CoreWeave Doubles GB300 NVL72 DeepSeek R1 Server Throughput in MLPerf 6.0

GB300 NVL72 results: 2x in server mode, 1.6x offline

According to CoreWeave, its MLPerf 6.0 submission with NVIDIA GB300 NVL72 systems shows clear gains over the MLPerf 5.1 round. Using DeepSeek R1 (671B, NVFP4), throughput improved 1.6x in offline mode and 2x in server mode over the last 6 months, the company says.

The listed figures for the GB300 NVL72 are 5,574 tokens per second per GPU in server mode and 9,821 in offline mode. CoreWeave states that the offline result was the highest of all submitters for DeepSeek R1.

The company also tested GPT-OSS-120B on the GB300 NVL72, reporting 14,946 tokens per second per GPU in server mode and 14,226 in offline mode. According to CoreWeave, this showed up to 28% higher throughput than the NVIDIA GB200 NVL72, and it says it led all submitters in improvement over the GB200 NVL72.

GB300 versus GB200 and continued GB200 lead

CoreWeave describes the NVIDIA GB300 NVL72 as a rack-scale system built for AI reasoning, with 1.5x more dense NVFP4 Tensor Core FLOPS and 2x higher attention performance than NVIDIA Blackwell GPUs. It uses the same architecture as the GB200 NVL72 but has more memory and other improvements, according to the company.

In its DeepSeek R1 runs, CoreWeave reports a 14% improvement over the GB200 in server mode and a 34% improvement in offline mode.

On the GB200 NVL72, CoreWeave reports 4,868 tokens per second per GPU in server mode and 7,323 in offline mode for DeepSeek R1. It says these results led all submitters in both modes. The company also notes it was the first AI-native cloud to offer the GB200 NVL72 in production at scale and was the GB200 NVL72 benchmark leader in MLPerf 5.0.

The footnotes add context. The tokens per second per GPU figure is not an official MLPerf metric. CoreWeave uses it to normalize submissions that use different numbers of GPUs. Its GB200 node used 64 GPUs, while other submitters' nodes used 72.

From benchmarks to inference products

CoreWeave ties the results to its inference offerings, which it describes as a path from fully managed to fully self-managed infrastructure.

• Serverless Inference through Weights & Biases (W&B) Inference offers autoscaling, high availability and integrated observability, with pay-per-token pricing.
• Dedicated Inference, recently announced and currently in preview, lets customers bring their own weights and pick GPUs and open runtimes, while CoreWeave handles production operations and cost visibility. Teams can request preview access.
• Inference on CKS (CoreWeave Kubernetes Service) lets developers self-host their models, runtimes and orchestration, with control over tuning and observability.

According to the company, these paths share a consistent architectural base, so teams can move from iteration to scaled production without replatforming.

What it means for customers

The results are self-reported by CoreWeave and rely on its own per-GPU normalization, so comparisons should be read with that in mind. For teams running reasoning models such as DeepSeek R1, the company's message is that throughput on GB300 NVL72 systems has risen markedly in six months and that GB300 offers measurable gains over GB200 in its tests. CoreWeave says customers include Cursor AI, Cline, Mercado Libre and 8 of the 10 leading foundation model providers.