AnswersPublished17 sources
From June launch to commercial service: inside Wuxi’s 2,112-GPU AI center
The center’s first phase launched in June, entered trial operation in July and began commercial service in October, with 2,112 Moore Threads MTT S5000 GPUs deployed. Reports say the cluster reached high utilization within a few months and was handling model training and inference, though published customer counts an...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- The center’s first phase launched in June, entered trial operation in July and began commercial service in October, with 2,112 Moore Threads MTT S5000 GPUs deployed.
- Reports say the cluster reached high utilization within a few months and was handling model training and inference, though published customer counts and ramp up timelines differ.
- The facility is planned to grow to 10,000 GPUs, but that is not its current installed capacity; performance claims and long term commercial prospects need more independently verifiable data.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: As reported by Zhidx on November 2, 2026, how did Moore Threads’ MTT S5000 domestic computing center in Huishan, Wuxi, progress from its Jun. Article summary: The reporting portrays Huishan as a working, heavily used domestic-GPU computing service—not merely a hardware demonstration—with 2,112 S5000 GPUs operating against a planned 10,000-GPU scale.[3] However, the available e. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clic
Wuxi’s Huishan computing center has moved quickly from equipment rollout to commercial service. Its first phase launched in June 2026, entered trial operation in July and began commercial operations in October. By early November, reports said its 2,112 Moore Threads MTT S5000 GPUs were running at or near capacity.2
3
That is a notable operational milestone, but it is not the same as reaching the center’s planned 10,000-GPU scale. The current figures describe the first phase; the larger capacity remains an expansion goal.2
From launch to commercial service
A report published on November 2 said the center connected 33 mainstream large language models during its trial period before beginning commercial operations in October.2 Its first phase consists of four GPU cluster modules, each with roughly 500 cards, for a total of 2,112 MTT S5000 GPUs.
2
Moore Threads used a modular, turnkey delivery model, integrating servers, networking and storage into an end-to-end cluster solution. A company executive cited in the report said the roughly 2,000-GPU cluster reached full load within three months.2 An October 15 company announcement gave a shorter figure—two months to full load for about 2,000 cards—and the available reports do not explain the difference in timing.
22 The cautious takeaway is that the cluster ramped up rapidly, while the precise timeline varies by account.
What “full” utilization tells us
One November 2 report said more than 20 customers were testing or running business workloads on the platform, with training accounting for over 62% of activity and inference making up the rest.3 Another report put utilization at 92.9% and counted 55 customers.
7 The sources do not explain why the customer totals differ or provide enough detail to compare how utilization was calculated.
A November 4 interview added that around 60% of customers were paying, while others were still testing and adapting domestic servers.19 So “full load” should not be taken to mean that every GPU was being used for paid production work.
Even with those qualifications, the figures point to a service handling real workloads—not just a hardware demonstration. The reported share of training activity suggests the facility is being used to train models as well as run inference.3 Customer counts and utilization alone, though, cannot establish profitability, long-term reliability or how its GPUs compare with alternatives across different workloads.
A 10,000-GPU plan, not today’s footprint
Huishan was planned from the outset with a 10,000-GPU scale in mind, but its first phase has 2,112 cards.2 Autonomous-driving company Neolix plans to conduct larger-scale model training at the center, with projected computing needs approaching 10,000 GPUs.
2
32 That demand aligns with the expansion plan; it does not mean a 10,000-GPU training run is already underway.
Moore Threads and JD Cloud have also outlined a separate plan for a 100,000-GPU domestic cluster, intended to support training, inference and embodied intelligence, among other uses.18 That proposed project should not be conflated with Huishan’s current deployment or its planned expansion to 10,000 GPUs.
How to read the performance claims
A report said MTT S5000 inference tests on DeepSeek V4 Flash reached 92% of the performance of a comparable H800 configuration.7 The available material does not provide enough information about test conditions, measurement methods or independent replication to conclude that the two GPUs perform similarly across models and workloads.
The clearest indication of Huishan’s progress is therefore operational: domestic GPUs have moved from deployment and adaptation into a regional computing service handling customer workloads, including model training and inference. Its fast ramp-up and expansion plans are worth watching. But claims about performance, a future 10,000-GPU footprint and commercial sustainability will require more verifiable operating data.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "From June launch to commercial service: inside Wuxi’s 2,112-GPU AI center"?
The center’s first phase launched in June, entered trial operation in July and began commercial service in October, with 2,112 Moore Threads MTT S5000 GPUs deployed.
What are the key points to validate first?
The center’s first phase launched in June, entered trial operation in July and began commercial service in October, with 2,112 Moore Threads MTT S5000 GPUs deployed. Reports say the cluster reached high utilization within a few months and was handling model training and inference, though published customer counts and ramp up timelines differ.
What should I do next in practice?
The facility is planned to grow to 10,000 GPUs, but that is not its current installed capacity; performance claims and long term commercial prospects need more independently verifiable data.