CoreWeave deploys Vera Rubin NVL72 in production, Cognition sees 4.8x token throughput, and Mission Control 3.0 makes power a first-class scheduling primitive.

VERA RUBIN IN PRODUCTION CoreWeave has deployed Vera Rubin NVL72 in limited availability. Cognition is the first customer, reporting a 4.8x increase in total token throughput. This is the first real-world validation of NVIDIA’s claim that Vera Rubin delivers 35x the token throughput of Hopper at the same power. The platform is a unified, liquid-cooled system with 72 Rubin GPUs connected via NVLink 6, and 36 Vera CPUs. It handles trillion-parameter models at a 45x lower token cost than Blackwell. AI FACTORY CONTROL PLANE NVIDIA Mission Control 3.0 introduces a layered, API-driven architecture. Multi-org isolation is now possible via VXLAN on Spectrum-X Ethernet and PKeys on Quantum InfiniBand. Domain power service makes power a first-class scheduling primitive, supporting MAX-P and MAX-Q profiles. In one deployment, MAX-Q allowed a data center to run at 85% power with only 7% throughput loss. Mission Control also adds AI-powered predictive anomaly detection, correlating telemetry across GPUs, networking, and schedulers. LOCAL AI HARDWARE NVIDIA DGX Spark 64GB ships October 23 at $4,999. It runs 100B-parameter models on device, clusters to 128GB for 200B-parameter support. Two units connected via QSFP cable pool memory and deliver up to 1.7x the performance of a single system. NVIDIA PAIR, a Personal AI Router, distributes inference across local PCs, routing requests to available capacity. The network must decide: do we standardize on Vera Rubin for high-throughput workloads, or wait for broader availability and pricing clarity?

researched · 4 sources

5 Octreaches everyone