While Google and OpenAI reassure users with low energy-per-query figures, climate scientist Zeke Hausfather's detailed analysis reveals a starkly different picture: AI agents require roughly 600 times more energy than simple chat prompts. His measurement data from eight weeks of Claude Code usage documents an efficiency problem with significant implications for businesses and climate impact.
Key Facts
- 170 kWh electricity consumption over 8 weeks; 3.2 billion tokens processed
- 1,138 user inputs triggered over 14,000 model calls – averaging 12 per prompt
- Google claims 0.24 Wh per Gemini prompt; OpenAI CEO Sam Altman estimated ChatGPT at 0.34 Wh
- 96 percent of tokens were cache reads; only 0.4 percent was actual model output
The Gap Between Published Claims and Measured Reality
Google and OpenAI published their energy consumption figures last year to address concerns. Google claims an average Gemini text prompt uses less energy than nine seconds of television. OpenAI CEO Sam Altman compared average ChatGPT queries to a 2009 Google search.
Yet these figures ignore central realities of modern AI use: Reasoning models generate far more tokens from a single query, multi-agent systems coordinate multiple model calls in parallel, and code generators like Claude Code trigger hundreds of internal processing steps per user input. Hausfather's concrete measurements document exactly this gap.
How the Measurement Worked
Hausfather used Claude Code over eight weeks, logging every step: His 1,138 typed prompts triggered over 14,000 model calls – averaging 12 per input. Each prompt processed an average of 2.9 million tokens. For comparison, a typical chat without reasoning or web search runs on about 1,000 tokens.
The total balance: 170 kWh of electricity over eight weeks, with an uncertainty range of 70 to 330 kWh. Per input, that's roughly 150 watt-hours – 600 times a simple chat prompt. Extrapolated to a full year, this usage intensity would according to Hausfather generate CO₂ emissions comparable to running an electric clothes dryer.
One critical finding: 96 percent of processed tokens were mere cache reads, because the agent re-reads its entire accumulated context at each of the 14,000 steps. The actual model output Hausfather sees on screen accounts for only 0.4 percent of all tokens.
The Scaling Problem
Hausfather emphasizes another critical point: AI labs are already developing autonomous agent systems that work independently on tasks for days, weeks, or even months. Should such systems move into production, energy consumption could grow exponentially – far beyond scenarios discussed today.
Hausfather identifies shifting data centers to renewable energy as the most important lever for emission reduction. Yet this doesn't address the fundamental problem: the efficiency of autonomous AI systems is currently far lower than publicly claimed.
What This Means for European Enterprises
For mid-market and industrial companies in Europe, this analysis is a warning. Anyone planning to deploy AI agents for automated processes – whether for code development, data analysis, or autonomous optimization – should factor in real energy costs, not vendor marketing figures. Simultaneously, a critical gap remains: nobody outside AI labs knows actual per-token costs for state-of-the-art models. Hausfather himself emphasizes his figures are informed estimates, not definitive measurements. This makes transparent measurement and benchmarking all the more important – especially for companies taking sustainability commitments seriously.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
