Sugon 8000: China's AI Computing Enters the 100,000-Card Era

On July 10, 2026, at the Sugon Ecological Organization's Intelligent Computing Application Conference, Sugon dropped a bombshell: Sugon 8000 (Dengfeng) — China's first fully domestic 100,000-card AI supercluster — is officially operational. This isn't another headline-grabbing benchmark run. It's a production system, and it's real.

From the "Sugon-1" built with a modest RMB 2 million seed fund in 1993, to 100,000 domestically-made accelerators humming in unison today, Sugon's journey spans four decades. But Sugon 8000 is more than a number. It marks the moment China's AI computing infrastructure moved past the "throw more hardware at it" era and entered a new phase defined by two pillars: supercomputing-intelligence fusion and full-stack self-reliance.

Breaking the Scale Wall: You Don't Just Add Cards

Scaling from 10,000 to 100,000 cards isn't linear — it's exponential pain. Communication complexity, failure rates, energy consumption — all go through the roof when you 10x the nodes. Li Bin, Senior VP of Sugon, put it bluntly: "The biggest challenge is making it deliver the performance and efficiency it's supposed to."

Sugon's answer is full-stack, fully domestic. From Hygon and other homegrown chips at the base, through scaleFabric's InfiniBand-like native RDMA high-speed interconnect, to the ParaStor distributed file system — which claimed dual #1 spots (full-node production and 10-node) on the June 2026 IO500 list. For cooling, the cluster uses world-leading immersion phase-change liquid cooling, supporting megawatt-level power density per rack with year-round free cooling via domestic refrigerants.

This is a closed loop of complete technical autonomy. No black boxes, no chokepoints.

Supercomputing Meets AI: Two Worlds, One Architecture

Sugon 8000's core innovation is its native supercomputing-intelligence fusion architecture. Historically, scientific computing and AI computing were parallel tracks — one chasing FP64 precision for climate simulation, the other optimizing FP16/INT8 throughput for LLM training. The conventional fix was to bolt a supercomputing zone next to an AI zone and call it a day.

Sugon 8000 tears down that wall. The same system natively handles full-precision computing from FP64 down to INT8. The immediate beneficiary is AI for Science — materials discovery, drug screening, protein folding — domains where researchers need both AI's massive parallelism and traditional HPC's numerical rigor.

Tan Guangming, Secretary-General of the CCF High-Performance Computing Committee, called Sugon 8000 an "essential and significant foundational platform" for AI for Science, filling a gap the industry has struggled with for years.

From "What If" to "Let's Do It": Compute as a Utility

A supercomputer's ultimate value isn't benchmark scores — it's productivity. Sugon 8000 is connected to the National Supercomputing Internet, making its capacity accessible to research institutions and SMEs like a utility.

As of launch, over 300 supercomputing-intelligence fusion applications have been optimized on the 100,000-card node, spanning 20+ domains including LLMs, robotics, automotive, drug discovery, new materials, quantum computing, and meteorology. More than 70 applications have scaled beyond 10,000 cards. The system has been battle-tested on protein folding simulations, trillion-atom molecular dynamics, and hundred-trillion-grid turbulence modeling.

Academician Li Guojie of the Chinese Academy of Engineering observed that as AI evolves from large models toward agents and embodied intelligence, and as AI for Science accelerates fundamental research, the resulting compute demands can no longer be met by single-precision systems — they require the kind of system-level innovation that supercomputing-intelligence fusion enables.

Beyond One Machine: Toward Mass Deployment

Sugon 8000 isn't the end of the story. Sugon simultaneously announced a strategic partnership with the Beijing Academy of Artificial Intelligence for Science to begin developing a second fully domestic 100,000-card supercomputing-intelligence fusion system.

The signal is clear: 100,000-card, full-precision computing centers are moving from demonstration projects toward mass replication. In a global AI industry that's increasingly "rational and grounded," Sugon 8000 sets a new template — where the evaluation of large-scale computing centers shifts from raw scale and peak performance toward comprehensive system capability and real-world application efficiency.

Forty years from Sugon-1 to Sugon 8000, compressed into a single machine. But what comes next is more interesting: when 100,000 cards is no longer the ceiling — it's the baseline.