Featured image of post JD Cloud and Mole and join forces to build a 100,000-GPU AI factory, aiming to create the world's largest physical world operations center
AI News

JD Cloud and Mole and join forces to build a 100,000-GPU AI factory, aiming to create the world's largest physical world operations center

JD Cloud and Mole join forces to build a 100,000-GPU domestic AI compute cluster with parallel progress in data and model infrastructure.

Core Event: 100,000-GPU AI Factory Announced

On September 9, 2026, JD Group unveiled its latest progress in AI infrastructure at the 2026 JD Global Tech Explorers Conference. JD Cloud, in collaboration with Mole and, has planned the construction of a 100,000-GPU AI factory, while a domestic 10,000-GPU cluster is already operational.

Key hard facts:

  • Announcement date: September 9, 2026
  • Partners: JD Cloud × Mole and
  • Currently built: Domestic 10,000-GPU cluster
  • Under planning: 100,000-GPU cluster
  • Target applications: Large model training, token generation, agent training, embodied intelligence services

Three-pillar AI Infrastructure: Compute, Data, Models

JD Cloud’s announcement covers three interdependent pillars of AI infrastructure—not a single breakthrough.

Compute: JD Cloud and Mole and have completed their domestic 10,000-GPU cluster and elevated the 100,000-GPU project as next priority. Mole and CEO Zhang Jianzhong emphasized the factory’s purpose beyond raw compute: delivering industry-grade agent training and physical-world operations.

Data: JD Cloud’s large-scale human data collection initiative—10 million hours—is reportedly progresses smoothly. The company positions this as the “largest-scale human data acquisition in the industry,” a scale rarely matched globally.

Models: JD has launched the JoyAI foundational model matrix, comprising:

  • Multimodal models (handling image, text, audio indiscriminately)
  • World models (simulating physical-world dynamics)
  • Embodied models (capable of environmental interaction and action)

Counterintuitive Insight: Scale of Domestic GPU Cluster Ambition

A surprising detail: 100,000 GPUs places this cluster among the top tier of global AI infrastructure. As of 2026, only a handful of publicly disclosed AI clusters exceed 10,000 GPUs; 100,000-GPU projects remain exceptionally rare.

Even more significant, JD did not disclose use of NVIDIA or other foreign GPUs—instead highlighting “domestic 10,000-GPU cluster” as completed. Given Mole and’s identity as a domestic GPU vendor, this suggests accelerated building of a self-contained, domestically controlled large-model training infrastructure.

JoyAI’s architecture further reflects differentiation: unlike pure language-model vendors, JD explicitly prioritizes “embodied models” and “world models,” aligning with its core e-commerce/logistics strengths—optimizing warehouse, delivery, and supply-chain operations demands models that perceive and act in physical space.

Practical Guidance: Who Should Act Now, Who Should Wait

  • Act now if you:

    • Operate logistics, supply-chain, or smart-manufacturing businesses, and seek embodied-model integration with JD’s physical-world运营center;
    • Require国产ized AI alternatives and want to assess JD’s service timeline and commercial terms.

  • Wait and watch if you:

    • Have non-strategic demand for 100,000-GPU capacity and want clearer pricing and availability;
    • Rely solely on multimodal models for generative content creation, and hope for API early access.

Final Thoughts

JD’s pivot from model-centric AI to “physical-world operations” viaHeyAI signals the third evolution stage of large models—where value shifts from model quality to model-environment interactivity. This physical-world embedded path may become China’s key differentiator in global AI competition.