Vendor keynotes announce silicon – CoreWeave’s September 16 announcement proved something harder: that you can actually run it
The company said it had brought up a multi-rack NVIDIA Vera Rubin NVL72 cluster on its cloud, connecting hundreds of Rubin GPUs into a single scale-out system, making it, by its own account, the first AI cloud provider to validate the architecture at production scale, months after Nvidia declared it in full production at GTC.
“With multi-rack Vera Rubin, we are connecting hundreds of Rubin GPUs as a single scale-out cluster,” said Chen Goldberg, CoreWeave’s executive vice president of product and engineering, framing the achievement around what it unlocks for customers building agentic AI systems: greater scale and faster iteration as models and agents continuously learn.
The engineering matters more than the marketing
A single Vera Rubin NVL72 rack already ties together 72 Rubin GPUs, 36 Vera CPUs, NVLink 6 fabric and BlueField-4 DPUs under liquid cooling. Connecting multiple racks over Spectrum-X Ethernet turns that into a distributed-systems problem, where — as CoreWeave’s own engineering team put it — a single underperforming GPU can become a cluster-wide straggler. CoreWeave paired the compute milestone with new storage capabilities, including cross-region write acceleration that Cohere’s director of internal infrastructure, Cécile Robert-Michon, said solved a specific pain point: training schedules that had been “dictated by cross-region retrieval delays.” CoreWeave shares gained as much as 4% on the announcement, clawing back a chunk of a rough week for the stock.
Bottom line: The gap between a chip being “in full production” and a customer being able to rent it at scale has historically run many months. CoreWeave just compressed that gap to weeks for Nvidia’s flagship post-Blackwell platform — and set the bar every competing neocloud now has to clear.
