Anyscale has scaled Ray Core to handle AI training and inference on clusters of 10,000 nodes. The company says the key improvements were faster actor launches and fewer bottlenecks that used to slow distributed jobs.
The coordination problem
AI workloads aren't just about raw compute—they're about getting thousands of machines to work together without tripping over each other. At 10,000 nodes, the coordination overhead can become the real bottleneck. Anyscale says Ray Core's latest changes tackle that directly.
Ray Core is the distributed compute engine at the center of Anyscale's platform. It breaks work into tasks and actors, each of which needs to be launched, tracked, and kept in sync. When a job spans five digits of nodes, those little pieces matter.
Faster actor launches
Actors are the units that hold state and perform work in Ray. Launching them quickly matters because the whole job stalls if every node has to wait on a slow actor setup. Anyscale says actor launches are now faster, which means less idling and more time spent actually computing.
The company also points to fewer bottlenecks. In a cluster that big, a single slow component can hold up an entire training run. By trimming the choke points, Anyscale says the system keeps data moving more steadily across the network.
What that means for AI teams
Training large models and running inference on massive clusters isn't new, but scaling to 10,000 nodes has been a hurdle. The milestone suggests that Ray Core can handle the kind of scale that big AI labs and enterprises are starting to want.
The improvement isn't about a single speedup. It's about the combination: quicker actors, fewer stalls, and a system that doesn't fall apart when you add more machines. Anyscale hasn't said when the 10,000-node capability will be available to everyone, but the work points to where Ray is headed.
