The frontier of artificial intelligence hardware advanced sharply this week, as NVIDIA unveiled performance results for its new Vera Rubin NVL72 platform in the MLPerf Training v6.1 benchmarks. The system, designed for complex AI model training and agentic inference workloads, delivered what NVIDIA claims is up to double the performance per GPU compared with the company’s earlier GB300 NVL72 system—setting a new technical pace that could reshape large-scale AI deployments.

Announced on November 16, 2026, the MLPerf results saw the Vera Rubin NVL72 outpace competitors and its own lineage across all seven MLPerf Training benchmarks. Highlighted was a striking result on the DeepSeek-V3 671B model: the NVL72 trained the model in 4.14 minutes using 1,024 GPUs, halving the training time recorded by the previous-gen system at similar scale. NVIDIA’s data also reported that scaling up from 256 to 1,024 GPUs on Vera Rubin maintained an efficiency above 97%, a critical metric for institutions investing in scalable, high-throughput AI infrastructure.

The Vera Rubin platform’s inaugural submission to the MLPerf Endpoints suite introduced groundbreaking agentic AI performance. In a provisional test with the Kimi K3 model—a large model purpose-built for complex automated workflows—the NVL72 achieved over ten times the throughput per watt of GB300 NVL72 at equivalent response speeds. Such power efficiency matters for data centers where operational costs and environmental impact are increasingly under scrutiny. While these specific results await MLCommons Association peer verification, the technical implications have already garnered industry attention.

Adoption of the Vera Rubin platform is underway beyond laboratory tests. SpaceXAI, a prominent AI innovator, disclosed in August its deployment of NVIDIA’s Vera CPUs to boost agentic AI applications, marking one of the earliest large-scale deployments of this new generation of processors. This underscores the platform’s relevance not only for benchmarks but in practical, mission-driven AI environments.

According to NVIDIA, ongoing improvement is central to the Vera Rubin value proposition. The company outlined plans to further optimize its CUDA-driven software stack—a commitment intended to push performance even higher post-deployment and offer buyers a long-term path to increasing return on their infrastructure investment.

Industry observers now watch to see how quickly competitors will respond and whether independent verification will substantiate NVIDIA’s provisional gains in agentic AI workloads. For now, the Vera Rubin NVL72 platform stands as the new reference point for training and inference performance in AI hardware at scale.