Behind Enflame Technology's IPO approval: training product revenue accounts for only 1.15%, with multiple executive changes last year

Source of this article: Era Weekly, Author: Zhu Chengcheng

Following Moore Threads, MetaX, Biren Technology, and Tianshu Zixin, another domestic AI chip company is set to list on the A-share market.

On June 15, Shanghai Enflame Technology Co., Ltd. (hereinafter referred to as 'Enflame Technology') received approval from the Shanghai Stock Exchange’s Listing Review Committee for its IPO application on the STAR Market.

For a long time, Enflame Technology has been grouped by outsiders into the so-called 'Four Little Dragons of Domestic GPUs.' However, in its IPO inquiry response, the company actively distanced itself from the general-purpose GPU (GPGPU) route. Enflame emphasized that its core products are based on Domain-Specific Architecture (DSA), not general-purpose GPU architecture, and specifically explained the differences between the two in terms of technical approach and application scenarios.

Behind this statement lies a development path distinct from most domestic GPU chip companies. According to the prospectus, in 2025, revenue from Enflame Technology's training and training-inference integrated products accounted for only 1.15% of its total revenue from AI accelerator cards and modules, with the vast majority coming from inference products.

"Enflame’s choice of the DSA approach essentially trades specialization for efficiency and ecosystem independence for controllability. Given the current boom in the inference market, this strategy has already demonstrated its commercial value," said Zhang Guobin, a veteran semiconductor expert, to Times Weekly. He added that the long-term potential of DSA depends on the maturity of its software ecosystem, the pace of integration with general-purpose GPUs (GPGPUs), and whether Enflame can move beyond its reliance on Tencent toward diversified growth.

Notably, during the IPO process, there was an adjustment to the executive role of Enflame Technology’s founder, Zhao Lidong.

According to the prospectus, Enflame Technology was founded in March 2018. Zhao Lidong is the company’s founder, chairman, and one of its actual controllers, and had served as general manager for an extended period. In September 2024, Zhao resigned as general manager and stepped down from the executive management team, with co-founder Zhang Yalin succeeding him as general manager.

From that point until just before the IPO filing, although Zhao continued to serve as chairman, he was absent from the company’s list of senior executives. It wasn’t until one month before submitting the IPO application that he rejoined the executive roster, taking on the role of board secretary.

For a company at a critical juncture of going public, changes in the roles of its founder and core management team often attract market attention. On June 16, a Times Weekly reporter repeatedly called Enflame Technology but was unable to get through. The reporter subsequently sent an interview request to the company’s public email address, but had not received a response by the time of publication.

Avoiding the General-Purpose GPU Path

Currently, AI computing chips primarily follow three technical approaches: general-purpose GPUs (GPGPUs), ASICs, and FPGAs. Among these, DSA is regarded as a key evolutionary direction for ASICs in the AI era.

In its response to regulatory inquiries, Enflame stated that its DSA architecture optimizes core algorithms, models, and application scenarios in AI during the chip design phase, hardwiring essential computational, interconnect, and storage optimization features directly into the chip architecture. This approach delivers both specialized acceleration efficiency and programmable iteration capabilities, offering significantly better adaptability than narrow-defined ASIC architectures and outperforming general-purpose GPGPUs in AI-specific scenarios.

From an industry landscape perspective, companies such as NVIDIA, Muxi Semiconductor, Biren Technology, and Tianshu Zhixin primarily follow the general-purpose GPU route, whereas Google’s TPU, Amazon’s Trainium, Huawei’s Ascend, Kunlunxin, Cambricon, and Enflame Technology represent key players pursuing the DSA approach.

As large AI models gradually transition from the training phase to real-world deployment, inference demand is growing rapidly, and market attention on DSA architectures is also rising. According to Goldman Sachs Global Investment Research's model forecast, the share of non-GPGPU architecture chips (primarily DSA-based) in AI servers is expected to increase from 38% in 2025 to 45% in 2027, further narrowing the gap with GPGPUs.

‘Over the next two to three years, the inference market will favor DSAs, but it will be difficult to capture a meaningful share in the high-end large-model training segment,’ Zhang Xiaorong, Dean of DeepTech Research Institute, told The Weekly Times reporter. ‘In theory, as inference demand continues to grow, the DSA approach—being more power-efficient and cost-effective—is well-suited for large-scale cloud deployments, has lower dependency on NVIDIA’s software stack, and aligns better with domestic procurement requirements. Although the inference market is projected to surpass the training market in scale, general-purpose GPUs still hold a significant advantage in pure training scenarios involving ultra-large trillion-parameter models, a space where DSAs are unlikely to gain traction in the near term.’

‘Under China’s national strategy for autonomous and controllable computing capabilities, the DSA path has a clearly defined commercial window of opportunity for at least the next three to five years,’ said Zhang Guobin. He added that customers require integrated ‘training-and-inference’ solutions, and DSA’s limited versatility in training scenarios could constrain its share in such integrated deployments. Enflame’s fourth-generation L600 chip was specifically designed to address this shortcoming.

Training product revenue accounts for only 1.15%

Unlike most domestic AI chipmakers, Enflame Technology currently derives nearly all its revenue from the inference market.

Training products typically command higher prices and gross margins than inference products. In 2024, Enflame’s second-generation training card had a unit price of RMB 36,100 and a gross margin of 73.67%; by comparison, its second-generation inference card was priced at RMB 18,400 with a gross margin of 58.76%. The third-generation inference card saw further declines in both price and margin, dropping to RMB 13,100 per unit in 2025 with a gross margin of 34.53%.

Enflame Technology explained that training cards are priced higher because they feature superior hardware configurations, higher bill-of-materials costs, and greater technical complexity compared to inference products. Additionally, pricing for domestic AI accelerator cards is primarily driven by overall deployment cost-effectiveness. Since individual inference cards have performance limitations that increase customers’ total deployment costs, clients tend to push for even lower unit prices.

More noteworthy than the differences in product pricing and gross margins is the company’s revenue composition. From 2023 to 2025, revenue from Enflame’s training and integrated training-inference accelerator cards and modules accounted for 17.65%, 0.54%, and 1.15% respectively of its total AI computing accelerator card and module revenue. In 2025, AI computing accelerator cards and modules represented over 86% of the company’s total revenue. In other words, over the past two years, the vast majority of the company’s revenue came from inference products.

In fact, this proportion is significantly lower than that of its peers. During the same period, Moore Threads reported that revenue from its training and integrated training-inference products accounted for 52.92%, 99.32%, and 98.73% respectively; while Tianshu Zhixin reported figures of 82.86%, 72.89%, and 63.27%.

Industry insiders believe this divergence stems not only from differing technical approaches but also from the current ecosystem dynamics of the AI industry.

"The primary reason is that the entire industry is struggling with sales, and Enflame's own strategy has exacerbated the issue," Zhang Xiaorong noted, adding that training cards are largely locked into NVIDIA's ecosystem, and switching to domestic alternatives would require code modifications, incurring substantial costs.

However, the inference market is not Enflame Technology's sole target. In its response to regulatory inquiries, the company stated that with the mass production of its fourth-generation training-and-inference integrated product L600 and the delivery of its hyper-node systems, it will continue expanding applications in the training segment. However, due to industry demand structures, supply chain dynamics, and the pace of technological iteration, inference scenarios will remain the core application focus for its future training-and-inference integrated products, while the proportion of training applications will steadily increase.

Multiple senior executive changes occurred last year

According to the prospectus, as of the date of signing the prospectus, the actual controllers of Enflame Technology are Zhao Lidong and Zhang Yalin. Together, they directly hold 17.93% of the company’s shares and indirectly control an additional 10.21% through employee stock ownership platforms, giving them combined voting control of 28.14%.

As disclosed, Zhao Lidong and Zhang Yalin, as co-founders, have served respectively as Chairman and Chief Executive Officer, and General Manager and Chief Operating Officer since the company's inception, consistently steering the company’s strategic direction and major decision-making.

However, a reporter from Time Weekly noted that Enflame Technology’s senior management underwent significant adjustments during the reporting period. From September 2024 to December 2025, Zhao Lidong did not appear on the company’s executive roster. Specifically, between September 2024 and March 2025, Zhang Yalin assumed the role of General Manager in place of Zhao due to work arrangements, while Zhao remained Chairman. It was not until December 2025 that Zhao Lidong was appointed Secretary of the Board and reinstated to the executive list.

Additionally, Enflame Technology experienced successive changes among its executives: in March 2025, Board Secretary Zhu Dapeng resigned for personal reasons; in July 2025, CFO Wang Hongli stepped down due to work arrangements. Notably, between July and November 2025, Zhang Yalin was the sole remaining executive listed in the company’s management roster.

According to the 'Administrative Measures for the Registration of Initial Public Offerings,' issuers applying for an initial public offering and listing on the STAR Market must demonstrate stable control and management teams, with no material adverse changes in core business operations or board and senior management personnel over the past two years, and core technical personnel must also remain stable without any material adverse changes during the same period.

Lu Hong, founder of M&A expert platform Binggou Daren, told Time Weekly, "Under regulatory requirements, successive departures of senior executives—including board secretaries and CFOs—directly raise red flags regarding financial compliance and constitute grounds for rejection during review."

In Lu Hong’s view, a company’s investment value hinges on its technical capabilities. Enflame’s DSA architecture, four generations of chips, and deployment in Tencent’s business scenarios have been key factors enabling it to reach the approval stage. Given the long cycles and heavy investments required in AI chip development, team stability and corporate governance will determine whether technological innovation can be continuously iterated and successfully commercialized. Executive instability risks triggering core R&D talent attrition, project delays, and loss of trust from major clients.

28K Views