From December 17 to 19, the 2025 Artificial Intelligence Innovation Conference (HAIC 2025), organized by the Photosynthesis Alliance, was held at the Kunshan International Convention and Exhibition Center.
The 5,000-square-meter exhibition area buzzed with activity, featuring booths from leading brands such as Sugon, Inspur, and SenseTime. Demonstration screens displayed training efficiency curves of the scaleX 10,000-card supercluster system alongside real-time updates on liquid-cooled cabinet temperature control data. Nearby, engineers explained the architectural breakthrough behind the "world's first single-cabinet 640-card system"—a microcosm of China's systemic innovation in AI computing.
"This is not just a product showcase but a systematic demonstration of China's full-chain computing capabilities," said Ren Jingyang, Secretary-General of the Photosynthesis Alliance. He emphasized that industry forces, represented by the alliance, are driving China's AI computing from fragmented technological breakthroughs toward a new phase of high-quality, ecosystem-driven development.
As AI transitions from "model-driven" to "engineering-driven," new frontiers like scientific large models, world models, and vertical-specific models are emerging. However, challenges such as limited high-end AI hardware supply, fragmented software-hardware ecosystems, and high computing costs persist, forming "performance walls" and "ecosystem walls" that hinder progress. The industry is now rallying around "open architecture" to dismantle these barriers, establishing unified technical standards and collaborative ecosystems to optimize performance from chips to applications. HAIC 2025 serves as a pivotal platform to validate and advance this open-collaboration approach.
**Hardware Innovation Breaks the "Performance Wall"** At the heart of the exhibition was Sugon’s scaleX 10,000-card supercluster, the most anticipated hardware innovation. As China’s first commercially viable 10,000-card AI cluster system, its technical capabilities rival or surpass those of global leaders.
Comprising 16 supernodes, the scaleX system supports 10,240 AI accelerator cards, delivering over 5 EFLOPS of total computing power. Its single-cabinet computing density is 20 times higher, with a power usage effectiveness (PUE) as low as 1.04. Designed on an open AI computing architecture, it supports multi-brand accelerators and mainstream computing ecosystems while optimizing compatibility with over 400 large and world models.
A Sugon representative noted strong market demand for supernodes, with pre-orders rising for 2026. As CPU/GPU power consumption grows, liquid cooling (cold plate/immersion) will become the dominant thermal solution.
"Domestic computing must advance beyond single-card performance to systemic architectural breakthroughs," said Ding Zhenlei, Server Center Director at UCloud. Over the past two years, China’s computing power has grown tenfold, but scaling further requires overcoming barriers in hardware integration, cooling efficiency, and network communication. By 2026, GPU power per unit will continue rising, making liquid cooling essential—a challenge addressed by high-density integration under open architectures.
Sheng Lebiao, Senior Engineer at Nanjing University’s High-Performance Computing Center, highlighted that scaling from single cabinets to 10,000-card clusters represents a systemic engineering breakthrough, shifting the AI computing race from chip competition to a broader contest of system integration, software synergy, and ecosystem building.
**Open Ecosystem Tackles Fragmentation** "Open" and "ecosystem" emerged as key themes. Despite hardware progress, software bottlenecks and accessibility for SMEs and research institutions remain hurdles.
Ren Jingyang identified three challenges: high application migration costs, resource mismatches, and ecosystem fragmentation. The Photosynthesis Alliance’s open architecture aims to unify interfaces and standards, enabling cross-platform compatibility and reducing fragmentation burdens.
Exhibitors showcased solutions based on unified software stacks, spanning AI frameworks, operator libraries, and scheduling platforms. Innovations like compute-storage-network tight coupling and supercluster intelligent scheduling engines reduced end-to-end latency, while digital twin O&M platforms achieved 99.99% availability for ultra-large clusters.
Zhang Huamin, R&D Director at Fuzhou Chuangshi Xunlian, highlighted their domestically developed industrial control computers, offering reliable edge computing platforms to support industry-wide intelligent transformation.
"AI computing has entered an era of systemic competition," said Li Jun, Director of the National Advanced Computing Industry Innovation Center. Sustained advantage requires open collaboration across chips, servers, networks, and software—a "value loop" spanning the entire supply chain.
**Ecosystem Collaboration Paints the "Big Picture"** HAIC 2025 also served as a springboard for ecosystem partnerships.
Sugon, SenseTime, Shanghai Daxiao Infinite Robotics, Kylin Software, Empyrean Technology, and Hygon Information Technology announced strategic collaborations in AI computing optimization, world models, and embodied intelligence applications. Joint initiatives like the Scientific AI Research Consortium and the Open AI Computing Architecture Lab were unveiled, targeting large models, heterogeneous computing, unified software stacks, and supercluster optimization.
The "Light Up 100 Cities 2.0" initiative aims to deploy open architecture solutions in manufacturing, energy, and research. "Our next focus is lowering barriers for SMEs," said a Photosynthesis Alliance representative.
An Dongming, Vice President of Longshine Technology, noted growing industry momentum for collaboration, with efforts to tackle open-architecture hardware-software adaptation and accelerate AI adoption in smart manufacturing and automotive sectors.
At the exhibition’s edge, a smart terminal powered by domestic AI hardware drew crowds. From edge devices to AI all-in-ones, and from computing clusters to industry model platforms, the event mapped China’s "full-stack AI computing ecosystem," embodying the logic of "open-collaborate-deploy."
"2025 marks a turning point from fragmented breakthroughs to systemic output," said Yuan Bo, an expert at ZhiCan Think Tank. While open architecture lowers entry barriers, long-term global competitiveness hinges on mastering standards, enriching application ecosystems, and maturing cross-platform toolchains.
Like a prism, HAIC 2025 reflected China’s AI computing industry striving for openness amid constraints and breakthroughs amid bottlenecks. As 10,000-card clusters evolve from blueprints to large-model training foundations, China’s open computing path is forging its own answers.
Comments