What is happening: data centre demand is climbing
Electricity demand from data centres grew 17% in 2025, and demand from AI-focused data centres grew about 50%, against 3% growth in global electricity demand, according to the International Energy Agency. The agency projects data-centre use to roughly double by 2030, to about 945 TWh a year — comparable to the total electricity use of Japan — with AI the main driver. That figure is a projection, not a measurement.
Water follows the same curve. Data centres cool their servers largely by evaporating water, and the power plants that feed them use water too. United Nations researchers project that data-centre water consumption will roughly double by 2030 as well, and the pressure lands on the specific watersheds where facilities cluster, not on the planet evenly.
One prompt is small
The best public figure comes from Google, which instrumented its own serving stack: in May 2025 the median text prompt to its Gemini apps used 0.24 Wh of energy, emitted 0.03 g of CO2-equivalent and consumed 0.26 mL of water — about 5 drops. The measurement counts idle chips and data-centre overhead, not just the active processors. Google also reports that energy per median prompt fell 33-fold over the previous 12 months.
OpenAI has stated a figure of about 0.34 Wh for an average ChatGPT query, without publishing a method. Other providers have published no comparable figure. For scale, 0.24 Wh runs a television for about 9 seconds. Per prompt, the cost is small; the total grows with volume and, as the next section shows, with habits.
What multiplies it
The same request can cost very different amounts depending on the mode, the model and the number of attempts it takes. These are the multipliers the published studies support, against a single simple prompt:
- A clear prompt that works the first time
- Energy against one simple prompt
- 1× — the baseline
- Evidence
- Measured medians above
- Sending a request 5 times until the answer is right
- Energy against one simple prompt
- 5× — every attempt repeats the full cost
- Evidence
- Arithmetic; goal-level accounting study, 2026
- Reasoning mode switched on
- Energy against one simple prompt
- About 30× on average across 40 open models; hundreds of times for some
- Evidence
- AI Energy Score benchmark, 2025
- A long prompt to the most energy-intensive models
- Energy against one simple prompt
- Above 29 Wh per prompt, over 65× the most efficient systems
- Evidence
- Cross-model benchmark, 2025
- An agentic workflow instead of a linear one
- Energy against one simple prompt
- About 4× per successful goal (888 J against 205 J)
- Evidence
- Goal-level accounting study, 2026
- An AI agent handling a complex request
- Energy against one simple prompt
- Up to 136.5× per request; 348 Wh on average
- Evidence
- Korean university study, 2026
The biggest lever: the mode
In the AI Energy Score benchmark, one open reasoning model used 9,462 Wh across the test set with reasoning on and about 18 Wh with it off. Reasoning helps on hard problems; on a simple question, most of that extra energy buys little.
Why prompting is a sustainability skill
A model spends energy on every word it writes, so the length of the answer drives the cost more than the length of the question. A study of prompt design on open models found that the task a prompt sets matters more than its length, and that open-ended words such as explain or analyse tend to trigger longer, costlier answers.
Asking for less works. In a study of response length, prompts that asked for concise answers cut energy use by 25 to 60% while preserving answer quality against reference answers. And retries are the hidden multiplier: a vague prompt that needs four more attempts costs five times as much as one that works the first time.
The Sustainable Prompting Method
Six habits, each drawn from the evidence above. None asks anyone to use AI less; they get to the right answer sooner.
- Be specific — state the goal, the audience, the format and the length in the first message.
- Give context up front — paste what the model needs instead of correcting it afterwards.
- Right-size the model — a fast, light model for simple tasks; reasoning modes and agents for genuinely hard problems.
- Ask for the length you need — two sentences, a table, or no explanation, when that is enough.
- Edit, don't regenerate — ask for a targeted fix instead of rolling the dice again.
- Ask whether it needs AI at all — a calculator, a search or a known source is sometimes the better tool.
What exists, and the gap
Browser extensions now estimate the energy, carbon and water of chatbot sessions, and the AI Energy Score rates open models with stars, the way appliances carry energy labels. Both show the cost after it is spent. Few tools teach the habits that avoid it, which is where this subject of the library will continue.
What would make this wrong
Per-prompt figures are medians or averages chosen and published by the providers themselves; only Google's comes with a method, and none is an independent audit. Multipliers from benchmarks describe open models on fixed hardware, and commercial services batch requests, cache answers and change models often. Efficiency is improving fast — Google's own figure fell 33-fold in a year — which could shrink every per-prompt number here, while growing use pushes the totals up. Each figure is dated by its source, and this article will be updated as they change.
This is research analysis of published sources, not an audit. Company figures are the companies’ own disclosures; lab results describe their test conditions, not every product; projections are not measurements. Figures in this field change fast, so each one is dated by its source.
References & data
- 1.International Energy Agency — Energy and AI — Data-centre electricity demand: growth of 17% in 2025 (about 50% for AI-focused data centres) and the projection of about 945 TWh a year by 2030.
- 2.Google — Measuring the environmental impact of delivering AI at Google scale (arXiv 2508.15734) — Median Gemini Apps text prompt, May 2025: 0.24 Wh, 0.03 gCO2e, 0.26 mL of water; 33-fold energy reduction over 12 months.
- 3.Sam Altman — The Gentle Singularity — OpenAI's stated figure of about 0.34 Wh per average ChatGPT query, June 2025; no method published.
- 4.AI Energy Score — Hugging Face and Salesforce — Standardised energy ratings for open models; the reasoning-mode comparison of 40 models, December 2025.
- 5.Fortune — The rise of AI reasoning models comes with a big energy tradeoff — Reporting on the AI Energy Score reasoning study: about 30 times the energy on average; 9,462 Wh against about 18 Wh for one model.
- 6.How Hungry is AI? (arXiv 2505.09598) — Benchmark of energy, water and carbon per prompt: the most energy-intensive models exceed 29 Wh per long prompt, over 65 times the most efficient.
- 7.Panigrahy and Tyagi — Energy per Successful Goal (arXiv 2605.22883) — Goal-level energy accounting: agentic workflows used 888.1 J against 205.3 J per successful goal; retries counted in full.
- 8.The Korea Times — KAIST study on AI agent energy use — Agents used up to 136.5 times the energy per query of standard generative AI; 348.41 Wh on average for a complex request, July 2026.
- 9.Brevity is the soul of sustainability (arXiv 2506.08686) — Prompt strategies for concise answers cut LLM energy use by 25–60% while preserving response quality.
- 10.Green Prompting (arXiv 2503.10666) — How prompt and response characteristics drive inference energy; the task and specific keywords matter more than prompt length.
- 11.Reuters — AI to double data centre power and water use by 2030, UN researchers say — United Nations University report on data-centre electricity and water demand, June 2026.
- 12.AI Wattch — Open-source browser extension estimating the energy and carbon of ChatGPT, Claude and Gemini sessions.