From the bottom up, it has not followed the path paved by NVIDIA. Produced by | Zhangtongshe
Today (September 11), Shanghai Enflame Technology Co., Ltd. (hereinafter referred to as "Enflame Technology") officially listed on the SSE STAR Market.
The offering price for Enflame Technology this time was CNY 142.18 per share, the highest among the "Four Little Dragons of Domestic GPUs." As of today's close, the company's total market capitalization reached CNY 170.9 billion.
In less than a year, Moore Threads and Muxi Semiconductor have successively listed on the A-share SSE STAR Market, and Biren Technology has been listed on the Hong Kong Stock Exchange. The bell-ringing of Enflame Technology means that the last piece of the puzzle for the "Four Little Dragons of Domestic GPUs" has officially been completed.
A Tough Path Incompatible with CUDA
In the domestic GPU track, Moore Threads, Muxi Semiconductor, Biren Technology, and Enflame Technology are collectively known in the industry as the "Four Little Dragons," but the technical routes of the four have long diverged.
Moore Threads focuses on full-featured GPUs, balancing graphics and general-purpose computing; Muxi Semiconductor highlights its full-stack GPU product line and multi-scenario coverage; Biren Technology focuses on high-performance GPGPU and AI computing systems; Enflame Technology centers on cloud AI chips, accelerator cards, and intelligent computing clusters, emphasizing hardware-software collaboration and large-scale deployment capabilities.
What makes Enflame Technology most "special" is that, from the bottom up, it has not followed the path paved by NVIDIA.
In terms of hardware, its AI chips do not adopt the currently mainstream NVIDIA GPGPU architecture, but instead use the original self-developed DSA (Domain-Specific Architecture) architecture; in terms of software, it also actively chooses to be incompatible with the CUDA ecosystem.
Let's start with hardware. The reason for choosing the DSA architecture is mainly based on AI computing needs and technological development directions. Instead of catching up in general-purpose architectures, it is better to focus on AI computing scenarios for collaborative hardware-software optimization. DSA combines dedicated acceleration efficiency with programmable iterative capabilities, enabling higher energy efficiency and computing efficiency under specific workloads.
Then let's look at software. The CUDA ecosystem is highly bound to NVIDIA hardware. Over years of accumulation, it has built a closed moat for NVIDIA and, with the expansion of its product layout, has spread to more and more industries.
Once switching to a new non-CUDA architecture, customers would almost have to rewrite the entire set of software, making the migration cost extremely heavy. For this very reason, the vast majority of AI developers globally still rely heavily on CUDA, and other GPU companies in the industry mostly choose to be compatible with CUDA.
Enflame Technology, however, chose a difficult path. The company has independently developed a full-stack AI computing and programming software platform, "TopsRider," which includes drivers, compilation languages and compilers, operator libraries, and toolchains, to lower the threshold for model development and migration based on its own hardware.
The impact of this "copy nothing" path is obvious. On the one hand, the entire hardware and software chain is truly independent and controllable; on the other hand, building a complete technical system from scratch requires high R&D investment and a long ecological cultivation cycle, making it difficult to seize the domestic existing market through the "domestic substitution of NVIDIA."
Enflame Technology has also not yet achieved profitability. In 2023, 2024, and 2025, Enflame Technology's operating revenues were CNY 301 million, CNY 722 million, and CNY 990 million, respectively, and its net profits attributable to the parent company were -CNY 1.665 billion, -CNY 1.510 billion, and -CNY 1.164 billion, respectively. In the first half of 2026, the company's operating revenue was CNY 1.12 billion, and its net profit attributable to the parent company was -CNY 632 million.
The prospectus attributes the losses to the need for continuous large-scale investments in cloud AI chips and software stacks, the long cycle of customer verification and adaptation, and the fact that the revenue scale has not yet covered rigid R&D expenditures.
However, the good news is that signals of a turning point have emerged. The company expects its revenue in the first three quarters of 2026 to be between CNY 2.3 billion and CNY 3 billion, a year-on-year increase of 325.78% to 455.36%; the net loss attributable to the parent company will narrow to between CNY 860 million and CNY 700 million. If the revenue is realized as scheduled and the gross profit margin maintains a medium level, the company can basically achieve a break-even on the consolidated statements in 2026, and there is also the possibility of achieving profitability only in 2027.
Continuous massive R&D investment has destined Enflame Technology's high dependence on capital—and this is precisely the fundamental reason why Shanghai state-owned capital and Tencent have been able to accompany it all the way to the end.
Behind Enflame: Two Major Shareholders
Looking back on Enflame's growth path, it is inseparable from two pillars.
One is Shanghai state-owned capital. From the first investment at the start-up to the last round before the IPO, Shanghai state-owned capital has appeared at almost every critical node.
In June 2018, only three months after the company's establishment, Shanghai Science and Technology Innovation Group under Shanghai State-owned Capital Investment, Jiaxing Zhenge, Yihe Capital, and others invested CNY 15.5 million, becoming one of the earliest investors; in 2022, Shanghai International Group led the CNY 2 billion Series D round; in the last round before the IPO, the IC Fund under Shanghai State-owned Capital Investment led the investment of more than CNY 300 million, conveying clear confidence to the market.
What is more valuable is not the funds themselves, but the industrial chain ties.
Relying on invested resources, the IC Fund assisted Enflame in connecting with scarce domestic supply chain capacity. Its layout in advanced process foundry, HBM (High Bandwidth Memory), Advanced Packaging, high-speed interconnection, and other projects exactly constitutes the key upstream supporting facilities for the domestic landing of Enflame's chips, helping the enterprise to connect the entire chain of "design—manufacturing, packaging and testing—application."
From an office in Lingang to a star on the SSE STAR Market with a market capitalization of over CNY 170 billion today, Enflame Technology has long become one of the most representative samples in Shanghai's cultivation of a strategic echelon of domestic AI chips.
The other is Tencent. As the largest external shareholder of Enflame Technology, its figure runs through multiple key rounds.
In November 2018, Tencent Technology invested CNY 286 million at once, becoming the largest single shareholder at that time with a 25% stake, and subsequently added investments in multiple consecutive rounds. In the Series E round in December 2024, 36 investors including Shanghai Industrial Investment and Tencent injected a total of over CNY 2.7 billion, at which time Enflame's pre-money valuation was about CNY 17.5 billion.
More importantly, Tencent is not just a shareholder, but also a core customer and an industrial partner with deep strategic synergy.
From 2023 to 2025, Enflame Technology's total revenue from direct sales to Tencent and the AVAP (designated end customer is Tencent) model were CNY 100 million, CNY 273 million, and CNY 830 million, respectively, accounting for 33.34%, 37.77%, and 83.79% of the revenue in the same period.
National-level applications such as WeChat and Tencent Cloud have provided Enflame with complete scenarios from early verification to large-scale deployment.
Since the small-batch pilot in 2019, the launch of the first-generation training chip "Suisi 1.0" on Tencent Cloud in 2020, to the release of the AI inference chip "Zixiao" jointly developed by both parties in 2021, Enflame has cumulatively delivered tens of thousands of cards in scenarios such as speech recognition, online conferencing, video recommendation, AI search, and game agents, quickly running through the "technology closed-loop—business closed-loop."
The dual boost of capital and scenarios must ultimately fall on the products and technologies themselves. And listing is just the midpoint.
After Listing, the Real Test
Over eight years, Enflame Technology has launched five chips across four generations, all of which were successfully taped out on the first try, building a core technology product system of three major categories: chips and hardware, software and programming platforms, and computing power and cluster solutions, and practicing a multi-generation parallel mechanism of "pre-research one generation, R&D one generation, mass production one generation."
In 2019, the first-generation training chip "Suisi 1.0" and Yunsui training cards were introduced, and a thousand-card liquid-cooled computing power cluster was built for Zhijiang Laboratory; the second-generation training and inference chips have cumulatively shipped over 30,000 units.
The third-generation S60 was released in 2024, deployed in the Gansu Qingyang Intelligent Computing Center, building the country's first 10,000-card domestic computing power inference cluster. Currently, the cumulative shipment of S60 has reached 160,000 units, and its landing scale ranks in the first echelon of domestic AI accelerator cards, supporting over 300 application scenarios.
In July 2025, at the World Artificial Intelligence Conference that year, Enflame released the latest generation of integrated training and inference product "Enflame L600" and the Yunsui OGX series products.
In the synergy of "cloud-model-chip-application," Enflame Technology has also carried out numerous explorations and innovations. When rapidly adapting to mainstream open-source models such as DeepSeek-V4 and Hunyuan Hy4 preview, it adopts a self-developed operator generation agent.
On August 28, 2026, Tencent Hunyuan released and open-sourced the new generation large model Hy4 preview, which fully considered Enflame's product characteristics at the initial stage of its design. The two parties conducted in-depth joint design with "chip-model synergy," enabling the new model structure to fully leverage Enflame's hardware advantages. Relying on its full-stack architecture, Enflame also achieved deep matching of model capabilities, inference performance, and computing power efficiency.
The hardest fruit borne from this closed loop of "shareholder as customer" precisely proves that Enflame's technical route can go all the way from a single chip to a model.
Today, when the autonomy of global computing power has risen to a national strategy, what Enflame Technology pursues is that domestic computing power does not have to fight head-to-head with NVIDIA in the existing battlefield of training, but to walk its own differentiated route with an independent architecture plus an ecological closed loop.
Text | Zhouyu Editor | Yida