Alon Horev’s Post

Still thinking about yesterday's panel at Mistral AI Now Summit in with NVIDIA's Rod Evans, Mistral AI's Yann Leger, "No Shortcuts: AI Infrastructure Is the New Critical Infrastructure." The throughline: AI factories aren't bigger clusters, they're a different kind of system. Compute, data, memory and models have to be designed as one thing. That's the bet behind what we announced with Mistral and NVIDIA, one of Europe's largest GB300 NVL72 deployments, with Mistral Compute running on the VAST AI OS as part of its data foundation. Rod said it best on stage when asked about the biggest challenge in building AI factories: data. That landed, because it's exactly what we see every day. Two themes stuck with me on the flight back. How data itself has changed in this AI era. The old assumptions broke. Performance has to scale linearly with the cluster. Scale moved from petabytes to exabytes across sites. Multi-tenancy became a hard isolation problem, not a billing construct. The shape of data exploded: files, objects, tables, vectors, KV cache, agent memory, all in one namespace, no copies, no stitched pipelines. And uptime expectations shifted from "five nines for the database" to "the factory cannot stop." What's changing with inference and agents. Training was bursty. Inference and agents are always-on, latency-sensitive, stateful. Kubernetes becomes the substrate, so data services have to be first-class there, not bolted on. Security and governance move from perimeter to per-request, every agent call carries identity, policy and audit. And the unit of output is now useful tokens per watt, per dollar, per second. Sovereignty, in the end, is exercised at the data layer. Where data lives, how it moves, how it's governed in production. That's what makes this build matter. Thanks to the Mistral team for the room, and to Rod and Yann for a sharp conversation.