Nebius Sets New Inference Performance Standards for Blackwell and Blackwell Ultra

As the industry shifts toward reasoning models and multimodal systems, raw hardware access is no longer enough. These results prove that Nebius has optimized its networking and software stack to extract maximum throughput from NVIDIA's newest silicon. The GB300 NVL72 configuration demonstrated linear scaling across 72 GPUs, ensuring performance does not degrade at scale.

You can now leverage these optimized configurations for latency-sensitive applications like agentic loops. Nebius provides these capabilities through its AI cloud platform and the Nebius Token Factory. If you are scaling production deployments of DeepSeek R1, these benchmarks offer a verified baseline for expected tokens-per-second performance.