Economic Observer reporter Zheng Chenye

On April 15th, Nvidia CEO Huang Renxun was interviewed by American technology podcast host Patel and said a sentence about Chinese AI: "If one day achievements like DeepSeek first appear on the Huawei platform, it will be a very bad result for our country. ?

Nine days later, what he was worried about happened.

On April 24th, DeepSeekV4 was officially released. This generation of models is divided into two versions: V4 Pro (1.6 trillion parameters, 49 billion activations) and V4 Flash (284 billion parameters, 13 billion activations). The context processing length has been extended from the previous generation's 128K to 1M (approximately one million words), achieving one of the best levels of open source models in terms of agent capabilities, mathematical reasoning, and code generation.

On the same day, Huawei announced that the entire range of Ascend supernodes supports DeepSeekV4, with inference latency of up to 20 milliseconds for V4 Pro and 10 milliseconds for V4 Flash. In the technical report of DeepSeekV4, Huawei Ascend NPU (Neural Network Processor) and Nvidia GPU (Graphics Processor) both appear in the adaptation solution.

On the pricing page of DeepSeek's official website, there is a small line of text that reads: "Due to high-end computing power, the service throughput of Pro is currently very limited. It is expected that after the mass launch of the Ascend 950 super node in the second half of the year, the price of Pro will be significantly reduced. ?That is to say, one of the best big model companies in China has directly tied the pricing rhythm of flagship products in the second half of the year to the shipment schedule of Huawei chips.

In fact, DeepSeekV4 has been deeply adapted with Huawei Ascend chips during the development stage, which is officially referred to as "close collaboration of chip and mold technology" by Huawei. This is very different from the approach of "first running the model on Nvidia and then migrating it to domestic chips" in the past few years.

According to data from well-known market research firm IDC, by 2025, Nvidia's market share in China's AI acceleration card market will decrease from 95% two years ago to 55%, and the domestic chip market share will exceed 41% for the first time. Among them, Huawei Ascend shipped 812000 units, accounting for nearly half of all domestically produced chips.

Of course, it's not just Huawei's fault that Ascend has come this far. Since 2019, Huawei has invested in a group of upstream companies that produce connectors, power chips, chip testing, and optical chips through its Hubble investment system, cultivating a domestic computing power industry chain covering key component links around Ascend.

This is a self owned ecosystem of domestic computing power.

Hubble and Huawei's computing power chain

In April 2019, Hubble Technology Venture Capital Co., Ltd. was established with a registered capital of 3 billion yuan, and is 100% owned by Huawei Investment Holding Co., Ltd. In 2021, Shenzhen Hubble Technology Investment Partnership Enterprise (Limited Partnership) was established, and after multiple rounds of capital increase, the total subscribed capital has increased to 9.48 billion yuan by February 2025, with Huawei Technologies Co., Ltd. accounting for 69% of the investment.

According to incomplete analysis by Economic Observer reporters, since 2019, the Hubble Group has invested in more than 112 projects, of which at least 15 invested companies have successfully gone public on A-shares or Hong Kong stocks; By the end of 2025, the two aforementioned Hubble investment entities will still appear on the top ten shareholder lists of at least 10 A-share listed companies.

The main business of these companies is concentrated in the upstream component fields such as connectors, power chips, chip testing, packaging materials, and optical chips, which are key links for the normal operation of Ascend servers. For example, Huafeng Technology (688629. SH), which makes high-speed connectors, Jiehuate (688141. SH), which makes power management chips, Qiangyi Corporation (688809. SH), which makes wafer testing probe cards, Tianyue Advanced (688234. SH), which makes silicon carbide substrates, and Changguanghua Semiconductor (688048. SH), which makes high-power laser chips, together constitute key nodes in the domestic supply chain that ensure the stable operation of Ascend servers.

Among them, Huafeng Technology is one of the companies with high market attention. Huafeng Technology mentioned in its 2025 annual report that the company "became an early global supplier to Huawei, ZTE, and Nokia", producing high-speed backplane connectors and high-speed line modules. Ascend supernodes connect hundreds of AI acceleration cards into a computing power cluster, and high-speed data transmission between cards requires the use of such connectors and cable components.

According to the first quarter report of Huafeng Technology in 2026, Shenzhen Hubble Technology Investment Partnership holds 2.95% of its shares, making it the sixth largest shareholder and has never reduced its holdings since its investment. Huafeng Technology disclosed in its 2025 annual report that Huawei is the company's largest customer, with sales to Huawei of 1.53 billion yuan that year, accounting for 60.52% of the annual revenue of 2.528 billion yuan. The annual report also emphasizes in the risk warning that the company has a high degree of dependence on Huawei.

According to the financial report, in 2025, Huafeng Technology's revenue increased by 131.50% year-on-year, with a net profit attributable to the parent company of 359 million yuan. In 2024, it will incur a loss of 17.75 million yuan; In the first quarter of 2026, Huafeng Technology's revenue was 633 million yuan, a year-on-year increase of 56.15%, and the net profit attributable to the parent company was 105 million yuan, a year-on-year increase of 230.43%.

On April 15th, Liu Taiguo, the general manager of Huafeng Technology, stated during an investor survey that the overall production capacity of high-speed line modules is in a saturated state, and the company has already started expanding production with its own funds in advance. He also mentioned that the technology threshold for high-speed line module is relatively high, and the product directly affects the core performance of the server. Customers are cautious about introducing suppliers, and "it takes a long time for new suppliers to enter stable mass production"; In addition, with the continuous iteration of server chips and the performance upgrade of AI computing cards, the connection scheme between cards has become increasingly complex, and the value of a single high-speed line module will increase.

Huafeng Technology covers the connection link, and within the computing power server, power supply and chip testing are also crucial positions that cannot be bypassed. Huawei has also made capital layout in these two links through Hubble.

Among them, Jiehuate is a listed company that is simultaneously held by two Hubble entities. The company's 2025 annual report shows that Hubble Technology Venture Capital Co., Ltd. holds 3.01% of the shares, and Shenzhen Hubble Technology Investment Partnership holds 0.80%, totaling 3.81%.

Jiehuate's main business is power management chips, which are applied in scenarios such as server motherboards and computing acceleration cards. In 2025, Jiehuate achieved a revenue of 2.655 billion yuan, a year-on-year increase of 58.15%, but the net profit attributable to the parent company still incurred a loss of 717 million yuan. This company is still in the stage of "burning money for the market".

Qiangyi Corporation's main business is wafer testing probe cards, which serve the electrical performance testing in the chip manufacturing process. In 2025, Qiangyi Corporation achieved a revenue of 1.012 billion yuan, a year-on-year increase of 57.81%, and a net profit attributable to the parent company of 395 million yuan, a year-on-year increase of 69.43%. In the first quarter of 2026, its net profit attributable to the parent company was approximately 112 million yuan, a year-on-year increase of 697.6%. The company attributed the reason to the explosive demand for AI computing power and the industry boom cycle.

Currently, Shenzhen Hubble Technology Investment Partnership holds a 4.80% stake in Qiangyi Shares. In addition, Qiangyi Corporation will be listed on the Science and Technology Innovation Board at the end of 2025. Its prospectus shows that the company has a client with the code name "Company B", which was widely speculated by the market to be HiSilicon Semiconductor, contributing over 80% of its revenue.

The performance of the three companies that are "capital bound by Hubble" mentioned above will increase synchronously in 2025: Huafeng Technology's revenue will double, Qiangyi Holdings' net profit will increase by nearly 70%, and Jewate's revenue will increase by 58%. This situation seems to reflect a common signal - Ascend's shipment volume is increasing, and demand is beginning to propagate upstream along the industry chain.

In addition, there is a group of upstream suppliers in the Ascend industry chain that have no equity relationship with Huawei but benefit from Ascend's business, such as the A-share new stock king - Yuanjie Technology (688498. SH). In the first quarter of 2026, Yuanjie Technology achieved a revenue of 355 million yuan, a year-on-year increase of 320.94%, which is nearly 60% of its total revenue of 601 million yuan in 2025; The net profit attributable to the parent company in the first quarter was 179 million yuan, a year-on-year increase of 1153%.

The data center CW (continuous wave) light source product is driving its performance growth. The CW laser is the core external light source of the silicon optical module solution, and the computing power cluster extensively uses optical signals for inter card communication. The Ascend supernode connects thousands of cards together, and the optical chip is an indispensable component.

During the performance briefing on April 29th, Zhang Xingang, Chairman of Yuanjie Technology, answered a question from an Economic Observer reporter and stated that currently, high-speed optical modules such as 400G, 800G, and 1.6T mainly require chip products such as 100GEML, 70mW CW, and 100mW CW. With the expansion of computing power clusters and the continuous upgrading of optical module speeds, the power and performance requirements for optical chips are also increasing synchronously.

Zhang Xingang also stated that from the perspective of terminal AI investment, the demand growth of domestic and foreign customers is relatively fast.

Another industry chain company, Tongfu Microelectronics (002156. SZ), is one of the top ten integrated circuit packaging and testing enterprises in the world. In 2025, the company achieved a revenue of 27.921 billion yuan, a year-on-year increase of 16.92%, of which advanced packaging revenue accounted for about 70%; The net profit attributable to the parent company was 1.219 billion yuan, a year-on-year increase of 79.86%.

At the performance briefing on April 28th, Shi Lei, Chairman of Tongfu Microelectronics, told the Economic Observer that the company's revenue target for 2026 is 32.3 billion yuan, with a total capital expenditure plan of 9.1 billion yuan for the year. The company has entered the stage of mass production and introduction of photonic packaging (CPO).

When asked by a reporter from the Economic Observer about the company's cooperation with Huawei, the company's secretary of directors, Jiang Shu, said, "Most of the world's top 20 semiconductor companies and the vast majority of well-known domestic integrated circuit design companies have become company customers

According to relevant financial report information, from Huafeng Technology, Jiehuate, Qiangyi Shares, to Yuanjie Technology and Tongfu Microelectronics, the performance of upstream component companies in Huawei's computing power chain has begun to be concentrated. However, the products of these companies are all installed inside servers, which are not easy for end customers to see.

The other group of companies that truly sell server machines to customers, although Huawei Hubble did not invest in them, their operating conditions can also reflect what stage Huawei Ascend has reached in this industry chain.

For example, Digital China (000034. SZ) will generate AI related business revenue of 33.03 billion yuan in 2025, a year-on-year increase of 47.7%. The self owned brand products mainly include the "Shenzhou Kuntai" series servers based on Huawei Kunpeng and Ascend processors, with a revenue of 7.44 billion yuan, a year-on-year increase of 62.4%. However, the after tax net profit corresponding to this 7.44 billion yuan revenue is only 77.536 million yuan, with a profit margin of just over 1%. In the first quarter of 2026, Shenzhou Digital achieved a revenue of 40.557 billion yuan, a year-on-year increase of 27.62%, and a net profit attributable to the parent company of 236 million yuan, a year-on-year increase of 8.62%. The profit growth rate is still far lower than the revenue growth rate.

Several other companies downstream of Huawei's computing power chain are also in a similar situation.

For example, Tuowei Information (002261. SZ) is a comprehensive strategic partner of Huawei in the fields of Kunpeng, Ascend, HiSilicon, and HarmonyOS. Its main business includes AI server manufacturing and industry software development based on the Ascend platform. According to financial report data, the company achieved a revenue of 3.171 billion yuan in 2025, a year-on-year decrease of 22.79%.

In addition, Chuanrun Co., Ltd. (002272. SZ) specializes in liquid cooling and temperature control systems, and is widely regarded as one of Huawei Ascend's liquid cooling suppliers in the A-share market. According to its annual report, the company's sales revenue of liquid cooling systems and products in 2025 was 225 million yuan, a year-on-year increase of 77.76%, but its net profit attributable to the parent company for the whole year was still a loss of 24.91 million yuan.

At the performance briefing of Chuanrun Co., Ltd. on April 29th, an investor asked, "Is the company the exclusive supplier of Huawei liquid cooling?" The company's secretary of directors, Rao Hong, replied, "No." The Economic Observer reporter then asked about its cooperation with Huawei Ascend. The company's chairman, Luo Yongzhong, said, "Please pay attention to the company's official and designated media disclosure for specific customer cooperation

Shenzhou Digital, Tuowei Information, and Chuanrun Co., Ltd., these three companies respectively specialize in whole machine, system integration, and liquid cooling and heat dissipation. Compared to the concentrated performance explosion of upstream component companies, these three companies are not doing so well, and it seems that the dividends of Huawei's computing power chain have not yet been truly transmitted to their financial reports.

The next step of ascension

According to public information, Huawei released the first generation Ascend 910A chip in 2019, using TSMC's 7nm EUV (extreme ultraviolet lithography) process. TSMC will no longer provide OEM services for Huawei, and Huawei will switch to SMIC's 7nm like process. In 2023, Huawei will launch the 910B chip and mass produce the 910C in early 2025. According to industry insiders, the 910C chip design is a combination of two 910Bs, and according to the original route, the next generation should be a combination of four 910Ds.

After the release of DeepSeekR1 in early 2025, the computing power demand structure of the AI industry has changed. Many companies have found that the cost-effectiveness of investing heavily in training large models is far inferior to directly using DeepSeek's open-source models for local deployment and inference services, because training requires a huge amount of computing power to run continuously for weeks or even months. Therefore, the mainstream demand in the market has shifted from training to inference.

Huawei also adjusted its product plan accordingly, and instead of launching the 910D chip, it jumped directly to the 950 chip.

Huawei's rotating chairman, Xu Zhijun, officially unveiled the Ascend 950 series at the Huawei Connect Conference in 2025 and announced the chip roadmap for the next three years: launching the 950 series in 2026, the 960 series in 2027, and the 970 series in 2028, "continuously evolving at a rate of almost doubling the computing power of one generation per year". Xu Zhijun also said, "Computing power has been and will continue to be the key to artificial intelligence, and even more so, the key to China's artificial intelligence

The 950 series is divided into two versions - the 950PR is designed for the pre filling stage in the inference process (the process where the model reads a large amount of data after the user inputs prompt words, requiring high computing power), and the 950DT is designed for the decoding stage (the process where the model generates answers word by word, requiring high memory bandwidth).

Separating these two stages and running them on different chips is a common solution in the current industry, and Nvidia's latest Rubin series also adopts the same idea.

According to data released by Xu Zhijun at the All Connected Conference, there is still a gap between the Ascend 950 and Nvidia Rubin based solely on chip parameters. Huawei's way of filling the gap between single cards is through supernodes, which connect a large number of chips into a whole through high-speed networks for collaborative computing. Xu Zhijun introduced that Huawei's previous generation Atlas900 supernode fully supports 384 Ascend cards and has deployed over 300 sets. The newly released Atlas950 supernode supports 8192 Ascend cards and consists of 128 computing cabinets and 32 interconnected cabinets, with all-optical interconnection between cabinets. The FP8 total power reaches 8E-FLOPS.

In a podcast on April 15th, when questioned by the host about Huawei's insufficient memory bandwidth, Huang Renxun responded by saying, "Huawei is a networking company that can bring them together... to connect all calculations into a huge supercomputer." Huang Renxun also stated, "Most of the progress in AI comes from algorithmic advancements, not just raw hardware. ?

In addition, advances in algorithms and software tools are also helping Shengteng narrow the gap.

The most criticized aspect of domestic AI chips in the past was not their hardware performance, but their software ecosystem.

After nearly 20 years of accumulation, NVIDIA's CUDA (Unified Computing Architecture) has become the default programming platform used by AI developers worldwide, and the vast majority of AI models and algorithms are developed based on CUDA. If developers want to migrate from Nvidia to Ascend, they need to re adapt the code originally written based on CUDA to Huawei's CANN platform (Huawei's self-developed heterogeneous computing architecture), which is a time-consuming and labor-intensive process in the early stages.

For example, Chen Wei, the chairman of Qianxin Technology, recently used the phrase "climbing snow capped mountains and crossing grasslands" to describe the intensity of early model adaptation work on Ascend.

This situation is improving.

Xu Zhijun announced at the above-mentioned fully connected conference that the CANN compiler and virtual instruction set interface are open, and other software is fully open source. "In the future, open source and product launch will be synchronized. This is a crucial step taken by Huawei in the Ascend software ecosystem, which means that external developers can participate more deeply in the development and optimization of Ascend's underlying tools.

In addition, a team from the School of Computer Science at Peking University also developed an open-source programming language called TileLang last year, which allows developers to directly perform underlying performance tuning on Ascend chips without relying on Nvidia's CUDA. Huawei has also launched a new programming tool PyPTO on the CANN platform.

According to Huawei's official DeepSeekV4 adaptation announcement, PyPTO has shortened the development cycle of the new generation model operator (the basic unit that executes specific computing tasks in AI models) to the level of days. Developers can write a set of code to run on different generations of Ascend chips.

The software ecosystem is improving, and hardware validation is also advancing.

DeepSeekV4 can already be deployed for inference on Ascend. The so-called reasoning refers to the process of receiving user questions and generating answers after the training of the large model is completed. Compared to training, reasoning requires higher response speed.

According to official data from Huawei, the A3 supernode (previous generation product) can achieve a decoding throughput of over 2000TPS (TPS, the number of tokens processed per second) with 64 cards deployed on V4 Flash. The inference latency of the Ascend 950 supernode can reach millisecond level. Huawei Ascend CANN has completed the continued training adaptation of V4 Flash based on A3 supernode (i.e. continuing to train with new data on the existing model), and provided training reference implementation in the open source community.

In fact, in the field of domestic computing power, there are also other routes besides Huawei Ascend.

Haiguang Information (688041. SH) and Zhongke Shuguang (603019. SH) are a highly regarded combination in the industry, known as the "two lights" in the market. Among them, Haiguang Information produces chips (CPU and DCU acceleration cards), while Zhongke Shuguang produces servers. The two companies have both equity ties and long-term collaboration. Unlike Ascend's self-developed architecture, Haiguang Information adopts the x86 compatibility route, and the cost of developers migrating from the NVIDIA CUDA ecosystem is lower, forming a differentiated complementarity between the two.

During the performance briefing on April 29th, Sha Chaoqun, General Manager of Haiguang Information, responded to a question from an Economic Observer reporter and stated that DCU (its self-developed AI acceleration computing chip) products have covered over 20 key industries and more than 300 application scenarios, and have been adapted to 365 mainstream large-scale models such as DeepSeek, Qwen3, Hybrid, and Zhipu, covering 99% of non closed source large-scale models worldwide.

Sha Chaoqun positions the compatibility route of Haiguang Information as "not a short-term transition, but a long-term differentiation path that fits the reality of the industry". He also stated that the concentration of AI inference deployment by major companies has brought new incremental opportunities for domestic computing power.

During the performance briefing on April 27th, Li Jun, General Manager of Zhongke Shuguang, responded to a question from an Economic Observer reporter and stated that in the first quarter of 2026, its overall operating revenue increased by 23.71% year-on-year. "Benefiting from the rapid development of domestic substitution and the AI industry, the current demand for general servers and AI servers is strong.

Although Ascend and "Two Lights" have different technological paths, they face common shortcomings. Sha Chaoqun believes that at present, what domestic computing power needs to supplement the most are "deep optimization of ultra large scale clusters, operator optimization, and vertical industry integration solutions".

Pan Helin, a member of the Information and Communication Economy Expert Committee of the Ministry of Industry and Information Technology, told the Economic Observer that computing chips can solve the problem of insufficient computing power through parallel computing and chip stacking, and Ascend has already achieved this. But he also pointed out that Huawei is not a manufacturer of chips, and advanced process chips have production capacity and technological shortcomings. "Supply chain tension and real demand are both present, and the problem of supply chain tension may be more serious.

In a research report released in April 2026, Huatai Securities stated that 2026 is the "first year of domestic super nodes" and estimated that the market space related to domestic super nodes in 2028 will be 341.4 billion yuan. Bernstein Research, a globally renowned equity research firm, has a more optimistic forecast, not only optimistic about Huawei's leading position in the Chinese market for AI chips, but also predicting that Huawei's AI chip market share in China is expected to exceed 50% by 2026.