How These Estimates Work

Every number in this tool is an estimate. Some are well-established — EPA appliance standards, ICAO flight emissions methodology, peer-reviewed food lifecycle studies. Others are educated extrapolations from limited data. The goal is not perfect precision. It is defensible estimation with transparent uncertainty.

This page explains where the numbers come from, what they include and exclude, and why some carry more confidence than others.

What we measure

Each activity is estimated across three parameters:

System boundaries — what's in and what's out

The hardest methodological question in any footprint comparison is where to draw the boundary. We use operational/direct use as the default — meaning we count the energy, water, and emissions during the activity itself, not the full lifecycle of the devices or infrastructure involved.

There are three explicit exceptions, each labeled on the relevant entries:

AI activities measure inference only — the energy used when you send a prompt and receive a response. They do not include the energy used to train the model in the first place (Patterson et al. 2022 estimates amortized training roughly doubles the per-prompt footprint), the carbon cost of manufacturing the hardware, or semiconductor manufacturing water use. This boundary is consistent with most peer-reviewed AI inference energy studies.

Food activities use field-to-plate lifecycle estimates, because the direct energy of cooking is near-zero and meaningless as a comparison point. Carbon figures cover the full supply chain from agricultural production through processing, transport, and retail, and are the most reliable parameter for food comparisons. Water figures are volumetric water footprints — the total water consumed across the entire supply chain, including feed crops and irrigation — following the methodology of Mekonnen & Hoekstra 2012. This is not the same as local water scarcity and is labeled accordingly on each entry. Energy figures for food draw from sources with different system boundaries and should be treated as order-of-magnitude estimates only; see the note on food energy confidence below.

Flights use ICAO methodology including a radiative forcing multiplier, which accounts for the additional warming effect of contrails and high-altitude emissions beyond CO₂ alone. This is labeled on each flight entry.

We do not mix boundaries silently. Where an entry uses a different boundary than the default, the notes field says so.

Confidence badges

Every bar in this tool carries a confidence badge — High, Med, or Low — assigned per parameter, not per activity. An activity can have High confidence energy data and Low confidence water data. That's more honest than a single blended score.

Confidence reflects four main factors: the quality and type of source (peer-reviewed study vs. company disclosure vs. extrapolation), how recent the data is, how closely the source matches the specific activity, and how much sources agree with each other.

High means the estimate comes from government data, standards bodies, or peer-reviewed studies, the sources agree closely, and the activity is directly matched. EPA appliance data, ENERGY STAR figures, and ICAO flight methodology all typically produce High confidence estimates.

Med means reasonable sources exist but there's meaningful variance — company sustainability disclosures, industry association data, or studies that require moderate extrapolation to match the activity.

Low means the estimate is heavily extrapolated, sources disagree significantly, or the underlying data is simply not well-characterized. Low confidence doesn't mean the number is wrong; it means treat it as an order-of-magnitude estimate, not a precise figure.

A note on AI confidence

AI water and carbon figures are Low confidence almost by definition. AI providers don't publish the data needed to calculate these directly — water figures are estimated by applying datacenter water efficiency ratios to energy estimates, which compounds two uncertain numbers. Energy figures can reach Med confidence when derived from peer-reviewed inference studies with a matching model class, but even those reflect a snapshot that may be months out of date. Source dates are always shown for AI entries for this reason.

A note on social media confidence

No peer-reviewed study has directly measured the per-hour energy footprint of any specific social media platform. Social media entries in this catalog are built from components — device power draw, network energy intensity, and datacenter figures — each sourced from peer-reviewed work, but combined with assumptions about usage patterns that introduce meaningful uncertainty. All social media entries carry Low confidence across all three parameters.

A note on food energy confidence

Food carbon and water figures are the most reliable numbers in the food category, drawn from well-established lifecycle studies. Food energy figures are less reliable — they draw from sources with different system boundaries and methodologies that are not fully consistent with each other or with the energy figures for non-food entries. They are included because the relative comparisons within the food category are still informative, but they should be treated as order-of-magnitude estimates rather than precise figures. Carbon and water are the more meaningful parameters for comparing food choices.

Why AI numbers are especially uncertain — and change over time

AI inference energy figures are uniquely time-sensitive, and the uncertainty runs in two directions simultaneously.

The downward trend: Each generation of hardware and software makes AI inference more efficient. Google reported a 33× drop in energy per Gemini text prompt between May 2024 and May 2025 alone. Hardware improvements (each GPU generation substantially outperforms the last), software optimizations like quantization and speculative decoding, and more efficient model architectures all compound. The IEA has described AI inference efficiency as improving at a rate unprecedented in energy history.

New model spikes: Each new frontier model release tends to be more energy-intensive per response than its predecessor at launch — larger, more capable models require more computation. These spikes are temporary; efficiency optimization follows rapidly. But they mean the overall picture is not a smooth curve.

Reasoning models — models like o1, o3, and Claude Extended Thinking that generate extended internal reasoning chains before responding — are 10–70× more energy-intensive per response than standard models of the same generation, not because they are larger but because they generate far more tokens. This catalog includes a reasoning model entry; see the AI category for details.

Data lag: There is typically a 6–18 month lag between a model releasing and peer-reviewed energy measurements becoming available. All AI figures in this catalog represent a 2025–2026 snapshot and are reviewed every six months.

A note on widely-cited figures: the "3 Wh per ChatGPT prompt" number that circulated from 2023 onward was based on a 2009 Google search estimate scaled up — not a direct measurement. By 2025, independent studies placed median text-prompt energy at 0.24–0.43 Wh, roughly 10× lower. The figures in this catalog reflect the current literature.

Bitcoin's water footprint — why it's so large

The water figure for a Bitcoin transaction (approximately 4,300 gallons) is not a typo. It reflects a more complete accounting than older estimates, which typically counted only the water used directly at mining facilities for cooling equipment.

The larger component is indirect: the power plants supplying electricity to those facilities consume substantial water. Coal, gas, and nuclear plants use water to generate steam and cool reactors — most of which evaporates and cannot be returned to local ecosystems. Hydroelectric power, often promoted as a "green" source for miners, has the highest water consumption per kWh generated due to evaporation from reservoir surfaces. When you include both the direct facility water and the indirect power plant water, the per-transaction figure rises to thousands of gallons.

This methodology comes from de Vries (2023, Cell Reports Sustainability) and represents the current academic consensus on how to account for Bitcoin's full hydrological footprint.

Sources

Estimates draw from publicly available research and institutional reports. Preferred sources in order of reliability: government agencies (EPA, DOE, EIA, NREL, ICAO, FAO), standards bodies (ENERGY STAR, WaterSense), peer-reviewed studies, credible preprints, company sustainability disclosures (labeled as company-reported), and industry associations.

Company-reported figures — such as Google's Gemini inference data — are included where no independent measurement exists, but are always labeled as such. Unsourced blog posts, viral comparisons without traceable methodology, and advocacy estimates without disclosed methods are not used as primary sources.

Key sources used across the catalog include:

• EPA GHG Emission Factors Hub (2023) and eGRID (2022)
• DOE ENERGY STAR appliance databases
• ICAO Carbon Emissions Calculator (2023)
• Poore & Nemecek 2018 — "Reducing food's environmental impacts" (Science)
• Mekonnen & Hoekstra 2012 — water footprint of food and livestock
• Luccioni et al. 2023 — "Power Hungry Processing" (EMNLP) — AI inference energy
• de Vries 2023 — Bitcoin's growing water footprint (Cell Reports Sustainability)
• Cambridge Bitcoin Electricity Consumption Index (CBECI) 2025
• Cambridge Centre for Alternative Finance — Ethereum post-merge energy data 2024–2025
• Lannelongue et al. 2023 — AlphaFold inference energy
• IEA — Electricity 2024, Data Centres & Networks
• NREL PVWatts Calculator (2023)
• APTA transit energy and emissions data

Each activity entry lists its specific sources. Click any bar to see the sources for that activity.

A note on meat alternative water figures

The water figures for meat alternative burgers (Beyond Meat, Impossible Burger) carry a specific caveat worth understanding. The source lifecycle assessments used different system boundaries — Beyond Meat's LCA used cradle-to-distribution, while Impossible Burger's used cradle-to-gate. Neither matches the field-to-plate boundary used for the beef burger entry. This means the water figures for meat alternatives are not directly comparable to the beef burger water figure, or fully comparable to each other. The carbon figures are more consistent across sources and are the more reliable basis for comparing meat alternatives to beef. The water figure is included with Low confidence and this caveat disclosed in the entry notes.

A note on what this tool doesn't do

This tool shows estimates for individual activities in isolation. It does not model cumulative annual footprints, regional grid mix differences, or lifecycle costs of devices. It does not tell you what to do — it shows proportions and lets you draw your own conclusions. The data makes the point; the tool doesn't editorialize.