---
title: "Moore Threads’ Wuxi S5000 Cluster Hits 92.9% Utilization for 55 Clients"
slug: "moore-threads-wuxi-s5000-cluster-hits-92-9-utilization-for-55-clients"
date: "2026-11-03"
category: "AI & Automation"
author: "Dr. Kai Wei"
correspondent: "Dr. Kai Wei"
correspondentRole: "Lead Frontier AI Correspondent"
readingTimeMinutes: 1
tags: ["AI","News"]
originalSourceUrl: "https://pandaily.com/moore-threads-wuxi-huishan-mtt-s5000-cluster-2112-gpus-92-9-percent-utilization-deepseek-v4-flash"
canonicalUrl: "https://aurav.ae/moore-threads-wuxi-s5000-cluster-hits-92-9-utilization-for-55-clients"
publisher: "Aura-V Intelligence (aurav.ae)"
---

# Moore Threads’ Wuxi S5000 Cluster Hits 92.9% Utilization for 55 Clients

> Moore Threads’ domestic GPU farm in Wuxi reached 92.9% utilization, serving 55 customers with 2,112 S5000 cards.

## Key Takeaways (TL;DR)

- The first phase of Moore Threads’ Huishan, Wuxi center runs 2,112 GPUs at 92.9% utilization.
- Training accounts for over 62% of workloads; inference tests hit 92% of an H800 benchmark.
- Expansion plans include a 10,000-card phase and a partnership with JD Cloud for a 100,000-card cluster.

Moore Threads’ MTT S5000 cluster in Huishan, Wuxi entered commercial service in October and is now operating 264 servers with 2,112 GPUs for 55 customers, achieving 92.9% utilization, according to a report by Zhidx [link](https://pandaily.com/moore-threads-wuxi-huishan-mtt-s5000-cluster-2112-gpus-92-9-percent-utilization-deepseek-v4-flash).

Inference testing with DeepSeek V4 Flash (7K input, 2K output tokens) recorded 994.97 tokens per second per machine and an effective 1,241 tokens per second, handling 48 concurrent requests at 92% of an H800 reference setup (these figures were reported by the cluster’s operators and have not been independently verified).

A mixed deployment pairing three S5000 cards with two H20 cards delivered throughput comparable to nine H20 cards, cutting first-token latency by roughly half.

Neolix plans to train its self-driving models at the site, with demand projected toward the 10,000-card target—about five times the current card count. Moore Threads has previously claimed the 10,000-card configuration delivers 10 EFLOPS with 90% effective training time and 95% linear scaling efficiency. Expansion will proceed in phases, and the company is collaborating with JD Cloud on a potential 100,000-card cluster.

---

*Original Think-Tank Source: [Moore Threads' 2,112-GPU Wuxi Cluster Reaches 92.9% Utilization After Going Commercial](https://pandaily.com/moore-threads-wuxi-huishan-mtt-s5000-cluster-2112-gpus-92-9-percent-utilization-deepseek-v4-flash)*