An AiGovOps Foundation published artifact · every fact carries 4+ sources and a trail to check it yourself

We already pay for the technologies we love. What makes AI different?

Phones, streaming, cars, warm houses, cool houses, burgers, steel, cement, flying — modern life is a set of resource trade-offs we largely made peace with. This calculator puts AI's costs on the same table, fairly: the same ledgers, honest units, math a 6th grader can follow, and behind every number a stack of four or more sources you can go check.

AI is the first technology whose costs concentrate faster than its benefits become legible. That is an argument for governance instrumentation — not an argument for or against adoption.

measured  metered or telemetry data modeled  peer-reviewed or institutional estimate scenario  projection, not a forecast vendor  commercially interested study derived  computed here, formula shown

Your year in technology

Set your own usage. Every module shows its math step by step — no hidden formulas — and every number has a "4+ sources" drawer telling you exactly where it came from and how to go check it. The ledgers stay separate: no composite score is computed, because weighing them is a values choice that belongs to you.

AI assistant prompts

Pick your evidence basis — the three published numbers use different models, boundaries and years, and are not directly comparable.

See the math, step by step
4+ sources · how to verify
  1. Google technical paper (Aug 2025) — 0.24 Wh, 0.03 g CO2e, 0.26 mL per median Gemini text prompt.
  2. arXiv:2508.15734 — same paper; shows the narrower "existing approach" would claim just 0.10 Wh.
  3. MIT Technology Review and The Verge — independent coverage; the Verge notes the water figure excludes power-plant water.
  4. Epoch AI (Feb 2025) + TechCrunch — ~0.3 Wh typical GPT-4o query.
  5. de Vries, Joule (2023) + Nature Comment (2025) — "at most 2.9 Wh per request", GPT-3.5 era; Nature brackets short queries at 0.3–0.43 Wh and long ones at 3.5 Wh.
Find it yourself: search arxiv.org for "Measuring the Environmental Impact of Delivering AI at Google Scale" and read Table 1 (comprehensive vs existing approach). Then open epoch.ai → Gradient Updates → "How much energy does ChatGPT use?" and read the footnote tracing the old 3 Wh figure to de Vries (2023).

Video streaming

Energy: 0.077 kWh/hr modeled. Carbon basis:

See the math, step by step
4+ sources · how to verify
  1. IEA commentary (Kamiya, updated Dec 2020) — 0.077 kWh and 36 g CO2 per hour (revised down from 82 g).
  2. Carbon Brief factcheck — the original venue for the same analysis, independently edited.
  3. Carbon Trust white paper (2021) + DIMPACT — ~55 g CO2e/hr in Europe; your screen is most of it.
  4. Netflix statement + Fortune — operator + press confirmation of ~55 g/hr ("boiling a kettle for six minutes").
Find it yourself: on iea.org open Commentaries → "The carbon footprint of streaming video" and search the page for "36" — the December 2020 update note is where 82 g became 36 g. The 36 g and 55 g numbers measure different boundaries (global grid vs Europe, device-inclusive), so don't treat them as competing.

Smartphone

Ericsson lifecycle study (2015-era phone, 3-year life): 19 kg CO2e/yr device only; 62 kg CO2e/yr with allocated networks + data centers. modeled

See the math, step by step
4+ sources · how to verify
  1. Ericsson LCA (2016) — 57 kg over 3 years (19 kg/yr); 62 kg/yr including networks + data centers.
  2. Apple iPhone 17 Pro environmental report (2025) — 64–134 kg CO2e per device lifetime; production dominates.
  3. Deloitte (2022) — new phone ≈ 85 kg in its first year, 95% from manufacturing.
  4. Fraunhofer/Fairphone 2 LCA — 43.9 kg over 3 years, 82% production.
  5. ADEME/ARCEP (France, 2025) — 80.2 kg per phone, 99% manufacturing.
Find it yourself: search ericsson.com for "life cycle assessment of a smartphone" and read the summary paragraph. Every source agrees on the direction — making the phone matters far more than using it — but the level depends on the phone and its assumed lifetime, so treat 19 kg/yr as a 2015-era mid-range example, not a universal constant.

Driving (gasoline car)

8,887 g CO2 per gallon; average car ~400 g/mile at 22.2 mpg (tailpipe only). measured

See the math, step by step
4+ sources · how to verify
  1. US EPA, typical passenger vehicle — 8,887 g/gallon; ~400 g/mile; ~4.6 t/yr.
  2. EPA equivalencies calculator references — same coefficient, traced to the 2010 Federal Register rule and IPCC guidelines.
  3. US EIA CO2 coefficients — 8,780 g/gal (1.2% lower; 8,100 g/gal if ethanol is counted as non-emissive).
  4. fueleconomy.gov (DOE/EPA) — derives ~20 lb CO2 per gallon from first principles: carbon atoms grab oxygen when fuel burns.
Find it yourself: open epa.gov/greenvehicles → "Greenhouse Gas Emissions from a Typical Passenger Vehicle" and read the bullets under "How much tailpipe CO2 is created from burning one gallon of fuel?" — then compare EIA's motor-gasoline row to see how the ethanol convention moves the number.

Home heating

Heating is the single biggest slice of US home energy — 42% measured. A typical gas-furnace home: ~2,500 kg CO2/yr (published range 2,000–4,000; big cold-climate homes reach 8,000 CO2e). modeled

See the math, step by step
4+ sources · how to verify
  1. EIA RECS 2020 — heating = 42% of US residential energy; heating + AC together = 52%.
  2. EPA household calculator assumptions — average gas household: 3.11 t CO2/yr for all gas, ~63% of it space heating (~2 t).
  3. Columbia University QSEL — 3.2 t CO2e per 1,000 sq ft; a 2,500 sq ft home ≈ 8 t CO2e/yr.
  4. UC Davis WCEC + Energy Policy (2022) — heat pump vs gas furnace: −38–53% CO2 (US average across 99 cities); −44–60% counting methane and refrigerants (100-yr GWP).
  5. NRDC + KQED — independent restatements of the same bands.
Find it yourself: at eia.gov search "RECS 2020 press release" and read the first paragraph for the 42%. Then at wcec.ucdavis.edu search "heat pump greenhouse gas forecasts" — the three reduction bands (38–53%, 53–67%, 44–60%) are listed in sequence; each counts different gases.

Air conditioning

AC uses 19% of US home electricity measured — about 2,050 kWh/yr for an average home, ≈ 790 kg CO2e on the US grid. derived

See the math, step by step
4+ sources · how to verify
  1. EIA FAQ — AC = 19% (254 billion kWh) of US home electricity in 2020.
  2. EIA FAQ — average US home bought 10,791 kWh in 2022.
  3. Our World in Data — world AC ≈ 2,100 TWh in 2022, ~7% of world electricity.
  4. IEA space cooling tracker — 9% of final electricity consumption (same year, different denominator) + Ember — US grid ≈ 384 g CO2e/kWh (2025), used in our derivation.
Find it yourself: at eia.gov search FAQ "how much electricity is used for cooling" for the 19%; multiply 0.19 × 10,791 kWh yourself to get ~2,050 kWh — then × 384 g/kWh from Ember's "major countries" chapter. Notice how 7% vs 9% for world AC depends on which electricity total you divide by.

Beef burgers

One quarter-pound US beef patty ≈ 4 kg CO2e (3.7–4.3 across US LCAs) — mostly methane from cattle, not electricity. modeled

See the math, step by step
4+ sources · how to verify
  1. University of Michigan CSS (2018) — 3.7 kg CO2e per quarter-pound US beef patty (Table ES1).
  2. Putman/Rotz/Thoma 2023 US beef benchmark (via LCA) — 4.26 kg per patty; 36.8 kg CO2e per kg of US ground beef.
  3. Our World in Data (Poore & Nemecek 2018) — global average ~60 kg CO2e per kg of beef (retail weight).
  4. FAO — 68 kg CO2e/kg carcass weight from beef herds vs 18 from dairy herds; global average 46.2.
Find it yourself: open the UMich CSS report CSS18-10 and read Table ES1, "beef patty" column. Warning we verified: you cannot multiply the global 60 kg/kg by a US patty's weight — that double-counts land-use change and gives ~6.8 kg; US-specific studies say 3.7–4.3 kg.
Bar scale
AI prompts
Streaming

source modeled
Smartphone

source modeled
Driving

source measured
Home heating

source modeled
Air conditioning

source derived
Beef burgers

source modeled
These seven choices together (not your whole footprint):

Boundaries: AI figures exclude model training and your device (per each source's scope). Streaming's footprint is mostly your own equipment — for a European streaming hour, devices carry 51%, the home router 38%, the network 10%, and data centers roughly 1% Carbon Trust 2021:

devices 51%
router 38%
10%
end-user deviceshome routernetworkdata centers ≈1%

Derived CO2 uses the global average grid intensity of 458 g CO2e/kWh (2025) for AI prompts and the US average of 384 g CO2e/kWh for AC Ember Ember (US) — except the Google basis, which reports its own market-based 0.03 g CO2e/prompt.

The giants — where the world's electricity really goes

Imagine all the electricity generated on Earth in a year as 100 squares. The world made about 31,000 TWh in 2024 Ember Energy Institute IEA Our World in Data — so each square is about 310 TWh.

Air conditioning & space cooling ≈ 7 squares~2,100 TWh (2022) · OWID IEA Carbon Brief
Steel industry ≈ 4 squares~1,230 TWh electricity (2019) — plus far more energy as coal · IEA roadmap
Aluminum smelting ≈ 3 squares~900–1,000 TWh · Ember EC JRC WEF
ALL data centers ≈ 1⅓ squares415 TWh (2024), and AI is only part of that · IEA Nature EU Commission S&P Global
Honest chart rules: the sectors use different data years (2019–2024) because that's when each was last measured well — and the squares don't all measure the same thing (cooling includes some fans; steel's electricity is only a quarter of its energy, the rest is mostly coal). Even so, the picture holds: everything on this chart is bigger than all of the world's data centers combined — and the IEA projects data centers roughly doubling to ~950 TWh (≈3 squares) by 2030 IEA 2026.

The stuff around you — heavy industry's share

These don't fit a personal slider, but they dwarf the digital world. Each fact below carries its 4+ sources.

Steel measured

~7–9% of global CO2 and ~8% of the world's final energy. Every tonne of steel ⇒ 1.92 t CO2 (2023). An average car holds ~900 kg of steel — so a new car carries roughly 1.7 t of CO2 in its steel alone derived: 0.9 t × 1.92.

4+ sources · how to verify
  1. worldsteel, World Steel in Figures 2025 — 1.92 t CO2 and 21.27 GJ per tonne (2023).
  2. worldsteel steelFacts — 7–9% of direct emissions from global fossil-fuel use.
  3. IEA Iron & Steel roadmap — 2.6 Gt CO2 = 7% of energy-system CO2; 8% of final energy (845 Mtoe ≈ 35 EJ).
  4. IRENA + EIA — ~1.9 t/t and ~7% share, independently.
  5. worldsteel fact sheet + WEF — 900 kg of steel per average vehicle.
Find it yourself: worldsteel.org → Data → "World Steel in Figures 2025" → find the CO2-per-tonne line. Trap we caught: 21.27 is gigajoules per tonne, not the sector's total energy — that total is ~35 EJ (IEA roadmap, p. 37).

Cement modeled

The most-used man-made material: ~7–8% of global CO2 (peer-reviewed range 5–8%) — over half of it from the chemistry of making clinker, which no clean electricity can fix.

4+ sources · how to verify
  1. Chatham House — ~8% of global CO2.
  2. GCCA (industry body) — "approximately 7%".
  3. Nature Communications (2023) — 5–8%; 2,059 Mt CO2 in 2018, 66% from process chemistry.
  4. Annual Reviews (2024) + BBC — independent 5–8% and ~8% restatements.
Find it yourself: search gccassociation.org for "net zero progress report" and find "7% of global CO2"; then read the Nature Communications introduction (s41467-023-43660-x) for the peer-reviewed 5–8% range and why the boundary (cement vs all concrete) moves the number.

Aluminum modeled

Smelting aluminum is electricity turned into metal: ~900–1,000 TWh a year, roughly 3–4% of world electricity — each tonne needs ~14 MWh, and in 2019 coal supplied 64% of smelter power.

4+ sources · how to verify
  1. Ember — 918 TWh for electrolysis in 2019, 64% coal-powered.
  2. European Commission JRC — 13.2–14.1 MWh per tonne (citing International Aluminium Institute data).
  3. World Economic Forum — ~4% of global power consumption.
  4. University of Auckland (2026) — ~14 kWh/kg, ~4% of world electricity demand.
  5. International Aluminium Institute — the primary statistical series (interactive chart).
Find it yourself: search ember-energy.org for "aluminium's dirty little secret" and read the sentence with "918TWh"; then open international-aluminium.org → Statistics → "Primary Aluminium Smelting Power Consumption", set region to World and read the latest annual value.

Sector scale — large, but not an outlier

Global data centers used about 415 TWh in 2024 — roughly 1.5% of world electricity IEA Energy and AI. On this axis AI is large but unremarkable. Read the boundary notes: these bars deliberately do not share a definition, which is exactly why single-bar comparisons mislead.

Read the boundaries before the bars. "Data centers" ≠ "AI" (all workloads, all operators). The GSMA 300 TWh covers operators' entire electricity — fixed + mobile networks, data centers, offices; mobile access networks alone are roughly 135–180 TWh GSMA Mobile Net Zero 2026. Cooling includes some fans; steel's electricity is a quarter of its energy (the rest is mostly coal); the dairy figure is on-farm only Mohsenimanesh et al. 2021; the aviation figure is a European-departures decarbonization scenario for 2050 — proof that clean-energy transitions compete with AI for the same clean electrons Destination 2050 (2025).

Where AI genuinely is different: density, concentration, velocity

A globally modest percentage can still be a local grid, water and ratepayer emergency. The defensible claim is not that AI uses uniquely vast energy — it is that AI's demand concentrates in specific places and grows faster than the institutions watching it.

7–8×
A generative-AI training cluster "might consume seven or eight times more energy than a typical computing workload" — Noman Bashir, MIT expert estimate
MIT News 2025
23%
Share of Ireland's metered electricity used by data centers in 2025 — up from 5% in 2015 measured
CSO RTÉ Irish Times FactCheckNI
5 states
US states above 10% of electricity going to data centers (2023): Virginia 25.6%, N. Dakota 15.4%, Nebraska 11.7%, Iowa 11.4%, Oregon 11.4% modeled
EPRI 2024 LBNL
+17%
Growth of global data-center electricity demand in 2025 alone — versus a projected doubling to ~950 TWh by 2030 scenario
IEA 2026 update
The balancing fact: even so, data centers account for less than 10% of global electricity demand growth between 2024 and 2030 in the IEA Base Case — air conditioning, industry and EVs each add more IEA Energy and AI. And mobile networks scaled alongside a felt consumer benefit; AI's infrastructure spending precedes its felt benefit by years. Identical energy numbers, very different politics.

The water ledger — one distinction carries the whole debate

Water withdrawn is taken from a source and mostly returned; water consumed is evaporated or otherwise removed from the local environment USGS glossary. Nearly every alarming AI-water headline blurs the two. Toggle below — same kilowatt-hour, ~36× apart.

13.9 gal
freshwater withdrawn per kWh of US electricity
36 drops withdrawn…
4+ sources · how to verify
  1. Argonne National Laboratory (2010) — 13.9 gal withdrawn / 0.39 gal consumed per kWh (2005 mix).
  2. Macknick et al., ERL (2012) + NREL review — the standard per-technology water factors and the withdrawal/consumption definitions.
  3. USGS (2019) — measured 2015 ratio ~38:1 withdrawal:consumption, matching Argonne's ~36:1.
  4. US EIA (2018) — withdrawal intensity fell from 15.1 (2014) to 13.0 gal/kWh (2017) and keeps falling.
Find it yourself: search publications.anl.gov for "ANL/ESD/11-2" and read the executive summary sentence beginning "On average, 13.9 gallons…". Note the vintage: it's a 2005-era mix; EIA's 2017 figure is ~13.0 gal/kWh. Watch the unit trap — most studies publish gallons per MWh (1,000× bigger numbers).
ClaimWhat the source actually says
4.2–6.6 bn m³ by 2027Global AI demand's projected water withdrawal — more than 4–6× Denmark's annual total. Often misquoted as "consumption". modeled Li et al., "Making AI Less Thirsty"
9.3 trillion L by 2030Water footprint of the electricity generation powering data centers — equal to the basic annual domestic water needs of Sub-Saharan Africa's 1.3 billion people. scenario UNU-INWEH 2026
0.26 mL per promptGoogle's measured median water consumption per Gemini text prompt (May 2025) — about five drops. Critics note it excludes the water behind the electricity. measured Google 2025 The Verge

The other ledgers — kept separate on purpose

Low-carbon electricity is not automatically low-water or low-land. Collapsing these into one score destroys exactly the information a governance conversation needs.

Land scenario

The land footprint of the electricity supplying data centers is projected to exceed 14,500 km² by 2030 — roughly twice metropolitan Jakarta UNU-INWEH.

E-waste modeled range

Generative-AI e-waste could reach 0.4–2.5 Mt/yr by 2030 with no mitigation — 2.5 is the aggressive top of the range, and circular strategies could cut it 16–86% Wang et al., Nature Comput. Sci.. The viral "16 Mt" figure came from an unrefereed preprint.

Avoided emissions scenario

Existing AI applications could cut 1.4 Gt CO2 in 2035 in the IEA's exploratory Widespread Adoption Case — but "there is currently no momentum" ensuring it, and AI "is not a silver bullet" IEA.

And the benefits column — for everything on the table

A fair ledger records why we chose these technologies at all. AI's benefit line is the only one still being written.

TechnologyThe positive we boughtThe cost we accepted
Mobile5.8 billion unique subscribers — 70% of humanity connected; ~$6.5 trillion (5.8%) of global GDP GSMA Intelligence~300 TWh of operator electricity (2024, all operations) GSMA
StreamingBeats the store-rental era it replaced — a 2014 LCA put store-rented DVDs at 0.71 kg CO2e/viewing-hour vs 0.42 for streaming Shehabi et al. 2014~36–55 g CO2e per hour, mostly in your own living room IEA
Warm & cool homesShelter and survivable summers — cooling alone is nearly 20% of buildings' electricity worldwide OWIDHeating ≈ 42% of US home energy EIA; world AC ≈ 2,100 TWh, ~1 Gt CO2 Carbon Brief
Beef & livestockLivelihoods of at least 1.3 billion people; 34% of global food protein FAOLivestock ≈ 6.2 Gt CO2e/yr (~12% of emissions); one US beef patty ≈ 4 kg CO2e FAO GLEAM UMich
Steel & cementEvery bridge, hospital, school and home — the built world itselfTogether ~14–17% of global CO2 IEA Nature Comms
Cars3.29 trillion vehicle-miles of American mobility in 2024 FHWA~4.6 t CO2 per typical vehicle per year EPA
AviationGlobal connectivity at 2.5% of CO2 emissions Our World in DataIts net-zero path needs 762 TWh of clean electricity by 2050 (Europe alone) Destination 2050
AIA scenario, not yet a receipt: up to 1.4 Gt CO2 avoided in 2035 if adoption is widespread IEA415→~950 TWh (2024→2030 Base Case), concentrated in a handful of grids IEA

Method — standards we inherit, not invent

This artifact adopts existing legitimacy chains rather than asking you to trust a new formula. And every fact was re-verified against its sources on 12–13 August 2026 by independent research passes.

SCI — ISO/IEC 21031:2024

The Software Carbon Intensity spec: SCI = (E × I + M) per R — energy × grid intensity + embodied carbon, per an explicitly chosen functional unit. Now an ISO standard. GSF spec ISO

MLPerf Power

The only measured, boundary-defined AI energy benchmark — "from microwatts to megawatts," training and inference, 1,841 reproducible measurements. arXiv:2410.12032 MLCommons

Grid data

EPA eGRID (US regional factors, 2023 data) EPA; global average 458 g CO2e/kWh, US 384 g (2025) Ember; per-workload telemetry via CodeCarbon GitHub.

Why the grid you plug into matters more than the query you type

Median published lifecycle emissions per kWh of electricity generation NREL LCA Harmonization:

Nuclear
12 g
Wind
13 g
Solar PV
43 g
Natural gas
486 g
Coal
1,001 g
Editorial rules of this artifact: six separate ledgers, mandatory functional units, withdrawal and consumption never conflated, scenarios visually distinct from observations, vendor studies labeled as vendor studies, no default composite score — and every fact backed by at least four sources, each with a "find it yourself" trail. Rights-related and irreversible harms are never netted against benefits. Weighing the ledgers is a values exercise — this page refuses to do it for you.
Data Centers, AI & Social Trade-offs — published by the AiGovOps Foundation in the Library · August 2026.
Companion: For more reading — every source, every caveat, and the most common misquotes · ← back to the Library
All figures verified against the linked primary sources on 12–13 August 2026. Corrections welcome — file an issue in the Library repo.