CoreWeave Forge targets continuous AI post-training

CoreWeave's announcement of CoreWeave Forge and its preview RL Rollouts capability points to where the bottleneck in AI post-training has moved. As enterprises continually refine AI agents, the limitation is less about how smart a model is and more about how efficiently infrastructure moves data and reloads updated models between training rounds. Corey Sanders described work on weight synchronization that brings weights in a hot start rather than starting cold, pulling them from nearby peers instead of object storage. CoreWeave AI Object Storage now supports cross-region writes, allowing post-training jobs to write results back for others to use.

You.com, which runs its own web index, joined CoreWeave's partner network to give agents a search layer. Saurabh Sharma framed the shift plainly: the ceiling is no longer model intelligence, and an agent's ability to use tools dictates its success. Working with Nvidia, the companies used RL Rollouts to post-train Nemotron 3.5 Lightning in eight hours with You.com's web search tools. Sharma said customers are seeing higher accuracy and lower total cost of ownership, and that an eight-hour post-training run is no longer something only frontier labs can do.

The stakes hinge on whether the tool-use layer, not raw model intelligence, becomes the deciding factor in agent performance, and on whether the eight-hour result holds across customers' production workloads rather than a single demonstration. For teams buying AI infrastructure, the pitch is faster iteration at lower cost; whether that translates into durable advantage likely depends on how broadly these capabilities move out of preview.