100,000 Homegrown GPUs: JD Cloud and Moore Threads to Build China's Largest Domestic-Chip AI Cluster

For years, the phrase “Chinese AI compute” carried an asterisk: the clusters existed, but the GPUs inside them were overwhelmingly American. That asterisk just got a little smaller. On September 9, 2026, at JD’s annual Global Technology Explorers Conference in Beijing, JD Cloud announced plans to build a 100,000-GPU intelligent-computing cluster whose compute base is Moore Threads’ domestically designed full-function GPUs — described by the partners as the first time Chinese silicon has entered a 100K-scale core cluster at a major domestic AI cloud provider.

The scale matters. A 10,000-GPU cluster is a serious training facility; 100,000 GPUs is hyperscale territory, the size of buildout normally associated with Nvidia-powered flagships from ByteDance, Alibaba, or the US hyperscalers. Doing it with Moore Threads — a company founded only in June 2020 by former Nvidia China head Zhang Jianzhong — moves domestic accelerators from “proven in niche deployments” to “load-bearing infrastructure for a top-tier AI cloud.”

What was actually announced

The cluster will be built on Moore Threads’ full-function GPU lineup (the company trades on the Shanghai STAR Market as 688795.SH, and its data-center accelerator line includes the MTT S4000 with 48GB of GDDR6 memory and MTLink interconnect). Three workload categories are explicitly targeted: large-model training, inference, and embodied intelligence — the “physical AI” bet that JD is pushing harder than almost anyone.

Critically, the computing capacity won’t be a private playground. JD Cloud says the cluster’s power will be opened to companies across industries, positioning itself as the domestic-supply answer for Chinese enterprises that want hyperscale AI compute without depending on restricted US export channels.

The announcement extends an earlier milestone: the two companies previously co-deployed a 10,000-GPU cluster, so this represents a 10x step-up in scale — and a statement that the earlier deployment delivered well enough to bet bigger.

The policy tailwind behind it

The timing is not accidental. China’s Ministry of Industry and Information Technology (MIIT) recently issued the “15th Five-Year Plan for Information and Communications Industry Development,” which explicitly calls for the orderly deployment of 10,000-card and 100,000-card-class intelligent-computing clusters, with increased effort to adapt domestic compute chips. In other words: the regulator is not merely tolerating domestic-GPU mega-clusters, it is actively scheduling them into national planning.

This aligns with the broader picture reported elsewhere this week — China’s compute ambitions now stretch to 9,800 EFLOPS by 2030 under a ¥3.8 trillion plan. JD Cloud’s 100K cluster is exactly the kind of concrete, corporate-side building block that such national targets are assembled from. And it lands amid a intensifying tech-security backdrop: just yesterday, US agencies formally accused six Chinese AI firms of large-scale model distillation against US frontier labs, and Beijing formally rejected those claims today. Every layer of the AI stack — models, data, and now demonstrably the silicon — is being pulled into separate, self-reliant ecosystems.

Why Moore Threads got the order

Yicai’s reporting frames the selection bluntly: Moore Threads is one of the few domestic vendors with genuinely training-capable GPUs, having commercially deployed thousand-card and ten-thousand-card clusters within a single network, and having completed what the report calls “hard breakthroughs” in core training scenarios including foundation models, embodied brains, and world models.

JD’s side of the logic comes from Cao Peng, Chairman of JD’s Technology Committee and President of JD Cloud: “We are fully advancing the physical AI strategy — high-performance, highly reliable domestic compute is the core support. Choosing deep cooperation with Moore Threads is precisely because of its scarce capability in domestic training-grade GPUs and its mature delivery experience in large-scale clusters. This cooperation is an important step in JD Cloud’s AI strategy, and an important starting point for embracing the domestic intelligent-computing ecosystem.”

Zhang Jianzhong, Moore Threads’ founder, chairman and CEO, called the order “the strongest validation of our technical strength and engineering delivery capability, and a key milestone for domestic intelligent computing moving from usable to commercialized at scale,” adding that the two sides will “jointly build a world-class AI factory, letting domestic compute power not only support intelligence in the digital world, but enter the physical world and drive industrial transformation across thousands of industries.”

The physical-AI angle

What makes this more than a chip-procurement story is JD’s embodiment agenda. The company has proposed building the “world’s largest physical-world operations center” and plans to construct the world’s largest embodied-intelligence data center within two years. Training embodied agents requires a different compute profile than chatbots — simulation, synthetic data generation, world-model training — and enormous amounts of it.

The partnership explicitly targets that loop: full-stack coordination from chips and cloud platform through model training, supporting the evolution of JD’s JoyAI model family and building a complete “data → training → simulation → deployment” closed loop. JD Cloud’s pitch is that its real-business depth across retail, logistics, healthcare, and industry gives AI thousands of genuine scenarios to be hardened in — and now, a domestic compute base to harden them on.

The honest caveats

Scale announcements are easier than scale deliveries. Moore Threads’ accelerators remain well behind Nvidia’s frontier parts on raw performance and software-maturity — CUDA’s ecosystem gravity doesn’t vanish because a cluster is large. The stock’s own volatility (shares hit new lows recently amid share-unlock pressure, even as the company narrowed losses and launched a Hong Kong listing process eight months after its A-share debut) is a reminder that domestic GPU economics are still being proven. And “拟建设” — “plans to build” — is the operative verb in the Chinese announcements: this is a committed roadmap, not a completed facility.

But the direction is unambiguous. Domestic Chinese GPUs have now been entrusted with a 100,000-card core cluster at a top AI cloud — the first entry of homegrown silicon into the big leagues of AI infrastructure. For everyone tracking whether China can actually substitute restricted US accelerators at frontier-training scale, this is one of the clearest data points yet. The era of domestic compute as science project is ending; the era of domestic compute as infrastructure has a launch date, and it is September 9, 2026.