The "Four Little Dragons of GPU" have all gathered on the STAR Market. How is the actual competitiveness of Enflame, the last one to get listed among them?
The last of the "Four Little Dragons" in the GPU sector has gone public.
On September 11, 2026, Enflame Technology was listed on the Sci-Tech Innovation Board (STAR Market) of the A-share market. It opened at a price of 410 yuan per share and closed at 397 yuan on the trading day, which is 2 to 3 times its issue price of 142.18 yuan.
In the previous subscription process, Enflame Technology was regarded as another "high-yield new share". Calculated based on the closing price, each lot of the share can bring a revenue of 128,000 yuan. It is worth noting that due to the excessive number of online subscriptions during the new share offering, Enflame Technology once launched the callback mechanism, transferring 3.443 million shares from the offline tranche to the online tranche. After the callback, the total number of shares issued online reached 10.3285 million, with the final winning rate standing at 0.02455315%.
At the close of the listing day, the market value of Enflame Technology reached 170.85 billion yuan. The company was originally founded in a temporary office of less than 100 square meters in Zhangjiang, Shanghai.
In the spring of 2018, Zhao Lidong and Zhang Yalin embarked on their entrepreneurial journey, with the original aspiration of "developing high-end chips and focusing on hard technology". Eight years later, this spark of entrepreneurship finally landed on the STAR Market. However, the 7.03 million investors who snapped up this new share may not know that each share they won contains a long journey from Silicon Valley to Shanghai, and from AMD to Enflame.
01 Veteran from AMD, Tsinghua Chip Dream and Tencent's Investment Support
After graduating from Tsinghua University, Zhao Lidong has been deeply engaged in the semiconductor industry in Silicon Valley for more than 20 years. He once served as Senior Director of AMD's Product Engineering Department, returned to China in 2014 to join Tsinghua-affiliated Unis Group, and once served as Vice President of the group before resigning to start his own business.
In March 2018, Zhao Lidong teamed up with his former AMD colleague Zhang Yalin to start the chip business, and registered and established Enflame Technology.
Since the establishment of the company, the two have always been partners standing side by side. When answering investors' questions before the IPO, Zhao Lidong referred to himself and Zhang Yalin as the actual controllers of Enflame Technology. According to the documents disclosed by Enflame Technology before listing, Zhao Lidong and Zhang Yalin each directly hold 8.96% of the company's shares, and jointly control about 28.14% of the voting rights through arrangements such as employee stock ownership platforms. The two have signed a concert party agreement.
At present, Zhao Lidong serves as Chairman and Secretary of the Board of Directors, and Zhang Yalin serves as CEO. According to the concert party agreement signed by the two, Zhang Yalin is mainly responsible for matters such as personnel appointment and removal, product technology, and R&D plans, while other matters such as investment and financing, and strategic planning are subject to Zhao Lidong's opinions.
Data source: Enflame Technology's prospectus
It is quite interesting that Zhao Lidong is a rare chairman who also serves as the secretary of the board of directors, which is due to the previous management changes. Former board secretary Zhu Dapeng resigned in March 2025, and the financial director Wang Hongli also left the company later. For a period of time, there was only one executive Zhang Yalin left in the management. In November 2025, the company internally selected new management, and then Zhao Lidong concurrently served as the secretary of the board of directors.
The fact that the founder personally takes charge of these work is not only due to the management changes, but also probably because he is the person who knows the industrial capital best in the company. Zhao Lidong graduated from the 1985 class of the Department of Electronic Engineering of Tsinghua University, which is the well-known "EE85 Class" in the industry. Many founders of listed semiconductor companies have emerged from this class, including Yu Renrong (founder of Will Semiconductor, the richest man in China's chip industry), Shu Qingming (co-founder of GigaDevice), Feng Chenhui (co-founder of Maxscend), Zhao Lixin (founder of GalaxyCore), etc.
Shu Qingming started his business at the age of 38, Yu Renrong at 41, while Zhao Lidong started his business at 51. Different from the trend of advocating "young geniuses" in AI entrepreneurship in recent years, the successful practitioners in the computing power hardware field are usually "veteran players" in the industry: due to the long return on investment cycle and high risks of the semiconductor industry, founders with strong industrial background are more likely to gain the favor of capital and industry resources.
In hard technology entrepreneurship, the value of industrial resources is often far more than funds. In the 1980s, Bill Gates' Microsoft, through cooperation with IBM, provided the DOS operating system for IBM's personal computers, thus establishing the standard of the software industry. In the 1990s, after Jensen Huang founded NVIDIA, relying on TSMC's advanced manufacturing process, the company was able to compete with industry giants in the GPU field.
Microsoft had IBM, NVIDIA had TSMC, and Enflame has Tencent. Since its establishment, Enflame has completed more than 10 rounds of financing, among which Tencent participated in the Pre-A, Series A, Series B, Series C, Series C+ and Series D rounds. In the customer list of the prospectus, Tencent is Enflame's largest customer. In 2025, 80% of Enflame's revenue came from Tencent, mainly from AI accelerator cards and module products.
Among them, the speech-to-text function of WeChat is supported by Enflame's inference cards. The cooperation between the two sides began in 2019. In the early stage, they adopted the AVAP model (supplying to system integrators as designated by end customers), and switched to direct sales in 2025. Enflame's three generations of AI accelerator cards have been widely deployed in hundreds of internal business scenarios of Tencent, including WeChat speech-to-text, Tencent Meeting minutes, content understanding, and advertisement recommendation.
02 The fastest-growing "Four Little Dragon" and the smallest DSA company
Among the four "Little Dragons", only Enflame's chips adopt the DSA architecture (Domain-Specific Architecture).
Compared with general-purpose GPUs, this architecture cuts out the graphics rendering features of GPUs, and focuses on general-purpose computing and tensor/matrix computing. Enterprises that also adopt the DSA architecture include Google, Cambricon, HiSilicon of Huawei, and PingTouGe of Alibaba. The other three "Little Dragons" adopt the general-purpose GPGPU route, and are mainly compatible with or adapted to NVIDIA's CUDA ecosystem.
The advantage of DSA lies in cost performance. On the premise of sufficient production volume, it has higher energy efficiency ratio and lower unit computing power cost. On the other hand, the DSA architecture does not have the complete ecosystem of the CUDA architecture, so it needs self-developed compilers and customized operator libraries, which usually needs to form a closed loop within the enterprise ecosystem. Therefore, enterprises adopting DSA architecture are usually large AI manufacturers themselves, or have a single large customer with sufficient scale.
The flip side of high cost performance is that DSA lacks universality, and its performance on AI training scenarios with high computing accuracy requirements and more diverse and complex processes is not as good as that of general-purpose GPUs represented by NVIDIA.
In other words, Enflame's chips are good at inference but relatively weak in training. The prospectus shows that more than 80% of Enflame's revenue from AI accelerator cards and modules from 2023 to 2025 came from inference products. Inference products pay more attention to low latency, high throughput and low power consumption, and customers are more sensitive to cost and cost performance, with lower pricing space. As a result, Enflame's gross margin is lower than that of its peers, which stood at 22.6%, 30.6% and 31.8% respectively from 2023 to 2025.
But on the other hand, inference better reflects the future demand of computing power, which is also faithfully reflected in Enflame's growth rate.
From 2023 to 2025, Enflame's revenue reached 301 million yuan, 722 million yuan and 990 million yuan respectively, with a three-year compound growth rate as high as 81%. In the first half of 2026, its revenue has reached 1.12 billion yuan, exceeding the total revenue of 2025. Among them, the growth in 2024 was mainly driven by price increase, with the average unit price rising by 60.8%; the growth in 2025 was realized by production expansion, with the sales volume increasing by 197.8%.
As a typical model of the chip industry that "invests for many years and realizes returns at one go", Enflame has not yet made a profit by the first half of 2026, and its R&D investment is several times higher than its revenue: the R&D expenses from 2023 to 2025 were 1.229 billion yuan, 1.312 billion yuan and 1.135 billion yuan respectively, accounting for 408.01%, 181.66% and 114.63% of the revenue.
According to the calculation of Tonghuashun, Enflame Technology's price-to-sales ratio (PS) is 94 times, which is higher than that of Cambricon (market value 654 billion yuan, PS 68 times), Moore Threads (market value 168.2 billion yuan, PS 66 times), and roughly consistent with that of Muxi Technology (market value 195.7 billion yuan, PS 95 times). The high expectation comes not only from the liquidity brought by the first day of listing, but also from its faster revenue growth.
03 Fierce Competition in the Chip Industry
At present, Enflame adopts the strategy of advancing multiple generations of products simultaneously and making breakthroughs in echelons.
The fourth-generation training and inference integrated chip L600 was released in July 2025, which is expected to be mass-produced in the second half of 2026 and delivered in batches in 2027. The fifth-generation and sixth-generation chips are in the late stage of R&D and early stage of R&D respectively. In addition, the AI software and hardware collaborative innovation project will run through the R&D process of these two generations of products, to make up for the shortcomings of the full-stack capability from four directions: the core technology of super ten-thousand-card cluster, AI software stack, core IP localization, and cloud AI chips based on domestic manufacturing processes.
The prospectus shows that 1.503 billion yuan (about 25%) of the raised funds from the company's listing will be used for the R&D of the fifth-generation chip, which is expected to complete tape-out and be launched in 2027; 1.197 billion yuan (about 20%) of the funds will be used for the R&D of the sixth-generation chip, which is expected to be launched in 2029; the remaining more than half of the raised funds will be invested in the software and hardware collaborative innovation project.
The first challenge in the future lies in the progress of R&D and mass production. Advancing multiple generations of products simultaneously means huge pressure. As of press time, L600 has not announced the realization of mass production yet. Whether it can achieve mass production as planned will directly affect whether the company can expand its market share in the training chip market.
Another challenge lies in the ecosystem construction.
Enflame's DSA route is not directly compatible with the complete CUDA ecosystem. When customers migrate their existing models to Enflame's platform, they usually need to make adaptations at the levels of framework, compiler, operator and cluster software. The cost of migration depends on the model type, business scenarios and the existing tool chain.
This is also the value of Enflame's cooperation with Tencent. Tencent not only provides orders for Enflame, but also offers opportunities for joint software and hardware adaptation in large-scale business scenarios. In the future, Enflame needs to prove whether it can replicate the product capability formed by this kind of joint verification to more customers, instead of relying only on one large customer.
At the same time, the model-computing power customer relationship in the industry is becoming more and more stable. In September 2026, DeepSeek planned to purchase 160,000 Huawei Ascend 950DT chips, and the two sides are expected to jointly build a large data center in Inner Mongolia. JD Cloud has in-depth cooperation with Moore Threads, and plans to purchase its GPUs to build a domestic computing power cluster of 100,000 cards. There are also reports that ByteDance has pre-purchased more than 200,000 chips from Cambricon before.
Large model manufacturers are ensuring supply chain security through "computing power binding", and the relationship between Enflame and Tencent is essentially a product of this trend.
From the perspective of broader demand, domestic chip R&D still relies on international suppliers in links such as advanced process design, EDA tools, and IP core authorization. At the same time, international trade frictions keep changing, and there is uncertainty in the supply rhythm and stability of key links in the supply chain. From the perspective of the whole industry, the entire market ultimately depends on whether AI applications can be implemented on a large commercial scale. If the progress of AI applications fails to meet expectations, the over-construction of computing power infrastructure may face the pressure of low return.
The story of Enflame Technology is a microcosm of China's hard technology entrepreneurship. From the AMD veteran to the Tsinghua chip dream, from Tencent's full support to landing on the STAR Market, Zhao Lidong and Zhang Yalin have spent eight years completing a journey of "spreading the fire of domestic chips".
But listing is not the end, but a new starting point: how to break through the ceiling of the training market while maintaining the energy efficiency advantage of the DSA architecture, how to transform the "dependence" on Tencent into real "symbiosis", how to increase the market share from 1.7% to a higher level in the trillion-yuan window period of domestic substitution, are the questions that Enflame must answer in the next eight years.
This article is from the WeChat official account "Bianbao", written by Zhang Jingwei, and published with authorization from 36Kr.