JD Cloud plans 100,000-GPU computing cluster powered by Moore Threads

Introduction

What Happened

During the conference, JD Cloud’s senior executives detailed a multi‑year roadmap to assemble a cluster that will eventually house 100,000 Moore Threads GPUs. The GPUs are marketed as “general‑purpose” accelerators, meaning they can handle a broad spectrum of tasks—from deep‑learning matrix operations to traditional high‑performance‑computing (HPC) simulations. JD Cloud emphasized that the cluster will be made available on a pay‑as‑you‑go basis, allowing companies of any size to tap into world‑class AI horsepower without the capital expense of building their own data centers.

The announcement also highlighted two immediate use cases: (1) large‑model training and inference for generative AI, such as foundation models with billions of parameters, and (2) embodied AI, which integrates AI algorithms into physical systems like autonomous robots, drones and smart factories. By providing a unified platform for both virtual and physical AI workloads, JD Cloud aims to become a one‑stop shop for enterprises seeking to embed intelligence into products and services.

Key Details

Moore Threads’ GPUs are built on a 7‑nanometer process and feature a unified memory architecture that promises high bandwidth and low latency—critical for scaling deep‑learning workloads. Each GPU is expected to deliver roughly 30 teraflops of FP16 performance, and the cluster’s aggregate compute power will exceed 3 exaflops, placing it among the world’s most powerful AI supercomputers. JD Cloud plans to distribute the GPUs across several geographically dispersed data centers to ensure redundancy, low latency access for regional customers, and compliance with data‑sovereignty regulations.

The rollout will occur in three phases. Phase 1 (2026‑2027) focuses on hardware procurement, rack‑level integration, and the establishment of a high‑speed interconnect fabric based on InfiniBand and custom optical links. Phase 2 (2027‑2028) introduces a suite of AI‑optimized software services, including a managed training platform, model‑as‑a‑service APIs, and tools for embodied‑AI development such as real‑time perception pipelines. Phase 3 (2028 onward) expands the ecosystem by partnering with AI framework developers, offering pre‑installed libraries for PyTorch, TensorFlow, and emerging multimodal toolkits, and launching a marketplace where third‑party vendors can sell AI models and data sets directly to JD Cloud users.

Background

JD Cloud has steadily expanded its AI portfolio over the past five years, launching services such as JD‑AI Studio, a low‑code environment for model building, and JD‑Vision, a suite of computer‑vision APIs for retail and logistics. The company’s strategy has increasingly emphasized vertical integration—controlling everything from silicon to software—to reduce reliance on foreign technology and to meet China’s policy push for self‑sufficiency in critical AI infrastructure.

Moore Threads, founded in 2020, quickly rose to prominence by delivering GPUs that rival the performance‑per‑watt of established Western vendors while being manufactured domestically. The partnership with JD Cloud represents Moore Threads’ first large‑scale deployment in a public cloud environment, giving the chipmaker a high‑visibility platform to validate its designs and to attract further enterprise customers.

Why It Matters

The scale of the proposed cluster signals a decisive shift in the global AI compute landscape. While the United States and Europe have traditionally dominated the supercomputing arena, China’s investment in a 100,000‑GPU farm underscores its ambition to close the gap in AI research capacity. For Chinese enterprises, the availability of a domestically sourced, massive‑scale GPU pool reduces exposure to export controls and supply‑chain disruptions that have affected other cloud providers.

Beyond geopolitics, the cluster’s focus on embodied AI is noteworthy. As industries move from purely digital AI applications toward physical automation—smart factories, autonomous delivery robots, and AI‑driven logistics—the need for tightly coupled compute and edge‑deployment capabilities grows. JD Cloud’s integrated offering, which couples massive training capacity with low‑latency inference services, could accelerate the commercialization of such technologies, giving early adopters a competitive edge.

What Happens Next

In the coming months, JD Cloud will begin the physical installation of the first batch of Moore Threads GPUs, targeting an initial capacity of 10,000 units by the end of 2027. Parallel to hardware deployment, the company will launch a beta program for select enterprise partners to test the managed training platform and provide feedback on performance, pricing models, and integration workflows.

Looking ahead, JD Cloud plans to open the platform to the broader market in early 2028, accompanied by a robust developer ecosystem that includes tutorials, certification programs, and a marketplace for AI models. The company also intends to publish benchmark results that compare the Moore Threads‑powered cluster against leading global AI supercomputers, aiming to demonstrate competitive performance while highlighting the benefits of a fully domestic supply chain.

Conclusion