Inside WAIC: How the Shuguang 8000 Supercomputer Hit Full Capacity on Day One

During the recent WAIC exhibition, the Shuguang 8000—China’s first fully domestic AI supercluster with 100,000 GPUs—was displayed in a dedicated showcase area and highlighted as a "treasure of the hall" by the event organizers.

The system’s operational data released on July 18 showed that in its first week it ran at full load, handling more than 150,000 jobs per day with a single‑day peak exceeding 500,000, indicating that the demand for massive, high‑quality compute power from research institutions, enterprises, and developers was already accumulated.

Full load does not mean congestion. The Shuguang 8000 relies on an intelligent scheduling engine and data‑affinity algorithms that dynamically allocate resources based on task type, data location, and system state, allowing it to sustain millions of concurrent user requests without queuing.

Performance is achieved through a combination of technologies: a globally first high‑density cabinet design that raises compute density by 20× per node, the self‑developed scaleFabric‑like IB‑native RDMA interconnect that links all 100,000 GPUs reliably, ParaStor distributed storage that won two global IO500 first‑place awards for 2026, and immersion phase‑change liquid cooling that ensures long‑term stability.

To lower the barrier to using this high‑end compute, the Shuguang 8000 offers several incentive programs—such as the "100k‑GPU Co‑creator Incentive Plan" and ecosystem partnership initiatives—allowing enterprises, research teams, and developers to access the supercomputer on demand without building their own clusters.

The article concludes that the first‑week full‑load operation, the near‑thousand application adaptations covering large‑model training, high‑throughput inference, and scientific computing, and the system’s stability under heavy load demonstrate that the 100k‑GPU AI infrastructure is no longer a concept but a practical, reliable service for real workloads.