How much energy does it actually take to send a message to an AI assistant? And how does that compare to boiling a kettle, driving a kilometre, or eating a beef burger?
This post collects the best available numbers on energy, carbon, and water across Claude, ChatGPT, Copilot, Gemini, and Grok — per query, at the individual scale, and globally. It also looks at what happens to those numbers when reasoning models enter the picture.
Key takeaways up front:
- For an individual, AI use is a rounding error compared to transport, diet, and heating
- At global scale it’s a real and growing electricity load, on track to double or triple by 2030
- The dominant uncertainty isn’t model efficiency — it’s whether reasoning and agentic workloads stay niche or become default
Per-query comparisons
Energy per query (Wh)
~100 input → 300 output tokens. Gemini’s figure is from Google’s own disclosure; the rest are external estimates from academic benchmarking (Jegham et al. 2025, Epoch AI).
Gemini’s 0.24 Wh looks dramatically better than Claude’s 0.84 Wh, but the comparison hides three things: Google’s figure uses their own fully-loaded methodology, Anthropic publishes nothing so Claude is an external estimate, and the 3.5× gap is still under one watt-hour. None of these models is energy-expensive per query — the question is what you scale them to.
CO₂e per query (grams)
Gemini uses Google’s renewable-PPA grid factor; others use the global-average 430 g CO₂/kWh (IEA 2025). The bars are not strictly comparable on the same grid — Claude and ChatGPT on Google’s mix would drop 5–10×.
The CO₂e ranking is mostly the energy ranking multiplied by grid factor — Gemini wins partly on efficient inference and partly on Google’s renewable PPAs. If Anthropic and OpenAI ran on the same renewable mix Google reports, their per-query CO₂e would drop 5–10× with no model change. This is a deployment story more than a model story.
Water per query (mL)
Self-reported figures (Google, OpenAI) cover only on-site evaporative cooling (scope-1). Including off-site power-generation water (scope-2) raises totals roughly 3–10×. Claude and Grok are scaled from energy ratios using Microsoft Azure WUE.
Most “AI water” headlines come from this metric, and 0.3–0.8 mL per query is genuinely small — a few drops at most. The catch: these are scope-1 figures only. Including off-site power-generation water raises every bar by roughly an order of magnitude (see the scope-2 section below).
The individual’s footprint
Annual footprint at 10 queries per day (kg CO₂e)
For context: one passenger seat on a one-way Sydney → Melbourne flight is roughly 90 kg CO₂e. A year of heavy AI use is comfortably under 2% of that.
A heavy individual user’s annual AI footprint is comfortably under 2% of a single Sydney → Melbourne flight’s emissions. For an individual, AI use is a rounding error against transport, diet, and heating.
Everyday energy comparison (Wh, log scale)
A single chat query sits between a Google search and a minute of YouTube. Boiling a kettle is two orders of magnitude more energy than your entire day of AI use.
A chat query sits between a Google search and a YouTube minute in raw energy. Boiling a kettle is two orders of magnitude more energy than your entire day of AI use; an EV mile is three. If you care about your individual energy footprint, AI is among the lowest-leverage things to optimise.
Everyday water comparison (mL, log scale)
Per-query water is a millilitre or less. A single beef burger requires roughly 3 million Claude queries’ worth of water to produce.
The cup of coffee in your hand required roughly 175,000 Claude queries’ worth of water to produce. A beef burger is around 3 million Claude queries. Agricultural water footprint of one meal swamps a year of heavy AI use by three orders of magnitude.
Everyday carbon comparison (g CO₂e, log scale)
A Claude query emits roughly the same CO₂e as 2 metres of driving a petrol car. A single beef burger ≈ 8,000 Claude queries.
A Claude query emits about as much CO₂e as driving a petrol car 2 metres. A Sydney → Melbourne flight ≈ 250,000 Claude queries per passenger. Transport and diet dominate individual emissions; AI is in the noise — and the comparison holds across all the major chat models.
The reasoning amplifier
Reasoning models dwarf the standard-chat gap (Wh per query, log scale)
Chain-of-thought tokens multiply inference cost. The 3× gap between Gemini and Claude is rounding error compared to switching on a reasoning model.
This is the chart that matters most going forward. The 3× gap between Gemini and Claude on standard chat is rounding error compared to switching on a reasoning model: o3 and DeepSeek-R1 consume 50–100× more per query. As reasoning becomes the default mode — GPT-5 routing, Claude extended thinking, Gemini Deep Think — per-query averages will rise faster than user counts.
At global scale
Across the five major commercial chat services, the aggregate numbers are already significant:
| Metric | Value | Context |
|---|---|---|
| Daily energy (5 services) | 1.76 GWh | ≈ 60,000 AU homes / day |
| Annual energy | 643 GWh | ≈ Hobart’s annual use |
| Daily water (scope-1) | 1.7 m³ | ~8 AU households |
| Annual water (scope-1) | 620 m³ | ~1/4 Olympic pool |
Daily energy by service (MWh / day)
Grok ranks second on energy despite only ~6% of ChatGPT’s user base — a consequence of its 2.5 Wh/query rate versus 0.34 Wh for GPT-4o.
ChatGPT’s daily energy is roughly equal to a small Australian city’s electricity use. Efficiency × scale is what matters; neither dimension alone tells the right story.
Daily water by service (L / day, scope-1)
On-site cooling only. Scope-2 (off-site power generation) multiplies these 3–10×.
Total scope-1 water across all five services is ~1.7 m³/day — about 8 Australian households’ daily usage. The scope-2 picture below tells a much larger story.
Daily queries served (millions)
Volume rank doesn’t match energy rank — a small inefficient service can outweigh a larger efficient one.
ChatGPT serves 10× the queries of Claude, but their daily energy ratio is only 4:1 because Claude’s per-query energy is higher. Volume × efficiency × user growth jointly determine the footprint trajectory — three separate dimensions, not one.
Trajectory and the bigger picture
Annual energy trajectory — historical and projected (GWh / yr)
Three scenarios for 2027–2030. Conservative assumes efficiency improvements outpace user growth; steady assumes ~30% user growth offset by ~15% efficiency gains; reasoning revolution assumes broad migration to agentic and reasoning workloads.
The conservative scenario assumes Gemini-style efficiency gains keep outpacing user growth. The steady scenario tracks the IEA’s projection that total data-centre electricity roughly doubles by 2030. The reasoning-revolution scenario reaches 7 TWh/year by 2030 — comparable to Tasmania’s annual residential electricity demand. Which of these we get depends almost entirely on whether agentic and reasoning workloads stay niche or become default.
The full global AI inference picture (MWh / day)
The five chat services above are only ~47% of global AI chat-inference energy. Adding Gemini-powered AI Overviews in Google Search (2 B monthly users), Meta AI (1 B MAU), DeepSeek, Perplexity, and direct API/enterprise traffic roughly doubles the total.
The largest hidden line is Gemini-powered AI Overviews inside Google Search — most users don’t think of themselves as “AI users” when they Google, but Gemini is running on roughly half of all search results. API and enterprise traffic built on OpenAI/Anthropic/Google APIs is the next largest hidden category.
Daily water including scope-2 (m³ / day)
Including off-site power-generation water using Lawrence Berkeley Lab’s 7.6 L/kWh grid-water intensity factor.
Including off-site power-generation water flips the story. Daily water across the five services jumps from ~1.7 m³ (scope-1) to ~13 m³ (scope-2). Reporting just scope-1, as most providers do, understates true water footprint by an order of magnitude.
Bottom line
For an individual
Don’t optimise here first. The energy of a year’s worth of heavy AI use is dwarfed by a single short flight, a beef-heavy weekly diet, or running a tumble dryer instead of line-drying. If you want to reduce your digital carbon footprint, AI is among the smallest levers available — choose a Gemini-style efficient model over Grok if you can, but don’t lose sleep over the choice.
For an organisation
Track aggregate AI usage if you’re running large deployments, but optimise transport, heating, and grid procurement first — that’s where the real numbers live. For research and HPC environments, the case for self-hosting comes down to data sovereignty, fine-tuning, and predictable cost, not energy savings; the per-query footprint of a well-tuned on-prem deployment is comparable to commercial cloud.
For a policymaker or analyst
Focus on grid decarbonisation, mandatory disclosure (scope-2 water, real PUE, real WUE), and watching the reasoning-model transition. The IEA projection of 945 TWh global data-centre electricity by 2030 is the right frame — what fraction of that is AI inference is the question that matters.
What’s genuinely uncertain
Anthropic publishes nothing about Claude’s per-query footprint; OpenAI publishes only on-site water; Google publishes the most but on its own renewable-PPA accounting. External research (Jegham et al. 2025; Epoch AI) gives defensible estimates, but ±30–40% per-query variance is real. Daily query volumes for Gemini, Claude, Copilot, and Grok are external estimates from market-share data; ±50% uncertainty applies. Anything past 2026 in the trajectory chart is a scenario, not a forecast.
What this isn’t
This compares inference — the energy and water of running a query against an already-trained model. Training is excluded entirely. Training GPT-3 reportedly required ~1,287 MWh and ~700,000 L of water, but amortised across the model’s lifetime of queries it’s typically a small fraction of total footprint. Inference now dominates roughly 80–90% of an AI model’s lifetime impact.
Sources: Google Cloud Blog (2025) — Gemini self-reported energy, water, CO₂e. OpenAI (Altman, June 2025) — ChatGPT water per query; OpenAI to Axios (July 2025) — ChatGPT 2.5 B prompts/day. Jegham, Abdelatti, Koh, Elmoubarki & Hendawi (2025), How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference, arXiv:2505.09598. Epoch AI (2025) — GPT-4o energy range. IEA (2025) — global grid carbon intensity, data-centre electricity projections. Lawrence Berkeley National Lab — 7.6 L/kWh grid-water intensity for scope-2.