A Major Shift in AI Hardware
JD Cloud and GPU manufacturer Moore Threads announced on Wednesday a strategic partnership to construct a 100,000-unit GPU cluster. This project marks the first instance of domestically engineered graphics processing units reaching such a massive scale within a major Chinese cloud platform. The initiative serves as a bridge for homegrown hardware, moving from niche pilot projects toward widespread commercial implementation.
This infrastructure buildout focuses on the Universal GPU architecture developed by Moore Threads. The hardware is designed to handle demanding tasks including large-model training, inference, and embodied intelligence. Moore Threads stated that the capacity will be offered to external clients across various industrial sectors. This move reflects a shift in China's technology sector toward massive, integrated computing resources rather than smaller, fragmented deployments.
Driving Domestic Infrastructure Growth
Strategic priorities within China have increasingly centered on the growth of intelligent computing. The Ministry of Industry and Information Technology recently released plans covering the 15th Five-Year period. These guidelines explicitly call for the orderly deployment of clusters containing 10,000 to 100,000 or more accelerator cards. Policy is now directing companies to prioritize infrastructure that relies on locally produced chips rather than imported alternatives.
JD Cloud aims to funnel these new computing resources into specific domains like supply-chain AI, simulation, and high-quality data generation. The company intends to integrate this power into its own JoyAI model pipeline. By connecting the full technology stack from silicon to cloud platform, the partnership seeks to create a continuous loop for training and model deployment. The speed of training cycles remains the primary bottleneck for many AI developers today.
Scaling the Technology Stack
Moore Threads has built clusters ranging from thousands to tens of thousands of cards over the past few years. Its systems support the training of foundation models and world models currently in development. Founded in 2020, the Beijing-based company achieved significant status when it listed on the Shanghai STAR Market in December 2025. This public listing granted the company additional resources to accelerate its hardware roadmap.
Building a 100,000-card cluster requires more than just raw chip count. The engineering challenges involve networking, thermal management, and power distribution across thousands of connected units. Previous efforts have struggled to match the efficiency of global leaders, but this new project intends to prove domestic hardware can handle large-scale concurrent workloads. The broader significance lies in the transition of domestic chips from theoretical alternatives to central components in professional cloud environments. Future industry performance will depend on whether this cluster hits its operational targets.