JD Cloud and Moore Threads have announced a plan that would test how far China’s domestic AI hardware can move beyond isolated deployments. At the 2026 JD Global Technology Explorer Conference on September 9, the partners said they intend to build a 100,000-card intelligent-computing cluster based on Moore Threads’ all-function GPUs. The proposed facility would be used for model training, inference, and embodied intelligence workloads, according to Moore Threads.
The most important word is “plan.” Neither company provided an opening date, installed capacity, or delivery schedule. The announcement does not describe a live 100,000-GPU system. It outlines a target for a future domestic-silicon platform, and its commercial importance will depend on whether the companies can assemble, network, operate, and sell access to the cluster at the announced scale.
For JD Cloud, the proposal connects infrastructure to its broader physical-AI ambitions. For Moore Threads, it offers a chance to place Chinese GPUs in a cloud environment large enough to matter for customers building models, simulators, and industrial applications. It also extends a trajectory EastFrontier tracked when Moore Threads moved its GPUs toward physical-AI training.
A Planned Domestic AI Factory at 100,000 GPUs
Moore Threads said the cluster would use its all-function GPUs as the computing base. The companies described the intended use cases as large-model training, inference, and embodied intelligence. JD Cloud also said the platform would support iteration of its JoyAI family of models and help establish a data-to-training-to-simulation-to-deployment loop.
That list matters because it places the project across several demanding workloads rather than positioning it as a conventional cloud expansion. Training large models requires distributed communication and failure management across many accelerators. Inference demands efficient scheduling and latency control. Embodied AI adds simulation and data-generation workloads that can be highly compute intensive. A cluster that meets all three requirements would need hardware, interconnects, orchestration software, and customer-facing cloud services to work together.
The partners have not released a breakdown of the hardware configuration, network topology, power requirements, capital expenditure, or procurement timetable. Those omissions make it impossible to compare the proposed platform directly with existing foreign-GPU clusters. They also mean that the 100,000-card number should be treated as an announced target, not as a verified measurement of operating capacity.
Moore Threads has presented the effort as a step toward larger commercial use of domestic computing hardware. The company says it has already delivered thousand-card and 10,000-card clusters. Its view that the new proposal marks a first for a Chinese GPU within a major cloud provider’s core cluster is a company assessment, not an independently established industry ranking.
Why JD Is Linking Compute to Physical AI
JD Cloud’s stated focus is notable because it ties the planned cluster to supply-chain AI, model training, simulation, and deployment in industry settings. JD has already made physical AI a strategic theme. In August, the company opened its physical-AI stack to global partners, putting its logistics, retail, and automation experience at the center of a broader software and hardware platform pitch.
The September 9 agreement suggests that JD sees domestic compute as a prerequisite for that strategy. A company deploying AI into warehouses, supply chains, and other physical systems needs more than a chatbot model. It needs pipelines for gathering data, training models, testing behavior in simulation, and moving systems into operational environments. The announced “data, training, simulation, deployment” loop maps directly onto that sequence.
This is also why the cluster proposal is different from an ordinary data-center story. JD’s argument is that GPU infrastructure can be integrated with the company’s own commercial scenarios. The potential advantage is access to deployment settings and data feedback. The challenge is that those settings require reliability that model benchmarks alone do not prove.
Domestic GPU demand has risen as Chinese cloud providers and AI developers look to reduce exposure to restricted foreign accelerators. EastFrontier previously examined how Chinese technology groups shifted purchases toward domestic GPUs. The JD Cloud proposal would convert that purchasing shift into a much larger test of domestic hardware at cloud scale.
Execution Will Decide the Cluster’s Significance
The announcement gives both firms a strategic narrative, but it does not remove the engineering risks. Moving from 10,000 to 100,000 accelerators increases the complexity of networking, software compatibility, power delivery, cooling, operations, and fault recovery. Moore Threads and JD Cloud will need to demonstrate that their system can sustain useful performance for customers rather than only reach a headline hardware count.
Customer access is another open question. The companies said the cluster would be open to industry, but they did not identify pricing, service levels, launch customers, or the proportion of capacity reserved for JD. Those details will help determine whether the project becomes a shared infrastructure platform or mainly an internal technology showcase.
The policy setting is favorable. China’s communications-industry planning has encouraged deployment of large intelligent-computing clusters and adaptation to domestic chips. Yet policy support does not establish throughput, reliability, or unit economics. The key evidence will come later: construction milestones, available capacity, named customers, training results, and operating metrics.
For now, the 100,000-GPU plan is a statement of direction. It shows that JD Cloud and Moore Threads want Chinese AI infrastructure to be measured not only by individual chips or small clusters, but by the ability to support large-scale industrial training and physical-world applications. Whether that ambition becomes a running platform remains the next question.