Nvidia Vera Rubin NVL72 Doubles GB300 GPU Training Speed

Nvidia's Vera Rubin NVL72 platform swept every benchmark in the MLPerf Training v6.1 suite at its debut, delivering up to twice the per-GPU throughput of the company's own GB300 NVL72 baseline. Officially released on November 17, 2026, the platform also posted the first results in MLPerf's new Endpoints category, where it claimed more than 10x generational gains in throughput per watt for agentic AI workloads. The results position Vera Rubin as Nvidia's most aggressive bid yet to accelerate the rollout of next-generation training infrastructure.
Twenty cloud and hardware partners submitted results on Nvidia platforms this round, including CoreWeave, Microsoft Azure, Oracle, Dell, HPE, Google Cloud, Lambda, Nebius, Crusoe, and Supermicro. CoreWeave reported the first multi-rack Vera Rubin cloud deployment in the same batch of submissions, while Azure matched the top GB200 scale at 10,240 GPUs. The Vera Rubin figures top out at 1,024 GPUs in the published results.
Release Date and Availability
Officially launched on November 17, 2026, Vera Rubin NVL72 debuted with full MLPerf Training v6.1 benchmark coverage and immediate availability through Nvidia's partner ecosystem. The platform arrived alongside the first MLPerf Endpoints submissions, a new category designed to measure agentic AI performance on real coding and enterprise workflows. Nvidia's accompanying blog post detailed the benchmark sweep but offered no executive quotes or third-party corroboration, flagging the Kimi K3 agentic figure as provisional.
DeepSeek-V3 Training Efficiency
On the DeepSeek-V3 benchmark, Vera Rubin NVL72 scaled from 256 to 1,024 GPUs at approximately 97% efficiency, cutting training time from 16 minutes to 4 minutes at the top configuration. The near-linear scaling across a 4x GPU increase demonstrates the platform's ability to maintain throughput as clusters grow, a critical factor for foundation model training where coordination overhead typically erodes gains beyond 512 accelerators. Nvidia attributed the efficiency to architectural improvements in inter-GPU communication and memory bandwidth over the GB300 generation.
Agentic AI Inference Gains
The most striking claim sits in the new MLPerf Endpoints category. On a provisional Kimi K3 agentic inference test, 32 Vera Rubin accelerators delivered more than 10x the throughput per watt of a 48-accelerator GB300 configuration at identical response speeds. The Endpoints benchmark evaluates systems on multi-step agentic coding and enterprise workflows, requiring the agent to complete tasks correctly while measuring energy efficiency. Nvidia framed the result as evidence that Vera Rubin can handle the spiky, context-heavy inference patterns emerging in production agentic systems, where long-context windows and iterative tool use dominate compute budgets.
Competitive Landscape and Market Context
The sweep arrives amid a dense run of Nvidia coverage, with 284 tracked stories in the past 90 days. No competing accelerator vendor appears in the MLPerf v6.1 submissions for the training or agentic categories, leaving the benchmark round as an Nvidia-versus-Nvidia comparison. The company positioned Vera Rubin as the successor to GB300, its previous Blackwell-generation platform, and emphasized the per-watt gains as the central argument for infrastructure upgrades. The lack of third-party validation and the provisional status of the Kimi K3 figure leave room for independent confirmation as deployments scale beyond the initial partner submissions.
What This Means
Vera Rubin NVL72's 2x per-GPU training gains and 10x energy efficiency improvement on agentic inference set a new high-water mark for Nvidia's internal generational progress. The DeepSeek-V3 scaling results demonstrate that large-scale training clusters can maintain near-linear efficiency past 1,000 GPUs, a threshold that has historically separated research prototypes from production infrastructure. For enterprises evaluating GPU refresh cycles, the energy equation matters as much as raw speed: a 10x throughput-per-watt gain translates directly to lower operating costs and smaller datacenter footprints. The absence of competing submissions in MLPerf v6.1 suggests Nvidia's lead in training hardware remains unchallenged for now, but the provisional Kimi K3 benchmark signals that agentic AI will become the next contested front as inference workloads grow more complex.
on Emergent today





