Moore Threads' 2,112-GPU Wuxi cluster goes commercial and reports 92.9% utilisation
The MTT S5000 cluster, planned to reach 10,000 cards, served 55 customers and 33 models in trials; operators say DeepSeek V4 Flash inference ran at 92% of an H800 setup under the same test, figures not independently verified.
Why it matters
The Wuxi build is early evidence of whether Chinese GPU makers can run commercial AI clouds at scale without Nvidia. Training makes up more than 62% of Wuxi's workloads, and the cluster supports more than 30 model families including Qwen, GLM, DeepSeek, Kimi and MiniMax. Autonomous delivery firm Neolix plans training demand near 10,000 cards, about five times today's count. All performance numbers come from the operators via a Zhidx report. Watch for the next expansion phase and any independent throughput tests.
Key facts
- 264 servers carrying 2,112 MTT S5000 cards
- Reported utilisation: 92.9%
- 55 customers and 33 models onboarded in trials
- Training is more than 62% of workloads
- DeepSeek V4 Flash inference at 92% of an H800 setup, per operators
- Commercial service began in October