Moore Threads: MTT S5000 and the KUAE 10K-Card Cluster

1. Introduction

1.1. Vendor Positioning

Moore Threads (688795.SH) is the "first stock of domestic GPUs": it listed on the Shanghai Stock Exchange's STAR Market on 2025-12-05. Its technical route stands apart within this group — the full-feature GPU route: rather than making a single-purpose AI inference accelerator (DSA), it self-develops full-feature GPUs (graphics rendering + AI compute + video processing), carried on the MUSA unified system architecture, benchmarking against the international mainstream GPU ecosystem. This route determines that its AI intelligent-computing business (94.73% of revenue in H1 2025) is built on general-purpose GPUs, not dedicated accelerators.

The narrative thread of this article: full-feature GPU route + commercial catch-up player in 10K-card clusters; all performance/utilization metrics are company figures, lacking third-party benchmarks (see Section 6.3 and the information-gap declaration).

1.2. Basic Information Card

1.3. Financial and Listing Performance

1.4. Position in the AI Harness System

Moore Threads is the only domestic vendor in this group that publicly discloses both training-efficiency metrics and a layered cluster-management platform: the layered structure of KUAE Platform (cluster management) + KUAE ModelStudio (LLM service) gives L3/L5 a productized form that can be assessed. Its Harness role: providing a complete silo from chip to cluster-scheduling platform for the "Xinchuang + intelligent-computing center" scenarios.

2. Glossary

3. Feature Description and Product Line

3.1. MTT S5000 Flagship Chip

Company figures (earnings-call sessions 2026-04/05, grade A):

1. "Pinghu" architecture, large-scale commercialization in 2025;
2. Up to 1000 TFLOPS of dense AI compute per card, full precision coverage from FP8 to FP64;
3. Day-0 adaptation to SOTA LLMs such as DeepSeek, GLM, MiniMax, Kimi, and Qwen;
4. Inference figures (first MUSA Developer Conference 2025-12, company figures): DeepSeek R1 671B full model on a single MTT S5000, Prefill throughput exceeding 4000 tokens/s and Decode exceeding 1000 tokens/s.

It must be emphasized: the above inference data and compute metrics are all self-reported by the vendor, with no MLPerf or third-party same-scenario comparison (comparative data vs. H100/H200/MI300X is missing; see the information-gap declaration).

3.2. KUAE 10K-Card Cluster

Company figures (annual report + earnings-call session, grade A · company figures):

3.3. MUSA Unified System Architecture

1. Full-stack self-developed: chip, instruction set, and software stack are all self-developed and can evolve independently;
2. Compatible with the international mainstream GPU ecosystem: lowers the migration cost for existing CUDA-ecosystem users;
3. Ecosystem activity: the first MUSA Developer Conference (2025-12) disclosed inference performance and ecosystem progress.

3.4. Roadmap

1. MTT C256: super-node architecture plan (disclosed in the company annual report, grade A);
2. "Huagang": technical research underway for 100K-card cluster architecture;
3. MTT AIBOOK: an edge-terminal product based on the "Changjiang" SoC.

4. Platform Architecture

4.1. Technology Stack, from Chip to Cluster

Figure 5-1 | Moore Threads: the technology stack from MTT S5000 to the KUAE 10K-card cluster

Data sources: Moore Threads 2025 annual report, earnings-call sessions, and disclosures at the first MUSA Developer Conference; the diagram is drawn based on this article's analysis.

4.2. KUAE Platform Layering

The productized layering of the KUAE system is the core of this article's architecture analysis: KUAE Platform manages the cluster (scheduling/operations/fault tolerance), while KUAE ModelStudio manages model services (inference-as-a-service). This layering maps clearly onto the structure of NVIDIA's "Mission Control + Dynamo/NIM" — domestic vendors have converged on the practices of leading international players in the form of cluster-management platforms.

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 / L4 Context and State Layers

The company annual report explicitly lists "KV Cache management, inference decoupling (PD separation), batch scheduling, and low-latency service orchestration" as technology layout directions (grade A · company figures) — note these are layout directions, not delivered capabilities, and no detailed specifications of corresponding product components appear in public materials. L4's 10K-card fault detection and elastic recovery are supported by cluster-stability metrics (effective training time above 90%).

5.3. L3 Orchestration and Control Layer

The dual-platform layering of KUAE Platform (cluster management) and KUAE ModelStudio (LLM service) provides productized entry points for cluster scheduling and model services. Compared with the open-source serving ecosystem of Article 09, Moore Threads chose to build its own platform layer rather than directly reuse the vLLM/K8s ecosystem — this is a pragmatic choice in the Xinchuang-delivery context, but the openness of the ecosystem remains to be seen.

5.4. L5 Evaluation and Observability Layer

What is valuable about Moore Threads is that it publicly discloses the training-efficiency indicator system itself: Dense MFU 60%, MoE MFU 40%, effective training time ratio 90%+, linear scaling efficiency 95% — this set of indicators is comparable to the MFU experience ranges of the training-framework ecosystem in Article 10 (70B Dense: Megatron 45—55%+). But it must be emphasized: all are company figures, with no MLPerf or independent third-party verification (weaker than the grade-A verification of NVIDIA in this group).

5.5. L6 Governance and Security Layer

Full-stack self-control (chip / instruction set / software stack) is its core governance selling point in Xinchuang and key-industry scenarios; successive wins of provincial intelligent-computing-center customers (with government backgrounds) constitute supporting evidence.

6. Practical Cases

6.1. KUAE 10K-Card Cluster: Company-Figure Metrics

The 10K-card cluster has been deployed and is in operation (company figure: 10 ExaFLOPS), with supporting metrics in Section 3.2. Case-writing anchors: two company-figure data points — the KUAE cluster MFU metrics and the MTT S5000 single-card R1 671B inference throughput — both lack third-party benchmarks, and are the disclaimer premise that must be carried when assessing Moore Threads' true level.

6.2. Intelligent-Computing Centers and Commercial Orders

6.3. Data Credibility Notes

Almost all performance and utilization data in this article are company figures (grade A · company disclosure ≠ third-party verification), following these three disciplines:

1. Company figures annotated at each occurrence: wherever metrics such as MFU, ExaFLOPS, and tokens/s appear, they are marked "company figures";
2. No comparison against competitors without third-party verification: there is no same-scenario third-party comparative data for MTT S5000 vs. H100/H200/MI300X, and this article makes no performance benchmarking claims;
3. Commercial orders follow company disclosure / annual report: the ¥6.6 亿 order comes from the annual report (grade A), while the ¥20 亿 intelligent-computing-centers total comes from a Baidu Baike citation (grade C); the credibility grading is annotated item by item.

7. Summary

Strengths:

1. Scarcity of the full-feature GPU route: the only company in China to go public with a full-feature GPU, and the combination of graphics + AI capabilities has associative value in intelligent-computing-center and Xinchuang scenarios;
2. Public indicator system: proactively discloses core training-side metrics such as MFU / scaling efficiency / effective training time, with transparency higher than most domestic peers;
3. Commercial validation: turned profitable in its first year after listing (2026 Q1), a ¥6.6 亿 10K-card order, and coverage of intelligent-computing centers in 6 provinces;
4. Platform productization: the dual KUAE platform layering provides a deliverable cluster-management form.

Weaknesses:

1. No third-party benchmarks: all performance metrics are company figures, MLPerf submissions are missing, and cross-vendor comparison is not feasible;
2. MUSA ecosystem scale yet to be validated: no official figures for operator coverage or open-source repository scale;
3. High R&D intensity: R&D expense is 86.68% of revenue, and profitability is sensitive to revenue growth.

Applicable boundaries: provincial/regional intelligent-computing-center construction (Xinchuang and self-control requirements), scenarios needing full-feature GPU (AI + graphics) mixed loads, and domestic efforts at 10K-card training clusters.

Selection recommendations: treat "company-figure" metrics as initial screening rather than acceptance criteria, and require measurement of MFU and scaling efficiency on your own models/loads at acceptance time; validate MUSA's operator coverage on the target training framework (Megatron-style API, see Article 10) before 10K-card deployment; and watch the landing pace of the MTT C256 super-node and the "Huagang" 100K-card architecture.

Information-Gap Declaration

1. Missing verification by third-party benchmarks (MLPerf, etc.): the KUAE cluster's MFU / scaling efficiency are all company figures with no independent review;
2. Missing same-scenario third-party comparative data between the MTT S5000 and international same-generation products (H100/H200/MI300X);
3. No official figures for MUSA ecosystem open-source repository scale and operator coverage;
4. The ¥20 亿 provincial intelligent-computing-centers total is based on a Baidu Baike citation of company disclosure (grade C); it is recommended to re-verify against company announcements.

8. References

1. KUAE Intelligent-Computing Center — Baidu Baike (cross-checked with company disclosure), 2025—2026. https://baike.baidu.com/item/KUAE（夸娥）智算中心/63862920
2. Moore Threads 2025 Annual and 2026 Q1 Earnings-Call Session — China Securities Network Roadshow Center, 2026. https://roadshow.cnstock.com/fbh/mexc2025
3. "First Stock of Domestic GPUs Is Born! Moore Threads Officially Listed" — Kuaikeji, 2025-12. https://news.mydrivers.com/1/1090/1090711.htm
4. "Domestic GPU Commercialization Speeds Up; Moore Threads Returns to Profit in Q1" — China Economic Net, 2026-04. https://www.ce.cn/xwzx/gnsz/gdxw/202604/t20260427_2932385.shtml
5. Moore Threads 2025 Annual Report — Eastmoney Data Center, 2026. https://data.eastmoney.com/notices/detail/688795/AN202604261821587780.html
6. Moore Threads Official Website (MUSA product line) — Moore Threads, 2025. https://www.mthreads.com/