NVIDIA: performance per watt as the key metric — Blackwell GB300 up to 25× more efficient than Hopper

What is “performance per watt” and why has it become the key metric?

Performance per watt denotes the amount of useful compute work — for example, the number of AI tokens generated — per unit of electrical energy consumed. As the costs of power and cooling in data centers grow alongside demand for AI inference, this metric increasingly determines the economic viability and environmental footprint of AI infrastructure. NVIDIA’s new blog post claims that the GB300 NVL72 has redefined the efficiency standard.

Blackwell GB300 NVL72 — concrete results

On the DeepSeek V4 Pro model, NVIDIA’s Blackwell GB300 NVL72 achieves up to 25× better performance per watt compared to the previous Hopper generation. On the GLM5.1 model the improvement reaches up to 20×, while on Kimi K2.6 it reaches up to 10×. These are hardware measurements covering several leading open models in production environments. The Blackwell architecture brings significantly improved memory bandwidth and the energy efficiency of the NVLink network topology in the NVL72 configuration with 72 GPUs per rack.

Software optimizations: 5× in one month

Notably, hardware is not the only driver of improvement. Software optimizations alone — without any hardware change — delivered up to a 5× efficiency improvement on DeepSeek V4 within a single month. NVIDIA’s tool ecosystem includes TensorRT-LLM, NVIDIA Dynamo, SGLang, vLLM, and NVIDIA DSX MaxLPS. This fact suggests that the inference orchestration stack is just as important as the silicon — an unusually candid admission from a company known for its focus on GPU hardware.

Production clients and industry implications

Some of the best-known AI companies are already using Blackwell systems in production: Anthropic, OpenAI, CoreWeave, Perplexity, and Fireworks AI. For comparison, the previous Hopper generation (H100/H200) was the de facto standard for AI training and inference from 2022 to 2024. The shift to Blackwell in the production environments of leading service providers signals that the GPU generation transition is proceeding faster than analysts predicted.