Moore Threads is using its latest results briefing to argue that a domestic GPU supplier can serve more than conventional model training. Eastmoney, carrying reporting from The Paper, said the Chinese company described a stack that stretches from large-model training and inference to simulation and deployment for embodied, or physical, AI systems.
The claims come as Chinese technology companies seek alternatives to imported AI hardware and as robotics developers look for computing platforms that can handle more than language-model workloads. Moore Threads says its MTT S5000 clusters have supported large-model training and physical-AI research, while its planned MTT C256 supernode is designed for larger training and inference installations.
The company also disclosed new financial figures. It reported first-half 2026 revenue of 1.736 billion yuan, an increase of 147.42% year on year, exceeding the 1.506 billion yuan it reported for all of 2025. Those are company-reported results, not independently audited measures of market position, but they offer a window into the scale the company says it is reaching as domestic demand for AI infrastructure expands.
MTT S5000 Is Presented as a Training Platform, Not Just a Chip
Moore Threads said MTT S5000 AI training clusters have been commercially deployed in Beijing, Wuxi, and Hangzhou. The company identified the S5000 with its PH100 chip and said the clusters had completed from-scratch training for an MoE-236B foundation model using 25 trillion tokens of data. It also cited full-process training for RoboBrain 2.5 and EvoPhys-World.
Those details are important because they define the company’s position as a systems supplier rather than a standalone accelerator vendor. Training a large model requires chips, network interconnects, system software, scheduling, storage, cooling, and operational support. A claim that a cluster completed a training run says something about the integration of those elements, even if it does not by itself establish performance against competing platforms.
The company said it has been developing a broader physical-AI toolchain around the hardware. That includes MT Lambda, which it describes as a simulation platform, and a MUSA backend for MuJoCo Warp. It also said it had completed Sim2Real verification work with bipedal and quadruped robots. The development path connects synthetic data and simulation to eventual hardware deployment, a workflow increasingly important to robotics teams that cannot collect all of their training material in the physical world.
Every performance claim in this section is Moore Threads’ own. The reporting does not provide a common benchmark against Nvidia, AMD, or another domestic competitor. It does, however, show that the company is trying to position its software ecosystem around the specific needs of embodied-AI development rather than only generic data-center inference.
The Planned C256 Supernode Raises the Scale Ambition
The MTT C256 is central to the next stage of that strategy. Moore Threads has described the supernode architecture as a 256-GPU high-speed interconnect system aimed at clusters used for large-scale training and inference. At the September briefing, company executives said the design is intended to serve installations ranging from ten thousand to one hundred thousand cards.
That is an ambitious target. The company describes sub-microsecond card-to-card latency as an architectural goal, but it is a company-provided specification rather than an independently tested measurement. Actual performance will depend on hardware availability, network design, software maturity, workload mix, and the scale of deployment.
The C256 is not entirely new to EastFrontier readers. In July, Moore Threads introduced its 256-GPU MTT C256 system. The new briefing adds a physical-AI and commercialization angle. It connects the supernode’s planned scale to a wider stack that includes model training, simulation, and agents intended to operate in physical settings.
That distinction matters because physical AI stresses different parts of a computing environment. It can require rendering, physics simulation, synthetic-data generation, multimodal training, and low-latency inference at the edge. A company that can run a language model may not automatically have the software support needed for robotics. Moore Threads is presenting its stack as an answer to that difference.
Domestic Compute Is Becoming a Full-Stack Competition
China’s domestic accelerator market is not a single-product race. Companies have to show that their hardware works with model frameworks, inference engines, networks, development tools, and the growing number of Chinese models that enterprises want to deploy. The more these layers work together, the more attractive a platform can become to customers trying to reduce integration risk.
EastFrontier recently reported that China’s AI chipmakers face a new training-compute test. That test is not only about producing more chips. It concerns whether domestic systems can run demanding workloads with the reliability, tooling, and scale that model developers and cloud operators need.
Moore Threads’ first-half revenue growth provides evidence of commercial movement, but it does not settle that competitive question. The company is still reporting a net loss, and its hardware and software roadmap will have to be evaluated through actual deployments rather than corporate presentations. Its statement that the S5000 has been sold at scale in several cities is more informative than a generic claim of readiness, but customers, workloads, and utilization levels have not been disclosed in detail.
The physical-AI emphasis is nevertheless strategically sensible. Robotics and simulation can expand the set of workloads that domestic GPUs support, and they create an opportunity for hardware companies to partner with research institutes and developers building embodied models. The company’s cited work with RoboBrain 2.5 and EvoPhys-World is an example of that approach.
For customers, the next evidence will come from reproducible training runs, independent performance data, and stable deployment across a wider range of models and robotics applications. For Moore Threads, the opportunity is to demonstrate that a domestic GPU platform can operate across the full path from model development to simulated testing and real-world deployment. The September briefing lays out that ambition. Its practical value will be determined by how much of the stack proves usable outside the company’s own presentations.