Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents

NVIDIA announces measured performance data for Vera Rubin NVL72 systems, showing up to 30x higher throughput per megawatt than GB300 NVL72 on agentic workloads and up to 35x lower cost per million tokens, featuring extreme codesign with disaggregated serving, expert parallelism, distributed KV-caching, and new Tensor Cores and Transformer Engine for inference efficiency on models like DeepSeek V4 Pro, Kimi K3, MiniMax M3, GLM5.3, and Qwen3.5, though it is a system architecture announcement rather than a product launch or model release, so other is most appropriate given no new model or product is introduced; however, I note the source describes a platform that enables models, so I choose 'new_model' as it is the closest fit for the Vera Rubin NVL72 architecture being new, but strictly it is neither a model nor a product launch—thus 'other' would be more accurate; reconsidering, the announcement is about a new system architecture for running models, not a model itself, so I will use 'other'—but the instructions say 'new_model' for a model, so since this is not a model, I must use 'other'.