briefing in Shanghai, and for the first time announced the detailed architecture diagram of the Kirin 9050 series of logic folding chips, which is the first Kirin chip built based on logic folding technology; the CPU large core energy efficiency of Kirin 9050 Pro is increased by 41%, and the NPU computing power is increased by 140%. On the AI side, the Ascend 950PR has been released and launched in the first quarter of 2026, targeting inference scenarios; the super node based on 950DT is mainly used for training, and large-scale supply is expected at the end of this year or early next year. Xu Zhijun said that "the training performance of 950DT is very good". Huawei also exhibited the Ascend 960 super node at the HC conference, and the maximum number of computing power cards it can carry has been upgraded from 1024 cards of the previous generation to 4096 cards.
SMIC: The Foundation of Full Production
SMIC is the cornerstone of domestic wafer manufacturing. In the second quarter of 2026, the company's single-quarter sales revenue reached 3.006 billion US dollars, a year-on-year increase of 36.1% and a quarter-on-quarter increase of 20.0%, hitting a record high; the attributable net profit was 479 million US dollars, a year-on-year increase of 261.7%; the gross profit margin was 25.3%, a quarter-on-quarter increase of 5.2 percentage points; the capacity utilization rate was as high as 93.7%. In that quarter, the proportion of revenue from China's region rose to 90.2%, and the trend of overseas orders returning and localized manufacturing strengthening is very clear.
The company's guidance for the third quarter given at the second quarter performance meeting is that revenue will increase by 2% to 4% quarter-on-quarter, and gross profit margin will be 26% to 28%. According to this calculation, the third quarter revenue falls in the range of 3.07 billion to 3.13 billion US dollars, and the consensus expectation of many brokerages is also around 3.1 billion US dollars. The complete financial data for the third quarter shall be subject to the company's official announcement.
SMIC's strategic choice is very clear: it will not compete head-on at the 3-nanometer and 2-nanometer nodes, but maximize production capacity, yield rate and cost at mature and sub-advanced nodes, while promoting improved processes such as N+1, N+2, N+3 through DUV multi-patterning to meet the foundry demand of local AI computing power chips.
The risks are also clear — the semiconductor industry is a strongly cyclical industry. Although the domestic production lines for advanced processes have achieved phased breakthroughs, the company is still restricted by the shortage of equipment, components and raw materials, and it will take time for the supply chain to achieve full independent controllability. As of now, the company's A-share market value is about 830 billion to 840 billion yuan.
Cambricon: From Loss to Explosive Profit Growth
Cambricon is the most eye-catching profit realization case of domestic computing power in 2026. It was spun off from the Institute of Computing Technology of the Chinese Academy of Sciences. In its early years, it started with terminal intelligent chips, and later focused all its resources on cloud AI acceleration cards.
After its listing in 2020, it suffered losses for consecutive years, and did not achieve its first full-year profit until 2025. The turning point came after the export of high-end GPUs from the United States was restricted, the domestic smart computing centers had an urgent demand for domestic substitution, and the Siyuan (MLU) series gradually achieved stable operation in inference and training scenarios. The three quarterly reports show that in the first three quarters of 2026, the revenue was 9.526 billion yuan, a year-on-year increase of 106.76%; the attributable net profit was 3.715 billion yuan, a year-on-year increase of 131.49%. Among them, the single-quarter revenue in the third quarter was 3.531 billion yuan, a year-on-year increase of 104.47%; the attributable net profit was 1.404 billion yuan, a year-on-year increase of 147.74%. Cloud products are absolutely dominant, and have been deployed on a large scale in core industries such as finance and the Internet.
Institutional expectations have diverged: on October 26, Capital Securities lowered its target price to 1200 yuan, but the average target price given by 15 brokerages in the past three months still reached 1522.20 yuan, with GF Securities giving the highest price of 1918.71 yuan. The hidden worries lie in the high customer concentration, the maturity of the software ecosystem still needs time, and the northbound funds reduced their holdings by more than 2 million shares in the third quarter, the number of shareholders increased by 23.57% compared with the end of June, and the chip structure tends to be decentralized. As of the beginning of November, the company's total market value is about 550 billion yuan.
Hygon Information: The Dual-Drive Engine of CPU and DCU
Hygon Information follows a high-growth path driven by both CPU and DCU. In its early years, it obtained x86 CPU related technologies through compliant authorization. Most of the systems in government, enterprise, finance and operator scenarios are based on x86, so Hygon can achieve almost seamless substitution; on this basis, it independently developed the DCU Deep Computing series, targeting NVIDIA acceleration cards.
The three quarterly reports show that in the first three quarters of 2026, the revenue was 14.955 billion yuan, a year-on-year increase of 57.58%; the attributable net profit was 2.847 billion yuan, a year-on-year increase of 45.12%; the comprehensive gross profit margin was 56.59%. The single-quarter revenue in the third quarter was 5.856 billion yuan, a year-on-year increase of 45.46%; the attributable net profit was 1.048 billion yuan, a year-on-year increase of 37.90%; the single-quarter gross profit margin was 58.73%, a quarter-on-quarter increase of 3.84 percentage points. DCU has been adapted to mainstream large models at home and abroad, and has been deployed in hundreds of scenarios in government departments, state-owned banks, operators and Internet manufacturers.
The pressure is also very real: operating cash flow is still in a net outflow state, the company continues to stock up materials and expand production in advance to undertake orders, and the short-term cash flow is under pressure. Once demand fluctuates, it will face the risk of inventory impairment. As of the beginning of November, the company's total market value is about 395.6 billion yuan, which has dropped significantly from about 487.5 billion yuan in early October.
Yangtze Memory: The Most Critical Storage Base Despite Supply Chain Restrictions
Yangtze Memory is not listed, and its STAR Market IPO has been accepted, but it is the only player in China's NAND flash field that can compete head-to-head with international giants. The bottleneck of AI servers is shifting from "slow computing" to "insufficient storage and slow transmission", and the strategic value of storage autonomy is no less than that of logic chips.
According to data from Counterpoint Research, in the third quarter of 2026, Yangtze Memory's NAND shipment share further rose to 17%, a quarter-on-quarter increase of 3 percentage points, firmly ranking third in the world, only behind Samsung's 25% and SK Hynix's 22%; the share was only 14% in the second quarter three months ago. The company judged in exchanges with the market that the global NAND supply shortage may continue for about three more years, memory prices have risen by 5 to 10 times, and the overall increase of NAND is between 3 and 5 times. In terms of capacity expansion, the third wafer fab in Wuhan is expected to start operation at the end of 2026, gradually increase production in 2027, and build two more wafer fabs on the basis of the completed new plant. After all are put into production, the total capacity will more than double, and the company plans to become the world's largest NAND flash manufacturer by the end of 2027.
In terms of technical routes, its self-developed Xtacking architecture manufactures the storage array and peripheral circuits separately and then bonds them together, which has been developed from the early 64 layers to mass production of more than 200 layers. Public information indicates that 267-layer products have been mass produced, and products with more than 300 layers are under development. In domestic complete machines such as Huawei, Yangtze Memory's particles are gradually replacing imported ones. But the more realistic gap lies in DRAM and HBM, which are still significantly behind Samsung, SK Hynix and Micron.
ChangXin Memory Technologies: The New Storage King with Sharp Market Value Correction
ChangXin Memory Technologies is the strongest card in China's DRAM battlefield. It landed on the STAR Market on July 27, 2026, and its market value once topped the A-share market.
The three quarterly reports show that in the first three quarters of 2026, the revenue was 269.024 billion yuan, and the attributable net profit was 131 billion to 132.5 billion yuan, a year-on-year increase of about 726% to 742%. However, the quarter-on-quarter growth momentum slowed down significantly: the revenue in the third quarter was 118.714 billion yuan, and the quarter-on-quarter growth rate dropped from about 96% in the second quarter to about 19.3%; the attributable net profit was 54.435 billion yuan, a quarter-on-quarter increase of only about 3%. The gross profit margin of main business is still at a very high level, and the proportion of overseas income is considerable.
At the end of September, ChangXin Memory Technologies announced at the 2026 World Manufacturing Conference that the fifth-generation DRAM technology platform G5 has achieved mass production, shrinking the half-pitch of the active area of the memory array to 11.95 nanometers, close to the process level of the world's top memory mass production platform. The self-developed LPDDR6 chip has achieved mass production and commercial use, and was first carried on the Xiaomi 18 Fold foldable flagship phone, becoming the world's first LPDDR6 product to be commercially deployed on flagship phones. In mid-October, the company further announced the R&D progress of the 4F² memory cell architecture, and it is expected to launch a new generation of DDR5 RDIMM memory products containing this technology at the end of this year. The news caused shocks in the international storage sector.
According to Counterpoint Research, ChangXin's share in global DRAM revenue has climbed from 4% in the second quarter of 2025 to about 10% in the second quarter of 2026, officially breaking through the double-digit threshold and firmly ranking as the world's fourth largest DRAM manufacturer. However, its market value experienced a significant correction in October: it fell by more than 23% cumulatively that month. As of the close on October 30, the total market value was about 2.84 trillion yuan, evaporating nearly 600 billion yuan compared with about 3.43 trillion yuan in early October. There is still a gap of about two to three generations in process technology compared with Samsung, SK Hynix and Micron, and the gap in the HBM field is even more obvious. The memory chip industry is a strongly cyclical industry. The sudden slowdown of the quarter-on-quarter growth rate in the third quarter has already hinted at the tension between the "price increase slope" and the implied expectation of market value.
JCET: The Hidden Champion of Advanced Packaging Under Short-Term Pressure
JCET occupies the key link for AI chip performance breakthrough — advanced packaging. It landed on the STAR Market on April 21, 2026, raising about 50 billion yuan, which is the largest IPO on the STAR Market since 2026.
In the field of chiplet multi-chip integrated packaging, its market share in Chinese Mainland is as high as 85%, and it is the only domestic enterprise that has achieved large-scale mass production of silicon-based 2.5D packaging; its 12-inch WLCSP revenue ranks first in China. Huawei's Ascend high-end AI chips are its core supply targets. However, the performance in the third quarter dropped significantly: the single-quarter revenue was 1.713 billion yuan, a year-on-year decrease of 8.33%; the attributable net profit was 186 million yuan, a year-on-year decrease of 45.96%. The revenue in the first three quarters was 5.120 billion yuan, a year-on-year increase of 1.45%; the attributable net profit was 636 million yuan, a year-on-year decrease of 18.48%. The company attributed the revenue reduction to the tight supply of core materials.
On October 12, at the Integrated Circuit Exhibition of the 26th China International Industry Fair, JCET officially released the self-developed SmartPoser®-BD high-performance computing packaging technology platform based on silicon bridge interconnection technology, and announced that the 2.5D silicon bridge technology with a single packaging size of 6 times the reticle has completed phased verification, and has the ability to undertake customer product import verification. Technological progress is still advancing, but the decline in performance superimposed on the market value correction has