CoreWeave says the next wave of AI infrastructure will need more than faster chips. It will require data centers designed around the demands of rack-level AI systems, especially as workloads grow larger and more complex.
In an exclusive interview, the cloud provider described two internal upgrades built specifically for Nvidia's Vera Rubin NVL72 platform, a 72-GPU system aimed at AI supercomputing. The improvements focus on cooling and control, two areas CoreWeave says are becoming critical as AI customers push hardware harder.
CoreWeave's senior vice president of product, Corey Sanders, said the company is shifting its thinking from individual machines to entire racks. That, he said, reflects a broader change in how AI infrastructure has to be built and operated.
One of CoreWeave's tools is a programmable valve system nicknamed Valvey. The system manages how coolant moves through server racks and uses software to track pressure, flow rate, and leaks. CoreWeave says the approach is designed to limit the impact of hardware problems and reduce downtime by preventing a single failure from spreading across a broader section of infrastructure.
The second upgrade, called Racky, is a controller mounted on top of each rack. It brings together power management, cooling, and environmental sensor data in one interface. CoreWeave says the goal is to give customers a more unified way to monitor and manage their racks, which could make it easier to expand capacity over time.
Jacob Yundt, senior director of compute architecture at CoreWeave, said these kinds of systems matter because AI infrastructure is becoming more integrated and more demanding. While the features may sound technical, CoreWeave argues they can translate into real operational benefits by reducing service interruptions and simplifying day-to-day management.
Sanders said the partnership with Nvidia changes what CoreWeave's customers can do with the hardware. He said more capable infrastructure gives them room to experiment, take risks, and build new applications. That is especially important, he said, as AI agents increase the amount of work these systems must handle.
CoreWeave said the rack has become a key unit of design because of the kinds of customers it serves. Sanders said the company supports nine of the 10 leading model providers, making rack performance and reliability directly relevant to large-scale AI workloads.
Looking further ahead, Sanders suggested that AI jobs may not stay confined to one data center or one cloud provider. He said future workloads could stretch across multiple facilities and potentially across several clouds, adding another layer of complexity to the infrastructure behind AI systems.
The interview also underscored Nvidia's influence across the AI stack. The chipmaker is already central to the GPU systems powering much of the industry, and CoreWeave's latest work shows how its technology is also shaping the design of the surrounding infrastructure.
As AI continues to stress everything from networking to power and cooling, companies building the hardware are trying to keep pace. CoreWeave's latest changes suggest that for AI factories, the competitive edge may depend as much on how racks are managed as on the chips they hold.