JD Cloud said on 9 September 2026 that it will build a computing cluster of 100,000 Chinese-made graphics processors. The China AI chip cluster would run on general-purpose GPUs from Moore Threads and would be the first at that scale at a major Chinese cloud provider, according to the company. JD Cloud gave no completion date.
What JD Cloud Announced
The announcement was made at JD’s 2026 Global Technology Explorers Conference, the annual technology event of the Chinese retail group JD.com.
JD Cloud, the group’s cloud computing arm, said the cluster will support large-model training and inference and what it calls embodied AI, meaning models that control robots and other physical machines. The company said the capacity will be sold to businesses across industries, and described the aim as building large-scale computing infrastructure on Chinese hardware.
The chips come from Moore Threads, a Beijing GPU designer founded by former Nvidia China staff. The two companies have worked together before on a cluster of 10,000 GPUs, so the new plan is a tenfold step up rather than a first partnership.
One word carries most of the weight in the announcement: plans. Nothing has been reported as built, and JD Cloud has not said when any of it will run.
What Beijing’s Plan Asks For
The announcement came two days after China’s government published a five-year plan that asks for exactly this kind of cluster.
The Ministry of Industry and Information Technology published its Information and Communications Industry Development 15th Five-Year Plan on 7 September 2026, covering 2026 to 2030. The plan calls for the orderly deployment of intelligent computing clusters at the 10,000-card scale and at scales of 100,000 accelerator cards or more, and for greater efforts to adapt infrastructure to domestically produced computing chips.
| Measure | Figure |
|---|---|
| AI computing capacity, end of June 2026 | 2,185 exaflops |
| Target for 2030 | 9,800 exaflops |
| Implied increase | About four and a half times |
| Information infrastructure investment, 2026 to 2030 | 3.8 trillion yuan, about $532 billion |
Those figures come from the South China Morning Post’s reporting on the plan, and the baseline is measured at FP16 precision. The plan document itself is dated 12 August 2026 and was released the following month.
Read together, the sequence is straightforward. Beijing set a target for 100,000-card clusters on home-made silicon on Monday, and a large cloud provider announced one on Wednesday.
What JD Cloud Has Not Disclosed
Almost every detail that would let an outsider judge the cluster is missing.
- GPU model: JD Cloud named Moore Threads’ general-purpose GPU line but not which chip.
- Completion date: no timeline, no phasing, no date for first operation.
- Performance: no benchmark, and no figure for the cluster’s total compute.
- Location: no site named.
- Investment: no cost or funding figure.
As of 10 September 2026, no part of this cluster has been reported as operating, and neither JD Cloud nor Moore Threads has published a benchmark for it. The claim of a first is also narrower than it looks: it is specific to a leading Chinese AI cloud provider. Chinese systems reaching 100,000 domestic accelerator cards have been announced before in other settings, including at the World Artificial Intelligence Conference in Shanghai in July 2026.
How Strong Are Moore Threads GPUs?
Moore Threads has made strong performance claims for its AI chips, and none of them have been independently benchmarked.
The company unveiled an AI training and inference chip line called Huashan, with mass production slated for 2026. Its founder, Zhang Jianzhong, has said Huashan outperforms Nvidia’s Hopper generation, including the H100 and H200, on compute, memory bandwidth and memory capacity, and that floating-point performance approaches Nvidia’s newer Blackwell B200.
Two cautions apply. Those are company statements, not measured results, and no published benchmark data supports them. And analysts have generally placed Moore Threads behind Huawei and Cambricon among Chinese AI chip suppliers, which makes its selection for a flagship cluster notable in itself.
JD Cloud has not confirmed that the 100,000-GPU cluster uses Huashan. Reports name only the general-purpose GPU line.
Why This Matters Beyond China
The cluster is a test of whether export controls have pushed China towards a working domestic alternative or merely an announced one.
United States restrictions have progressively cut Chinese buyers off from Nvidia’s most capable accelerators, and Washington has moved to close routes through third countries. The response inside China has been to fund domestic designers and to instruct operators to build on their chips, which is what the five-year plan says in plain terms.
Whether that substitution works is an engineering question, not a policy one. Running 100,000 accelerators as a single training system is difficult regardless of who makes the chips: interconnect bandwidth, failure rates and software maturity decide whether a cluster trains a large model or merely exists. Nvidia’s advantage has rested as much on its software stack as on its silicon. Until JD Cloud publishes a benchmark or a customer reports training a model on the system, the announcement establishes intent rather than capability.
The demand side is not in doubt. Compute scarcity is shaping company strategy worldwide, including in Europe, where Mistral AI raised €3 billion partly to buy compute at Chinese scale.
Frequently Asked Questions
When Will the China AI Chip Cluster Be Finished?
JD Cloud has not said. No completion date, phasing schedule or first-operation date was given in the 9 September 2026 announcement.
Which Chips Will the Cluster Use?
Moore Threads general-purpose GPUs. JD Cloud has not named a specific model, and has not confirmed that the Huashan AI chip line is involved.
Is This the Largest Domestic GPU Cluster in China?
JD Cloud claims it is the first 100,000-GPU core cluster using Chinese GPUs at a major domestic AI cloud provider. Other Chinese systems at 100,000 domestic accelerator cards have been announced in different settings, so the claim is narrower than a general record.
Are Moore Threads GPUs as Fast as Nvidia’s?
Moore Threads says its Huashan line beats Nvidia’s H100 and H200 on compute and memory. Those are company claims with no published benchmark data behind them, so the real gap is unverified.
What Is China’s Overall AI Computing Target?
9,800 exaflops of intelligent computing capacity by 2030, against 2,185 exaflops at the end of June 2026, backed by 3.8 trillion yuan of information infrastructure investment across 2026 to 2030.
