NVIDIA published performance-per-watt benchmarks for GB300 NVL72 against Hopper and the gap is large: 25x on DeepSeek V4 Pro, 20x on GLM5.1, 10x on Kimi K2.6, all sourced from SemiAnalysis InferenceX. The numbers matter less than the framing. NVIDIA is not arguing it has a faster GPU. It is arguing that performance per watt is the correct denominator for every AI infrastructure decision because it cannot be gamed: a fixed power budget at the building level sets a hard ceiling on AI factory revenue, and the only way to move that ceiling is tokens per watt, not tokens per benchmark run.
The architecture story behind the numbers is the domain-size shift. Hopper ran an 8-GPU NVLink domain. GB300 NVL72 runs a 72-GPU domain. Every frontier model in production today uses a mixture-of-experts (MoE) architecture where inference routes activations to a subset of expert weights on each forward pass. In an 8-GPU domain, the experts are split across domain boundaries and cross-boundary traffic is a bandwidth tax on every token. In a 72-GPU domain, the full expert set fits inside the NVLink fabric. The 25x performance-per-watt improvement is mostly that tax going to near zero, not a silicon-level transistor gain.
NVIDIA also ships DynoSim, a simulation tool that models the Pareto frontier between latency and throughput for a given workload before spending a single GPU-hour on empirical testing. That is the tooling signal inside the benchmark release: infrastructure buyers can now simulate their optimal operating point before signing a rack contract, rather than discovering it after deployment. The metric frame plus the simulation tool together shift AI infrastructure procurement from "trust the benchmark sheet" to "model your workload, then buy."
Vendors selling Hopper-generation hardware on raw throughput benchmarks have 12-18 months before the renewal conversation resets around per-watt efficiency. Hyperscalers running fixed power budgets at the building level are already doing that math. The question for mid-market AI infrastructure buyers is whether they want to run that same calculation before their next hardware cycle or after it.