The Electricity Problem at the Heart of Modern AI
Every AI data centre runs on a fixed power budget. The company that squeezes the most useful work out of each watt wins. Here is what that means in plain English.
Key points
- NVIDIA's GB300 NVL72 system delivers up to 25 times more useful AI work per watt than the previous Hopper generation, according to SemiAnalysis InferenceX analysis cited in NVIDIA's own blog.
- Software improvements alone boosted performance per watt on one model by up to 5 times within a single month in 2025.
- NVIDIA claims its power-management software lets operators run up to 40 percent more chips within the same electricity budget.
- Anthropic, OpenAI, and Perplexity currently run production workloads on Blackwell hardware, according to NVIDIA.
- The figures here come from an NVIDIA promotional post, not an independent peer-reviewed study.
AI is an electricity problem. Every time a chatbot answers your question, or an AI agent (software that carries out multi-step tasks on its own) handles a task for you, a data centre burns power to do it. That facility has a fixed electricity budget. The company extracting the most useful output from that budget makes the most money; the one that can't, can't scale.
NVIDIA calls this "performance per watt": how much AI output you get for each unit of electricity consumed. The higher that figure, the more tokens (the small text chunks AI systems process) a facility can produce before hitting its power ceiling.
The numbers NVIDIA published are striking. Its GB300 NVL72 system, a rack-scale server platform built around its latest Blackwell chips, reportedly delivers up to 25 times better performance per watt than its older Hopper generation on DeepSeek V4 Pro, one of the current open AI models. GLM5.1, another frontier model, shows a gain of up to 20 times. Kimi K2.6, built for longer automated tasks, reaches up to 10 times.
Those figures come from SemiAnalysis InferenceX, a third-party benchmarking group, cited in NVIDIA's own blog post. They're not from a peer-reviewed study or independent audit.
Does any of this affect ordinary people?
Yes, indirectly but meaningfully. The speed and cost of AI tools people use every day depend on how efficiently data centres run. Perplexity says it serves millions of queries daily on Blackwell hardware. More efficient infrastructure means the services built on top of it can get faster and cheaper.
NVIDIA also highlights a less obvious drain: cooling. In a typical AI facility, only around 60 percent of electricity drawn from the grid actually reaches the chips and does useful work. NVIDIA's DSX MaxLPS software, which shifts power between chips in real time and supports liquid cooling, aims to recover some of that waste, letting operators run up to 40 percent more chips within the same power budget.
For Anthropic and OpenAI, which NVIDIA says run inference on Blackwell NVL72, this is about economics as much as engineering. More output per watt means lower cost per query, which feeds directly into margins.
Software matters as much as hardware. On DeepSeek V4, updates to NVIDIA's inference software stack improved performance per watt by up to 5 times within a single month, a large gain without replacing a single chip. Our 15 July story found the same pattern emerging across the industry: companies quietly shifting focus from raw model size toward cheaper, more efficient systems.
All of this lays ground for NVIDIA's next platform, Vera Rubin, which the company says will push rack-scale efficiency further.
The core tension is real even if the marketing is selective. Electricity is finite. AI appetite isn't. What to watch: whether independent researchers replicate these efficiency gains, and whether the cost savings actually reach the people paying for AI services or stay inside the data centre.
