JD Cloud plans China's first 100,000-GPU cluster built on domestic chips. Moore Threads says it'll hit 95% scaling efficiency -- a number nobody outside the company has checked

JD Cloud, the cloud-computing arm of JD.com, announced plans at its 2026 Global Technology Explorers Conference on September 9 to build a computing cluster of 100,000 GPUs using chips from Moore Threads, the Shanghai-listed Chinese GPU maker. JD Cloud is calling it the first deployment of domestically developed Chinese GPUs at 100,000-GPU core-cluster scale by a major domestic AI cloud provider -- a real milestone in China's push to build large-scale compute infrastructure that doesn't depend on Nvidia hardware, which US export controls have made increasingly hard for Chinese buyers to obtain at the newest tiers. The cluster is meant to support large-model training and inference plus embodied-AI workloads, and JD Cloud says it plans to eventually rent the capacity to outside companies, the same way it already resells Nvidia-based cloud compute today.

The two companies aren't starting from zero -- JD Cloud and Moore Threads have already run a 10,000-GPU cluster together, so this is a declared tenfold scale-up of a partnership already tested at a smaller size, not an untried pairing. What's new, and what carries almost all of the real technical risk, is the jump in scale itself: interconnect and cooling problems that don't show up in a 10,000-GPU cluster routinely appear once a system crosses into six figures, which is exactly the regime where Moore Threads' own performance claims get hardest to take at face value.

The claim: 95% scaling, 60% utilization. The question: says who?

Moore Threads is claiming 95% linear scaling efficiency and 60% model-FLOPs utilization (MFU) for dense models at 100,000-GPU scale, connected via its own MTLink 4.0 interconnect at up to 1,314 GB/s. Neither figure has been verified by anyone outside the company, and the cross-node fabric architecture that would actually connect all 100,000 GPUs hasn't been disclosed -- which matters, because at this scale, the interconnect and software stack's ability to minimize collective-communication overhead is precisely what determines whether a claim like this is achievable or aspirational.

That comparison isn't a gotcha so much as a scale problem every large training cluster runs into. When OpenAI trained GPT-4 on roughly 25,000 A100 GPUs, MFU fell to just 32-36% because inter-node communication overhead came to dominate at that scale, compounded by hardware failures forcing checkpoint restarts -- and that was on Nvidia's most mature training architecture, four years and multiple generations into large-cluster tuning. JD Cloud's planned cluster is four times larger than that GPT-4 run, running on newer, less-proven hardware, with a communications fabric nobody outside Moore Threads has seen described. A 95%/60% claim at that combination of scale and hardware immaturity isn't impossible -- but it would represent a genuinely exceptional engineering result, not an incremental one, and the company making the claim is also the company selling the chips.

The company's underlying business, separate from the unverified cluster claims, is real and growing fast. Moore Threads reported RMB 1.505B in 2025 full-year revenue, up 243% year-on-year, with R&D spending of RMB 1.305B -- nearly 87% of revenue -- funding a business that only turned its first profitable quarter in Q1 2026, posting RMB 29M in net profit on RMB 738M of quarterly revenue, up 155% year-on-year. That growth is what's financing the AI-chip ambitions behind this cluster; it says nothing on its own about whether the cluster will hit the efficiency numbers Moore Threads is claiming for it.

The deal fits a pattern this newsroom has tracked across Huawei, CXMT and now Moore Threads: Chinese cloud and chip companies pairing up to prove out domestic hardware at production scale, driven by US export controls that have made the newest-tier Nvidia chips harder for Chinese buyers to obtain. Moore Threads didn't start as an AI-first chipmaker -- it built its name on gaming GPUs, and its next-generation gaming architecture claims a 15x performance jump and 50x faster ray tracing over its predecessor. Its in-development AI GPU is separately said to land somewhere between Nvidia's Hopper and Blackwell generations in raw capability -- another company-stated figure, on a chip that hasn't shipped yet. That same pattern -- confident, unverified performance claims paired with genuinely fast revenue growth -- is exactly why this specific cluster matters as a test case: it's the first chance for an outside party to check Moore Threads' numbers against a deployment big enough that the gap between claim and reality would be hard to hide.

None of this makes the 100,000-GPU plan fake or the underlying push toward domestic Chinese compute insincere -- Moore Threads' revenue growth, its STAR Market listing, and its existing 10,000-GPU deployment with JD Cloud are all real and independently reported. What isn't established yet is the specific number every reader will remember from this story: that a cluster four times larger than OpenAI's GPT-4 training run, on newer and less-proven silicon, will run nearly twice as efficiently. That claim belongs to Moore Threads alone until someone who isn't Moore Threads measures it.