CoreWeave Pushes GPU Utilization as AI Storage Becomes Production Priority

CoreWeave Forge connects model deployment, evaluation and improvement, according to Corey Sanders, senior vice president of product at CoreWeave. The RL Rollouts capability, currently in preview, supports repeated cycles of generating training responses and updating models. Sanders said CoreWeave has worked on weight synchronization to bring weights in a 'hot start' instead of starting cold every time. Instead of always pulling weights from object storage, the system can get weights from nearby peers, he said. CoreWeave AI Object Storage now supports cross-region writes, letting post-training jobs write results back for others to use. Sanders described the storage approach as letting customers write like it is a local machine while CoreWeave treats it as a global system.

You.com Inc., which runs its own web index, has joined CoreWeave's partner network to provide a search layer for agents. 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. Saurabh Sharma, chief product officer of You.com, said the ceiling is no longer model intelligence; models are getting more intelligent, but their ability to use tools dictates an agent's success. Sharma added that customers are able to run workloads with higher accuracy and lower total cost of ownership, and that doing this in eight hours means it is no longer something only frontier labs can do.

At the Supermicro Open Storage Summit, theCUBE Research's Rob Strechay said organizations succeeding today are focusing less on models and more on operationalization. They are aligning data teams, infrastructure teams and business stakeholders around measurable outcomes and repeatable deployment strategies, he said. The interviews examined how storage architectures are evolving to meet production demands for enterprise AI.

Greg DiFraia, senior vice president of AI and alliance partnerships at Scality Inc., said keeping GPUs productive requires storage that serves active workloads quickly while accommodating growing volumes of less frequently accessed data. An effective AI storage strategy balances those demands across performance and capacity tiers. DiFraia said utilization must be driven as high as possible, and that for customers with tens or hundreds of petabytes or even exabytes, not all data will live in flash. Expanding agent contexts are also creating demand for storage that holds intermediate calculations used during inference. When key-value caches outgrow GPU memory, additional tiers must balance capacity with fast access. Anat Heilper, director of AI architecture at Vast Data Inc., said key-value cache optimization can replace compute with storage; high KV cache hit rates save compute and reduce latency significantly. Solidigm Inc. flash drives, Super Micro Computer Inc. integrated systems and Vast Data's platform support this evolving hierarchy, she said.

Production AI requires workload-specific systems and operational controls, according to the summit discussions. In financial services, the speed of data processing can affect risk assessment and capital availability, noted Moiz Kohari, vice president of enterprise AI and data intelligence at DataDirect Networks Inc., who spoke alongside Vince Chen, senior director of solutions architecture at Supermicro. Kohari said a company like BlackRock or State Street looking at a trillion dollars on its books with 6% to 12% capital lockup faces a big deal, and that moving data back and forth quickly can unlock 6% of that capital to be leveraged elsewhere. Chen said Supermicro works with partners to build vertically integrated and pre-validated T-shirt size options so enterprise customers can build an AI factory as purpose-built infrastructure. Ruhi Sehgal, agentic AI solutions marketing lead at Nutanix Inc., said as pilots become production services, highly constrained resources must be shared by a massive set of new users. Infrastructure admins managing multi-tenancy, ensuring security and optimizing performance are doing exactly what AI needs, she said.

The summit also identified access, orchestration and governance as the means by which enterprise data becomes useful.