Nvidia and CoreWeave Validate Vera Rubin NVL72 Rack Platform

CoreWeave announced on June 1, 2026 that it is the first AI cloud provider to bring up and complete system-level validation of NVIDIA's Vera Rubin NVL72 rack-scale platform on its cloud (coreweave.com press release). Each NVL72 rack combines 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs, connected by a 260 TB/s NVLink 6th-generation fabric, and CoreWeave says the platform delivers up to 10x better inference per watt, roughly a quarter fewer GPUs, and one-tenth the cost per million tokens compared with NVIDIA Blackwell. Chen Goldberg, CoreWeave's EVP of Product and Engineering, said the milestone reflects full-stack innovations built specifically for Vera Rubin at production scale, including a programmable liquid-cooling control system (Valvey) and a unified rack-control appliance (Racky), both patent-pending (coreweave.com, SiliconANGLE). NVIDIA's Ian Buck called Vera Rubin "the most capable AI platform NVIDIA has ever built." NVIDIA's own technical blog and press materials describe a larger Vera Rubin POD architecture built on third-generation MGX racks, citing POD-scale figures of up to 1,152 Rubin GPUs, roughly 60 exaflops of FP8 performance, and 10 PB/s of aggregate bandwidth across five rack-scale systems.

The published specs point to three engineering shifts that matter for deployment and benchmarking. Extremely high intra-rack fabric bandwidth, 260 TB/s per NVL72, reduces cross-host traffic, making very long-context inference and persistent agent sessions more viable at rack locality. Integration of NVIDIA BlueField-4 DPUs and Vera CPUs signals that IO, storage caching, and sandboxed CPU tasks are increasingly collocated at rack scale for multi-tenant security. CoreWeave's software-defined cooling and unified rack-control tools also move reliability and serviceability decisions up to the rack level rather than the individual node, per CoreWeave and NVIDIA's technical materials.

Rack-scale co-design shifts where performance and operational risk concentrate for large models and agentic systems: treat rack- and POD-level metrics, interconnect bandwidth, DPU/CPU pairing, cooling and power headroom, as primary capacity and latency constraints for always-on reasoning sessions and million-token contexts, rather than per-GPU FLOPS alone. When planning migrations to NVL72-class infrastructure, factor in that software and orchestration layers need updates to exploit rack-local resources and avoid cross-rack penalties, based on patterns from earlier rack-scale transitions such as Grace, Blackwell, and GB200.

• •Cloud availability and price-performance data for NVL72 instances beyond CoreWeave's initial validation.
• •Independent, third-party benchmarks measuring end-to-end latency for multi-step agent workloads on Vera Rubin.
• •Broader ecosystem support for rack-level management stacks, including telemetry, firmware coordination, and non-disruptive servicing.
• •NVIDIA's Vera Rubin production ramp; NVIDIA's own press materials describe production timelines as forward-looking.