CoreWeave expands full-stack AI cloud push as inference demand grows

CoreWeave is leaning hard into the part of AI infrastructure that happens after the model is trained. The company is presenting itself less as a place to rent GPU muscle and more as a full-stack cloud built for inference, agents and the messy work of putting AI into production.

The latest proof point is its claim that it completed the industry’s first bring-up and validation of Nvidia’s Vera Rubin NVL72 on CoreWeave Cloud. That matters because the company and its backers are treating Vera Rubin as more than a faster box. They frame it as a unified platform for agentic AI, with compute, networking, storage, software, data, security and operations tied together instead of sold as separate pieces.

That pitch lands at a moment when the market is moving away from experimentation and toward deployment. CoreWeave cites TheCUBE Research figures showing that 30% of organizations still have an operational readiness gap between AI experimentation and production, and nearly 88% of AI pilots never make it to production. At the same time, TheCUBE’s research says 86% of enterprises prioritize data unification over compute, which is a neat way of saying the problem is bigger than a pile of accelerators.

CoreWeave’s answer is to sell reliability, faster access to capacity and better token economics. TheCUBE analysts say the new system could cut cost per million tokens to one-tenth of Nvidia’s previous releases. That’s a very practical kind of promise, and one that fits the broader argument here: if AI is supposed to do real work, then the plumbing matters as much as the model.

The company also has the backlog to back up its confidence. As of June 30, CoreWeave said its revenue backlog was about $104 billion, and it has added new AI services this year, including tools for deploying self-improving AI agents and a Physical AI Field Engineering service. It is also building for heat and density, with Vera Rubin racks reaching up to 250 kilowatts per rack, which is why liquid cooling and purpose-built facilities are part of the story now. The old general-purpose cloud was never built for this kind of load.

And there’s a bigger strategic bet underneath all of it: that neoclouds can challenge hyperscalers by being better at the hard parts of AI operations, not just cheaper at renting silicon.