AI energy and water calculator

Start with the example below, then change an input. Results update automatically and carry into your report.

Your yearly estimate

25 text requests/day.

3.1 kWh
Electricity/year
1.2 kg CO₂e
Emissions/year
2.9 L
Text water/year

Rounded totals assume the same daily activity for 365 days. The report includes more decimal places. Energy and water may describe different evidence cases and must not be presented as the measured footprint of one system.

Choose your inputs

Energy and water use separate evidence cases. The grid factor converts electricity into estimated emissions.

Energy evidence case

Selected boundary: Sam Altman’s June 2025 reported average; model mix, task mix and accounting method were not disclosed. Not a current per-model measurement.

Response length (teaching multiplier)
Text requests per day 25
Water evidence case

Selected boundary: 0.000085 US gallons per query, reported by Sam Altman in June 2025. Accounting boundary and method were not disclosed; do not assume direct-only or life-cycle coverage.

Advanced: images, video & vibe coding Optional · off

Add only activity not counted in text requests above. Include retries and discarded generations. Use a daily average; zero excludes an activity.

Image source & limits
○ Study or modeled estimateimage.wh

Luccioni et al., FAccT 2024, Table 2: mean 2.907 kWh per 1,000 image generations = 2.907 Wh/image across tested open models.

Historical benchmark mean, not a current product average. Model, image size and sampling settings vary; the reported standard deviation is 3.31 Wh, not a confidence interval. Inference hardware energy only; no facility overhead, training or water added. Newer GPU-only ML.ENERGY v3 measurements of 1024 × 1024 images on one H100 run from 0.116 Wh (SANA 1.5 1.6B, 20 steps) to 2.208 Wh (FLUX.1-dev, 50 steps); they are separate cases and are not averaged into this value.

image.wh: Reported source value. Supports this numerical value

Open original source ↗Power Hungry Processing: Watts Driving the Cost of AI Deployment?luccioni-facct-2024 · Luccioni, Jernite & Strubell · ACM FAccT · 2024Locator: Table 2 image generation: mean 2.907 kWh/1,000 inferences, standard deviation 3.31; section 3 methodologyBoundary: CodeCarbon inference hardware energy for selected open models; excludes facility overhead, training and waterGeography: Test hardware and grid in paperReview status: peer-reviewed publication; AI Footprint Lab extraction not independently reviewedChecked: 2026-09-05 · review by 2027-03-05
Open original source ↗ML.ENERGY v3: text-to-image, text-to-video, chat and GPQA measurementsmlenergy-v3-tasks-2026 · ML.ENERGY research team · ML.ENERGY · 2026Locator: text-to-image.json, text-to-video.json, lm-arena-chat.json and gpqa.json at commit 5d899ea (unchanged through 2bdd007, 23 Sep 2026); H100 rows; energy_per_image_joules, energy_per_video_joules, energy_per_request_joulesBoundary: Steady-state GPU energy measured with Zeus; no CPU, memory, facility overhead or water. Reasoning comparison uses each model’s lowest-energy configuration, so prompt sets and batch settings differ.Geography: Laboratory hardware; grid selected separatelyReview status: Benchmark framework published at NeurIPS 2025 Datasets & Benchmarks; v3 dataset extraction not independently reviewedChecked: 2026-10-12 · review by 2027-04-12

Reference clip: about 6 seconds at 720 × 480, without audio. Other models and clip sizes can differ greatly.

Video source & limits
○ Study or modeled estimatevideo.wh

Delavande, Pierrard & Luccioni (NeurIPS 2025 workshop), Table 4, CogVideoX-5B: GPU 21.6 + CPU 2.4 + RAM 1.3 = 25.3 Wh/clip. One H100; 49 frames at 8 fps, 480 × 720, 50 steps; mean of 50 prompts × 5 runs after 2 warmups.

One open model at its default settings, not a product average. Energy grows roughly quadratically with resolution and frame count, so this value does not scale to other clips: the same table reports 0.14 Wh for AnimateDiff (16 frames, 512 × 512) and over 415 Wh for WAN2.1-T2V-14B (81 frames, 720 × 1280), and GPU-only ML.ENERGY v3 runs range from 7.95 Wh (CogVideoX-2b) to 323 Wh (HunyuanVideo). Excludes facility overhead, training and water. Does not estimate Sora, Veo, audio, upscaling or arbitrary clip lengths.

video.wh: Reported source value. Supports this CogVideoX-5B clip specification only; NeurIPS 2025 workshop paper, not a product average

Open original source ↗Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Modelsdelavande-video-neurips-2025 · Delavande, Pierrard & Luccioni · NeurIPS 2025 workshop paper (arXiv) · 2025Locator: Section 4.3 protocol (50 prompts, 2 warmups, 5 runs); Table 2 CogVideoX-5b defaults; Table 4 (PDF p. 9) CogVideoX-5B rowBoundary: CodeCarbon GPU (NVML) + CPU (pyRAPL) + RAM heuristic on one H100 SXM; 49 frames, 8 fps, 480 × 720, 50 steps; warmups excluded from the mean. No facility overhead, training or water.Geography: Laboratory hardware; selected grid applied separatelyReview status: Accepted as an oral at the NeurIPS 2025 NextVid workshop; not a journal article; AI Footprint Lab extraction not independently reviewedChecked: 2026-10-12 · review by 2027-04-12
Open original source ↗ML.ENERGY v3: text-to-image, text-to-video, chat and GPQA measurementsmlenergy-v3-tasks-2026 · ML.ENERGY research team · ML.ENERGY · 2026Locator: text-to-image.json, text-to-video.json, lm-arena-chat.json and gpqa.json at commit 5d899ea (unchanged through 2bdd007, 23 Sep 2026); H100 rows; energy_per_image_joules, energy_per_video_joules, energy_per_request_joulesBoundary: Steady-state GPU energy measured with Zeus; no CPU, memory, facility overhead or water. Reasoning comparison uses each model’s lowest-energy configuration, so prompt sets and batch settings differ.Geography: Laboratory hardware; grid selected separatelyReview status: Benchmark framework published at NeurIPS 2025 Datasets & Benchmarks; v3 dataset extraction not independently reviewedChecked: 2026-10-12 · review by 2027-04-12

Code-completion proxy, not a measured agent session. Excludes your computer, builds, tests and hosting.

Coding sources & calculation
○ Study or modeled estimatecodingCases.large.joulesPerToken

Sessions/day × output tokens/session × 19.601410817113816 J/token ÷ 3,600 = GPU Wh/day. Qwen3-Coder 480B FP8 · 4 B200s. Code completion, maximum batch 8, mean output about 261 tokens; GPU energy only.

Measured short code completions are used as a proxy for a larger session. Input processing is implicit in the benchmark workload; long-context and agent-tool overhead are not separately modeled. No facility overhead or water is added. No provider-specific agent estimate is claimed.

codingCases.large.joulesPerToken: Reported source value. Supports this code-completion configuration only; not a measured vibe coding session

Open original source ↗ML.ENERGY v3: Qwen3-Coder code-completion measurementsmlenergy-coding-2026 · ML.ENERGY research team · ML.ENERGY · 2026Locator: Qwen__Qwen3-Coder-30B-A3B-Instruct__sourcegraph-fim.json and Qwen__Qwen3-Coder-480B-A35B-Instruct-FP8__sourcegraph-fim.json; configurations with B200, max_num_seqs=8, num_gpus=1 and 4 respectively; energy_per_token_joulesBoundary: Steady-state GPU energy on Sourcegraph fill-in-the-middle code completion. GPU joules divided by generated tokens includes the benchmark input-processing workload; no separate input-token term. Session scaling is an authored extrapolation, not a measured coding agent.Geography: Laboratory hardware; grid selected separatelyReview status: Benchmark framework published at NeurIPS 2025; v3 dataset extraction not independently reviewedChecked: 2026-09-05 · review by 2027-03-05

Adds electricity and estimated emissions. Reported water remains text-only. These benchmarks cover different hardware boundaries; their sum is a scenario, not a complete footprint.

Choose a region for a stated scenario or a known serving location. Your location does not identify the server’s location. 2023 annual total-output electricity-generation CO₂e rate. Not hourly or marginal emissions, a provider contract, or a full life-cycle factor. Transmission losses are not added.

EPA table and units · Check the subregion map. Editing the number removes the regional attribution.

Grid carbon intensity 390 gCO₂e/kWh

Per text request

These values describe one text request. Your yearly total above also includes any images, video and coding you added.

0.34 Wh
○ Text-query energy
0.13 g CO₂e
◇ Emissions per text request
0.32 mL
○ Water per text query
Excluded or simplified: input tokens, model routing, reasoning effort, caching, batching, hardware utilization, training, model development, networking, chip manufacture, data-centre construction, e-waste, land use, and local watershed effects.
Everyday comparisons

These examples provide scale. AI serving electricity, vehicle operation, and food life cycles have different accounting boundaries. Similar CO₂e totals do not establish equal overall impacts or which activity is worthwhile.

Carbon volume comparisons

Each comparison uses grams of CO₂e. Equivalence does not imply equal social value or environmental context.

One text query: 0.13 g CO₂e

💡◇ 1/29 of an hour of LED light3.9 g each
🔋◇ 1/35 of a full phone charge4.7 g each
📺◇ 1/271 of a streaming hour36 g each
🧊○ 1/2,901 of a refrigerator-day385 g each
🚗◇ 1/3,017 of a driving mile400 g each
🍔◇ 1/22,624 of a beef burger3.0 kg each