IR note says Cognition ran the first customer Vera Rubin inference benchmarks on CoreWeave, under the same ops model as GB200/GB300 NVL72 fleets — watch for measured throughput claims vs GB200 baseline.
CoreWeave’s 30 Sep 2026 IR release says NVIDIA Vera Rubin NVL72 is available on its cloud with Cognition as the first production customer, running under the same tooling as GB200/GB300 NVL72 fleets. Cognition’s engineers ran the first customer-executed Vera Rubin inference benchmark against a GB200 NVL72 baseline and reported up to 4.8× total-token throughput on SWE-2 workloads, plus 3.8× output-token throughput on reinforcement-learning jobs — vendor/customer-reported until third-party replication lands. Ops buyers: same Kubernetes/Mission Control model as prior NVL72 fleets; treat the multiplier as a workload-specific claim, not a generic rack upgrade.
Source: CoreWeave IR — NVIDIA Vera Rubin NVL72 at production scale (starting with Cognition) (2026-09-30)