On June 15, 2026, Suiyuan Technology received approval to list on the Science and Technology Innovation Board, marking another milestone in what has become a major event in China's semiconductor sector. With this approval, China's domestic GPU "four dragons" have now completed their market debut.

Last December, Moore Threads and Muxi Technology both listed on the Science and Technology Innovation Board, each with market capitalizations exceeding 300 billion yuan. In January of this year, Biren went public on the Hong Kong stock exchange, with a current market capitalization exceeding 120 billion yuan. Behind these high valuations lie accelerating financial results: in 2025, these three listed companies achieved revenue growth rates of 243.37%, 121.26%, and 207.12% respectively.

Suiyuan Technology is also experiencing rapid business growth, with revenue growth of 140% in 2024 and 37% in 2025, and projected growth of 258.8% to 289.13% for the first half of 2026.

The key difference lies in their technology choices. Moore Threads, Muxi, and Biren have all pursued a general-purpose GPU route, emphasizing compatibility and adaptation with the CUDA ecosystem, while Suiyuan is committed to a DSA architecture combined with self-developed software stacks—deliberately incompatible with CUDA.

Zhao Lidong graduated from Tsinghua University's EE85 class, often called the "Whampoa Academy" of China's semiconductor industry, whose alumni have founded or led enterprises including GigaDevice, Chip King Star, Verisilicon, and Yangtze Memory. Zhao spent over twenty years working in Silicon Valley, rising to senior vice president of AMD's Computational Products division, where he oversaw the planning and commercialization of GPU and AI accelerator chips.

Zhang Yalin graduated from Fudan University's Department of Electronic Engineering and held positions at AMD including senior chip manager and technical director of AMD's China R&D center. As one of AMD's principal global chip R&D leaders, he successfully led the development and mass production of multiple critical products.

When these former colleagues reunited for a startup in Shanghai in 2018, they faced three technology choices: CPUs, GPUs, and AI chips. They chose AI. "An important reason," one account notes, "is that CPUs and GPUs have already developed for decades, and whether in technology ecosystem or industry ecosystem, the integration between customers and original equipment manufacturers is extremely strong, making separation very difficult."

The AI chip landscape was different—with massive market potential, it offered opportunities for cloud-side chips based on innovative architectures. Suiyuan chose AI and directly targeted the highest-difficulty category: cloud-side AI training chips.

Suiyuan has maintained this strategic direction throughout its product development. Just eighteen months after founding, the company released its first-generation cloud training chip Kylinx 1.0 and the Suiyuan T10 training accelerator card, achieving successful tape-out on the first attempt. It released its first-generation inference product in 2020, its second-generation training product in July 2021, and its second-generation inference product in December of that year.

By the end of 2025, Suiyuan had independently iterated through four generations of architecture and five cloud AI chips, building a complete product ecosystem spanning chips, accelerator cards, AI computing clusters, and software platforms.

Among these, accelerator card sales are the revenue driver. In 2025, AI accelerator cards and modules contributed 856 million yuan, or 86.83% of total revenue. Intelligent computing systems and clusters, while generating 56% of revenue in 2024, saw their scale decline by roughly two-thirds in 2025, accounting for only 13% of revenue—the systems-level business is only just beginning.

However, the commercialization of AI accelerator cards and modules has not showcased Suiyuan's first-mover advantage in training. Instead, the product mix is heavily weighted toward inference.

Suiyuan pursues DSA architecture—Domain-Specific Architecture. Unlike general-purpose GPUs that are compatible with the CUDA ecosystem and allow customers zero-cost migration, DSA architecture provides hardware-directed optimization for specific AI computing operations like matrix multiplication and convolution. Simply put, it sacrifices generality to achieve higher energy efficiency and lower per-unit compute costs.

Training requires processing massive data, large-scale parallel computation, and extremely high cluster communication efficiency—precisely the areas where CUDA's ecosystem maturity and generality are essential. These strengths form NVIDIA's core competitive moat. Inference, by contrast, targets specific business scenarios with single-pass computation, more fixed and fragmented tasks where cost sensitivity and specialization matter far more than generality.

From 2023 to 2025, Suiyuan's AI accelerator cards and modules generated revenue of 186 million yuan, 308 million yuan, and 856 million yuan respectively, with growth rates of 65.6% in 2024 and 178.5% in 2025.

However, the drivers of growth shifted. In 2024, Suiyuan released its third-generation inference chip Kylinx S60, achieving a breakthrough in domestic ten-thousand-card inference clusters. That year's growth was primarily driven by price increases, with the average unit price of AI accelerator cards and modules rising 60.8%. In 2025, growth came from volume, with sales volume rising 197.8%.

Looking further ahead, research firm Zhuo Shi Consulting projects that the global AI accelerator card inference market will grow from $47.611 billion in 2024 to $325.618 billion in 2028, a compound annual growth rate of 61.71%. The domestic market is projected to reach 808.5 billion yuan, representing over 70% of the global market.

Data shows that under AVAP accounting (where the company sells AI accelerator cards or modules to designated server vendors at prices negotiated with internet customers), Suiyuan's revenue from Tencent increased from 100 million yuan in 2023 to 270 million yuan in 2024, then further to 830 million yuan in 2025. This represented 33.34%, 37.77%, and 83.8% of total revenue in those respective years.

Tencent is not only the largest customer but also a major shareholder of Suiyuan, with Tencent and its affiliates collectively holding 20.26% of the company's shares.

Suiyuan's own assessment is: "The continuous increase in the proportion of sales to Tencent and its affiliates represents the result of the company adopting a 'single-point breakthrough first, followed by line extension, then gradual area expansion' strategy based on its current development stage and limited resources, and is commercially reasonable."

The DSA architecture plus self-developed software stack means that Suiyuan's customer onboarding costs exceed those of general-purpose GPU vendors. Even though other vendors have also failed to diversify their customer base, Suiyuan's disadvantages in customer expansion are more pronounced. Therefore, commercial rationality dictates concentrating resources to penetrate core customers.

Tencent has massive AI computing power demands. Products like Yuanbao, HunYuan, WeChat AI, Tencent Meeting, and Enterprise WeChat all consume tokens and require computational capacity and cost-per-token calculations. Tencent has sufficient patience and capital to support a domestic chip supplier, and has strong incentive to build a second supply chain for computing power beyond NVIDIA.

In its first round of inquiry responses, Suiyuan noted that Tencent lacks both the motivation and realistic conditions to substantially replace the company's products through alternative suppliers at this stage. In other words, Suiyuan's hardware and software ecosystem has become embedded in Tencent's AI infrastructure.

This "sunk cost" resulting from technical architecture and software adaptation ensures high certainty in Tencent's continued purchasing.

A set of figures in the second round of responses provides useful reference: in 2025, the gross margin on Suiyuan's third-generation AI accelerator cards sold to Tencent was 33.95%, while the gross margin for third-generation cards sold to non-internet customer B was 43.05% (quasi-AVAP model), and the gross margin for third-generation cards sold to internet customer A was 39.46% (AVAP model, with Tencent as the end customer).

Combined with the fact that inference card gross margins are lower than training card margins, and Suiyuan's revenue is heavily driven by inference cards, this results in Suiyuan's gross margins falling well below industry averages.

Suiyuan expects to achieve profitability in 2026 or 2027. Whether this materializes depends on two factors: the achievement rate of operating revenue, and the status of gross margins.

On the revenue side, according to a Bernstein report, Tencent's AI computing power demands continue to climb, with sustained increases in AI-related capital expenditures. Its projected direct and indirect AI-related capital expenditures for 2026–2028 are estimated at $11 billion, $13 billion, and $14 billion respectively. Under this demand pressure, Tencent's purchases from Suiyuan are unlikely to cease in the near term.

In 2026, the reality of customer expansion is limited: three internet-sector potential customers are "expected to make small-scale deliveries by end of 2026," and one non-internet customer is "expected to generate revenue in 2026." Clearly, Suiyuan has not yet built a stable pyramid of repeat-purchase customers, and the "point-to-line" expansion effect still requires further improvement.

In May 2026, at Tencent's shareholders meeting, Zhang Xiaolong commented that although Tencent has boarded the AI ship, it is struggling to find a seat—the ship also isn't fast enough. This sentiment equally describes Suiyuan's situation. Suiyuan and Tencent are "on the same ship," but the ship's speed is not under Suiyuan's control.

On one hand, if remaining CUDA-compatible forever, even with access to advanced process nodes, it would be impossible to catch up to NVIDIA in engineering implementation capability and systems optimization within a short timeframe. Moreover, CUDA compatibility would further strengthen the CUDA ecosystem, meaning the company would forever remain in NVIDIA's shadow.

On the other hand, currently Google, Microsoft, and Amazon all insist on self-developed DSA architecture AI chips, and in China, the major domestic manufacturers with higher shipment volumes in 2025 primarily employ DSA architecture.

This means that achieving truly independent and controllable computing capacity requires starting from scratch. Suiyuan's bet is that only through complete self-sufficiency can it possibly gain the power to define the next generation of AI computing power landscape.

Therefore, from chip and hardware technology to computing software and programming platform technology to AI computing cluster solutions, Suiyuan develops everything in-house.

From 2023 to 2025, Suiyuan accumulated R&D investment of 3.676 billion yuan, amounting to 1.8 times its revenue over the same period. Its 643 R&D personnel constitute 76.73% of total headcount. The four generations of architecture and five chips are the fruits of these R&D efforts.

Going forward, Suiyuan's fundraising proceeds will continue to be directed toward research and development of fifth and sixth-generation chips and software-hardware co-innovation projects, enabling continued product and ecosystem iteration.

Within Tencent, Suiyuan's hardware and software have been deeply embedded in core business lines like Yuanbao, forming a closed yet vibrant internal ecosystem. At this stage, the cost of ecosystem isolation is temporarily absorbed by the scale of the major customer. However, when Suiyuan attempts to promote to other manufacturers, they may lack Tencent's strategic patience—each integration becomes a difficult battle.

As noted earlier, from a business rationale standpoint, Tencent has neither the motivation nor the conditions to replace Suiyuan, but commercial rationality is not the same as strategic rationality. Google developed TPU, Meta developed proprietary ASICs—superplatforms keep computing power under their own control. The possibility that Tencent will follow this path remains ever-present.

Even absent direct replacement, adjustments to collaboration depth and prioritization could materially impact Suiyuan's business structure. The crucial point is that decision-making authority for such adjustments does not rest with Suiyuan.

From 2023 to 2025, Suiyuan accumulated losses exceeding 4.3 billion yuan. For the first half of 2026, the company projects losses of 577 million to 608 million yuan, a modest improvement. Whether profitability is realized depends heavily on achieving revenue targets and improving gross margins—points that require no further elaboration here.

Market positioning presents an equally difficult dilemma. If viewed as a core asset within Tencent's computing power ecosystem, the valuation becomes tied to Tencent's AI investment pace and could rise and fall with Tencent's AI capital expenditures. If viewed as a high-concentration-risk asset, the valuation could suffer from single-point dependency concerns.