On September 9, JD.com held the JDDiscovery-2026 Global Technology Explorer Conference in Beijing under the theme “JoyAI: Leap into the Physical World.” At the event, JD announced the latest progress across three AI infrastructure lines—compute, data, and models: its cloud business has partnered with Moore Threads and others to build a domestic 10,000-GPU cluster and plans a 100,000-GPU cluster; it is simultaneously advancing 10 million hours of real-world human-scenario data collection, with the first open-source dataset, EgoLive, now publicly available; and on the model side, it released JoyAI-Echo WM, a real-time interactive world model that scored 81.6 on the public WBench Navigation benchmark.
Taken together, these moves point to one judgment: for physical AI to leave the lab, a smarter model alone is not enough. Robots must understand environments, make decisions, and execute reliably over long periods. That requires continuously supplied compute, data drawn from real sources, and an engineering system that connects the two.
JD.com’s answer is a “super AI supply chain,” using six modules—cloud, data, model, device, scenario, and chain—to connect compute, data, models, terminals, and scenarios into a closed loop: scenarios generate data, data trains models, models drive terminals, and task feedback drives further iteration. This kind of flywheel is not new; the hard part is making every link run inside real business operations rather than stopping at demonstrations.
The real bottleneck often appears at the “putting it to use” stage. Building a 100,000-GPU cluster is only the starting point. How to schedule training, inference, and various business tasks within the same resource system, and how to make chips from different vendors and architectures work together, determine how much output compute can actually generate. JD.com’s choice to partner with domestic GPU vendors to build clusters also shows that self-controllability is not just a slogan; it must be implemented in runnable engineering details.
This is precisely the direction StarWar Technology continues to invest in on GPU compute platforms and scheduling: organizing scattered, heterogeneous computing resources into a unified resource pool, so that model training, world-model inference, and upper-layer business tasks can be efficiently orchestrated by priority and real-time load. The flywheel JD.com feeds with supply chain scenarios and what a compute platform aims to solve are essentially the same thing—making compute