JD Cloud 100,000-GPU Cluster

JD Cloud, the cloud-computing division of Chinese e-commerce group JD.com, announced plans at JD's Global Tech Explorers Conference on September 9, 2026 to build a 100,000-GPU cluster using domestically designed Moore Threads GPUs (trade press including tech news outlet coverage reports the part as the MTT S5000; a smaller predecessor JD Cloud/Moore Threads cluster of roughly 10,000 GPUs, reported as the earlier MTT S4000 part, was already operational per the same announcement). The cluster uses Moore Threads' proprietary MTLink GPU-to-GPU interconnect fabric and is intended to support large-model training and inference, embodied-AI/robotics workloads, and JD's own JoyAI foundation-model family, alongside a commercial GPU-hour cloud service JD Cloud sells to outside enterprises. JD Cloud and Moore Threads describe it as the first 100,000-GPU-class cluster built on domestically designed GPUs at a major Chinese cloud provider. Moore Threads has separately claimed 95 percent linear scaling efficiency and roughly 10 ExaFLOPS of aggregate performance for the cluster, but coverage of the announcement notes no independent organization has verified either figure against a controlled workload, so no Rmax/Rpeak figure is recorded here and the tier stays reported rather than verified. As of the September 2026 announcement the cluster had not yet been built, so status is recorded as planned with no first-operational date; the exact site within JD Cloud's China data center footprint (which spans Beijing, Guangzhou, Shanghai and Suqian) is undisclosed.