CoreWeave (Nasdaq: CRWV), in collaboration with NVIDIA and IBM, delivered the largest-ever MLPerf® Training v5.0 submission on NVIDIA Blackwell, using 2,496 NVIDIA Blackwell GPUs running on CoreWeave's AI-optimized cloud platform. This submission is the largest NVIDIA GB200 NVL72 cluster ever benchmarked under MLPerf, 34x larger than the only other submission from a cloud provider highlighting the large scale and readiness of CoreWeave's cloud platform for today's demanding AI workloads.

The submission achieved a breakthrough result on the largest and most complex foundational model in the benchmarking suite–Llama 3.1 405B–completing the run in just 27.3 minutes. When compared against submissions from other participants across similar cluster sizes, CoreWeave's GB200 cluster achieved more than 2x faster training performance. This result highlights the significant performance leap enabled by the GB200 NVL72 architecture and the strength of CoreWeave's infrastructure in delivering consistent, best-in-class AI workload performance.

"AI labs and enterprises choose CoreWeave because we deliver a purpose-built cloud platform with the scale, performance, and reliability that their workloads demand," said Peter Salanki, Chief Technology Officer and Co-founder at CoreWeave. "These MLPerf results reinforce our leadership in supporting today's most demanding AI workloads."

These results matter because they translate directly to faster model development cycles and an optimized Total Cost of Ownership. For CoreWeave customers, that means cutting training time in half, scaling workloads efficiently, and training or deploying their models more cost-effectively by leveraging the latest cloud technologies, months before their competitors. With leading submissions for both MLPerf Inference v5.0 and Training v5.0 benchmarks and the sole cloud provider ranked in the Platinum tier of SemiAnalysis' ClusterMAX, CoreWeave sets the standard for AI infrastructure performance across the entire cloud stack.