NVIDIA claims GB300 NVL72 systems deliver up to 25x performance per watt over Hopper

NVIDIA said on July 14, 2026, that its GB300 NVL72 systems deliver up to 25 times more tokens per watt than the prior Hopper generation, citing SemiAnalysis InferenceX benchmark data. The company published a technical argument that tokens per watt has become the decisive metric for AI infrastructure because power is the binding constraint in 2026. Transformer lead times stretch five years and switchgear is sold out through 2028, limiting new capacity. NVIDIA said a 200-megawatt site earns materially different revenue depending on token-per-watt efficiency. The GB300 NVL72's 72-GPU NVLink domain at 130 terabytes per second eliminates communication bottlenecks for mixture-of-experts models such as DeepSeek V4 Pro, GLM5.1 and Kimi K2.6. Benchmarks show up to 25 times more tokens per watt on DeepSeek V4 Pro, up to 20 times on GLM5.1 and up to 10 times on Kimi K2.6. NVIDIA presents these as Pareto-curve comparisons that vary with serving configuration and quantization. Software improvements alone delivered a 5x efficiency increase on DeepSeek V4 over a single month on the same hardware. SemiAnalysis InferenceX is an independently operated benchmark platform; AMD's developer blog cites the same dataset to show MI355X delivers lower cost-per-token at high-concurrency FP8 configurations without multi-token prediction.