CoreWeave just brought multi-rack Nvidia Vera Rubin NVL72 clusters online, stitching together hundreds of Rubin GPUs into a single scale-out system built specifically for agentic AI. That’s the headline number everyone will repeat. The part actually worth understanding is what CoreWeave paired it with: two storage upgrades built to stop those GPUs from sitting idle while data crawls in from somewhere else.
Agentic AI workloads are unusually sensitive to lag, because delays compound every time an agent calls a model, retrieves context or hands work to a tool. A slow storage fetch that barely registers in a normal application can quietly stack up across thousands of repeated calls until an entire training run drags. CoreWeave’s answer is cross-region write acceleration, which lets a job write data locally while the system replicates it to a second region in the background, so a team never sits around waiting on a manual copy job just to keep training moving.
Cécile Robert-Michon, who directs internal infrastructure at Cohere, described the practical difference this makes: their datasets span multiple regions, and a training schedule dictated by cross-region retrieval delays simply isn’t something they can tolerate. She said the change gives Cohere a unified dataset footprint with reads cached locally, so nothing sits waiting on the network, calling it the difference between planning around data and just training.
CoreWeave also added a new Archive storage tier, aimed at a habit most AI teams already recognize in themselves: quietly deleting data they’d rather keep, because storing it costs more than the risk of needing it later feels worth. Archive removes that trade-off with no fee to retrieve data, no fee to delete it early, and no fee for reading directly from that tier.
None of this works in isolation, though. Each Vera Rubin rack pairs 72 Rubin GPUs with 36 Vera CPUs and Nvidia’s latest networking hardware, and CoreWeave links multiple racks into one coordinated cluster through automated validation and a fabric built to expand without redesign at every step. The GPUs get the attention. The storage layer is what actually determines whether all that compute spends its time working or waiting.
