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.
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.
Pick your evidence basis — the three published numbers use different models, boundaries and years, and are not directly comparable.
Energy: 0.077 kWh/hr modeled. Carbon basis:
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
8,887 g CO2 per gallon; average car ~400 g/mile at 22.2 mpg (tailpipe only). measured
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
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
One quarter-pound US beef patty ≈ 4 kg CO2e (3.7–4.3 across US LCAs) — mostly methane from cattle, not electricity. modeled
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:
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.
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.
These don't fit a personal slider, but they dwarf the digital world. Each fact below carries its 4+ sources.
~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.
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.
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.
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.
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.
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.
| Claim | What the source actually says |
|---|---|
| 4.2–6.6 bn m³ by 2027 | Global 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 2030 | Water 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 prompt | Google'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 |
Low-carbon electricity is not automatically low-water or low-land. Collapsing these into one score destroys exactly the information a governance conversation needs.
The land footprint of the electricity supplying data centers is projected to exceed 14,500 km² by 2030 — roughly twice metropolitan Jakarta UNU-INWEH.
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.
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.
A fair ledger records why we chose these technologies at all. AI's benefit line is the only one still being written.
| Technology | The positive we bought | The cost we accepted |
|---|---|---|
| Mobile | 5.8 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 |
| Streaming | Beats 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 homes | Shelter and survivable summers — cooling alone is nearly 20% of buildings' electricity worldwide OWID | Heating ≈ 42% of US home energy EIA; world AC ≈ 2,100 TWh, ~1 Gt CO2 Carbon Brief |
| Beef & livestock | Livelihoods of at least 1.3 billion people; 34% of global food protein FAO | Livestock ≈ 6.2 Gt CO2e/yr (~12% of emissions); one US beef patty ≈ 4 kg CO2e FAO GLEAM UMich |
| Steel & cement | Every bridge, hospital, school and home — the built world itself | Together ~14–17% of global CO2 IEA Nature Comms |
| Cars | 3.29 trillion vehicle-miles of American mobility in 2024 FHWA | ~4.6 t CO2 per typical vehicle per year EPA |
| Aviation | Global connectivity at 2.5% of CO2 emissions Our World in Data | Its net-zero path needs 762 TWh of clean electricity by 2050 (Europe alone) Destination 2050 |
| AI | A scenario, not yet a receipt: up to 1.4 Gt CO2 avoided in 2035 if adoption is widespread IEA | 415→~950 TWh (2024→2030 Base Case), concentrated in a handful of grids IEA |
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.
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
The only measured, boundary-defined AI energy benchmark — "from microwatts to megawatts," training and inference, 1,841 reproducible measurements. arXiv:2410.12032 MLCommons
Median published lifecycle emissions per kWh of electricity generation NREL LCA Harmonization: