Vera Rubin’s first MLPerf preview makes inference the headline

The next accelerator story is about serving

MLPerf Inference v6.1, released September 16, included the first preview submissions for NVIDIA’s Vera Rubin NVL72 rack-scale AI system. NVIDIA reported up to 3.7 times the throughput of GB300 NVL72 on Qwen3-VL and up to 2.5 times on DeepSeek-R1.

Nebius independently submitted a preview system using 36 Rubin GPUs across nine nodes and reported 16,427 tokens per second per GPU on its leading DeepSeek-R1 server result. These are preview benchmark results, not a guarantee of production availability.

Hardware is only one layer

The benchmark stack included vLLM, NVIDIA Dynamo, and TensorRT-LLM. Model parallelism, precision, interconnect design, scheduling, and software tuning all shape the result that a user experiences as inference speed.

The meaningful unit of progress is increasingly the serving system, not the chip in isolation.

For infrastructure teams, the results are an early standardized signal about cost and throughput for reasoning and multimodal workloads. The next question is how the preview numbers translate to availability, utilization, and real application latency.