Supernodes emerging as new standard for computing power

A wall of black server racks dominated Hall H2 of the 2026 World Artificial Intelligence Conference, or WAIC, which concluded in Shanghai last week — the first public global debut of China's first domestically produced 100,000-card AI supercluster, the Dawning 8000 (Dengfeng), which has been named one of WAIC 2026's "Treasures of the Pavilion".

The exhibit signals a systemic shift in China's computing power landscape. Almost every major computing vendor at the conference rolled out supernode products with varying card counts: Chinese tech company Dawning Information Industry, also known as Sugon, showcased its Dawning 8000, Lenovo displayed its supernode products and Huawei unveiled its Atlas 950 SuperPoD.

The race has moved from single-card performance to holistic supernode systems — a fundamental paradigm shift in domestic computing power competition.

Li Liu, vice-president of Sugon, said at a WAIC forum that supernodes are evolving from high-end custom equipment to standardized computing units, with modularization becoming key to large-scale development. "In the future, supernodes will further lower the threshold for high-end computing power, providing foundational support for the large-scale development of the AI industry," he said.

Zhou Tao, general manager of the server business unit at Lenovo China's infrastructure business group, told China Daily that "this year will be a big year for AI application implementation."

Lenovo is pursuing a two-pronged strategy: seizing rapid growth in internet business while deepening penetration into traditional non-internet industries. With AI-driven price increases in infrastructure and storage components, both unit prices and sales volumes are rising.

Zhou expressed confidence that Lenovo's China infrastructure business can approach 100 billion yuan ($14.77 billion) in annual revenue next year, leveraging its comprehensive AI product portfolio.

To address fragmentation caused by the growing diversity of graphics processing unit and central processing unit combinations, Lenovo has introduced an innovative"1+3+N" universal architecture. This platform accommodates multiple mainstream CPUs and various GPU configurations — including 8-card, 16-card and supernode setups — eliminating duplicate development efforts.

"This strategy significantly reduces research and development costs and shortens development cycles, while giving customers the flexibility to swap hardware components and rapidly iterate products," Zhou said.

He highlighted the efficiency gains of supernodes: traditional GPU servers only interconnect cards within the server; when linking to another server, latency is introduced through network and CPU connections. In supernodes, all GPUs are directly interconnected. "Our supernode solution can improve efficiency by more than 30 percent compared with traditional servers," Zhou said.

Noting that training has been shifting toward inference over the past two years, Zhou said supernodes — with their improved computing power and reduced energy consumption — will gain wider adoption in the next one to two years. "This year marks the debut of supernode products from major manufacturers — it's the first year," he declared.

Dong Longfei, senior vice-president of Moore Threads, another Chinese computing company, described the supply-demand imbalance: "Computing power is in short supply. The AI industry is much larger in scale than the internet industry. We're building the computing foundation for the next ten to twenty years — it's like building railways and highways in the past."

He said: "A supernode is like an orchestra — each card plays well individually, but they must collaborate harmoniously for the whole piece," he said.

Gao Linli, also a senior vice-president at Moore Threads, emphasized ecosystem integration. "We cannot let developers adapt to different toolchains for each vendor — the cost is too high," she said. The industry is pushing for a unified approach where "one set of code adapts to all domestic hardware. "Only by solving this can the domestic computing power industry unite and thrive, she added.

As WAIC 2026 showcases, China's computing power industry is racing to build the robust, scalable infrastructure that will underpin the AI era — and the supernode is emerging as the new standard.

masi@chinadaily.com.cn