On October 31, Moore Threads signed a strategic cooperation agreement with Wuxi's Huishan Investment Promotion Group and Neolix. The same day, Neolix's East China headquarters project for intelligent driverless vehicles landed in the Huishan Economic Development Zone, with planned total investment of 1 billion yuan. The agreement rests on the Wuxi (Huishan) domestic AI data center, whose first phase of 2,112 Moore Threads MTT S5000 cards has been in service for three months and is running essentially at full load.

On November 2, Securities Times and Chipinsights each visited the facility, and its operating data was disclosed for the first time: 264 domestic GPU servers, 2,112 compute cards, 92.9% utilization of computing power, 59.7% of that computing power paid for, 55 customers engaged, and more than 20 already running tests or live workloads.

On a Fully Loaded Cluster, 60% of the Work Is Training

The AI data center was powered on in June, began trial operation in July, and entered commercial service in October, hosting 33 mainstream large models during the trial period. The first phase is planned on a 10,000-card scale; what has actually been deployed is four GPU cluster modules of roughly 500 cards each. The site also serves as the Huishan node of the southern Jiangsu computing cluster within the Yangtze River Delta hub under East Data West Computing.

The most counterintuitive figure is the workload mix: training accounts for more than 62%, combined training and inference 14.7%, and inference 23.1%. The claim that domestic cards are good only for inference does not hold up on this machine.

The customer list matters, though. Zhang Tao, deputy general manager of Huishan Investment Promotion Group, told Securities Times that the center currently serves local universities, research institutes, and medical and financial institutions, "running existing foundation models to meet demand for agent applications in local vertical sectors." The "training" here is more likely fine-tuning of industry models than trillion-parameter pre-training. On the same cluster, Langjing Innovation is running aviation simulation work and Neolix is training autonomous driving models, both multimodal, simulation-heavy workloads.

That also explains the choice of full-function GPUs. Zhou Ling, vice president of Langjing Innovation, said flight-perspective work places extreme demands on low latency and parallel multi-view rendering, spanning 3D rendering, three-dimensional reconstruction, and simulation computing: "Apart from NVIDIA, Moore Threads is the only domestic supplier of this class of GPU."

Forty Percent of the Computing Power Earns Nothing Yet

With utilization at 92.9% and paid computing power at 59.7%, roughly a third of the load is testing and trials. Zhang Tao put it more plainly: "The entire cluster is now at 100% utilization, including test and production users, and we still have prospective customers queuing. Paying users account for about 60%."

Read the two measures together: the cluster is genuinely full, but by compute, only six-tenths of it is generating revenue. Of 55 customers engaged, just over 20 are actually using the system, less than half. Chipinsights reported that dozens more organizations are waiting to begin testing on the cluster.

Zhang Tao left room for maneuver on expansion: "If the vendors have further usage needs, we'll work together to find investors for a second and third phase." The 10,000-card build is not committed investment; the money has to be found first.

92% of an H800, on the Operator's Own Numbers

According to operator Huishan Investment Promotion Group and co-investor China Mobile Cloud, tests using DeepSeek V4 Flash on the center's MTT S5000 resource pool showed, at 7K input and 2K output, single-machine output throughput of 994.97 tokens/s, effective output throughput of 1,241 tokens/s, and a maximum of 48 concurrent streams meeting SLA, reaching 92% of NVIDIA's H800, with 1.6 Tbps of memory bandwidth and memory priced at less than half that of the H800.

On the hardware side, the MTT S5000 is built on the fourth-generation MUSA "Pinghu" architecture, with up to 1,000 TFLOPS of dense AI computing power per card, 80GB of memory, 1.6TB/s of memory bandwidth, and 784GB/s card-to-card interconnect. It supports FP8 through FP64 and is compatible with PyTorch and vLLM.

One caveat: the "1,000 TFLOPS" figure comes with no stated precision. The same number can differ twofold between FP16 and FP8, and that is where cross-vendor computing power comparisons most often go wrong. The "92% of an H800" result, likewise, is a measurement on a single machine, with a single model, at a specified input and output length, which is not the same as cluster-level parity.