In August 2024, Suiyuan Technology, which was established only six years ago, officially submitted an A-share IPO application with a valuation of 16 billion yuan, aiming to become the "first domestic AI chip stock".
Behind this company stands Tencent, which has invested more than 2 billion yuan in six consecutive rounds, and a group of "chip veterans" from Tsinghua EE85 class. Founders Zhao Lidong and Zhang Yalin, two former AMD executives, pushed Suiyuan to the forefront of the industry with a record of 18-month successful tape-out.
Its "Sui Si" series chips have penetrated Tencent Conference voice-to-text and the Wanka cluster of the National Computing Hub in Qingyang, Gansu. Its performance is comparable to NVIDIA A100, and the energy efficiency utilization rate of the liquid-cooled computing power cluster reaches 92%.
But under the halo, the challenges are equally sharp: TSMC's 12nm process lags behind the international market by two generations, there are only 87,000 CUDA ecosystem developers, and global expansion is difficult. This IPO is not only the fruit of 7 billion yuan of capital, but also a key battle for the breakthrough of domestic computing power.
From AMD to Tsinghua Unigroup
---the six-year rapid development of two "Silicon Valley veterans"
In March 2018, in an office in the Lingang New Area of Shanghai, Zhao Lidong and Zhang Yalin signed the founding agreement of Suiyuan Technology.
These two "Silicon Valley veterans" who had worked together for many years at AMD had already left a deep mark in China's semiconductor industry.
Zhao Lidong, a member of the EE85 class of Tsinghua University, led AMD's Shanghai R&D Center and participated in the R&D of Ryzen processor cores; Zhang Yalin led the team to mass-produce Microsoft Xbox main chips and held 12 patents.
Their cooperation is regarded by the industry as a strong combination of "technology definers" and "mass production operators".
Since joining AMD in 2007, Zhao Lidong has been responsible for CPU/GPU core research and development and managed a team of more than 1,000 people.
In 2014, he returned to China to join Tsinghua Unigroup, led the acquisition and integration of RDA Microelectronics, and promoted Tsinghua's layout in the field of communication chips.
However, he always had the ambition to "define global chip products in China." In 2018, Zhao Lidong left Tsinghua Unigroup and co-founded Suiyuan Technology with Zhang Yalin, aiming directly at the AI training chip market monopolized by Nvidia.
Only one month after the company was founded, Suiyuan Technology received seed round financing from Yihe Capital, Zhen Fund and other institutions, verifying the industry appeal of the founding team.
Five months later, Tencent led the Pre-A round with 340 million yuan. This move was called a "hot start strategy" by Zhao Lidong: Tencent not only invested, but also opened up cloud service scenarios, so that Suiyuan chips were deeply coupled with actual needs such as Tencent Conference voice-to-text and OCR recognition during the research and development stage.
In the following six years, a capital frenzy surged in - Sequoia China, National Big Fund Phase II, CPE Yuanfeng and other institutions entered the market one after another, with a total financing of nearly 7 billion yuan. Tencent participated in six rounds of investment, holding 20.494% of the shares, and remained the largest shareholder.
Under the push of capital, the Suiyuan team expanded rapidly.
R&D personnel account for 70%, and most of the core members come from AMD, NVIDIA, and HiSilicon, forming a "Silicon Valley gene + localized innovation" model.
In June 2019, the first AI training chip "Sui Si 1.0" was successfully taped out, taking only 18 months, setting an industry record. This chip, which uses TSMC's 12nm process, has a computing power of 128TFLOPS (FP16), and it is the first time that domestic training chips have been commercialized in Tencent Cloud.
In 2020, Yunsui T10 training accelerator card was launched, with computing power doubled to 256TFLOPS, adapted to trillion-parameter large model training, and landed in Hubei Yichang Intelligent Computing Center, with a shelf rate of over 80%.
The business map is expanding simultaneously.
In 2023, Suiyuan Technology won the bid for China Mobile (Gansu Qingyang) Data Center Project, built the first domestic 10,000 card inference cluster, with a single cluster computing power of 40EFLOPS, and was selected as a national computing network hub node.
Internet customers include Tencent and iQiyi, and a real-time anti-fraud system has been implemented in the financial field, with the response time reduced to milliseconds; autonomous driving partner car companies have entered the road test stage.
As of August 2024, before the IPO was launched, Suiyuan Technology's valuation reached 16 billion yuan, but specific profit data was not disclosed.
The three strategies to fight Nvidia: chips, clusters and ecosystems
Suiyuan Technology's technological breakthrough began with a precise attack on Nvidia's monopoly on the market. Its "three-axe" strategy - high-performance chips, liquid-cooled computing power clusters, and independent software ecosystems - directly targets the core barriers of international giants.
"Suisi 1.0" released in 2019 uses TSMC's 12nm process, supports mixed-precision computing, and has an energy efficiency ratio of 1.2TOPS/W, which is 15% behind Nvidia V100 in the same period, but the price is only 60% of the latter.
The 2021 iteration of the Yunsui T20 training card has increased its computing power to **TFLOPS, supports FP8 precision, and is suitable for trillion-parameter model training such as Llama 3, achieving a 20-fold DAU growth in Tencent Yuanbao AI tools.
The Yunsui i20 inference card launched in 2024 has an INT8 computing power of **TOPS and an energy efficiency ratio of 1.5TOPS/W. The liquid-cooled version of S60 consumes only 150W, supports high-density deployment, and is applied to the Qingyang Wanka cluster, with a computing power utilization rate of 92%.
Suiyuan Technology's liquid-cooled computing power cluster uses a cold plate heat dissipation solution, with a single cabinet power density of 40kW, which is 40% lower than the traditional air-cooled energy consumption. Its dynamic resource scheduling algorithm can automatically balance the computing power load. In the Qingyang project in Gansu, the latency of the 10,000-card cluster was less than 5 microseconds, supporting real-time reasoning of models with hundreds of billions of parameters.
This technology makes Suiyuan one of the only two companies in China (the other is Huawei Ascend) to achieve commercial implementation of 1,000-card AI clusters.
Software ecosystem is the biggest gap between Suiyuan and NVIDIA
The "Yusuan" programming platform launched in 2021 supports the migration of TensorFlow and PyTorch frameworks, pre-installs 200+ operator libraries, and improves model training efficiency by 50%.
The supporting "Jiansuan" inference engine realizes model quantization compression, and the ResNet-50 inference delay is reduced to 5ms, which is 60% optimized compared with the open source solution.
However, compared with the 3 million developers in the CUDA ecosystem, "Yusuan" has only 87,000 registered developers, and the operator coverage rate is less than 60%. To this end, Suiyuan plans to invest 20% of the IPO funds to develop a CUDA-compatible tool chain, with the goal of achieving 90% operator compatibility in 2025, and the migration cost will be reduced to 40 working hours per person.
In the Tencent ecosystem, Suiyuan chips support voice-to-text response time reduced to milliseconds, and OCR recognition accuracy increased to 99.3%.
The "Suiyuan Yaotu" Wenshengtu tool released in 2024, based on self-developed chips, generates 20 4K images per second, becoming the first AI creation platform in China that supports real-time rendering.
In the financial field, a leading bank uses Yunsui i20 to complete anti-fraud analysis of 10,000 transactions per second.
In the cooperation of autonomous driving, the computing power density of multi-sensor fusion algorithm is 2 times higher than that of GPU solution.
The carnival and hidden worries of domestic substitution
A microcosm of the computing power game between China and the United States
According to data from the Ministry of Industry and Information Technology, the proportion of domestic AI chips in the purchase of intelligent computing centers has jumped from 18% in 2023 to 47% in 2025, and the market size has exceeded 60 billion yuan.
Behind this carnival, there are not only milestones of technological breakthroughs, but also hidden concerns about the supply chain and ecology.
1. Carnival: Dual catalysis of policy dividends and technological breakthroughs
In 2024, the Ministry of Finance and other four departments jointly issued the "Action Plan for High-quality Development of Computing Infrastructure", which clearly requires that the proportion of domestic chips used in newly built intelligent computing centers should not be less than 50%, and the purchase subsidy should be up to 30%. Policies directly drive a surge in orders:
Suiyuan Technology won the bid for the China Mobile (Gansu·Qingyang) Data Center Wanka Cluster Project, with a contract value of 800 million yuan. It uses Yunsui i20 liquid-cooled inference cards, with a computing power utilization rate of 92%;
Huawei Ascend won the State Grid Intelligent Inspection System Upgrade Project, deployed more than 50,000 Ascend 910B chips, and increased inference efficiency by 40%;
Cambrian and the Hefei Municipal Government jointly built a 200PFlops computing power center, and the number of Siyuan 590 chips deployed exceeded 100,000, mainly used for traffic monitoring low-light scene optimization.
Technological breakthroughs are accelerating simultaneously.
Suiyuan Technology's Yunsui T20 training chip supports a 20-fold increase in DAU in Tencent Yuanbao AI tools, and the voice-to-text response time is reduced to 0.8 seconds;
In the Llama3-400B model training, the single-card performance of Biren Technology's BR104 chip is 72% of NVIDIA A100, and the cost is only 50% of the latter.
The cost-effectiveness of domestic chips has enabled them to gradually penetrate into market scenarios such as the Internet and finance: after a leading bank adopted the Suiyuan solution, the cost of real-time anti-fraud analysis was reduced by 60% and the processing speed increased by 3 times.
2. Hidden worries: the double shackles of technology dependence and ecological shortcomings
The carnival of domestic substitution cannot hide the underlying crisis.
Process dilemma: Suiyuan Yunsui T20 relies on TSMC's 12nm process, and the transistor density is only 33.8MTr/mm², which is two generations behind NVIDIA H200's 4nm process (138MTr/mm²). If it is included in the entity list and switches to SMIC's 14nm, the chip area will increase by 58%, power consumption will soar by 45%, and performance loss will reach 35%.
HBM supply interruption risk: The HBM (high-bandwidth memory) of domestic chips is completely dependent on SK Hynix and Samsung. The reserve of Suiyuan Qingyang project is only enough for 12 months, and NVIDIA has increased the memory bandwidth to 4.8TB/s through its self-developed HBM3E, forming a generational suppression.
Ecological island: NVIDIA's CUDA ecosystem has 3 million developers and supports one-click deployment of 2,000 models; there are only 87,000 registered developers on the Suiyuan "Yusuan" platform, with an operator coverage rate of 58% and an average migration cost of 120 man-hours per person. Although Huawei Ascend has opened hardware authorization, the CANN software adaptation efficiency is still 30% behind CUDA.
Geopolitics further exacerbates uncertainty.
In 2024, the U.S. Department of Commerce will update the Entity List, adding five Chinese chip companies including Biren Technology, restricting their access to process technology below 14nm.
The performance of NVIDIA's special version of the H20 chip for China has shrunk by 85%, but the unit price is as high as 110,000 yuan, forcing domestic companies to fall into the dilemma of "high price and low efficiency" and "technical bottleneck".
3. Game: Strategic transformation from "replacement" to "coexistence"
Faced with suppression, the domestic camp turned to differentiated competition:
Hybrid architecture breakthrough: BiRen Technology launched a heterogeneous GPU collaborative training solution (HGCT), which supports mixed use with NVIDIA chips, and the training efficiency loss is controlled within 15%;
Edge computing positioning: Cambrian Siyuan 590 has a market share of 35% in the smart camera market, and the accuracy of low-light scene recognition has increased to 92%;
Policy market bundling: Suiyuan Technology and Tencent Cloud launched the "Meta Computing Power" subscription service, and customers can call domestic computing power on demand, with a cost 30% lower than NVIDIA's solution.
But the counterattack of international giants has been upgraded simultaneously.
NVIDIA will launch the Blackwell chip in 2025, and the FP8 precision training efficiency will be increased by 2 times compared with H200. It will also jointly establish the "AI Computing Alliance" with Meta and Google to consolidate the CUDA ecological barriers through framework optimization.
AMD cut the price of its MI300X chip by 20%, grabbing market share from Suiyuan and Huawei in Chinese Internet companies' orders.
The replacement wave of domestic AI chips is the result of policy will, capital push and technological breakthrough.
Companies such as Suiyuan Technology and Huawei Ascend have opened a gap in the $154 billion global market through the construction of intelligent computing centers, edge scene positioning and hybrid architecture innovation.
However, the triple shackles of process generation gap, ecological shortcomings and geopolitical stranglehold will continue to restrict its development for a long time.
The essence of this game is a paradigm shift from "replacement" to "coexistence".
Domestic chip companies need to continue to increase their investment in 3nm process research and development, CUDA compatible tool chain construction and supply chain redundancy construction in order to hold their ground under Nvidia's annual $4.7 billion R&D firepower.
The IPO of Suiyuan Technology is not only a highlight moment for domestic computing power, but also the starting point of a new round of competition - only by crossing the triple rift of technology, ecology and commercialization can we truly achieve a qualitative change from "usable" to "easy to use".
Research on science and technology investment, the author's in-depth
