Moore Threads signed a strategic agreement with Wuxi Huishan Investment Group and autonomous-driving firm Neolix on October 31, multiple outlets including ZhiDongxi and The Paper reported. The Huishan domestic computing center's first phase — 2,000 Moore Threads GPUs — hit full utilization within three months of launch, while Neolix's training demand approaches the 10,000-card scale planned for the center's expansion.
More than 20 customers now run testing or production workloads on the phase-one cluster, with training accounting for over 62% of business — a training-plus-inference mix has taken shape. On October 30 alone the center signed compute agreements with ten enterprises and universities covering embodied AI, video generation, spatial intelligence and life sciences. Its operator says the project is China's first commercially operated, all-domestic computing cluster, spanning hardware deployment, model adaptation and commercial delivery.
The training workloads run on Moore Threads' MTT S5000 training-inference card. Public specs list the fourth-generation MUSA "Pinghu" architecture, peak dense AI compute of 1,000 TFLOPS per card, 80GB memory with 1.6TB/s bandwidth and 784GB/s interconnect, full FP8-to-FP64 support and compatibility with PyTorch and vLLM. Neolix has validated its autonomous-driving training on the card and plans to scale toward the 10,000-card mark.
Demand-side projects landed the same day: Neolix's East China headquarters for autonomous unmanned vehicles broke ground in Huishan with planned investment of 1 billion yuan, housing sales, settlement, dispatch and R&D functions. Driving workloads span AI computation, graphics rendering and physics simulation — the single-chip versatility of full-function GPUs is the main draw.
Phase one full, training over half, 10,000 cards on the table — the question for domestic compute has shifted from "can it run" to "can it carry production scale." Watch the post-expansion rack cadence and customer renewals.