AI impact calculator
Everyday use, a project or your own computer: a low, typical and high range for energy, water and CO₂e, with the source and confidence of each number. It runs in your browser, and what you enter is never sent.
Version 1.0 ·
A plug-in power meter gives the best reading: measure the machine during real work, then enter that number in watts.
Filled in with Hydro-Québec’s Rate D first tier. Replace it with the price on your bill.
This calculator runs in your browser. What you enter is never sent. Your entries are kept on this device only, and Reset erases them.
Your estimate
Enter a number above 0 to see a range.
Energy
Typical:
Confidence:
Water
Typical:
Confidence:
Greenhouse gases
Typical:
Confidence:
Electricity cost
In everyday terms
What is counted
What is not counted
How this calculator works
Coefficients as of November 2, 2026.
What the labels mean
- Published: a company or peer-reviewed paper states this figure, with the boundary we quote. Publishing a figure does not make it independent or audited.
- Estimated: we worked it out from published measurements of comparable systems, or from a stated assumption. It can be off by several times.
- Unknown: we found no usable figure, so we show no number.
The formulas
Everyday use energy (Wh) = prompts per day × days × Wh per prompt, added up for each type of use water (mL) and CO₂e (g), text = prompts per day × days × published mL or g per prompt water and CO₂e, image, video, research = energy × mL per Wh; energy ÷ 1000 × g per kWh A project energy (Wh) = (input tokens × Wh per input token + output tokens × Wh per output token) × model factor water and CO₂e = energy × mL per Wh; energy ÷ 1000 × g per kWh Your own computer energy (Wh) = watts × hours per day × days CO₂e (g) = energy ÷ 1000 × g per kWh of your grid cost ($) = energy ÷ 1000 × cents per kWh ÷ 100 Comparisons kettle boils = Wh ÷ 93.02; phone charges = Wh ÷ 19; glasses = mL ÷ 250; car km = g ÷ 244
Worked example: 20 prompts a day for a year at 0.24 Wh is 20 × 365 × 0.24 = 1,752 Wh, or 1.752 kWh. At 0.26 mL it is 20 × 365 × 0.26 = 1,898 mL, or 1.898 L.
A month counts as 365 ÷ 12 days. Low, typical and high are computed separately with the matching coefficients, and each range is a span of published or estimated values, not a statistical confidence interval.
Who has published what
| Who | Energy | Water | CO₂e | Status | Notes |
|---|---|---|---|---|---|
| 0.24 Wh | 0.26 mL | 0.03 g | Published | Google (Elsworth et al.) · Median Gemini Apps text prompt, May 2025, measured in production. Counting active chips only: 0.10 Wh, 0.12 mL, 0.02 g. | |
| OpenAI | 0.34 Wh | 0.32 mL | Unknown | Published | Sam Altman, OpenAI · An average query, stated in one sentence of a blog post by its chief executive, with no method, boundary or date of measurement. Water is given as 0.000085 gallons. |
| Mistral AI | Unknown | 45 mL | 1.14 g | Published | Mistral AI, with Carbone 4 and ADEME · Marginal impact of a 400-token Le Chat response, by a life-cycle method reviewed by outside firms. The company calls it a first approximation and gives no energy figure. |
| Microsoft | 0.31 Wh (0.16–0.60) | 0–0.067 mL | Unknown | Estimated | Microsoft Research (Oviedo et al.) · A modelled research estimate for large models in general, not a measurement of Copilot or any product. Water figure from the company’s blog, on-site cooling only. |
| Epoch AI | ≈ 0.3 Wh (0.1–4) | Unknown | Unknown | Estimated | Epoch AI (Josh You) · An independent calculation from assumed hardware, for one assumed model and 500 tokens. Not a company disclosure. |
| Anthropic, Meta, xAI, Amazon, Apple, DeepSeek, Alibaba | Unknown | Unknown | Unknown | Unknown | We found no per-prompt disclosure from these companies as of the date above. Unknown means not found, not low or high. |
Per prompt or per response. We present no figure for a specific model unless its maker published it, and we apply the same test to every company. The AI Energy Score ranks open models on GPU energy alone, so it is useful for comparing models but is not used as a coefficient here.
Why published figures differ by more than 100 times
Water shows it best. Google’s narrow figure is 0.12 mL per prompt and Mistral’s is 45 mL, a gap of about 375 times. Greenhouse gases differ by about 57 times: 0.02 g against 1.14 g. Much of that gap is what each one counts, and the two are different services, so neither figure proves the other wrong.
Boundaries. Counting only active chips gave Google 0.10 Wh; adding host servers, idle machines and cooling gave 0.24 Wh. Mistral’s life-cycle method adds upstream stages such as making the hardware, which on-site cooling figures leave out.
Median or mean. Google reports a median, the prompt in the middle, and says a few long prompts skew the mean. OpenAI says “average” without saying how it was computed. The two words can describe quite different numbers for the same service.
Market or location carbon. Google’s fleet emitted 345 g per kWh by location in 2024, but 94 g after counting its clean-energy purchases, a difference of about 3.7 times from accounting alone. Its 0.03 g figure uses the lower one.
Measured or modelled. Google measured its production fleet. Epoch AI and Microsoft’s researchers calculated from assumptions. Microsoft’s paper says high-profile estimates built on non-production assumptions overstate energy by 4 to 20 times.
Long inputs. Epoch AI estimates about 2.5 Wh for 10,000 input tokens and almost 40 Wh for 100,000, against about 0.3 Wh for a short question. A single word, “prompt”, covers all of them.
Why training is not per prompt
Training happens once, before anyone asks anything. To turn it into a per-prompt number, you would need the model’s total prompts over its whole life, and companies do not publish that. Divide by a billion or a hundred billion and the answer changes by 100 times.
So the calculator leaves training out and says so. Mistral reports 20.4 kilotonnes of CO₂e and 281,000 cubic metres of water for Mistral Large 2 training plus 18 months of use, a useful total, but not one that can be divided fairly by a single question.
Why location matters
The same kilowatt-hour carries about 2.6 g of CO₂e in Quebec and about 628 g in Nova Scotia in the 2024 Canadian tables, a gap of more than 200 times. That is why your own computer depends so heavily on the region you choose.
For cloud AI, you usually cannot tell which data centre answered you, so the calculator uses company-wide or world figures instead of your province. Water is also local: a litre drawn from a stressed basin is not the same as one drawn where water is plentiful, and per-prompt averages cannot show that.
Canadian figures leave out emissions from land flooded for hydro dams, and they are not on the same boundary as the Ember life-cycle figures for other regions.
How fast these numbers age
Quickly. Google reported a 33-fold drop in energy per median prompt in one year. The IEA says energy per AI task has fallen at least tenfold a year recently. New models also use more tokens, and video and agents use far more per request, so a lower figure per token does not mean a lower total.
Treat every figure here as dated. The date each source was read is in the tables above, and the date of the whole set is at the top of this section.
Every coefficient
| Coefficient | Low | Typical | High | Unit | Label | Boundary | Source, date read |
|---|---|---|---|---|---|---|---|
| Text prompt, energy | 0.1 | 0.24 | 0.6 | Wh/prompt | Published | Low: Google, AI chips only. Typical: Google median, May 2025, with chips, host servers, idle machines and cooling. High: top of the middle range in Microsoft’s research estimate for large models. | Google (Elsworth et al.) (read 2026-11-02) Microsoft Research (Oviedo et al.) (read 2026-11-02) |
| Text prompt, water | 0.12 | 0.26 | 45 | mL/prompt | Published | Low and typical: Google, water consumed for on-site cooling. High: Mistral, marginal impact of a 400-token response, by a life-cycle method including upstream stages such as hardware. | Google (Elsworth et al.) (read 2026-11-02) Mistral AI, with Carbone 4 and ADEME (read 2026-11-02) |
| Text prompt, CO₂e | 0.02 | 0.03 | 1.14 | gCO₂e/prompt | Published | Low and typical: Google, market-based, counting clean-power purchases, plus embodied hardware. High: Mistral, marginal impact of a 400-token response, by a life-cycle method. | Google (Elsworth et al.) (read 2026-11-02) Mistral AI, with Carbone 4 and ADEME (read 2026-11-02) |
| Image, energy | 0.63 | 2.907 | 11.49 | Wh/image | Estimated | Open models run by researchers, not commercial services. Low: Stable Diffusion 3 Medium, GPU energy doubled for overhead (2,282 J). Typical and high: mean and least efficient model in a 2024 benchmark. | MIT Technology Review (read 2026-11-02) Luccioni, Jernite and Strubell, ACM FAccT 2024 (read 2026-11-02) |
| Video clip, energy | 3.79 | 109.47 | 944 | Wh/clip | Estimated | Clips of about 5 seconds from open models. Low: LTX-Video. Typical: WAN2.1 14B. Both count GPU, CPU and memory, not the building. High: CogVideoX, 3.4 million joules. Commercial services publish nothing comparable. | Delavande, Hugging Face (read 2026-11-02) Delavande, Pierrard and Luccioni (read 2026-11-02) MIT Technology Review (read 2026-11-02) |
| Deep research task, energy | 4.32 | 24 | 240 | Wh/task | Estimated | No service publishes a figure. Low: one long reasoning query, research estimate. Typical and high: our reading of the IEA’s “hundreds or thousands of times” a simple query, as 100 and 1,000 × 0.24 Wh. | Microsoft Research (Oviedo et al.) (read 2026-11-02) International Energy Agency (read 2026-11-02) Google (Elsworth et al.) (read 2026-11-02) |
| Water per electricity | 0.36 | 1.15 | 5.67 | mL/Wh | Estimated | Water per unit of electricity, used wherever no water figure is published. Low: US average on-site, 2023. Typical: Google on-site, 2024. High: US on-site at Google’s rate plus 4.52 L/kWh for generating the electricity. | Lawrence Berkeley National Laboratory (read 2026-11-02) Google (Elsworth et al.) (read 2026-11-02) |
| Carbon per electricity | 94 | 345 | 458 | gCO₂e/kWh | Estimated | Carbon per unit of electricity, used wherever no carbon figure is published. Low: Google fleet, market-based, 2024. Typical: Google fleet before clean-power purchases. High: world average, life cycle, 2025. Hardware manufacture is not included. | Google (Elsworth et al.) (read 2026-11-02) Ember (read 2026-11-02) |
| Output token, energy | 0.00053 | 0.00103 | 0.002 | Wh/output token | Estimated | Our division: 0.16, 0.31 and 0.60 Wh per query for large models, over the study’s median of 300 output tokens. Full server and building, with the study’s assumptions. No company publishes energy per token. | Microsoft Research (Oviedo et al.) (read 2026-11-02) Microsoft Research (Oviedo et al.) (read 2026-11-02) |
| Input token, energy | 0.00005 | 0.00022 | 0.0004 | Wh/input token | Estimated | Typical and high: our division of Epoch AI’s estimates for 10,000 and 100,000 input tokens. Low: our assumption for efficient serving. Cached input is not counted. | Epoch AI (Josh You) (read 2026-11-02) |
| Small model factor | 0.3 | 0.3 | 0.3 | × large model | Estimated | Ratio only, from the EcoLogits fit for GPU energy per token: about 7 billion active parameters against 200 billion. | GenAI Impact (read 2026-11-02) Mistral AI, with Carbone 4 and ADEME (read 2026-11-02) |
| Medium model factor | 0.5 | 0.5 | 0.5 | × large model | Estimated | Ratio only, from the EcoLogits fit: about 70 billion active parameters against 200 billion. | GenAI Impact (read 2026-11-02) Mistral AI, with Carbone 4 and ADEME (read 2026-11-02) |
| Large model factor | 1 | 1 | 1 | × large model | Estimated | Reference class: models above 200 billion parameters, as in the Microsoft study. | Microsoft Research (Oviedo et al.) (read 2026-11-02) |
| Short response | 300 | 300 | 300 | output tokens | Published | Median output of a standard query in the Microsoft study, about 225 words. | Microsoft Research (Oviedo et al.) (read 2026-11-02) |
| Medium response | 1,000 | 1,000 | 1,000 | output tokens | Estimated | Our round midpoint, about 750 words. | Microsoft Research (Oviedo et al.) (read 2026-11-02) |
| Long response | 5,000 | 5,000 | 5,000 | output tokens | Published | Median output of a long reasoning query in the Microsoft study, about 3,750 words. | Microsoft Research (Oviedo et al.) (read 2026-11-02) |
| Short session | 20 | 20 | 20 | model calls | Estimated | Our estimate of a short agent or coding session. Each call: 2,000 new input tokens and 500 output tokens. | International Energy Agency (read 2026-11-02) |
| Working session | 100 | 100 | 100 | model calls | Estimated | Our estimate of a working session, same tokens per call. The result falls in the IEA’s “hundreds of times” a simple query. | International Energy Agency (read 2026-11-02) |
| Long session | 400 | 400 | 400 | model calls | Estimated | Our estimate of a long session, same tokens per call. | International Energy Agency (read 2026-11-02) |
| Session input | 2,000 | 2,000 | 2,000 | new input tokens per call | Estimated | Our assumption. Context already cached is not counted, since it is not processed again. | Epoch AI (Josh You) (read 2026-11-02) |
| Session output | 500 | 500 | 500 | output tokens per call | Estimated | Our assumption, matching Epoch AI’s typical response length. | Epoch AI (Josh You) (read 2026-11-02) |
| Kettle boil | 93.02 | 93.02 | 93.02 | Wh/boil | Published | Physics: 1 kg × 4.186 kJ/(kg·K) × 80 K = 334.9 kJ, 1 litre from 20 to 100 °C, with no heat lost. A real kettle uses somewhat more. | NIST Chemistry WebBook (read 2026-11-02) |
| Phone charge | 19 | 19 | 19 | Wh/charge | Published | EPA’s energy for one full smartphone charge. | U.S. Environmental Protection Agency (read 2026-11-02) |
| Glass of water | 250 | 250 | 250 | mL/glass | Estimated | A common glass of water, by definition here. | Our assumption |
| Gasoline car | 244 | 244 | 244 | gCO₂e/km | Published | EPA’s average US gasoline passenger vehicle: 393 g per mile, 1.609 km to the mile. | U.S. Environmental Protection Agency (read 2026-11-02) |
| Range around your watts | 0.8 | 1 | 1.2 | × watts entered | Estimated | Our allowance for how draw moves during a session. A plug-in power meter narrows it. | Our assumption |
| Electricity price | 7.065 | 7.065 | 7.065 | ¢/kWh | Published | Default only: Hydro-Québec Rate D, first 40 kWh a day on average, April 1, 2026. Enter your own price. | Hydro-Québec (read 2026-11-02) |
Machine presets
| Machine | Low (W) | Typical (W) | High (W) | Boundary | Source, date read |
|---|---|---|---|---|---|
| Laptop | 15 | 30 | 65 | Our round estimate at the wall under load. | Our assumption |
| Compact desktop (Mac mini M4 type) | 4 | 30 | 65 | Low and high: Apple’s idle and maximum at the wall. Typical: our estimate. | Apple (read 2026-11-02) |
| Desktop with a 450 W graphics card (RTX 4090 type) | 200 | 550 | 850 | Typical: NVIDIA’s card power plus our 100 W for the rest. High: NVIDIA’s required system power, a ceiling. | NVIDIA (read 2026-11-02) |
| Desktop with a 575 W graphics card (RTX 5090 type) | 250 | 700 | 1000 | Typical: NVIDIA’s card power plus our 125 W for the rest. High: NVIDIA’s required system power, a ceiling. | NVIDIA (read 2026-11-02) |
Grid intensity by region
In g CO₂e per kWh. Canadian tables come from Environment and Climate Change Canada’s national inventory. Figures for 2024 are preliminary.
| Region | Low | Typical | High | Source, date read |
|---|---|---|---|---|
| Quebec | 1.6 | 2.6 | 2.6 | Environment and Climate Change Canada (read 2026-11-02) |
| Ontario | 42.6 | 73.3 | 73.3 | Environment and Climate Change Canada (read 2026-11-02) |
| British Columbia | 15.7 | 18.1 | 18.1 | Environment and Climate Change Canada (read 2026-11-02) |
| Alberta | 346.1 | 346.1 | 521.5 | Environment and Climate Change Canada (read 2026-11-02) |
| Manitoba | 1.4 | 3 | 3 | Environment and Climate Change Canada (read 2026-11-02) |
| Saskatchewan | 580.7 | 580.7 | 700.5 | Environment and Climate Change Canada (read 2026-11-02) |
| Nova Scotia | 613.8 | 628.3 | 702 | Environment and Climate Change Canada (read 2026-11-02) |
| New Brunswick | 234 | 375.5 | 375.5 | Environment and Climate Change Canada (read 2026-11-02) |
| Prince Edward Island | 176 | 264.6 | 264.6 | Environment and Climate Change Canada (read 2026-11-02) |
| Newfoundland and Labrador | 16.7 | 17 | 17.9 | Environment and Climate Change Canada (read 2026-11-02) |
| Yukon | 74 | 133.6 | 133.6 | Environment and Climate Change Canada (read 2026-11-02) |
| Northwest Territories | 180 | 490 | 490 | Environment and Climate Change Canada (read 2026-11-02) |
| Nunavut | 760 | 760 | 830 | Environment and Climate Change Canada (read 2026-11-02) |
| Canada, average | 106 | 110.6 | 115.2 | Environment and Climate Change Canada (read 2026-11-02) |
| United States, average | 383.8 | 384.4 | 392.9 | Ember (read 2026-11-02) |
| European Union, average | 209.9 | 209.9 | 235.8 | Ember (read 2026-11-02) |
| World average | 458.3 | 458.3 | 483.2 | Ember (read 2026-11-02) |
Consumption intensity, g CO2e per kWh consumed. Typical: 2024, preliminary. Low and high: lowest and highest of 2021 to 2024. Counts generation emissions and losses. Does not count flooded land behind hydro dams or upstream fuel supply.
Life-cycle intensity, g CO2e per kWh generated, so not directly comparable with Canadian figures. Typical: 2025. Low and high: lowest and highest of 2023 to 2025.
Related articles
- How much energy and water does using AI really take?
- What are the environmental costs of AI?
- How does an AI data center actually work?
- How do AI data centers stay cool, and how much water does that really use?
All sources
- Google (Elsworth et al.) (read 2026-11-02): Measuring the environmental impact of delivering AI at Google Scale
- Sam Altman, OpenAI (read 2026-11-02): The Gentle Singularity
- Mistral AI, with Carbone 4 and ADEME (read 2026-11-02): Our contribution to a global environmental standard for AI
- Microsoft Research (Oviedo et al.) (read 2026-11-02): Energy use of AI inference, efficiency pathways, and test-time scaling (Joule)
- Microsoft Research (Oviedo et al.) (read 2026-11-02): Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute (preprint)
- Microsoft (read 2026-11-02): Scaling AI with 8 to 20x energy efficiency
- Epoch AI (Josh You) (read 2026-11-02): How much energy does ChatGPT use?
- GenAI Impact (read 2026-11-02): EcoLogits methodology: LLM inference
- Hugging Face and partners (read 2026-11-02): AI Energy Score
- International Energy Agency (read 2026-11-02): Key Questions on Energy and AI: Executive summary
- Luccioni, Jernite and Strubell, ACM FAccT 2024 (read 2026-11-02): Power Hungry Processing: Watts Driving the Cost of AI Deployment?
- MIT Technology Review (read 2026-11-02): We did the math on AI’s energy footprint
- Delavande, Hugging Face (read 2026-11-02): How Much Power does a SOTA Open Video Model Use?
- Delavande, Pierrard and Luccioni (read 2026-11-02): Video Killed the Energy Budget
- Lawrence Berkeley National Laboratory (read 2026-11-02): 2024 United States Data Center Energy Usage Report
- Environment and Climate Change Canada (read 2026-11-02): National Inventory Report 1990–2024, Annex 7: Electricity in Canada (intensity tables A7-1 to A7-14)
- Ember (read 2026-11-02): Lifecycle carbon intensity of electricity (Ember data, via Our World in Data)
- U.S. Environmental Protection Agency (read 2026-11-02): Greenhouse Gas Equivalencies Calculator: calculations and references
- Hydro-Québec (read 2026-11-02): Electricity Rates, effective April 1, 2026
- Apple (read 2026-11-02): Mac mini power consumption and thermal output
- NVIDIA (read 2026-11-02): Compare GeForce graphics cards
- NIST Chemistry WebBook (read 2026-11-02): Water: condensed phase thermochemistry data (specific heat of liquid water, about 75.3 J/(mol·K), or 4.18 kJ/(kg·K))