NVIDIA highlights Blackwell rack-scale AI efficiency gains

NVIDIA has published a blog post arguing that power efficiency is the defining constraint for AI factories and highlighting the rack-scale performance of its Blackwell and Vera Rubin platforms.

• The post says NVIDIA GB300 NVL72 delivers up to 25x performance per watt versus Hopper on leading open models, with additional examples of up to 20x on GLM5.1 and 10x on Kimi K2.6.
• It also says NVIDIA DSX MaxLPS can help operators run up to 40% more GPUs within the same power budget, and cites production use by Anthropic, OpenAI, CoreWeave, Perplexity, and Fireworks AI.
• This is a commentary-style NVIDIA blog, not a standalone product launch announcement, though it references existing deployments and platform capabilities.