On January 22, according to information disclosed by the Shanghai Stock Exchange, the IPO application of Shanghai Suiyuan Technology Co., Ltd. (hereinafter referred to as Suiyuan Technology) for the STAR Market has been accepted, with CITIC Securities acting as the sponsor. This signifies that another leading domestic AI chip company is poised at the threshold of the secondary market, following recent moves by Moore Thread, MetaX Integrated Circuit, and Biren Technology.
The prospectus reveals that the company plans to issue no less than 43.0352 million and no more than 68.35 million new shares, accounting for no less than 10% and no more than 15% of the total share capital after the issuance. It aims to raise approximately 6 billion yuan, with the funds earmarked for R&D and industrialisation projects based on its fifth-generation AI chip series, its sixth-generation AI chip series, and an advanced AI hardware-software collaborative innovation project. According to equity transfer information disclosed in the prospectus, Suiyuan Technology was valued at 18.2 billion yuan as of August 2025.
Suiyuan Technology's founding team bears a distinct "AMD lineage"; the company's co-founder and CEO, Zhao Lidong, previously served as Vice President of Tsinghua Unigroup and had a seven-year tenure at AMD earlier in his career, holding core positions such as Senior Director of the Computing Business Unit, with experience traceable back to the 1990s at the veteran graphics chip company S3 Graphics. Another co-founder, Zhang Yalin, also hails from AMD, where he served as a Senior Chip Manager and Technical Director of the China R&D Center.
Regarding the equity structure, Suiyuan Technology has no single controlling shareholder; Zhao Lidong and Zhang Yalin are the joint actual controllers. The prospectus shows that through direct shareholding and an employee持股 platform, the duo collectively control 28.1% of the company's voting rights.
Suiyuan Technology is widely regarded as a "Tencent-backed" AI chip unicorn. Tencent not only maintains long-term and deep operational synergy with Suiyuan but is also its core capital supporter. The prospectus indicates that Tencent Technology and its concert party, Suzhou Paiyi, collectively hold a 20.26% stake in Suiyuan Technology, making them the largest shareholder. Yao Leiwen, Managing Director of Tencent Investment, is a member of Suiyuan Technology's board of directors.
As a significant shareholder and a major customer, Tencent's dual role provides substantial support for the commercial deployment of Suiyuan Technology's AI chips.
Collaboration between Suiyuan Technology and Tencent can be traced back to 2020. At that time, Suiyuan's first-generation training chip, "Suisi 1.0," had already completed internal testing and was deployed on Tencent Cloud. In 2021, the AI inference chip "Zixiao" released by Tencent was a joint development outcome of the two parties. This "customized" cooperation model allowed Suiyuan's products to be refined and validated within Tencent's business scenarios, such as WeChat voice recognition and advertising recommendation algorithms.
During the reporting period, Suiyuan Technology's sales revenue from the Tencent ecosystem accounted for a relatively high proportion. For example, in the first three quarters of 2025, the top five customers collectively contributed 96.41% of revenue, with direct sales to Tencent Technology (Shenzhen) accounting for 57.28%. When combining revenue where the end customer was designated as Tencent under models like AVAP, the company's related-party sales to Tencent constituted 71.84% of its total operating revenue. Regarding accounts receivable, as of the end of September 2025, the balance from Tencent Technology (Shenzhen) accounted for 29.92% of the top five customer receivable balances. Suiyuan Technology stated that it expects the situation of high sales concentration to Tencent to persist for a certain period in the future. If new customer expansion falls short of expectations, Tencent's procurement strategy undergoes significant changes, or the cooperative relationship is replaced by other suppliers, it could lead to reduced purchases by Tencent, adversely affecting the company's business development and operating performance.
Furthermore, Suiyuan Technology has recently begun actively expanding into government-led intelligent computing center projects.
Since 2023, Suiyuan Technology has intensively deployed projects at "East Data West Computing" nodes. The company signed strategic cooperation agreements with the Qingyang Municipal People's Government of Gansu Province, China Telecom Gansu Company, and others to jointly build a 10,000-card cluster. In 2024, the Qingyang 10,000-card cluster was officially launched; in January 2025, the Taihu Yixin (Wuxi) Intelligent Computing Center, based on Suiyuan's S60 computing power cluster, became operational. The prospectus shows that in 2024, revenue from Suiyuan's intelligent computing systems and cluster business once reached 56.24% of its total.
Despite the safety net of Tencent's orders and the boost from government projects, Suiyuan remains far from profitability due to high R&D investment. From 2022 to 2024, Suiyuan Technology's R&D expenses were 988 million yuan, 1.229 billion yuan, and 1.312 billion yuan respectively, with the three-year total reaching 3.529 billion yuan, exceeding 316% of its revenue during the same period. This led to net losses attributable to owners of the parent company after deducting non-recurring损益 of 1.150 billion yuan, 1.567 billion yuan, and 1.503 billion yuan for 2022-2024, respectively. Net cash flow from operating activities for 2022-2024 was -987 million yuan, -1.209 billion yuan, and -1.798 billion yuan, respectively.
For the first three quarters of 2025, Suiyuan Technology's revenue, net loss after deducting non-recurring items attributable to parent company owners, and net cash flow from operating activities were 540 million yuan, -912 million yuan, and -770 million yuan, respectively. Assuming no major changes in the external environment, such as international trade frictions, Suiyuan Technology anticipates that the earliest it could reach its break-even point is 2026.
Unlike domestic AI chip companies like Moore Thread and MetaX, which follow the GPGPU technical route and are compatible with NVIDIA's CUDA ecosystem, Suiyuan emphasizes a dedicated AI accelerator architecture for training and inference, opting to build its own chip ecosystem and supply chain.
Suiyuan states in its prospectus that in core technology areas, after years of accumulation, the company has formed a comprehensive, three-dimensional core technology system comprising chips & hardware, software & programming platforms, and computing power cluster solutions. At the hardware level, based on its own instruction set, the company has developed the original GCU-CARE acceleration computing unit and GCU-LARE high-speed inter-chip interconnect technology,对标ing NVIDIA's Tensor Core acceleration computing units and NVLink inter-card connection technology. The corresponding architecture not only offers programming flexibility but also provides deep support for highly parallel accelerated computing required by large AI models.
At the software platform level, Suiyuan claims it did not follow the NVIDIA-dominated CUDA ecosystem but instead developed its own full-stack AI computing and programming software platform, "Yusuan TopsRider," which includes drivers, a compiler language and compiler, operator libraries, and toolchains. This platform links the company's hardware with AI applications, significantly reducing the programming difficulty and migration costs for mainstream AI models based on its hardware, enabling better performance release in application scenarios.
Zhao Lidong once told reporters that in China, the dominant NVIDIA is not without "weaknesses": the vast majority of AI vendors rely on NVIDIA's hardware and software ecosystem, limiting their bargaining power; furthermore, developing AI functionalities based on NVIDIA's non-open-source software makes deeper customization difficult. Regarding customer support and service teams, it's challenging for NVIDIA to treat Chinese clients like BAT on par with US companies like Google and Amazon. Therefore, he argued for the necessity of building a self-owned chip ecosystem and supply chain.
Consequently, Suiyuan chose to build its own development toolchain, tailored to its hardware, promoting the integration of its hardware and software to complete its ecosystem. According to Zhao Lidong, while the vast majority of domestic chipmakers focused on AI rely on or are compatible with NVIDIA's CUDA software ecosystem to avoid its barriers initially, he believes it's difficult to evaluate this approach from an integrated hardware-software perspective, and surpassing NVIDIA GPUs while being CUDA-compatible might raise intellectual property issues.
The prospectus shows that according to IDC data, the total shipment scale of AI accelerator cards in China in 2024 exceeded 2.7 million units, with NVIDIA shipping approximately 1.9 million units, capturing about 70% of the market share. Suiyuan Technology's sales of AI accelerator cards and modules reached 38,800 units that year, corresponding to a market share of approximately 1.4% in the Chinese AI accelerator card market. Looking ahead, while there is some uncertainty regarding NVIDIA's sales in the domestic market, it is highly probable that the shipment share of domestic AI chip manufacturers in the Chinese market will continue to increase.
On the product front, Suiyuan Technology is also iterating its products to meet the demands of the large model era. The fourth-generation training-inference integrated chip, Suisi 400, released in July 2025, natively supports FP8 low-precision computing power. The L600 training-inference integrated acceleration module built on Suisi 400 supports a topology with full interconnection for up to 128 cards in a single layer. Suiyuan claims that its Yunzui ESL super-node system, built with L600, can be applied to ultra-large-scale model pre-training and high-parallelism inference, meeting the computational needs for training and inference of models with hundreds of billions of parameters.