Suiyuan Tech's Eight-Year Journey: From AI Chips to 10,000-Card Clusters

The report explains that comparing AI chips by peak computing power, memory capacity, and bandwidth is no longer sufficient. As large models move into scale training and inference, and as workloads such as MoE and agents grow, the ability to interconnect dozens, hundreds, or even tens of thousands of chips has become decisive. Suiyuan’s eight-year technical route, based on its own instruction set and DSA architecture, was designed to address this system-level challenge. Unlike companies that adopted GPGPU architectures or CUDA compatibility, Suiyuan built its own stack from chips to software to clusters.

Suiyuan’s system is organized in three layers. At the bottom are chips and hardware, including the self-developed GCU-CARE acceleration unit, now in its fourth generation with native FP8 support, and the GCU-LARE interconnect technology. Above that is the TopsRider software platform, covering drivers, compilers, operator libraries, tools, and deep learning frameworks. The company says its hardware has been adapted to nearly 1,000 AI models and more than 300 application scenarios. The top layer consists of intelligent computing systems and large-scale clusters. This full-stack approach is capital-intensive: Suiyuan spent 3.676 billion yuan on R&D from 2023 to 2025, and 643 of its 838 employees as of the end of 2025 were researchers.

The report also details Suiyuan’s deep relationship with Tencent. Tencent holds a 20% stake and is Suiyuan’s largest customer. Their collaboration began in 2019, with products now deployed at scale in multiple Tencent AI businesses, including national-level applications. The article argues that a real-world customer like Tencent provides more than orders: it acts as a long-term validation ground where engineering issues invisible in labs are exposed and feed back into next-generation chip and software design. Financially, Suiyuan’s revenue rose from 301 million yuan in 2023 to 990 million yuan in 2025, and reached 1.12 billion yuan in the first half of 2026, already exceeding the full-year 2025 figure.

The recent launch of the Yunxun ESL64-O supernode with ZTE is a key move in Suiyuan’s cluster strategy. The supernode uses an OEX orthogonal backplane architecture to achieve "zero-cable" interconnection, which replaces flat cables with vertical spatial crossing. This design, according to public information, supports 10,000-card cluster networking and is considered an example of system-level innovation for domestic intelligent computing infrastructure.

The report concludes that the first ticket for domestic AI chip companies was designing and mass-producing a chip with sufficient performance. The next ticket is turning that chip into a scalable computing system. Suiyuan’s eight-year effort has built a computing system spanning chips, hardware, software, and clusters, and it is now at the point where its commercial value faces a real test beyond its anchor customer.