Dell shipped the first Nvidia Vera Rubin NVL72 rack to CoreWeave. Each cabinet packs 72 Rubin GPUs, 36 Vera CPUs, 3.6 exaFLOPS FP4 inference, 75 TB HBM4 memory, and 260 TB/s NVLink bandwidth [According to @rohanpaul_ai].
Key facts
- Dell shipped first Nvidia Vera Rubin NVL72 rack to CoreWeave.
- 72 Rubin GPUs, 36 Vera CPUs per rack.
- 3.6 exaFLOPS FP4 inference per cabinet.
- 75 TB fast memory per rack.
- 260 TB/s NVLink bandwidth per cabinet.
Dell Technologies has delivered the world's first Nvidia Vera Rubin NVL72 rack to cloud GPU provider CoreWeave, marking the first public deployment of Nvidia's next-generation architecture succeeding the Blackwell B200 and Grace Hopper lines [According to @rohanpaul_ai].
The Vera Rubin platform, named after astronomer Vera Rubin, represents Nvidia's 2026 flagship data-center GPU. Each NVL72 rack integrates 72 Rubin GPUs paired with 36 Vera CPUs in a 1:2 CPU-to-GPU ratio, interconnected via fifth-generation NVLink providing 260 TB/s bisection bandwidth. Memory totals 75 TB of fast memory — likely HBM4, though Nvidia has not confirmed the generation — per cabinet.
The unique take: the Vera Rubin NVL72 delivers 3.6 exaFLOPS at FP4 precision per rack, roughly 2.5x the FP4 inference throughput of a Blackwell B200 NVL72 cabinet (which offered ~1.4 exaFLOPS FP4). This implies a generational leap in dense compute density, though inference-optimized workloads benefit most from the FP4 boost. For training at FP8/FP16, the delta is likely smaller; Nvidia has not published BF16 or FP8 peak figures for Rubin.
CoreWeave's early access is strategic. The GPU-as-a-service provider has been Nvidia's preferred launch partner for high-end hardware, previously securing early Blackwell B200 allocations. CoreWeave has not disclosed the number of racks ordered or the total contract value [per the source].
Dell's role as first integrator is notable. The company has been aggressively expanding its AI factory business, competing with Supermicro and HPE for hyperscaler and cloud provider rack-scale deals. Dell did not comment on the timeline for volume production or additional customers.
The Vera Rubin architecture is expected to enter volume production in the second half of 2026, with Nvidia targeting data-center deployments for both training and inference across major cloud providers including AWS, Google Cloud, and Microsoft Azure. Pricing per rack has not been disclosed; the B200 NVL72 previously carried estimated system pricing above $3M per cabinet.
Key Takeaways
- Dell delivered the first Nvidia Vera Rubin NVL72 rack to CoreWeave.
- Each rack packs 72 Rubin GPUs, 36 Vera CPUs, 3.6 exaFLOPS FP4 inference, 75 TB memory, and 260 TB/s NVLink bandwidth.
What to watch
Watch for Nvidia's Vera Rubin pricing disclosure at GTC 2026 (March). Also track CoreWeave's Q2 2026 earnings call for rack count and utilization metrics. Dell's next earnings will reveal whether additional cloud customers have placed Vera Rubin orders.
