On 30 September CoreWeave put NVIDIA Vera Rubin NVL72 into limited availability, with Cognition "the first customer anywhere running production workloads". Its engineers measured two multiples, per GPU at matched interactivity against GB200 NVL72: 4.8x total token throughput for SWE-2 inference, 3.8x output token throughput for reinforcement learning. A post fifteen minutes later took 3.8x as more Devin sessions per GPU. Cognition's Silas Alberti quoted 4.8x for inference; 3.8 ÷ 4.8 leaves that claim 21% low.
Both are real and CoreWeave's; only the workload differs. Rack sizes match at 72 GPUs a side, so per GPU and per rack are the same multiple.
Whether the gain reaches the buyer is not computable. CoreWeave's price card has no Vera Rubin row: GB200 NVL72 is $42.00 an hour for four GPUs, so $42.00 ÷ 4 = $10.50 a GPU-hour. At $10.50 × 4.8 = $50.40 a Rubin GPU-hour the inference gain cancels per dollar, and only a Rubin price says which side it lands on. "Hundreds of NVIDIA Rubin GPUs deployed across multiple regions" is three to twelve racks, beside Cognition's fleet in the thousands: the first-cloud window we dated to 2H 2026 opened on schedule, still small.
Cognition, first Vera Rubin NVL72 customer — 30 September 2026
Both multiples, the rack size and the availability wording are CoreWeave's, 30 September 2026; Cognition ran the benchmarks on its own SWE-2 configuration. The rates and the missing Rubin row are its price card, read that day. The restated 3.8x is a third-party post. Both divisions are ours.
