The A-share market continues its upward momentum, with the Shanghai Composite Index surpassing 3,800 points, setting a new ten-year high.
Domestic computing power, represented by chips, has become the mainstay supporting this wave of the technology bull market.
Cambricon Technologies reached the daily limit of a 20% increase, with a market capitalization exceeding 500 billion yuan and a share price surpassing 1,200 yuan.
All chip stocks on the STAR Market rose, with the STAR Chip Index surprisingly increasing by 10.05%, leading the major chip indices in the market; the STAR Chip ETF (588200) surged by 11.92%.
Under strong industrial trends, multiple catalysts from the industrial sector have accumulated upward momentum, which today fully ignited the chip sector's rally.
01. Surge of Limit-Up Stocks!
On the market, sectors such as chips, securities, precious metals, education, software, and rare earth permanent magnets led in gains; while concepts like fertilizers, textiles and apparel, mining, banking, and agriculture underwent corrections.
Today, the most "crazy" sector undoubtedly belongs to semiconductors, which opened high and continued to rise.
By the close, Haiguang Information, Cambrian, and Shengmei Shanghai all reached the daily limit, while Xinyuan Micro, Jiahua Te, Hengshuo Co., and SMIC saw gains exceeding 10%. The Hong Kong semiconductor sector also performed very strongly, with Huahong Semiconductor rising over 15%.
In particular, the two chip giants Cambrian and Haiguang Information, both worth hundreds of billions, reached the daily limit of 20%. Such a situation is rare.
What exactly happened?
The most direct catalyst comes from DeepSeek.
DeepSeek mentioned in its official announcement of the "official release of DeepSeek-V3.1" that DeepSeek-V3.1 uses the parameter precision of UE8M0FP8Scale. V3.1 has made significant adjustments to the tokenizer and chat template, showing clear differences from DeepSeek-V3.
In addition, DeepSeek's official WeChat account stated in a pinned message that UE8M0FP8 is designed for the upcoming next-generation domestic chip.
A new architecture and next-generation domestic chips, all conveyed in fewer than 20 characters, yet containing immense information.
UE8M0 is a real technology that has been implemented in engineering, primarily used as a scaling factor for FP8 computation in the next-generation AI chips, achieving extreme optimization for reduced bandwidth, lower power consumption, and higher throughput.
DeepSeek's previously open-sourced project, the 5.6k-star FP8GEMM kernel DeepGEMM, already supports UE8M0, although this project is mainly tailored for NVIDIA chips and the CUDA ecosystem.
However, most of the domestically produced AI accelerators that have already been mass-produced still rely on FP16/BF16+INT8 computing pathways and have not integrated complete FP8 multiply-accumulate units such as E4M3/E5M2.
Nevertheless, new domestic chips such as the Moores Threads MUSA3.1 GPU and the Chipone VIP9000 NPU, set to launch in the second half of 2025, have already listed support for 'native FP8' or 'BlockFP8' in their promotional materials and have jointly verified the UE8M0 format with DeepSeek and 15 other manufacturers.
Based on this information, the market has developed some speculations and expectations that DeepSeek may leverage the knowledge and technologies gained from cutting-edge practices (such as the application of UE8M0) to empower the research and development of domestic chips, advancing the process of achieving autonomy in computing power.
To extend further, it may represent that domestic AI is transitioning towards a phase of hardware-software synergy, which can substantially reduce dependence on foreign computing power from companies like NVIDIA and AMD. This resembles the "Wintel" alliance of the past, where the technology integration of Microsoft and Intel constructed a protective moat for the PC ecosystem.
As for which domestic chip has been optimized, no further information has been disclosed. However, this range is quite limited, and it includes publicly listed companies, explaining the significant reactions in stock prices of companies like Cambricon and Haiguang Information, both of which have products supporting FP8 computing.
The domestic AI computing power sector has recently experienced a concentrated catalytic upward breakout, with chips being the most directly benefited sector, notably exemplified by "Hanwang."
A series of market rumors widely circulated online have caused Cambricon's market value to skyrocket; since its low point in July, Cambricon's stock price has more than doubled, with an annual increase exceeding 80%.
For instance, last week, the market rumored that Cambricon had placed substantial orders with Taiwan's major substrate manufacturer, Kingboard, suggesting a surge in order volume. There are even rumors that by 2026, planned production capacity could reach 500,000 units, with the company's annual revenue expected to exceed 10 billion yuan.
Although the company later clarified that this information was inaccurate, since the second half of the year, the successful placement of Cambricon's private placement, the merger of Haiguang Information and Zhongke Shuguang, as well as the acceptance of initial public offerings for Muxi Co. and Moore Threads on the Science and Technology Innovation Board, seem to convey a clear signal.
With the power of the secondary market, the localization of AI computing power is already accelerating. Only when a complete closed loop is formed—from domestic chips to large models and then to the AI application ecosystem—can the domestic industry grasp the acceleration of its own AI industry development.
02. Why are innovative technology chips leading the charge?
Since the establishment of the Science and Technology Innovation Board in 2019, the popularity of the main themes has completed a switch from new energy and pharmaceuticals to artificial intelligence. The Science and Technology Innovation Board covers a variety of mainstream companies related to AI, especially with its complete industrial chain of chip "design-manufacturing-packaging and testing," which serves as the backbone of domestic computing power.
When the current wave of market activity hit the A-shares on September 24, 2024, the Science and Technology Innovation Board, represented by semiconductor companies, became the leading sector in this AI-driven bull market.
This performance is reminiscent of the ChiNext market from 2013 to 2015, which also supported a number of core "Internet+" concept companies at that time.
The Science and Technology Innovation Chip Index has accumulated a rise of 46.62% since its low point on April 8, outperforming the CSI Semiconductor Index (37.80%), the National Semiconductor Index (36.11%), and the chip industry (35.60%). The Science and Technology Innovation Chip ETF (588200) increased by 49.28% during the same period, beating the index by 2.66 percentage points.
This is because the high beta elasticity of innovative technology chips is tightly bound to the domestic computing power chain. Its constituent stocks cover the chip ecosystem (design/manufacturing/packaging and testing/materials and equipment) that is most closely aligned with the AI industry chain, including both cloud training chips and edge computing, as well as advanced manufacturing processes, all of which gather a leading group of top companies in the current domestic computing power chain.
The emergence of DeepSeekR1 at the beginning of the year has sounded the clarion call for the revaluation of domestic computing power. Several chip companies have proactively adapted to large models to enhance the utilization of domestic chips in the inference segment.
As the third quarter brings another round of intensive catalysts, the upward effect on leading companies becomes evident, making it clear that the narrative of revaluation for domestic chips is not constrained by the current semiconductor cycle framework.
From the demand side, starting in 2025, the iteration of domestic large models will accelerate, significantly enhancing product diversity. The usage of tokens is rapidly increasing, with domestic AI products represented by the Alibaba Quark APP occupying most positions on the active user list, second only to ChatGPT, confirming the upward trend in domestic computing power demand.
Half a year has passed, and as the overseas computing power performance explodes, it has completely dispelled doubts about computing power demand. The performance and capital expenditures of tech giants have established a verifiable closed loop for the growth of computing power demand driven by AI commercialization. This framework can also be applied to domestic computing power, replicating the overseas path of 'Tokens → Capital Expenditures → Industry Chain Expansion.'
As the largest component of capital expenditures, chips not only support the foundational infrastructure of AI in China but also represent a critical segment that urgently needs to transition its role to occupy a dominant position in the domestic substitution.
On August 13, Tencent's earnings call also mentioned that Tencent has various options on the supply side for inference chips. The market may interpret this as indicating that the current trade policies regarding high-end overseas chips still carry significant uncertainty, but it is expected that domestic suppliers such as Huawei, Cambricon, and Haiguang will be able to meet the relevant computing power demands.
In terms of performance, the warming of the cycle's bottom has provided support for the valuation recovery of component stocks.
Specifically, the year-on-year net profit attributable to the parent company of the innovative technology chips is generally in a state of bottom oscillation in 2023 and 2024. However, since the second half of 2024, the innovative technology chips have begun to experience continuous improvement from the bottom. By the first quarter of 2025, the year-on-year net profit attributable to the parent company of the innovative technology chip index is expected to rise from -6.9% at the end of 2024 to 83.2%.
The "growth indicators" (year-on-year revenue growth rate, year-on-year net profit growth rate) of the innovative technology chip index also significantly outperform similar indices. According to Wind data, the expected year-on-year revenue growth rate for the constituent stocks of the innovative technology chip index in 2025 is 24.93%.
In addition, the IPO process for domestic semiconductor giants is accelerating, and the domestic key chip sector is expected to receive more policy and financial support in the capital market, while the trend of AI chip localization continues to strengthen.
Moore Threads and Muxi's IPO on the STAR Market was accepted by the Shanghai Stock Exchange at the end of June, entering the public review stage. Changxin Storage also completed the IPO listing counseling filing with the Securities Regulatory Commission on July 7, officially starting the listing process.
In the future, the STAR Market may gather more leading AI chip companies, further attracting capital allocation.
Speaking of capital allocation, the current level of fund allocation to the STAR Market still has significant room for improvement compared to the growth enterprise market in 2013-2014.
As of January 2025, the allocation ratio of the STAR Market, based on the sample of ordinary stock funds in equity funds and the equity-mixed funds in hybrid funds, has reached 16%. In comparison, as of March 2013, the allocation ratio for the ChiNext was approximately 14%, indicating that there is still room for growth before reaching the high allocation levels observed in 2015.
03. Conclusion
The computing power sector is collectively heating up; in fact, comments are merely a catalyst, while the underlying trend is the further revaluation of Chinese technology stocks.
Since the market trend beginning on September 24 of last year, semiconductor stocks may be in the second wave of revaluation.
Moreover, companies in the STAR Market related to semiconductors exhibit notable growth potential, with a single-day increase of 20% allowing them to maintain remarkable elasticity in their price increases during the main market trends; this advantage is difficult for semiconductor companies in other sectors to match. On September 30 and October 8 of last year, the STAR semiconductor ETF (588200) experienced consecutive trading days with a '20CM' limit-up.
However, the entry threshold for the STAR Market is relatively high; if one wishes to participate in this wave of market activity, ETFs present a better option.
The STAR semiconductor ETF (588200) tracks the Shanghai Stock Exchange STAR Market Semiconductor Index, covering the entire semiconductor industry chain and focusing more on critical bottleneck areas (manufacturing, equipment, materials), aligning closely with the domestic substitution theme. The largest constituent stock by weight in the index is Cambricon, recognized as the 'first AI chip stock', with the latest weight at 13.89%, representing the highest concentration in the semiconductor thematic index.
The latest scale of the Sci-Tech Innovation Chip ETF (588200) is 31.271 billion yuan, with an average daily trading volume of 2.012 billion yuan and a net inflow of 1.953 billion yuan year-to-date, ranking first among similar products. This ETF also has an off-exchange linked fund (Class A: 017469, Class C: 017470).
The marginal changes brought about by the DeepSeek update serve as a reminder that software can drive hardware iteration.
The narrative surrounding domestic chips holds significant potential for the future.
