Suiyuan Technology has listed on Shanghai’s STAR Market, capping an eight-year journey that began in the winter of 2017 with two founders, a long table in a Shanghai villa office, and a roughly 20-page PowerPoint deck.
The company opened at RMB 410 per share, up 188.37% from its IPO price of RMB 142.18, for a total market capitalization of RMB 176.4 billion. According to the source article, that made it the company with the highest market value at opening among this year’s listings so far.
From a cold meeting room to a public listing
The first key meeting took place at 392 Jianguo West Road in Shanghai’s Xuhui district, at the office of Delta Capital. Seated on one side of the table were Delta Capital founding managing partners Li Quansheng and Ye Weigang. Across from them sat Zhao Lidong and Zhang Yalin, who had not yet formally started the company.
Zhao, born in 1966, had worked at S3, Juniper Networks and AMD. During seven years at AMD, he was involved in product planning for multiple products, core IP development, and the establishment of AMD’s China R&D center. After returning to China in 2014, he served as president of RDA Microelectronics and vice president of Tsinghua Unigroup.
Zhang, 12 years younger than Zhao, was also a veteran chip engineer. He moved from Shanghai Qima Digital to AMD and later served as a senior chip manager and technical director at AMD’s China R&D center. He worked on the development and mass production of chips including the main chip used in Microsoft’s Xbox One series, and helped build and manage several AMD R&D teams in Shanghai and Beijing.
In November 2017, both men left their previous companies to build a domestic cloud AI chip startup. Before meeting Delta Capital, they had already spoken with several investment firms. Most of those conversations did not lead anywhere. Some investors were openly dismissive, asking, 「Just the two of you, and you want to challenge Nvidia?」
Ye Weigang, who had previously worked at Cadence as a senior business development director in marketing, knew both the technical and commercial sides of semiconductors well. At that meeting, the Delta team did not focus on slogans. They asked Zhao to pull up the architecture diagram for the company’s planned first chip, then went layer by layer through the design: how the compute core would work, how memory bandwidth would be matched, how inter-chip interconnects would be handled, which IP blocks would be developed in-house and which would be purchased, how the team planned to deal with the software ecosystem already built around CUDA, and where supply chain and tape-out resources would come from.
Ye later recalled that he looked at the structure “layer by layer” and asked about it “module by module.” The point was to see how much thinking actually sat underneath those 20 pages.
After that meeting, Delta Capital moved quickly and decided to invest. The deal gave Zhao and Zhang more confidence. In March 2018, they formally established Suiyuan Technology, with Zhao as CEO and Zhang as COO. Delta Capital joined the seed round as a co-lead investor.
Choosing the harder path from day one
Investors interviewed by Jiaziguangnian said Suiyuan’s distinctiveness was already visible in that early architecture diagram.
At the time, China’s AI chip startup wave was just beginning. Companies such as Cambricon, Horizon Robotics and DeePhi were entering investors’ field of view, and many startups chose to begin with inference chips or edge scenarios. Zhao and Zhang aimed their first product at cloud AI training instead.
That market carried higher barriers and required more capital. A cloud AI training chip has to handle massive data volumes and intensive parallel computing, but raw compute is only the start. A startup also has to solve memory bandwidth, chip interconnect, software compilation, model adaptation and large-scale cluster deployment. Above all of that stands Nvidia’s software and hardware stack, built over many years.
Delta Capital said this was the first time it had heard a domestic startup team clearly state that it wanted to build a cloud training chip for large-scale compute and compete directly in a market long dominated by Nvidia. Li and Ye said they did not find that easy to believe at first.
What changed their minds was the level of detail. Zhao and Zhang had not only sketched the chip architecture. They had also mapped out a planned 20-person core team, the backgrounds required for each role, where those people might be recruited from, which modules had to be developed internally, which IP could be bought, and which suppliers might be used based on prior cooperation.
For Delta Capital, the question was no longer whether there was a market for AI chips. It was whether these two founders could actually build one, and who would use it if they did.
Shortly after the company was founded, Suiyuan completed its seed round with backing from Delta Capital, ZhenFund, Oriza FoFs-affiliated Wuyuefeng Kechuang, Yunhe Capital and Shanghai Sci-Tech Venture Capital, according to the source article. Over the next eight years, the company completed 10 financing rounds before listing. Its shareholder roster at different stages included Delta Capital, ZhenFund, Yunhe Capital, Wuyuefeng Kechuang, Tencent Investment and Sunshine Ronghui Capital.
Tencent and the idea of a “hot start”
When Suiyuan first started operating, the whole company had fewer than 10 employees. The office was borrowed from a friend of Zhao’s and could be used free for three months.
At the end of April 2018, Tencent Investment managing director Yao Leiwen and investment director Xiao Hongda reached out to Zhao. Zhao later told Jiaziguangnian that the first conversation took place over video and lasted nearly two hours, after which Tencent’s team came to Suiyuan’s office in Shanghai.
What stood out to Zhao was a line from Xiao: Tencent could help Suiyuan with a “hot start.”
The “cold start” problem is familiar in the chip industry. A company can define a product based on public information and broad industry trends, spend years building the chip, and then discover that actual customer workloads do not match those early assumptions. By then, changing the hardware architecture may no longer be practical.
A “hot start” meant that from the product-definition and R&D stage, Suiyuan could already engage with the real needs of a large internet company and understand what training jobs, model frameworks and data-center environments would demand from the chip.
In August 2018, Tencent led Suiyuan’s RMB 340 million Pre-A round. The source article said the investment agreement did not make Tencent procurement a precondition, and the business relationship was not exclusive. Tencent’s investment arm assessed company value, while business units would still decide whether to use the product based on performance, stability and cost.
Zhang had also told Jiaziguangnian that one reason Suiyuan chose to work with Tencent was Tencent Investment’s relatively hands-off style. He said it gave portfolio companies room to operate, did not interfere in micro-management, and offered support on key decisions.
Tencent’s “hot start” gave Suiyuan early help in shaping R&D. At the same time, Tencent’s relatively restrained relationship with the company created a real and demanding test ground. For Suiyuan, Tencent was both an early shareholder and the most important potential customer for its first product. Before any customer order could arrive, though, the company had to survive the most dangerous stage in a chip startup’s life: tape-out.
Successful tape-outs and a maturing product roadmap
In the second half of 2018, Zheng Yan joined Sunshine Ronghui Capital, a private equity fund manager affiliated with the insurance sector. His doctoral research focused on design for testability, or DFT, and fault tolerance in chips, giving him strong familiarity with integrated circuit design.
He said he reviewed most of the domestic AI chip startup projects at the time and eventually narrowed his attention to Suiyuan. By early 2019, through a younger Tsinghua alumnus, he got in touch with Zhao. Because of his technical background, his due diligence quickly moved inside the engineering organization.
At that point, Suiyuan’s first chip had not yet reached RTL Freeze, meaning the chip’s function and logic design were still not fully fixed. There were still several critical steps before the design could be sent to a foundry for tape-out. Zheng and his team interviewed the heads of compute core, on-chip network, SoC, DFT and back-end design. He was not simply checking resumes. He wanted to know whether all of those modules could actually come together.
Chipmaking is a deeply collaborative engineering effort. One or two standout engineers can raise the ceiling of a local design block. Tape-out success depends on whether architecture, front-end, back-end, verification, test and supply chain teams can all deliver on the same schedule.
For a conservative insurance-backed investor, putting money into a chip startup before tape-out and before the commercial picture had become clear was far from routine. Zheng said the investment came from both a professional assessment of Suiyuan’s team and tape-out risk, and Sunshine Ronghui’s effort to test whether long-term capital could be deployed into strategic emerging industries.
A single tape-out for an advanced chip can cost more than RMB 100 million. If it fails, the company loses not only the development budget but also the time needed for redesign, verification and manufacturing rescheduling, often six months or longer. In AI chips, missing six months can mean missing an entire market generation.
Suiyuan did not miss. In December 2019, it released its first-generation cloud AI training chip, “Suisi,” later referred to in the prospectus as Suisi 1.0. In the same period, it launched the YunSui T10 training accelerator card based on Suisi 1.0. In December 2020, it introduced its first-generation inference accelerator card, YunSui i10.
The company then kept iterating:
- In July 2021, it released the second-generation training chip Suisi 2.0 and the YunSui T20 and T21 cards.
- In December 2021, it launched the second-generation inference chip Suisi 2.5 and the YunSui i20 card.
- In 2024, it rolled out the third-generation inference chip Suisi 320 and the Suiyuan S60.
- In July 2025, it released the fourth-generation integrated training-and-inference chip Suisi 400 and the Suiyuan L600.
According to the prospectus, by the end of 2025 Suiyuan had completed four generations of computing architecture and five cloud AI chips, and all five chips had succeeded on the first tape-out.
To understand the company’s route, the article draws a distinction between two mainstream approaches in cloud AI acceleration. One is the GPGPU route represented by Nvidia, built around general-purpose parallel compute and covering AI, scientific computing and graphics rendering, supported by the mature CUDA ecosystem. The other is the AI-focused accelerator path, which includes products from Google TPU, Huawei Ascend and Suiyuan.
These products usually rely on DSA, or domain-specific architecture, and are implemented as ASICs. By stripping away more general-purpose hardware that is not central to AI workloads, they can devote more chip area, bandwidth and compute resources to matrix operations, aiming for stronger performance, better efficiency and a more attractive cost structure in specific AI tasks.
Suiyuan uses a self-developed GCU-CARA computing architecture with multiple programmable GCU-CARE acceleration units inside. Around the chip, it built the TopsRider software stack. That gives the company more room for hardware-software co-optimization, but it also means Suiyuan cannot simply inherit Nvidia’s existing CUDA ecosystem.
When a customer migrates a model that originally runs on Nvidia’s platform to Suiyuan chips, code, operators and business systems usually need to be adapted again. Suiyuan has to tune models one by one and verify workloads scene by scene, then persuade customers to bear the migration cost. Tape-out success, in that sense, is only the start of market entry.
The long middle stretch
Yunhe Capital was one of Suiyuan’s core seed-round backers and kept adding capital across five rounds. Founder Zhao Yun said his team believed early on that integrated hardware and software for domestic cloud training chips would become the core long-term moat.
Zhao recalled that in Suiyuan’s early days, Zhao Lidong and Zhang Yalin were deeply grounded in low-level chip technology, but weaker in commercial presentation and market storytelling. It was difficult for investors, industrial customers and industrial parks to quickly grasp the value of Suiyuan’s full compute system. For a period, Zhao said he would sit with them in hotels to review their pitch deck presentations. He had them record the presentation, then listen back together section by section, cutting redundant technical detail, adding the commercialization path, and clarifying what benefits customers would gain by replacing overseas chips.
Zhao told Jiaziguangnian that they reached a common understanding at the seed stage: training chips are not a contest of a single hardware component, but of a complete “hardware + software + toolchain” system. Whether Suiyuan could efficiently support mainstream AI frameworks and materially reduce the cost of migration and adaptation for customers was both the key to breaking Nvidia’s ecosystem lock-in and the basis for scaled commercialization.
As Suiyuan’s chips entered the market, the discussion changed. The issue was no longer only the chip itself. It became how to match the company with its first customers for scaled testing, and how to build a thousand-card compute cluster as the carrier for moving from sample cards to large commercial deployments.
In that long middle stretch, industrial resources mattered. Zhao said his team used its network to connect Suiyuan with leading research institutes, local government compute platforms and regional intelligent computing centers. One signature result was helping Suiyuan establish deep cooperation with Zhijiang Lab, jointly building a research center and deploying a thousand-card AI compute cluster.
For an early-stage company, a thousand-card cluster is not just a sales contract. Single-card performance is only the starting point. High-speed interconnect between chips, cluster software scheduling and stable 24/7 operation across the stack can only be validated in a large cluster environment. Those projects become benchmark orders, but they are also test fields for continued iteration.
After the Zhijiang Lab project landed, Yunhe Capital kept expanding Suiyuan’s industrial reach by connecting it with compute resources in the Yangtze River Delta and northern China, while also bringing in upstream supply chain and downstream AI application ecosystem partners.
That process showed what commercialization really looks like for an AI chip company: the chip has to light up, the software has to run, customers have to be willing to test, testing has to turn into procurement, and procurement has to turn into repeat orders at scale. Every step requires a different level of capability.
Fundraising also became harder after the first-generation AI chip launch. Li Quansheng and Ye Weigang described hard-tech financing as having three periods. The first round is mostly about the team and direction. Near listing, products, revenue and customers are relatively visible. The toughest stages are the middle rounds.
At Delta Capital, there was internal debate. Younger investors were more focused on Suiyuan’s details and numbers: sales growth was not fast enough, losses were widening year after year, and the next valuation kept rising above the last. Senior partners paid more attention to whether the products were being delivered according to plan and whether Zhao and Zhang had done what they had said they would do.
Li and Ye said Delta ultimately looked at two things: the direction of the industry and the people. In their view, the two founders had largely delivered on what they promised. Delta chose to keep investing, while acknowledging that the B and C rounds were the hardest to underwrite.
Suiyuan continued raising money:
- In May 2020, it completed a RMB 700 million Series B led by Wuyuefeng Capital, with Tencent following.
- In January 2021, it closed an RMB 1.8 billion Series C led by CITIC Industrial Fund, funds under CICC Capital and Primavera Capital.
- In July 2022, it completed a C+ round backed by the National Integrated Circuit Industry Investment Fund.
- In 2023, it completed Series D and D+ rounds.
- In 2024, it completed D++ and Series E rounds. The Series E financing totaled about RMB 2.72 billion at a pre-money valuation of RMB 17.5 billion.
From the seed round to Series E, Suiyuan completed 10 equity financing rounds before going public.
Export controls, cash pressure and the limits of financing
The external environment changed quickly. In October 2022, the United States tightened export controls on advanced computing chips and semiconductor manufacturing equipment. Zheng said he was traveling when the rules were released. That evening, after leaving his hotel, he called several domestic AI chip companies, including Suiyuan, asking how design, tape-out and supply chains might be affected.
There were no ready answers on the other end of the line. The new rules had to be interpreted jointly by lawyers and technical teams, and chip companies had to reassess parameters, suppliers and future product timelines.
The article notes that a few lines in a regulatory document can alter a chip company’s R&D plan for years. That is one reason investing in domestic AI chips differs from investing in ordinary internet projects. Beyond product and market risks, these companies also face uncertainty from supply chains, advanced process technology and shifting international rules.
Delta Capital said Suiyuan went through several moments when funds were especially tight. Cash on hand was shrinking while the next round was still under negotiation. Money arriving a few weeks earlier or later could change a company’s fate.
Still, continuous fundraising is not a substitute for commercialization. Investors can buy time for a chip company. They cannot replace the need for products and customers. Zhao Yun told Jiaziguangnian that Yunhe Capital had positioned itself over the past eight years as a long-term companion and industrial resource connector, while insisting on “support without overstepping.” He also said the foundation of every technical and commercial milestone came from Suiyuan’s own technical depth and market development capabilities.
Revenue rose, losses narrowed, and Tencent remained central
Suiyuan’s prospectus shows the company generated RMB 990 million in revenue in 2025, up from RMB 301 million in 2023, more than tripling over that period. Net loss narrowed from RMB 1.665 billion to RMB 1.164 billion.
Based on IDC data and sales volume disclosed in the prospectus, China’s AI accelerator card market shipped about 4 million units in 2025. Nvidia shipped about 2.2 million of them, or roughly 55% of the market.
Suiyuan sold 66,000 AI accelerator cards and modules in 2025, corresponding to roughly 1.7% of China’s AI accelerator card market, placing it among the leading domestic AI chip vendors cited in the article.
Those numbers suggest that Suiyuan had moved from product validation to scaled delivery. Another set of numbers shows it was still some distance from profitability.
From 2023 to 2025, the company’s cumulative R&D spending reached RMB 3.676 billion, equal to 182.55% of total revenue over the same period. By the end of 2025, it had 838 employees, 643 of them in R&D, or 76.73% of the workforce. Accumulated uncovered losses reached RMB 4.441 billion.
For a company still iterating its chip line, heavy R&D spending has a clear rationale. The capital market is more focused on whether that spending can turn into a broader customer base and sustainable revenue. At present, that question points in large part to Tencent.
From 2023 to 2025, Suiyuan’s direct and indirect sales revenue from Tencent was RMB 100 million, RMB 273 million and RMB 830 million, accounting for 33.34%, 37.77% and 83.79% of revenue in each respective period.
Tencent was not only Suiyuan’s largest customer, but also its biggest shareholder. Tencent Technology and its affiliates held 20.26% of the company.
That dual role is also why the market sees a tension here. In its first round of inquiries before the listing, the Shanghai Stock Exchange asked Suiyuan to explain whether its existing AI products were highly tailored to Tencent, whether it had the capability to win and support other customers, and what the market space, competitive landscape and future development trends looked like across different application areas.
In its response, Suiyuan acknowledged that its current AI chips were already highly adapted to many Tencent models and business scenarios. Internet applications must handle high concurrency, low latency, traffic volatility and 24/7 operation. Without targeted software optimization, companies can run into poor efficiency, conflicts in multi-card scheduling and insufficient stability.
At the same time, Suiyuan argued that “highly adapted to Tencent” does not mean “usable only by Tencent.” The company said the adaptation experience accumulated while serving Tencent had been absorbed into the TopsRider software platform as reusable compilers, operator libraries and development tools. By the end of 2025, its hardware and software platform had adapted to nearly 1,000 AI models across more than 300 application scenarios.
Ye Weigang said spending in the AI compute market is highly concentrated to begin with. In both China and the United States, only a small number of customers can sustain large long-term investment in compute infrastructure. For a startup, building a deep relationship with one top-tier customer first is a practical route into real workloads and scaled delivery.
The article cites a forecast from U.S. independent sell-side research firm Bernstein that by 2028, ByteDance, Alibaba and Tencent together will account for nearly 50% of China’s AI capital expenditures. That underscores the importance for accelerator vendors of getting into the supply chains of major internet companies.
From chips to systems
As the compute demand and data throughput required for large-model training and inference continue to climb, competition is moving from the performance of a single chip or a single compute node to system clusters built around high-speed interconnect.
At the 2026 World Artificial Intelligence Conference, or WAIC 2026, held in July 2026, more than a dozen chip and compute companies presented their own “super node” products. The names varied, but the underlying issue was the same: how to make a growing number of AI chips work together efficiently.
Suiyuan presented two super-node products at WAIC 2026. One was the “YunSui ESL64-C,” which uses a mature cable-tray route. The other was the “YunSui ESL64-O,” an orthogonal architecture solution built jointly with ZTE.
The two products reflect different engineering trade-offs. The cable-tray design links compute nodes and switching nodes through high-speed cables. Its technical system is relatively mature, and its deployment is closer to existing data-center practice. Its strengths are flexible networking, easier expansion and better compatibility with different server and data-center environments.
The orthogonal architecture aims for higher integration density. Compute boards and switch boards are vertically interconnected through orthogonal connectors, reducing high-speed cabling and intermediate connection points inside the rack. That shortens communication distance and lowers potential failure points.
The shift also tracks Suiyuan’s own product evolution. Early products such as the YunSui T10 and T20 were mainly training accelerator cards. The third-generation S60 captured demand in large-model inference. The fourth-generation L600 again covered both training and inference, and became the compute base for the OGX and ESL32/64 super-node products.
Suiyuan, in other words, is no longer trying to deliver only a standalone AI chip. It is trying to deliver a compute system that can run models reliably.
The prospectus also shows that as of the end of 2025, the OGX and ESL32/64 super-node products had not yet generated mass-market revenue.
Suiyuan said in the prospectus that, assuming revenue growth and gross margin meet relevant assumptions, it may achieve profitability on a consolidated basis in 2026 or 2027. The company also cautioned that this calculation is subject to major uncertainty and does not constitute a profit forecast or commitment.
After the bell, the next test begins
A few weeks before the listing, Zhao Lidong and Zhang Yalin returned to the old house at 392 Jianguo West Road in Shanghai and had drinks with Li Quansheng, Ye Weigang and several senior Delta Capital partners. This time, they did not spend much time revisiting old stories. They talked about post-listing plans. Li and Ye told Jiaziguangnian, “We’re already working on the next step.”
Looking back on Delta Capital’s eight years with the company, the two investors used three short phrases: “precise bet, long-term companionship, mutual achievement.” At the bell-ringing ceremony itself, the article says, there was little to add. They simply embraced Zhao and Zhang.
Eight years ago, Suiyuan did not yet exist, and its founders had to persuade investors with a 20-slide deck. Eight years later, the company has completed four generations of architecture, developed five chips, built a team of more than 800 people, and entered the public market.
The listing does not end the questions. It changes who is asking them. Part of what was imagined in that 20-page deck has now been realized. The rest will be tested by each new product generation, each customer batch and each set of financial results.