NVIDIA has outlined new claims about the energy efficiency of its Blackwell-based AI infrastructure, saying the figures relate to serving frontier AI models under fixed power budgets.
The announcement centres on the idea that performance per watt has become a defining metric for AI data centre economics, as model operators try to raise token output without breaching electricity and cooling limits.
Newer frontier models are increasingly built on mixture-of-experts architectures, NVIDIA said, spreading workloads across specialist parts of a model. The shift, it argued, favours larger GPU domains linked by high-speed scale-up interconnects over smaller systems built around fewer processors.
According to NVIDIA, its GB300 NVL72 rack-scale system delivers up to 25 times the performance per watt of the Hopper generation on DeepSeek V4 Pro, up to 20 times on GLM5.1 and up to 10 times on Kimi K2.6. It attributed those gains to the move from an eight-GPU domain in Hopper systems to a 72-GPU rack-scale design in Blackwell NVL72.
NVIDIA also argued that a single performance figure does not capture the full range of inference trade-offs. Some deployments are tuned for low latency, while others prioritise throughput and cost, and operators often shift between those settings depending on demand.
To address that, it presents Pareto curves for each model rather than a single benchmark point. NVIDIA also pointed to simulation tools intended to help customers identify different operating points before using live GPU capacity.
Rack design
NVIDIA tied the efficiency improvements to what it described as a tightly integrated rack-scale design spanning chips, interconnects and inference software. It said the NVLink Switch in its systems is built specifically for large GPU domains used in AI inference, rather than adapted from broader networking equipment.
That design approach extends into software. NVIDIA said its inference stack includes Dynamo, TensorRT LLM, SGLang and vLLM, with support for techniques such as NVFP4 quantisation, disaggregated serving, expert parallelism, KV-aware routing and KV cache offloading.
Software tuning alone can materially alter energy efficiency over time, NVIDIA said. On DeepSeek V4, performance per watt improved by up to five times in a single month through software advances, according to the company.
Power limits
Beyond processors and model serving, NVIDIA also focused on data centre power use outside the compute layer. It said cooling losses and rack-level inefficiencies can leave only about 60% of electricity drawn from the grid available for AI processing.
NVIDIA said its DSX MaxLPS software is designed to reduce that gap by shifting power between GPUs and racks in real time, while supporting warm-water liquid cooling and power steering. In its account, that can let operators run up to 40% more GPUs within the same power budget.
The broader message reflects a growing concern across the AI industry that access to electricity, rather than access to chips alone, may become the main constraint on scaling inference. As more companies deploy agentic AI systems that generate higher token volumes over longer sessions, the operating cost of each token has become more important.
Production use
NVIDIA also sought to anchor its efficiency claims in production deployments rather than synthetic tests. It said large rack-scale systems introduce failure modes not seen in smaller single-node installations, making operational reliability a central issue for AI factories.
Anthropic and OpenAI use Blackwell NVL72 systems for inference workloads, according to NVIDIA. It also cited deployments by CoreWeave, Perplexity and Fireworks AI as evidence that Blackwell systems are being used to serve open models under live traffic.
CoreWeave has deployed Kimi K2.6 on GB300 NVL72, NVIDIA said, using NVFP4 quantisation and speculative decoding to lift inference performance. Perplexity is running Qwen3 235B and post-trained Qwen3.5-397B-A17B on GB200 NVL72 for its AI agent platform, while Fireworks AI is deploying GLM 5.2 on Blackwell systems for customers including Cursor and Factory AI, according to the company.
NVIDIA framed that production experience as a base for its next platform, Vera Rubin, which it said will extend the same rack-scale focus on energy efficiency. It added that the sixth generation of NVLink Switch in Vera Rubin is designed for AI-specific workloads including SHARP, which performs in-network computing in the switch rather than on the GPU.
In making those claims, NVIDIA is arguing that AI infrastructure buying decisions are shifting from raw chip comparisons to whole-rack economics. For cloud providers, model developers and inference specialists, the winning system, in its view, will be the one that turns the largest share of a fixed power budget into useful AI work.