Sources & method

Inspect how each number is used, then follow it to the original evidence.

Review status

AI Footprint Lab is evidence-led but not yet independently validated. Use it as a draft teaching tool and source-audit exercise—not as a primary scientific authority.

ImplementedLinked evidence registryPublished equations and exclusionsCalculation/provenance tests
PendingEnergy/LCA specialist reviewHydrology/water-footprint reviewMulti-teacher classroom pilot

Method 1.7.0

Dataset release 2026.10.12.2; oldest recorded individual check 2026-09-05. Individual records retain their own check dates. Source metadata is licensed CC BY 4.0; source documents retain their own licenses.

Text-query energy

Wh/query = selected evidence case × response sensitivity

The response multipliers are AI Footprint Lab-authored sensitivity inputs. They are not measured universal factors.

Location-based carbon

gCO₂e = Wh ÷ 1,000 × grid gCO₂e/kWh

The serving location is unknown, so the user must choose a grid factor.

Water volume

annual mL = selected mL/query × request count

Water is not multiplied by response length or energy. No water is added for images or video.

Outside the calculator boundary: input-token effects, model routing, reasoning effort, caching, batching, utilization, training, model development, networks, chip manufacture, construction, e-waste, land use, and local ecological effects.

How we choose and use sources

  • Source selection: prefer original studies, methods and official datasets. Label first-party company disclosures and the scope of any assurance. A peer-reviewed venue does not independently validate AI Footprint Lab’s extraction.
  • Exact claims: a reference must substantiate the particular value, unit and boundary. Background citations do not turn authored inputs into measured facts.
  • Disagreement: compare methods, years, systems and uncertainty before describing sources as conflicting. Give weight according to evidence quality; do not manufacture equal credibility for unsupported claims.
  • Interpretation: separate reported evidence, arithmetic, inference and values. Students may reach different recommendations if their evidence and priorities are explicit.
  • Examples: everyday equivalents are optional scale illustrations with different boundaries. They do not determine whether an activity is worthwhile. No virtue scores or company rankings are assigned.
  • Missing information: not disclosed and not estimated never mean zero. Environmental costs and claimed benefits both require evidence and a defined comparison.
  • Accountability: publish accepted corrections, rejected proposals with reasons, conflicts, review scope and unresolved questions. Source updates must describe numerical and boundary implications.

Review framework: NAAEE Environmental Education Materials: Guidelines for Excellence. This reference is not an endorsement of AI Footprint Lab.

Numerical factor ledger

Reported values identify an original number and locator. Authored inputs have no exact empirical derivation asserted; their linked sources provide background only. Independent review is pending for every row. User-selected grid and request counts override defaults in the report. The download also includes the regional grid factors and model benchmark records with their separate boundaries.

grid.value

390 gCO₂e/kWh · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 390; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
390 gCO₂e/kWh is an adjustable, rounded teaching default informed by EPA eGRID—not a measurement of the data centre serving your request.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

eGRID 2023 Summary Data, Revision 2

scenarios.chatgpt.wh

0.34 Wh · Reported source value

Derivation and source links
Original
0.34 Wh per reported average ChatGPT query
Conversion
0.34 Wh × 1
Locator
Paragraph beginning ‘As datacenter production gets automated’ (June 2025)
Boundary
Sam Altman’s June 2025 reported average; model mix, task mix and accounting method were not disclosed. Not a current per-model measurement.
Evidence role
First-party statement; no published measurement method
Review
Source extraction requires independent review

The Gentle Singularity — ChatGPT resource-use statement

scenarios.measured.wh

0.24 Wh · Reported source value

Derivation and source links
Original
0.24 Wh per median Gemini Apps text prompt
Conversion
0.24 Wh × 1 = 0.24 Wh
Locator
Abstract: median text prompt energy
Boundary
Google's median Gemini Apps text prompt; not a generic AI request
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Measuring the environmental impact of delivering AI at Google Scale

scenarios.microsoftStudy.wh

0.31 Wh · Reported source value

Derivation and source links
Original
0.31 Wh median per query (IQR 0.16–0.60)
Conversion
0.31 Wh × 1 = 0.31 Wh
Locator
v2 Summary and section 2 (Equation 1; 8×H100 node power; PUE sampled 1.05–1.40; 500 input tokens; median 300 or 5,000 output tokens)
Boundary
Modeled median by Microsoft researchers (Joule, 2026) for three open models above 200B parameters on busy, batched 8×H100 servers, including whole-server power and PUE, for answers of about 300 output tokens. Not a measurement of any product, provider or account.
Evidence role
Peer-reviewed modeled estimate for three open models; not a measurement of any product or provider telemetry
Review
Source extraction requires independent review

Energy use of AI inference, efficiency pathways, and test-time scaling

scenarios.modeled.wh

1 Wh · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 1; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
An AI Footprint Lab-authored round number for sensitivity analysis; not a published consensus
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

EcoLogits: Evaluating the Environmental Impacts of Generative AI

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference

scenarios.publishedHigh.wh

4 Wh · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 4; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
An AI Footprint Lab-authored sensitivity value, not an exact published model/task result; not a worst case and not suitable for image, video, reasoning, or agentic work
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

EcoLogits: Evaluating the Environmental Impacts of Generative AI

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference

waterCases.chatgpt.ml

0.32176000164 mL · Reported source value

Derivation and source links
Original
0.000085 US gallons per reported average ChatGPT query
Conversion
0.000085 × 3785.411784 = 0.32176000164 mL
Locator
Paragraph beginning ‘As datacenter production gets automated’ (June 2025)
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.
Evidence role
First-party statement; water boundary unspecified
Review
Source extraction requires independent review

The Gentle Singularity — ChatGPT resource-use statement

waterCases.geminiDirect.ml

0.26 mL · Reported source value

Derivation and source links
Original
0.26 mL per median Gemini Apps text prompt
Conversion
0.26 mL × 1 = 0.26 mL
Locator
Section 3.3 (water allocation), section 3.4 (aggregation), and Table 2 (0.26 mL reported result)
Boundary
Google’s allocated direct cooling-water estimate for its median Gemini Apps text prompt, calculated from IT serving energy and annual fleet-average water-use effectiveness; excludes electricity-supply and broader life-cycle water.
Evidence role
Reported allocated estimate: IT energy × annual fleet-average consumptive WUE; not a prompt-level meter reading
Review
Source extraction requires independent review

Measuring the environmental impact of delivering AI at Google Scale

waterCases.teaching10.ml

10 mL · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 10; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
AI Footprint Lab-authored sensitivity value. It is not a measured generic AI-water factor and is kept separate from the energy case.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Measuring the environmental impact of delivering AI at Google Scale

EcoLogits: Evaluating the Environmental Impacts of Generative AI

lengths.options.0.mult

0.4 × · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 0.4; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
The model omits input length, reasoning effort, caching, batching, model routing, hardware, and utilization. Long-input/short-output tasks may not follow this simple output-length sensitivity model.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Power Hungry Processing: Watts Driving the Cost of AI Deployment?

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference

lengths.options.1.mult

1 × · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 1; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
The model omits input length, reasoning effort, caching, batching, model routing, hardware, and utilization. Long-input/short-output tasks may not follow this simple output-length sensitivity model.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Power Hungry Processing: Watts Driving the Cost of AI Deployment?

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference

lengths.options.2.mult

2.5 × · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 2.5; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
The model omits input length, reasoning effort, caching, batching, model routing, hardware, and utilization. Long-input/short-output tasks may not follow this simple output-length sensitivity model.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Power Hungry Processing: Watts Driving the Cost of AI Deployment?

How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference

image.wh

2.907 Wh · Reported source value

Derivation and source links
Original
2.907 kWh per 1,000 images
Conversion
2.907 × 1000 Wh/kWh ÷ 1000 images = 2.907 Wh/image
Locator
Table 2 image generation: mean 2.907 kWh/1,000 inferences, standard deviation 3.31; section 3 methodology
Boundary
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.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Power Hungry Processing: Watts Driving the Cost of AI Deployment?

ML.ENERGY v3: text-to-image, text-to-video, chat and GPQA measurements

video.wh

25.3 Wh · Reported source value

Derivation and source links
Original
GPU 21.6 + CPU 2.4 + RAM 1.3 Wh/clip (mean of 50 prompts × 5 runs)
Conversion
Sum of reported components = 25.3 Wh/clip
Locator
Section 4.3 protocol (50 prompts, 2 warmups, 5 runs); Table 2 CogVideoX-5b defaults; Table 4 (PDF p. 9) CogVideoX-5B row
Boundary
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.
Evidence role
Supports this CogVideoX-5B clip specification only; NeurIPS 2025 workshop paper, not a product average
Review
Source extraction requires independent review

Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models

codingCases.small.joulesPerToken

0.8442158286231319 J/output token · Reported source value

Derivation and source links
Original
0.8442158286231319 J/output token
Conversion
Generated tokens × J/token ÷ 3600 = GPU Wh; scaling sessions is an authored approximation
Locator
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_joules
Boundary
Code completion, maximum batch 8, mean output about 250 tokens; GPU energy only.
Evidence role
Supports this code-completion configuration only; not a measured vibe coding session
Review
Source extraction requires independent review

ML.ENERGY v3: Qwen3-Coder code-completion measurements

codingCases.large.joulesPerToken

19.601410817113816 J/output token · Reported source value

Derivation and source links
Original
19.601410817113816 J/output token
Conversion
Generated tokens × J/token ÷ 3600 = GPU Wh; scaling sessions is an authored approximation
Locator
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_joules
Boundary
Code completion, maximum batch 8, mean output about 261 tokens; GPU energy only.
Evidence role
Supports this code-completion configuration only; not a measured vibe coding session
Review
Source extraction requires independent review

ML.ENERGY v3: Qwen3-Coder code-completion measurements

mile.g

400 gCO₂e · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 400; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
Rounded US average gasoline passenger-vehicle carbon equivalency; fleet and driving conditions vary.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Greenhouse Gas Equivalencies Calculator: Calculations and References

burgerG.g

3000 gCO₂e · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 3000; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
Illustrative life-cycle carbon volume for a quarter-pound beef patty; production systems vary widely.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Reducing food's environmental impacts through producers and consumers

stream.g

36 gCO₂e · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 36; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
36 g/hour is retained as an illustrative input pending exact extraction review of the IEA 2019-condition estimate. Device, network, and grid choices can shift the result substantially.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

The carbon footprint of streaming video: fact-checking the headlines

fridge.whDay

986.3013698630137 Wh/day · Reported source value

Derivation and source links
Original
About 360 kWh/year for a certified top-freezer refrigerator
Conversion
360 × 1000 ÷ 365 Wh/day
Locator
PDF page 14, Refrigerator & Freezer Configurations
Boundary
ENERGY STAR’s 2024 guide: about 360 kWh/year for a certified top-freezer refrigerator, divided by 365. A representative efficient type, not an average of all refrigerators.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

2024 Quick Reference Guide — Refrigerators

phone.wh

12 Wh · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 12; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
Order-of-magnitude full-charge comparison; battery capacity and charging losses vary.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Greenhouse Gas Equivalencies Calculator: Calculations and References

bulb.w

10 W · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 10; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
Declared 10 W appliance assumption.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

home.kwhYr

10356 kWh/year · Reported source value

Derivation and source links
Original
863 kWh/customer/month in 2024
Conversion
863 × 12 = 10,356 kWh/customer/year
Locator
U.S. Total row: 143,144,185 customers; average monthly consumption 863 kWh
Boundary
EIA 2024 (Table T5.a): 863 kWh/month average consumption per US residential customer × 12 = 10,356 kWh/year. Grid purchases per customer account; homes with rooftop solar can use more. National average; actual household use varies.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

2024 Average Monthly Bill – Residential (Table T5.a)

golf.dayL

232840.818101436 L/day · Reported source value

Derivation and source links
Original
68.9 acre-feet/year median applied water per US facility in 2024
Conversion
68.9 × 1233481.83754752 L/acre-foot ÷ 365 days (international foot)
Locator
PDF page 8 National Water Use and Table 2: 2024 median
Boundary
GCSAA 2024 US facility median: 68.9 acre-feet/year, converted to liters and divided by 365. Annualized daily volume, not a growing-season day or an 18-hole-only average. Includes recycled irrigation water; applied water is not all consumed.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Golf Course Environmental Profile, Phase IV — Water Use and Management

shower.lpm

9.46352946 L/min · Reported source value

Derivation and source links
Original
2.5 US gallons/minute standard showerhead
Conversion
2.5 × 3.785411784 L/US gallon
Locator
EPA Showerheads: About WaterSense
Boundary
EPA standard showerhead: 2.5 US gallons/minute, converted to liters. Delivered water, not consumptive water; heating energy is separate.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Showerheads

aircon.kwhYr

612 kWh/year · Reported source value

Derivation and source links
Original
612 kWh/year
Conversion
Source Table 1, ENERGY STAR column; no conversion
Locator
Table 1, ENERGY STAR column; Energy Use and Cost Assumptions
Boundary
DOE Table 1: a 12,000 Btu/h ENERGY STAR room AC uses 612 kWh over 750 operating hours/year. Not whole-home central AC; actual climate and usage vary.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Purchasing Energy-Efficient Room Air Conditioners

aircon.hoursYr

750 hours/year · Reported source value

Derivation and source links
Original
750 operating hours/year
Conversion
Source calculation assumption; no conversion
Locator
Table 1, ENERGY STAR column; Energy Use and Cost Assumptions
Boundary
DOE Table 1: a 12,000 Btu/h ENERGY STAR room AC uses 612 kWh over 750 operating hours/year. Not whole-home central AC; actual climate and usage vary.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Purchasing Energy-Efficient Room Air Conditioners

electricityWater.litersPerKwh

1 L/kWh · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 1; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
An adjustable 1 L/kWh electricity-generation water-consumption illustration. NREL documents variation by technology and cooling; it does not establish this number as a grid average. Applied to electricity on both sides; excludes on-site AI cooling and all other water.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

A Review of Operational Water Consumption and Withdrawal Factors for Electricity Generating Technologies

toilet.gallonsFlush

1.28 US gal/flush · Reported source value

Derivation and source links
Original
1.28 US gallons/flush maximum
Conversion
1.28 × 3.785411784 = 4.84532708352 L/flush
Locator
Flush Facts and specification: WaterSense maximum 1.28 US gallons/flush
Boundary
WaterSense 1.28 US gallons/flush certification limit, converted to 4.845 L. Five flushes/day is an editable authored frequency; this is delivered water, not consumptive use. Supply and wastewater energy are excluded.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Residential Toilets

toilet.flushesDay

5 flushes/day · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 5; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
WaterSense 1.28 US gallons/flush certification limit, converted to 4.845 L. Five flushes/day is an editable authored frequency; this is delivered water, not consumptive use. Supply and wastewater energy are excluded.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Residential Toilets

visualBottle.ml

500 mL · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 500; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
A chosen 500 mL container for illustrating water volume only. Bottle manufacture, filling and transport are excluded.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

driving.milesDay

30 miles/day · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 30; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
Editable examples: 30 miles/day, 30 kWh/100 miles from the EV charging supply, and 30 miles/US gallon. These are not measured fleet averages. EV emissions use the selected grid; gas emissions use tailpipe CO2. Vehicle/battery manufacture, fuel production and infrastructure are excluded.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Learn More About the Fuel Economy Label for Electric Vehicles

Greenhouse Gas Emissions from a Typical Passenger Vehicle

driving.evKwh100Miles

30 kWh/100 miles · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 30; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
Editable examples: 30 miles/day, 30 kWh/100 miles from the EV charging supply, and 30 miles/US gallon. These are not measured fleet averages. EV emissions use the selected grid; gas emissions use tailpipe CO2. Vehicle/battery manufacture, fuel production and infrastructure are excluded.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Learn More About the Fuel Economy Label for Electric Vehicles

Greenhouse Gas Emissions from a Typical Passenger Vehicle

driving.gasMpg

30 miles/US gal · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 30; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
Editable examples: 30 miles/day, 30 kWh/100 miles from the EV charging supply, and 30 miles/US gallon. These are not measured fleet averages. EV emissions use the selected grid; gas emissions use tailpipe CO2. Vehicle/battery manufacture, fuel production and infrastructure are excluded.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Learn More About the Fuel Economy Label for Electric Vehicles

Greenhouse Gas Emissions from a Typical Passenger Vehicle

gasoline.co2Gallon

8887 g CO₂/US gal · Reported source value

Derivation and source links
Original
8,887 g CO2/US gallon
Conversion
Gallons burned × 8887 = tailpipe CO2 in grams
Locator
How much tailpipe carbon dioxide is created from burning one gallon: 8,887 grams CO2/US gallon
Boundary
EPA gasoline combustion: 8,887 g CO2/US gallon. Tailpipe CO2 only; not full life-cycle CO2e.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Greenhouse Gas Emissions from a Typical Passenger Vehicle

householdSize.people

2.53 people · Reported source value

Derivation and source links
Original
2.53 persons/household
Conversion
ACS 2020–2024 US average; no conversion
Locator
See source record exact locator
Boundary
Census ACS 2020–2024 US average: 2.53 people per household. Editable; household membership is not AI participation.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

QuickFacts: United States

domesticWater.gallonsDay

82 US gal/person/day · Reported source value

Derivation and source links
Original
82 US gallons/person/day
Conversion
Multiply by 3.785411784 for L/person/day
Locator
See source record exact locator
Boundary
EPA cites USGS 2015: 82 US gallons of domestic water per person/day. Delivered water, not all consumed.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Statistics and Facts: Household water

dishwasher.kwhYr

240 kWh/year · Reported source value

Derivation and source links
Original
240 kWh/year
Conversion
Certification example including electric water heating
Locator
See source record exact locator
Boundary
Certification-ceiling example: 240 kWh over 215 annual cycles, including electric water heating; 3.2 US gallons/cycle. Actual models and cycles vary.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Purchasing Energy-Efficient Residential Dishwashers

dishwasher.cyclesYear

215 cycles/year · Reported source value

Derivation and source links
Original
215 cycles/year
Conversion
DOE annual test-use assumption
Locator
See source record exact locator
Boundary
Certification-ceiling example: 240 kWh over 215 annual cycles, including electric water heating; 3.2 US gallons/cycle. Actual models and cycles vary.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Purchasing Energy-Efficient Residential Dishwashers

dishwasher.gallonsCycle

3.2 US gal/cycle · Reported source value

Derivation and source links
Original
3.2 US gallons/cycle maximum
Conversion
Multiply by 3.785411784 for L/cycle
Locator
See source record exact locator
Boundary
Certification-ceiling example: 240 kWh over 215 annual cycles, including electric water heating; 3.2 US gallons/cycle. Actual models and cycles vary.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Dishwashers Key Product Criteria

laundry.kwhCycle

0.5 kWh/cycle · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 0.5; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
14 US gallons/load is published ENERGY STAR context. 0.5 kWh/wash and 208 washes/year (4/week × 52) are authored sensitivity inputs. Includes assumed electric washing/water heating; dryer excluded.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Clothes Washers

laundry.gallonsCycle

14 US gal/cycle · Reported source value

Derivation and source links
Original
14 US gallons/load
Conversion
Multiply by 3.785411784 for L/load
Locator
See source record exact locator
Boundary
14 US gallons/load is published ENERGY STAR context. 0.5 kWh/wash and 208 washes/year (4/week × 52) are authored sensitivity inputs. Includes assumed electric washing/water heating; dryer excluded.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Clothes Washers

laundry.cyclesYear

208 cycles/year · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 208; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
14 US gallons/load is published ENERGY STAR context. 0.5 kWh/wash and 208 washes/year (4/week × 52) are authored sensitivity inputs. Includes assumed electric washing/water heating; dryer excluded.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

Clothes Washers

school.kwhSqft

9.395874623797988 kWh/ft²/year · Reported source value

Derivation and source links
Original
128 billion kWh / 13,623 million ft²
Conversion
128000 / 13623 = approximately 9.396 kWh/ft²/year
Locator
See source record exact locator
Boundary
128 billion kWh / 13,623 million ft² gives about 9.396 kWh/ft²/year for education buildings (EIA 2018). A proxy including colleges; not a K–12-only average. 50,000 ft² is an editable example.
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

CBECS 2018 Table C13: Electricity consumption

school.sqft

50000 ft² · AI Footprint Lab-authored teaching input

Derivation and source links
Original
No exact source extraction asserted
Conversion
Author-selected value: 50000; no empirical conversion asserted.
Locator
Not applicable to an authored input
Boundary
128 billion kWh / 13,623 million ft² gives about 9.396 kWh/ft²/year for education buildings (EIA 2018). A proxy including colleges; not a K–12-only average. 50,000 ft² is an editable example.
Evidence role
Background context only; does not substantiate this exact value
Review
Exact empirical derivation not established; retain as an assumption

CBECS 2018 Table C13: Electricity consumption

dcEnergy.twh2025

485 TWh · Reported source value

Derivation and source links
Original
485 TWh in 2025
Conversion
485 TWh × 1 = 485 TWh
Locator
Executive summary: data-centre electricity consumption in 2025
Boundary
All global data centres; not AI alone
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Key Questions on Energy and AI

dcEnergy.twh2030

950 TWh · Reported source value

Derivation and source links
Original
About 950 TWh in 2030
Conversion
950 TWh × 1 = 950 TWh
Locator
Executive summary: 2030 electricity outlook
Boundary
All global data centres; modeled central projection, not a measured outcome
Evidence role
Supports this numerical value
Review
Source extraction requires independent review

Key Questions on Energy and AI

Source review queue

Counting review dates against today’s date. These counts are measured in your browser, never baked into the page.

Evidence registry

Every registered source has a stable ID, direct link, locator, boundary, geography, review status, verification date, and next-review date.

delavande-video-2025How Much Power does a SOTA Open Video Model Use? ↗

Julien Delavande · Hugging Face community blog · 2025

Historical source retained for archived scenarios; not used by current calculator values.

Source details and review dates
Supports
21.71 + 2.91 + 1.31 = 25.93 Wh for this benchmark clip only
Locator
Experimental Setup and Energy Use tables: CogVideoX-5b row
Boundary
CodeCarbon GPU + CPU + RAM with amortized warmup; H100, 49 frames, 8 fps, 480 × 720, 50 steps. No facility overhead or water.
Geography
Serving location not established; selected grid applied separately
Review status
Exploratory benchmark; not peer reviewed; article has inconsistent run counts
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
mlenergy-coding-2026ML.ENERGY v3: Qwen3-Coder code-completion measurements ↗

ML.ENERGY research team · ML.ENERGY · 2026

Source details and review dates
Supports
0.8442158286231319 and 19.601410817113816 J/output token for two distinct configurations. Does not establish typical sessions, long-context or proprietary agent energy.
Locator
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_joules
Boundary
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 separately
Review status
Benchmark framework published at NeurIPS 2025; v3 dataset extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
epa-toilets-2026Residential Toilets ↗

US EPA WaterSense · US EPA · 2026

Source details and review dates
Supports
1.28 US gallons/flush reference; frequency remains an authored input
Locator
Flush Facts and specification: WaterSense maximum 1.28 US gallons/flush
Boundary
Fixture delivered-water volume; certification limit, not a measured average or consumptive footprint. Supply and wastewater energy excluded.
Geography
United States
Review status
Government guidance; extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
epa-gasoline-co2-2026Greenhouse Gas Emissions from a Typical Passenger Vehicle ↗

US EPA · US EPA · 2026

Source details and review dates
Supports
8,887 g CO2/US gallon; MPG and daily distance are authored examples
Locator
How much tailpipe carbon dioxide is created from burning one gallon: 8,887 grams CO2/US gallon
Boundary
Gasoline tailpipe CO2 only. Excludes other exhaust GHGs, fuel production, vehicle manufacture and infrastructure.
Geography
United States
Review status
Government guidance; extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
doe-ev-labelLearn More About the Fuel Economy Label for Electric Vehicles ↗

US DOE and EPA · FuelEconomy.gov · 2026

Source details and review dates
Supports
Units and interpretation; 30 kWh/100 miles is an authored example, not a fleet average
Locator
Section 5: fuel consumption in kWh/100 miles
Boundary
EV charging electricity and selected-grid generation emissions. No vehicle/battery manufacturing, infrastructure or full life-cycle estimate.
Geography
United States
Review status
Government guidance; extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
census-population-2025Vintage 2025 city and town population estimates ↗

US Census Bureau · US Census Bureau · 2026

Source details and review dates
Supports
SUMLEV 162 whole incorporated places and functioning SUMLEV 061 subdivisions; POPESTIMATE2025
Locator
SUMLEV 162 whole incorporated places and functioning SUMLEV 061 subdivisions; POPESTIMATE2025
Boundary
July 1, 2025 resident population estimates; overlapping geographies are not summed
Geography
US cities and towns
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
census-states-2025Vintage 2025 state population estimates ↗

US Census Bureau · US Census Bureau · 2026

Source details and review dates
Supports
SUMLEV 040; POPESTIMATE2025
Locator
SUMLEV 040; POPESTIMATE2025
Boundary
July 1, 2025 population estimates; states, DC and Puerto Rico
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
census-gazetteer-20252025 Gazetteer Files ↗

US Census Bureau · US Census Bureau · 2025

Source details and review dates
Supports
Places, county subdivisions and ZIP Code Tabulation Areas; INTPTLAT/INTPTLONG
Locator
Places, county subdivisions and ZIP Code Tabulation Areas; INTPTLAT/INTPTLONG
Boundary
Internal-point proximity for suggestions only; ZCTAs are not city or postal boundaries
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
census-household-size-2024QuickFacts: United States ↗

US Census Bureau · US Census Bureau · 2025

Source details and review dates
Supports
Persons per household, 2020–2024: 2.53
Locator
Persons per household, 2020–2024: 2.53
Boundary
ACS average household size, not a literal fractional household or number of AI users
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
epa-domestic-water-2015Statistics and Facts: Household water ↗

US EPA WaterSense · US EPA WaterSense · 2026

Source details and review dates
Supports
Why Save Water: 82 US gallons/person/day, citing USGS 2015
Locator
Why Save Water: 82 US gallons/person/day, citing USGS 2015
Boundary
Historical delivered domestic water; not consumptive use
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
doe-dishwasher-2023Purchasing Energy-Efficient Residential Dishwashers ↗

US Department of Energy · US Department of Energy · 2023

Source details and review dates
Supports
Table 1 ENERGY STAR column and assumptions: 240 kWh/year, 215 cycles/year, electric-resistance water heating
Locator
Table 1 ENERGY STAR column and assumptions: 240 kWh/year, 215 cycles/year, electric-resistance water heating
Boundary
Illustrative qualified dishwasher including electric water heating; not an observed household mean
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
energy-star-dishwasher-2023Dishwashers Key Product Criteria ↗

US EPA ENERGY STAR · US EPA ENERGY STAR · 2023

Source details and review dates
Supports
Standard dishwasher: at most 240 kWh/year and 3.2 US gallons/cycle
Locator
Standard dishwasher: at most 240 kWh/year and 3.2 US gallons/cycle
Boundary
Certification ceiling example, not average consumption
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
energy-star-washerClothes Washers ↗

US EPA ENERGY STAR · US EPA ENERGY STAR · 2026

Source details and review dates
Supports
Get a shower of water savings: full-size certified washers use 14 US gallons/load
Locator
Get a shower of water savings: full-size certified washers use 14 US gallons/load
Boundary
Wash water only; electricity is an independently authored 0.5 kWh/load illustration; dryer excluded
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
eia-school-electricity-2018CBECS 2018 Table C13: Electricity consumption ↗

US Energy Information Administration · US Energy Information Administration · 2022

Source details and review dates
Supports
Page 1, Education: 128 billion kWh / 13,623 million square feet
Locator
Page 1, Education: 128 billion kWh / 13,623 million square feet
Boundary
Derived 9.396 kWh/ft²/year for all education buildings; proxy for schools, not K–12-only or total fuel energy
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
lbnl-data-centres-20242024 United States Data Center Energy Usage Report ↗

Lawrence Berkeley National Laboratory · Lawrence Berkeley National Laboratory · 2024

Source details and review dates
Supports
Facility electricity, power usage effectiveness and water usage effectiveness methodology
Locator
Facility electricity, power usage effectiveness and water usage effectiveness methodology
Boundary
Context for facility accounting; capacity, utilization, PUE and WUE defaults in this calculator are authored examples
Geography
United States
Review status
Published reference; site extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
altman-chatgpt-2025The Gentle Singularity — ChatGPT resource-use statement ↗

Sam Altman · Sam Altman’s blog · 2025

Source details and review dates
Supports
Reported 0.34 Wh and 0.000085 US gallons per ChatGPT query; historical statement, not current telemetry
Locator
Paragraph beginning ‘As datacenter production gets automated’ (June 2025)
Boundary
Reported average query; energy and water boundaries unspecified
Geography
Not disclosed
Review status
not peer reviewed; methodology undisclosed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
energy-star-fridge-20242024 Quick Reference Guide — Refrigerators ↗

ENERGY STAR · US EPA · 2024

Source details and review dates
Supports
About 360 kWh/year top-freezer, 560 bottom-freezer, 630 side-by-side; not the mean of all households
Locator
PDF page 14, Refrigerator & Freezer Configurations
Boundary
Certified refrigerator operational electricity; excludes manufacturing and refrigerant leakage
Geography
United States
Review status
government product guidance
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
doe-room-ac-2024Purchasing Energy-Efficient Room Air Conditioners ↗

Federal Energy Management Program · US Department of Energy · 2024

Source details and review dates
Supports
612 kWh/year over 750 operating hours; implied 0.816 kWh/operating hour
Locator
Table 1, ENERGY STAR column; Energy Use and Cost Assumptions
Boundary
12,000 Btu/h room AC test example; operational electricity only
Geography
United States
Review status
government engineering example
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
gcsaa-water-2025Golf Course Environmental Profile, Phase IV — Water Use and Management ↗

GCSAA · Golf Course Superintendents Association of America · 2025

Source details and review dates
Supports
68.9 acre-feet per facility/year median; daily view divides annual volume by 365
Locator
PDF page 8 National Water Use and Table 2: 2024 median
Boundary
Applied irrigation water per US golf facility, all hole counts; recycled water included
Geography
United States, 2024 survey
Review status
industry survey; extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
nrel-electricity-water-2011A Review of Operational Water Consumption and Withdrawal Factors for Electricity Generating Technologies ↗

Macknick, Newmark, Heath & Hallett · NREL · 2011

Source details and review dates
Supports
Technology-dependent variation; background for an authored 1 L/kWh sensitivity input, not evidence for a national average
Locator
NREL/TP-6A20-50900, operational consumption and withdrawal review
Boundary
Electricity-generation operational water; consumption differs from withdrawal
Geography
Primarily United States power-generation technologies
Review status
national laboratory literature review
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2028-09-05
ai-energy-score-2025AI Energy Score text-generation benchmark cohort ↗

Luccioni, Gamazaychikov, Strubell, Hooker, Jernite, Mitchell & Chamberlin · Hugging Face · 2025

Source details and review dates
Supports
27 model energy records; imported CSV kWh per 1,000 tasks converted to Wh
Locator
February 2025 Class A rows with one GPU in parameter table; pinned revision c37e01a
Boundary
GPU-only energy, published text-generation protocol; not whole-service energy
Geography
Test cluster; grid not assigned
Review status
public benchmark; AI Footprint Lab extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
dalles-water-profile-2023Data Center Water Profile: The Dalles ↗

Google · Google · 2023

Source details and review dates
Supports
Almost 275 million gallons consumed in 2021; historical classroom case
Locator
Page 2, 2021 consumption
Boundary
Facility water, all services; not AI-only
Geography
The Dalles, Oregon
Review status
first-party historical disclosure
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
dalles-water-quality-20252025 Drinking Water Quality Report ↗

City of The Dalles · City of The Dalles · 2026

Source details and review dates
Supports
Historical place-based water-supply context; no scarcity score
Locator
Where does our drinking water come from?
Boundary
Municipal supply; not a facility source allocation
Geography
The Dalles municipal watershed and The Dalles Pool aquifer
Review status
municipal drinking-water report
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
google-gemini-2025Measuring the environmental impact of delivering AI at Google Scale ↗

Elsworth et al. (Google) · arXiv preprint · 2025

Source details and review dates
Supports
0.24 Wh serving energy and 0.26 mL allocated direct cooling-water estimate for the median Gemini Apps text prompt
Locator
v1 section 3.2 (serving energy), section 3.3 (water = IT energy × annual fleet-average WUE), section 3.4 (aggregation), and Table 2 (0.26 mL)
Boundary
Median Gemini Apps text prompt; allocated serving-stack energy, market-based carbon, and direct cooling-water estimate from fleet-average WUE
Geography
Google production fleet; locations not disclosed
Review status
preprint; first-party measurement
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
ecologits-joss-2025EcoLogits: Evaluating the Environmental Impacts of Generative AI ↗

GenAI Impact contributors · Journal of Open Source Software · 2025

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
Illustrative modeled ranges, not direct measurements
Locator
Model methodology and software documentation
Boundary
Modeled inference impacts; architecture, tokens, hardware, provider, and region are inputs
Geography
User-selected or modeled region
Review status
software peer review; energy-model validity not independently validated by JOSS
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2028-07-17
jegham-2025How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference ↗

Jegham et al. · arXiv preprint · 2025

Source details and review dates
Supports
Large task- and model-dependent inference variation
Locator
v6 abstract: most energy-intensive models exceed 29 Wh per long prompt, over 65× the most efficient; 0.42 Wh short-query example
Boundary
Inference estimates across selected models and tasks
Geography
Varies by modeled infrastructure
Review status
preprint (v6; headline values changed between v1–v6); closed-model results are modeled
Publication date
2025-11-24
Freshness
checked 2026-10-12 · review by 2027-04-12
luccioni-facct-2024Power Hungry Processing: Watts Driving the Cost of AI Deployment? ↗

Luccioni, Jernite & Strubell · ACM FAccT · 2024

Source details and review dates
Supports
Historical benchmark mean 2.907 Wh/image; not a current commercial product estimate
Locator
Table 2 image generation: mean 2.907 kWh/1,000 inferences, standard deviation 3.31; section 3 methodology
Boundary
CodeCarbon inference hardware energy for selected open models; excludes facility overhead, training and water
Geography
Test hardware and grid in paper
Review status
peer-reviewed publication; AI Footprint Lab extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
epa-egrid-2023eGRID 2023 Summary Data, Revision 2 ↗

US Environmental Protection Agency · US EPA · 2025

Source details and review dates
Supports
27 regional 2023 total-output CO₂e factors, lb/MWh × 0.45359237 to g/kWh; not inferred serving locations
Locator
US and subregion output emission rates
Boundary
Electricity-generation output rates; not the serving data centre's measured grid
Geography
United States
Review status
government statistical dataset
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
epa-equivalenciesGreenhouse Gas Equivalencies Calculator: Calculations and References ↗

US Environmental Protection Agency · US EPA · 2026

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
Passenger-car and electricity volume comparisons
Locator
Passenger vehicles and electricity
Boundary
US average equivalencies; not marginal emissions
Geography
United States
Review status
government methodology
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2027-07-17
eia-residentialHow much electricity does an American home use? ↗

US Energy Information Administration · US EIA · 2024

Historical source retained for archived scenarios; not used by current calculator values.

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
Annual US-home electricity comparison
Locator
Annual residential-customer average
Boundary
Purchased grid electricity per residential customer
Geography
United States
Review status
government statistical summary
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2027-07-17
energy-star-refrigeratorRefrigerators product guidance ↗

ENERGY STAR · US EPA / US DOE · 2026

Historical source retained for archived scenarios; not used by current calculator values.

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
Order-of-magnitude refrigerator comparison
Locator
Product energy-use guidance
Boundary
Qualified refrigerators; actual models vary
Geography
United States
Review status
government efficiency program
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2027-07-17
iea-streaming-2020The carbon footprint of streaming video: fact-checking the headlines ↗

George Kamiya · International Energy Agency · 2020

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
36 g CO2/hour central estimate and correction lesson
Locator
Central estimate and correction discussion
Boundary
2019 streaming conditions; device and grid dominate variation
Geography
Global-average scenario
Review status
IEA technical commentary; not journal peer reviewed
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2028-07-17
poore-nemecek-2018Reducing food's environmental impacts through producers and consumers ↗

Poore & Nemecek · Science · 2018

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
Illustrative beef carbon comparison
Locator
Food greenhouse-gas dataset
Boundary
Life-cycle food impacts; production systems vary
Geography
Global meta-analysis
Review status
peer-reviewed publication; AI Footprint Lab extraction not independently reviewed
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2028-07-17
mekonnen-hoekstra-2012A Global Assessment of the Water Footprint of Farm Animal Products ↗

Mekonnen & Hoekstra · Ecosystems · 2012

Historical source retained for archived scenarios; not used by current calculator values.

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
Beef water-category volume comparison; not scarcity-weighted impact
Locator
Product and production-system tables
Boundary
Green, blue, and grey supply-chain water footprint
Geography
Global with country/production variation
Review status
peer-reviewed publication; AI Footprint Lab extraction not independently reviewed
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2028-07-17
fulton-almond-2019Water-indexed benefits and impacts of California almonds ↗

Fulton, Norton & Shilling · Ecological Indicators · 2019

Historical source retained for archived scenarios; not used by current calculator values.

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
12 L total volume per almond and contextual caveat
Locator
2004–2015 water-footprint results
Boundary
Modeled green, blue, and grey water per almond
Geography
California
Review status
peer-reviewed publication; AI Footprint Lab extraction not independently reviewed
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2028-07-17
epa-watersenseShowerheads ↗

WaterSense · US EPA · 2026

Source details and review dates
Supports
Shower flow-rate comparison
Locator
About WaterSense: standard 2.5 gpm and WaterSense maximum 2.0 gpm
Boundary
Fixture flow rate; duration varies
Geography
United States
Review status
government product standard
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
usga-waterWater Resource Center ↗

United States Golf Association · USGA · 2026

Historical source retained for archived scenarios; not used by current calculator values.

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
Context for golf-course irrigation; no universal daily value
Locator
Water-use surveys and regional guidance
Boundary
Irrigation varies by climate, season, turf, and recycled-water use
Geography
United States
Review status
industry surveys and technical resources
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2027-07-17
masanet-2020Recalibrating global data center energy-use estimates ↗

Masanet et al. · Science · 2020

Earlier releases inherited a dataset date; an individual check date was not recorded.

Source details and review dates
Supports
2010–2018 electricity plateau and workload context
Locator
2010 and 2018 estimates
Boundary
Bottom-up global operational data-centre electricity
Geography
Global
Review status
peer-reviewed publication; AI Footprint Lab extraction not independently reviewed
Publication date
Not recorded
Freshness
checked individual date not recorded · review by 2028-07-17
iea-ai-2026Key Questions on Energy and AI ↗

International Energy Agency · IEA · 2026

Source details and review dates
Supports
485 TWh in 2025 and about 950 TWh central projection for 2030
Locator
Executive summary, ‘Our updated data centre electricity demand outlook’: 485 TWh (2025), 950 TWh (2030)
Boundary
All data centres; AI-focused facilities are a subset
Geography
Global
Review status
intergovernmental analysis and modeled projection
Publication date
Not recorded
Freshness
checked 2026-09-05 · review by 2027-03-05
google-environment-2026Google 2026 Environmental Report ↗

Google · Google Sustainability · 2026

Source details and review dates
Supports
Annual and hourly matching, absolute data-centre electricity and water, replenishment, location- and market-based Scope 2
Locator
PDF p. 4 (100% global annual renewable match); p. 72 (about 65% CFE in 2025, 66% in 2024); p. 90 (GHG table); p. 93 (electricity: data centres 42,415,800 MWh); p. 96 (water: data centres 13,562 million gallons withdrawn, 10,523 consumed); p. 6 (7.7 billion gallons replenished, about 78% of freshwater consumption)
Boundary
Calendar 2025. Electricity and water split into data centres and offices; hourly CFE covers data centres and offices; Scope 2 is company-wide, not data-centre-only
Geography
Global company operations
Review status
first-party disclosure; 2025 metrics marked for third-party limited assurance
Publication date
2026-06-30
Freshness
checked 2026-10-12 · review by 2027-07-17
microsoft-environment-2026Microsoft 2026 Environmental Sustainability Report ↗

Microsoft · Microsoft · 2026

Source details and review dates
Supports
FY2025 annual matching definition, data-centre PUE and WUE
Locator
PDF p. 6 (FY25 100% annual renewable match); p. 64 note 2 (match counts contracted projects plus “grid mix” supplied through utility rates and portfolio standards; excludes spot RECs); p. 23 (FY25 PUE 1.17); p. 24 and p. 64 note 12 (FY25 WUE 0.27 L/kWh; cooling and humidification water ÷ energy). Publication date from the 9 July 2026 announcement URL
Boundary
Fiscal year ended 30 June 2025. WUE and PUE cover owned data centres; matching covers global electricity
Geography
Global company operations
Review status
first-party disclosure
Publication date
2026-07-09
Freshness
checked 2026-10-12 · review by 2027-07-17
microsoft-data-fact-sheet-20262026 Environmental Data Fact Sheet ↗

Microsoft · Microsoft · 2026

Source details and review dates
Supports
Location- and market-based Scope 2, company electricity and water, data-centre values by location
Locator
PDF p. 3 Table 1A (FY25 Scope 2 location-based 12,030,556 and market-based 2,707,428 tCO₂e; total 21,121,000); p. 6 Table 5 (electricity 37,026,353 MWh; matched contracted/generated renewable electricity 34,534,392 MWh, excluding grid mix); p. 7 Table 8 (withdrawal 13,266 ML; consumption 8,170 ML); p. 25 Table 15 (FY25 electricity and withdrawal by owned data-centre location)
Boundary
Company-wide FY2025 except Table 15, which lists owned data-centre locations and excludes commissioning and locations that together make up less than 1% of the total; no global data-centre sum is published
Geography
Global company operations; Table 15 by metro location
Review status
first-party disclosure; assurance scope listed in Table 16
Publication date
Not recorded
Freshness
checked 2026-10-12 · review by 2027-07-17
amazon-sustainability-2025Amazon 2025 Sustainability Report ↗

Amazon · Amazon Sustainability · 2026

Source details and review dates
Supports
Annual match, data-centre water withdrawal, WUE and water-stress distribution, market-based Scope 2
Locator
PDF p. 10 (100% renewable match for the third year; PUE 1.14); p. 14 (data centres withdrew 9.4 billion litres; WUE defined as litres withdrawn per kWh of IT load); p. 15 (WUE 0.12 L/kWh; 22% of withdrawals in high or extremely high water stress); p. 46 (Scope 2 3.74 MMT CO₂e); p. 8 (2025 emissions use a market-based method). Publication date from the 1 July 2026 Amazon News announcement
Boundary
Calendar 2025. Water covers leased, owned and shared data centres; Scope 2 is company-wide and market-based
Geography
Global company operations
Review status
first-party disclosure; assurance varies by metric
Publication date
2026-07-01
Freshness
checked 2026-10-12 · review by 2027-07-17
meta-sustainability-2025Meta 2025 Sustainability Report ↗

Meta · Meta Sustainability · 2025

Source details and review dates
Supports
Annual clean-energy match and water-restoration disclosure
Locator
Highlights: matched 100% of electricity use with clean and renewable energy; projects restored over 1.6 billion gallons in 2024. Publication date from the 12 September 2025 launch post. No 2026 report was published when checked
Boundary
Calendar 2024; corporate operations
Geography
Global company operations
Review status
first-party disclosure; independent review scope linked by publisher
Publication date
2025-09-12
Freshness
checked 2026-10-12 · review by 2027-07-17
meta-environmental-data-index-2025Meta 2025 Environmental Data Index ↗

Meta · Meta Sustainability · 2025

Source details and review dates
Supports
Absolute data-centre electricity, water and Scope 2 for 2024
Locator
PDF p. 4 (Scope 2 data centres total 2024: market-based 135, location-based 5,862,614 tCO₂e); p. 6 (electricity, data centres total 2024: 18,061,781 MWh); p. 9 (water withdrawal, data centres total: 4,145 ML); p. 11 (water consumption, data centres total: 2,974 ML)
Boundary
Calendar 2024; “data centers total” equals owned sites plus leased and other data-centre-related facilities
Geography
Global company operations
Review status
first-party disclosure
Publication date
Not recorded
Freshness
checked 2026-10-12 · review by 2027-07-17
delavande-video-neurips-2025Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models ↗

Delavande, Pierrard & Luccioni · NeurIPS 2025 workshop paper (arXiv) · 2025

Source details and review dates
Supports
21.6 + 2.4 + 1.3 = 25.3 Wh for this clip specification; same table reports 0.14 Wh (AnimateDiff) and over 415 Wh (WAN2.1-T2V-14B) at their own default settings
Locator
Section 4.3 protocol (50 prompts, 2 warmups, 5 runs); Table 2 CogVideoX-5b defaults; Table 4 (PDF p. 9) CogVideoX-5B row
Boundary
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 separately
Review status
Accepted as an oral at the NeurIPS 2025 NextVid workshop; not a journal article; AI Footprint Lab extraction not independently reviewed
Publication date
2025-09-23
Freshness
checked 2026-10-12 · review by 2027-04-12
mlenergy-v3-tasks-2026ML.ENERGY v3: text-to-image, text-to-video, chat and GPQA measurements ↗

ML.ENERGY research team · ML.ENERGY · 2026

Source details and review dates
Supports
Images (1024 × 1024, batch 1): 0.116 Wh SANA 1.5 1.6B at 20 steps to 2.208 Wh FLUX.1-dev at 50 steps. Video (batch 1): 7.95 Wh CogVideoX-2b to 323 Wh HunyuanVideo. Qwen3-8B: 0.013 Wh per chat response (about 645 output tokens) vs 0.494 Wh per GPQA response (about 8,050).
Locator
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_joules
Boundary
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 separately
Review status
Benchmark framework published at NeurIPS 2025 Datasets & Benchmarks; v3 dataset extraction not independently reviewed
Publication date
Not recorded
Freshness
checked 2026-10-12 · review by 2027-04-12
oviedo-inference-2026Energy use of AI inference, efficiency pathways, and test-time scaling ↗

Oviedo, Kazhamiaka, Choukse, Kim, Luers, Nakagawa, Bianchini & Lavista Ferres (Microsoft) · Joule 10(8):102430 (DOI 10.1016/j.joule.2026.102430); arXiv:2509.20241v2 · 2026

Source details and review dates
Supports
Median 0.31 Wh/query (IQR 0.16–0.60); 3.91 Wh (IQR 2.15–7.05) for test-time-scaling queries about 15× longer, a 13× rise
Locator
v2 Summary and section 2 (Equation 1; 8×H100 node power; PUE sampled 1.05–1.40; 500 input tokens; median 300 or 5,000 output tokens)
Boundary
Modeled, not provider telemetry. Full-node electricity × PUE for saturated, batched serving of three open models above 200B parameters; idle capacity, networking and water excluded
Geography
Generic H100 deployment; no grid assigned
Review status
Peer-reviewed in Joule (August 2026); authors are Microsoft employees. AI Footprint Lab checked the arXiv v2 text; the journal text was not accessible
Publication date
2026-06-09
Freshness
checked 2026-10-12 · review by 2027-04-12
ai-energy-score-v2-2025AI Energy Score v2: Refreshed Leaderboard, now with Reasoning ↗

Luccioni & Gamazaychikov · Hugging Face community blog · 2025

Source details and review dates
Supports
Reasoning on vs off: DeepSeek-R1-Distill-Llama-70B 49.53 → 7,626.53 Wh/1,000 queries (154×); Phi-4-reasoning-plus 514×; SmolLM3-3B 697×
Locator
“Reasoning comes at a cost” section: reasoning models use on average 30× more energy; table of GPU energy per 1,000 queries with reasoning off and on
Boundary
GPU energy on the AI Energy Score protocol; not whole-service energy
Geography
Test cluster; grid not assigned
Review status
public benchmark write-up; not peer reviewed
Publication date
2025-12-04
Freshness
checked 2026-10-12 · review by 2027-04-12
lbnl-data-centres-2025United States Data Center Energy Usage Report: 2025 Update ↗

Smith, Hubbard, Newkirk, Ganeshalingam, Holecek, Sartor, Mills & Shehabi · Lawrence Berkeley National Laboratory · 2026

Source details and review dates
Supports
192 TWh (4.7%) in 2024; 649 TWh reference and 521–843 TWh range for 2030
Locator
Abstract (2030 Reference Case 649 TWh, 11.8% of US electricity; compounded range 521–843 TWh) and PDF p. 11 (192 TWh in 2024, 4.7%)
Boundary
Bottom-up model of all US data centres from equipment shipments, device power and cooling simulations; not AI-only; projections are scenarios, not forecasts
Geography
United States
Review status
national laboratory report (LBNL-2001758); modeled projections
Publication date
2026-06-18
Freshness
checked 2026-10-12 · review by 2027-06-18
mistral-lca-2025Our contribution to a global environmental standard for AI ↗

Mistral AI · Mistral AI · 2025

Source details and review dates
Supports
1.14 gCO₂e, 45 mL water and 0.16 mg Sb eq for one 400-token response; not comparable with direct-cooling-only water values
Locator
“Marginal impacts of inference” paragraph: a 400-token Le Chat response, excluding users’ terminals
Boundary
Life-cycle inference impacts including upstream hardware manufacturing (AFNOR Frugal AI; GHG Protocol Product Standard; ISO 14040/44); no energy value published
Geography
Mistral AI infrastructure; locations not disclosed
Review status
First-party life-cycle analysis with Carbone 4 and ADEME, reviewed by consultancies Resilio and Hubblo; the announcement links no full study and says results will later appear in ADEME’s Base Empreinte
Publication date
2025-07-22
Freshness
checked 2026-10-12 · review by 2027-04-12
eia-residential-20242024 Average Monthly Bill – Residential (Table T5.a) ↗

US Energy Information Administration · US EIA, Electric Sales, Revenue, and Average Price · 2025

Source details and review dates
Supports
863 kWh/month × 12 = 10,356 kWh per residential customer in 2024
Locator
U.S. Total row: 143,144,185 customers; average monthly consumption 863 kWh
Boundary
Utility-reported residential retail sales per customer account; grid purchases, not total use in homes with rooftop solar
Geography
United States
Review status
government statistical table (EIA-861, 861S, 861U); EIA lists the series as discontinued
Publication date
2025-10-07
Freshness
checked 2026-10-12 · review by 2027-04-12

Classroom package

Choose from three activities with printable worksheets, worked answers and a scoring rubric.

Governance, corrections, and conflicts

Transparency about review status is part of the method.

Authorship

Built by Chris Meehan, an AP Environmental Science educator and academic technology director. The site currently lists no sponsor, paid data provider, or institutional scientific endorsement.

Independent review

Energy/LCA reviewer: not appointed. Water/hydrology reviewer: not appointed. Review reports and unresolved disagreements will be published here when completed.

Freshness policy

Every evidence record has a review-by date. A review date is a scheduled check, not proof that no newer source exists. Unknown extraction details stay marked unresolved. Annual company disclosures should be checked within 30 days of publication. A dataset-wide date never substitutes for claim-level freshness.

Earth imagery

Blue Marble: a Suomi NPP / VIIRS satellite composite from January 4, 2012. Credit: NASA/NOAA/GSFC/Suomi NPP/VIIRS/Norman Kuring. Used in the social preview; local derivatives are retained for historical designs. The current homepage does not display the image. Original high-resolution image and credits ↗. Imagery use does not imply NASA endorsement.

Privacy and school use

AI Footprint Lab uses no account, localStorage, sessionStorage, analytics, or advertising code in this repository. Activities do not collect student answers. The electricity lab keeps entered readings and descriptions only in the open page until you download a record; they are not added to share URLs.

  • Investigation settings, including group size, participation and active days when selected, are stored in the page URL and browser history. Group place identity is omitted from links but included in user-downloaded JSON when a Census population was chosen. Facility inputs persist during in-app navigation and can be downloaded separately; they are not in scenario URLs.
  • Copying, sharing, or downloading is initiated by the user; the operating system or chosen share service may then receive the URL.
  • No student name, email address or school identity is requested. Optional ZIP entry or browser location can suggest cities/states. Census reference files are loaded from this site; coordinates and ZIP are processed in the browser, not transmitted to an external lookup service or stored in share URLs. Location permission is requested only after clicking Use my location.
  • Source Sans 3 and Satoshi are served by this site. No third-party font service is contacted.
  • Schools should still review hosting logs and deployment configuration before adoption.

Glossary

The same dictionary powers these definitions and the dotted terms across the site.

additionalityThe test an offset must pass: would the climate benefit have happened anyway without your money? Many cheap credits fail it.
bnBillion (1,000,000,000).
bn m³Billion cubic metres. One m³ = 1,000 litres, so 1 bn m³ = a trillion litres, lake-scale water.
CACMCommunications of the ACM: a leading peer-reviewed computer-science journal; where the key AI water-footprint paper was published.
CFECarbon-Free Energy. “24/7 CFE” means matching consumption with clean power every HOUR in the SAME region, a much stronger claim than annual REC matching.
CO₂eCarbon-dioxide equivalent: a common yardstick that converts all greenhouse gases (CO₂, methane, N₂O…) into the amount of CO₂ with the same warming effect.
CSRDThe EU's Corporate Sustainability Reporting Directive: a law forcing large companies to report emissions, including Scope 3 (which pulls in AI and cloud use).
DOEUS Department of Energy: source for appliance figures like refrigerator energy use.
EIAUS Energy Information Administration: official US energy statistics (e.g., typical home electricity use).
EPAUS Environmental Protection Agency: source of our vehicle, grid, and equivalency factors (public domain, revised annually).
gCO₂Grams of carbon dioxide.
gCO₂eGrams of CO₂-equivalent: greenhouse-gas warming effect expressed in grams, with gases like methane converted to their CO₂ equivalent.
gCO₂e/kWhGrams of CO₂-equivalent per kilowatt-hour: a location- and time-dependent electricity emissions factor. AI Footprint Lab's presets are adjustable teaching inputs, not measurements of the data centre serving a request.
gpmGallons per minute: a flow rate; modern showerheads run ~2.1 gpm (≈8 L/min).
GPUGraphics Processing Unit: the chip family that powers modern AI; thousands run in parallel to train and serve models.
GWGigawatt = 1,000 MW of POWER (a rate, not an amount), roughly one large nuclear plant's output.
GWhGigawatt-hour = 1 million kWh.
hyperscaleThe giant, highly automated data centres run by companies like Google, Amazon, and Microsoft, the most energy-efficient class of facility.
IEAInternational Energy Agency: the intergovernmental body whose data-centre and streaming figures we cite.
inferenceUsing a trained AI model to answer a query (vs. TRAINING, the one-time cost of building it). Your chats are inference.
Jevons paradoxThe 1865 observation that efficiency can INCREASE total consumption: cheaper per-use invites more uses. Why per-query gains haven't stopped total AI demand from rising.
JOSSJournal of Open Source Software: a peer-reviewed journal for research software; EcoLogits' methodology was reviewed there.
kgCO₂eKilograms of CO₂-equivalent: 1,000 grams of greenhouse-gas warming effect, with other gases (like methane) converted to their CO₂ equivalent.
kWhKilowatt-hour = 1,000 Wh, running a 1,000 W appliance for one hour. A US home uses ~29 kWh a day.
LLitre: the metric volume unit; a bit more than a quarter of a US gallon. 1,000 L = 1 m³.
m³Cubic metre = 1,000 litres (about 264 gallons), roughly five full bathtubs.
mLMillilitre: 1/1,000 of a litre. A teaspoon ≈ 5 mL.
MtMegatonne = 1 million metric tons (1 billion kg).
MWMegawatt = 1 million watts of POWER (a rate, not an amount). A big data centre draws tens to hundreds of MW continuously.
MWhMegawatt-hour = 1,000 kWh, roughly a month of electricity for a typical US home.
potableSafe to drink. Much data-centre cooling uses potable municipal water, part of why its “blue water” draws scrutiny.
PPAPower Purchase Agreement: a long-term contract to buy electricity from a specific plant, often used to finance new clean-energy projects.
PUEPower Usage Effectiveness: total data-centre energy ÷ energy that reaches the computers. Perfect = 1.0; Google reports ~1.09, meaning only ~9% overhead for cooling etc.
RECRenewable Energy Certificate: a tradable credit for 1 MWh of renewable generation. Buying RECs equal to your annual use is how companies claim “100% renewable” without using clean power every hour.
Scope 1 / 2 / 3Emissions accounting categories: Scope 1 = a company's own fuel burning; Scope 2 = purchased electricity; Scope 3 = everything upstream/downstream (construction, chips, suppliers), usually the biggest and hardest to cut.
SMRSmall Modular Reactor: a factory-built nuclear reactor (~50–300 MW), designed to be faster and cheaper to deploy than traditional plants. First US commercial units expected ~2030.
tokenThe chunks of text AI models read and write, roughly ¾ of a word each. Energy scales mostly with tokens generated.
TPUTensor Processing Unit: Google's custom AI chip (its answer to the GPU).
TWhTerawatt-hour = 1 billion kWh. Country-scale energy: ~415 TWh ran every data centre on Earth in 2024.
WhWatt-hour, a small unit of energy: a 10 W LED bulb burning for 6 minutes uses 1 Wh.

Public change record

Oldest recorded individual check 2026-09-05; dataset release 2026.10.12.2. Release notes describe both completed and still-pending work.

URLs and link previews · October 12, 2026 — Gave every page its own URL, title, description and link preview, and prerendered them so a crawler or a shared link does not need JavaScript. Each classroom activity now has its own address. Links shared before this keep working and are rewritten in place with their settings intact. Source freshness is measured in the browser so a build date can never stand in for a claim-level check.

Dataset 2026.10.12.2 · October 12, 2026 — Added a selectable electricity case, scenarios.microsoftStudy.wh = 0.31 Wh: the peer-reviewed modeled median by Microsoft researchers (Joule, 2026) for three open models above 200B parameters, including whole-server power and PUE. It is a modeled study value, not a measurement of any product, and has no paired water case. Mistral’s 45 mL life-cycle water value remains context only because its boundary includes hardware manufacturing and electricity generation, which the everyday water comparisons do not.

Dataset 2026.10.12.1 · October 12, 2026 — Rechecked company disclosures against the reports themselves and recorded their publication dates; all four were checked more than 30 days after publication. Corrected the company table: Google and Meta publish absolute data-centre electricity, Google reports about 65% hourly carbon-free energy, Microsoft reports a 0.27 L/kWh WUE and defines its 100% match to include utility “grid mix”, and Amazon reports 9.4 billion litres withdrawn by data centres. Replaced video.wh with the NeurIPS 2025 workshop version of the CogVideoX-5B benchmark (25.3 Wh; previously 25.93 Wh from the exploratory blog post). Updated home.kwhYr to EIA’s final 2024 average (863 kWh/month; previously 10,791 kWh/year for 2022). Pinned Jegham et al. to v6. Added cited context on reasoning energy (Microsoft/Joule 2026, AI Energy Score v2, ML.ENERGY v3), newer image and video measurements, Mistral’s life-cycle water value and LBNL’s 2025 US data-centre update; none adds a new selectable case. EPA had not published eGRID2024 when checked, so grid factors are unchanged.

1.7.0 · September 5, 2026 — Unified all monthly comparisons at 30 days; relabeled Gemini as a reported median. Added starting/alternative scenarios, independent experiments and complete comparison sharing. Source drawers link real factor IDs to the ledger; source checks and missing metadata are explicit. Retired unused Story factors; historical source records remain labeled for old snapshots. Improved navigation, accessibility and loading; independent review remains pending.

1.6.0 · September 5, 2026 — Moved image, video and new vibe coding inputs into collapsed advanced options on Home and Calculator. Replaced image.wh with the FAccT 2024 Table 2 mean (2.907 Wh); replaced video.wh with the CogVideoX-5b hardware-energy row (25.93 Wh) in Delavande 2025, retaining its exploratory status and run-count inconsistency. Added pinned ML.ENERGY v3 codingCases J/token values and explicitly authored session/token scaling. Hardware boundaries differ and are not a life-cycle total or confidence range. Added arithmetic, URL, export, interaction and accessibility checks. Water remains ordinary text only; independent scientific review pending.

1.5.0 · September 5, 2026 — Added toilet delivered-water comparisons (toilet.gallonsFlush), editable EV/gas driving scenarios and gasoline tailpipe CO2 (gasoline.co2Gallon), plus original SVG bottle, toilet and car pictograms. Equivalent counts respond to period, group, water basis and edited reference inputs. Fractional and grouped symbols show an explicit scale. EV generation CO2e and gasoline tailpipe CO2 retain distinct boundaries; gasoline electricity and water are not invented. Bottle size and driving efficiencies remain authored inputs.

1.4.0 · September 5, 2026 — Added individual/group scopes, participation and active-day assumptions, offline Census Vintage 2025 population lookup with consent-based device location and ZIP-area suggestions, household utility/dishwasher/laundry/school comparisons, and a separate facility calculator. Location never changes the AI serving grid. Added calculation, population, privacy and interaction coverage. Group scenarios are hypothetical extrapolations; official population does not establish AI use.

1.3.0 · September 5, 2026 — Rebuilt Home around daily-use inputs and visible electricity, water and CO₂e comparisons. Added the explicitly method-undisclosed 2025 ChatGPT report (scenarios.chatgpt.wh; waterCases.chatgpt.ml). Replaced fridge, golf and shower illustrations with traceable conversions (fridge.whDay; golf.dayL; shower.lpm), added the DOE room-AC example and optional electricity-supply water sensitivity model. Annualized periods, unmatched boundaries and unknown values remain explicit. Added comparison arithmetic, source and interaction tests; scientific review remains pending.

1.2.0 · September 5, 2026 — Added 27 documented model benchmarks with separate comparison exports, all 27 EPA eGRID2023 regional CO₂e factors with report provenance, and a historical water-supply case for The Dalles. Missing benchmark metadata and unmatched water periods remain explicit; no provider telemetry or life-cycle total is implied.

v1.1 · 2026-09-05 — Added a measurement guide and a separate electricity-observation lab with downloadable inputs, boundaries and calculations. Corrected Gemini water wording after checking section 3.3: a calculated allocation from serving energy and annual fleet-average WUE, not an individual-prompt meter reading. Replaced display/monospace UI fonts with self-hosted Source Sans 3 and Charter/Georgia headings; simplified headings, decorative panels and comparisons. No provider telemetry connection is claimed.

v1.0 · 2026-09-05 — Added student activities and a dedicated teacher package with worked answers, rubric, and pilot/reviewer protocols. Unified calculator/report inputs and added versioned links and downloadable result snapshots. Relabeled 4 Wh text and approximate image factors as authored illustrations; added a factor derivation ledger with unresolved evidence checks visible. Made everyday comparisons optional and published editorial rules. No external scientific review or classroom pilot is claimed. Rechecked the Gemini abstract and IEA executive-summary numbers; other source-check dates remain unchanged. Approximate comparison factors are now explicitly authored pending exact extraction review.

v0.9 — Classroom-readiness overhaul: added linked claim-level evidence metadata and downloads; separated energy and water cases; renamed normative scenarios; documented exclusions; updated IEA and corporate-report context through July 17, 2026; normalized the company disclosure table; added governance, privacy, classroom, and review-status sections; removed the X sharing shortcut and external font requests; added calculator and provenance tests. Independent scientific review and classroom pilots remain explicitly pending.

v0.8 — Improved chart interaction and labels, fixed navigation/report issues, and unpublished the deterministic claim checker because its wording analysis risked conclusions stronger than its evidence.

v0.7 — Introduced the editorial homepage, reproducible report URLs, discovery metadata, and a labeled API concept preview.

v0.6 — Design and copy pass; no environmental factors changed.

v0.5 — Added the sitewide glossary engine and matching regression tests.

v0.4 — Added the industry timeline and company-disclosure section.

v0.3 — Rechecked core factors and corrected almond, golf, and beef-water values.

v0.2 — Added explicit green, blue, and grey water categories.

v0.1 — First public factors.