CoreWeave targets GPU utilization in continuous AI post-training

CoreWeave is pushing harder on a problem that rarely gets the spotlight: not how smart an AI model is, but how fast it can be updated and put back to work. At the Fully Connected event, company executive Corey Sanders said the company is building a full-stack AI cloud around the agent lifecycle, where repeated post-training cycles matter as much as the initial model run.

That focus shows up in CoreWeave Forge, which ties together deployment, evaluation and improvement. Its Rollouts feature is in preview now, and it is designed for repeated loops of generating training responses and updating models. Sanders said CoreWeave has also worked on weight synchronization so new runs can start from nearby peers instead of pulling everything cold from object storage every time.

The same thinking extends to data movement. CoreWeave AI Object Storage now supports cross-region writes, so post-training jobs can write results back for others to reuse. Sanders described the goal as getting data into the GPU as fast and as easily as possible. In his words, the company wants storage that behaves like a local machine even though it is really a global system.

You.com fits into that plan as the search layer for agents. The company, which runs its own web index, has joined CoreWeave’s partner network, and the two companies worked with Nvidia to post-train Nemotron 3.5 Lightning in eight hours using You.com’s web search tools. Saurabh Sharma, You.com’s chief product officer, said model intelligence is no longer the ceiling; what matters now is whether models can use tools well enough to succeed. He also said customers are seeing higher accuracy and lower total cost of ownership, and that doing the work in eight hours takes this sort of job beyond the frontier labs.