How we know

What the verdicts mean

Each answer to a claim gets one of four verdicts. A verdict can’t be stronger than its evidence: if only a company says so, the answer is labelled “company figures”.

✓ YES

The evidence supports the claim as people usually say it.

◐ PARTLY

True in part, in some places, or only for older figures. The answer says which part.

✕ NO

The evidence contradicts it, even at the highest published estimate.

? NOBODY KNOWS

No one has published enough to say. The answer says what’s missing.

Labels on every figure: Measured · Estimated · Calculated by Hum (our arithmetic, shown under “Hum’s sums”, and checked by the build) · Company says so (not independently checked) · Forecast · Data row (copied from a published dataset and checked against the file).

Both sides: every answer that says your own use is small also gives the total, because one question is tiny and all of them together are not. Figures that measure different things are never ranked against each other.

Kept fresh: every Monday an automatic check downloads every page Hum quotes and confirms each quote is still there. A quote that changed or vanished is flagged for a person to re-read, and the answers that use it are updated.

Your privacy on the cards: your guesses and score stay on your device. Nothing you choose is sent to us.

The numbers

Nobody outside the AI companies can measure a single question exactly. So we start from what companies have published, add independent research, and always keep a low, middle and high estimate. Every number on the main page shows the whole range.

AI's share of all data-centre electricity is the United States' 2024 share, 26.5–38%, used as our estimate for the world: nobody publishes a world figure in words. The real world share is probably lower.

Constants version 11 · cf084a94471f042d · built 2026-10-25

How it grows, year by year

Sources publish only a few years: what data centres used in 2024 and 2025, and forecasts for 2030 and 2035. We fill in the years between with plain rules and mark every filled-in number as calculated. Where a source gives no low estimate, we say so instead of inventing one. Years after 2025 are forecasts.

Electricity used by the world's data centres (TWh/yr)

• 2017–2023: Calculated back from 2024 at about 12% growth a year
• 2024–2025: Published figure (single estimate)
• 2026–2029: Calculated: steady growth between the 2025 and 2030 figures; range drawn as a straight line between published ranges
• 2030: Published forecast; range drawn as a straight line between published ranges
• 2031–2034: Calculated: steady growth between the 2030 and 2035 figures; range drawn as a straight line between published ranges
• 2035: Published forecast

6 sources

• IEA, Energy and AI (2025): "data centre electricity consumption has grown by around 12% per year since 2017" · low,mid,high · 2026-10-03
• IEA, Energy and AI (2025): "Data centres accounted for around 1.5% of the world's electricity consumption in 2024, or 415 terawatt-hours (TWh)." · low,mid,high · 2026-10-03 (single published estimate)
• IEA, Key Questions on Energy and AI (2026): "electricity consumption from data centres roughly doubling from 485 TWh in 2025 to 950 TWh in 2030" · low,mid,high · 2026-10-03 (single published estimate)
• IEA, Key Questions on Energy and AI (2026): "from 485 TWh in 2025 to 950 TWh in 2030" · mid · 2026-10-03 (base case; the IEA gives no 2030 range in text)
• IEA, Energy and AI (2025): "Our Base Case sees global data centre electricity consumption rising to around 1 200 TWh by 2035." · mid · 2026-10-03
• IEA, Energy and AI (2025): "By 2035, the range of data centre electricity demand across our cases spans from 700 to 1 700 TWh." · low,high · 2026-10-03

CO2 from the electricity data centres use (Mt CO2/yr)

• 2017–2023: No published figure
• 2024: Published figure (single estimate)
• 2025–2034: Calculated: steady growth between the 2024 and 2035 figures; no published low
• 2035: Published forecast; no published low

3 sources

• IEA, Energy and AI (2025): "Emissions from electricity use by data centres grows from 180 million tonnes (Mt) today to 300 Mt in the Base Case by 2035, and up to 500 Mt in the Lift-Off Case." · low,mid,high · 2026-10-03 ("today" in a 2025 report; dated 2024 here, single published estimate)
• IEA, Key Questions on Energy and AI (2026): "The emissions associated with data centres double in IEA projections, reaching around 350 million tonnes in 2035" · mid · 2026-10-03
• IEA, Energy and AI (2025): "... and up to 500 Mt in the Lift-Off Case." · high · 2026-10-03 (no low case published in text)

Series ec5d0aaaabaf90fc · built 2026-10-03

The countries we compare with

Each country's total electricity use in 2025 (demand, not generation), copied from one dataset file. A check on every build downloads the file again, confirms it is the same file, and compares every number.

• Ember yearly electricity data (global) · CC-BY-4.0 · retrieved 2026-10-31
• File: release_generation_yearly_global.csv · sha256 c3bd8b3d18824dd5… · rows where "Electricity source" is "Demand"

Grids in the 2030 game

On the 2030 page you choose which grids power the world's data centres. Each grid's carbon intensity is a published figure for one year; real grids keep changing, mostly getting cleaner.

Data-centre places

Regions with many data centres, and their power capacity. Published figures measure different things, and we say which for each one.

Named cases and claims

Some effects have no worldwide number yet, so the main page explains them with real, named cases. Every case quotes its source, read on the page itself, and every number in our summary appears in one of those quotes (the build checks this). The "claims to read twice" name practices, never accuse a company.

Strain on the grid · Ireland and the eastern United States

In Ireland, data centres used 22% of all metered electricity in 2024, up from 5% in 2015. In the eastern US, a 5,250 MW increase in the PJM grid's demand forecast, almost entirely from data centres, helped drive the results of its latest capacity auction.

2 sources

• Central Statistics Office Ireland, 10 June 2025: "The percentage share of metered electricity consumption used by data centres rose to 22% in 2024 from 5% in 2015" · case · 2026-10-11
• Utility Dive, 18 December 2025: "In part, the latest auction results were driven by a 5,250-MW increase in PJM's demand forecast — almost entirely driven by data centers" · case · 2026-10-11

Local air · Memphis, Tennessee

In Memphis, the county's permit lets xAI operate 15 generators at its data centre, up to 247 megawatts; TechCrunch describes them as natural gas generators. According to their manufacturer, the turbines emit nitrogen oxides, a key contributor to smog, and formaldehyde.

2 sources

• TechCrunch, 3 July 2025, ‘xAI gets permits for 15 natural gas generators at Memphis data center’: "The permit issued by the Shelby County Health Department says that xAI can operate 15 Solar SMT-130 generators with certain emissions controls, which can generate up to 247 megawatts." · case · 2026-10-11
• NBC News, 17 June 2025: "The turbines emit nitrogen oxides, a key contributor to smog, and formaldehyde, among other pollutants, according to their manufacturer." · case · 2026-10-11

Chips and e-waste · Worldwide

A peer-reviewed study of generative AI's e-waste finds it could potentially reach 1.2 to 5.0 million tons in total over 2020 to 2030.

1 source

• Nature Computational Science, 28 October 2024, ‘E-waste challenges of generative artificial intelligence’: "this e-waste stream could increase, potentially reaching a total accumulation of 1.2–5.0 million tons during 2020–2030" · case · 2026-10-11

Land · Prince William County, Virginia

The planned Digital Gateway would span 34 buildings and 1.7 gigawatts of data-centre capacity on 2,139 acres near Manassas National Battlefield; WTOP reported it would be the largest data-centre corridor in the world. A court found the county's advertised notice did not comply with either the state or county code.

3 sources

• Bisnow, 8 April 2026: "Within the 2,139-acre Digital Gateway, Blackstone subsidiary QTS and Compass Datacenters have planned to develop large data center campuses that at full build-out would span 34 buildings, 22M SF and 1.7 gigawatts of capacity." · case · 2026-10-11
• WTOP, 9 August 2025: "would be the largest data center corridor in the world, with 37 data centers spread out across over 2,100 acres in western Prince William, near Manassas National Battlefield" · case · 2026-10-11 (Its building count (37) differs from Bisnow's newer 34; Hum uses Bisnow's.)
• WTOP, 9 August 2025, quoting the court: "This Court finds that the advertised notice provided by the County did not comply with either the State or County code." · case · 2026-10-11

Noise · Chandler, Arizona and Prince William County, Virginia

Residents near data centres in Prince William County say the constant drone, especially at night, causes anxiety and deprives them of sleep. Chandler, Arizona requires data centres to be built so their sound doesn't exceed the background level measured in a baseline study.

2 sources

• Prince William Times, 27 February 2023: "But residents say the constant drone, especially at night, causes anxiety and deprives them of sleep." · case · 2026-10-11
• City of Chandler, Ordinance No. 5033, 8 December 2022: "The Data Center must be designed and built to incorporate sound mitigation methods sufficient to prevent the sound levels emanating from the Data Center ... from exceeding the ambient noise levels that were observed in the baseline study." · case · 2026-10-11 (The ... replaces "(as determined by a third-party acoustic engineer)", which runs across a page break.)

Claims to read twice

Buying certificates, not clean power

Renewable-energy certificates are used to calculate "market-based" emissions, the official figures firms report. Researchers estimate that if the trend continues, 42% of committed scope 2 (electricity) emission reductions will not result in real-world mitigation.

2 sources

• Nature Climate Change, 9 June 2022: "If this trend continues, 42% of committed scope 2 emission reductions will not result in real-world mitigation." · case · 2026-10-11
• The Guardian, 15 September 2024: "Recs are used to calculate "market-based" emissions, or the official emissions figures used by the firms." · case · 2026-10-11

Matching over a year, not hour by hour

A peer-reviewed model found that matching a year's electricity use with clean power has minimal impact on long-run emissions. Matching it hour by hour keeps contributing significantly to cutting them, even as grids get cleaner.

2 sources

• Joule, 21 February 2024 (Princeton): "procurement strategies that match participants' demand with carbon-free generation on an annual basis have minimal impact on long-run system-level CO2 emissions" · case · 2026-10-11 (The page writes CO2 with a subscript ("CO 2").)
• Riepin and Brown, TU Berlin, arXiv 2023: "Even as grids become cleaner over time, the hourly matching strategy contributes significantly to system-level emissions reduction." · case · 2026-10-11

"Carbon neutral" through offsets

A review estimates that less than 16% of the carbon credits issued to the projects it investigated were real emission reductions. A new EU directive will ban product claims of a neutral, reduced or positive environmental impact based on offsetting.

2 sources

• Nature Communications, 14 November 2024: "We estimate that less than 16% of the carbon credits issued to the investigated projects constitute real emission reductions" · case · 2026-10-11
• European Parliament, 17 January 2024: "the directive will ban claims that a product has a neutral, reduced or positive impact on the environment because of emissions offsetting schemes" · case · 2026-10-11

"Water positive"

One company's water stewardship goal left out the water consumed by the power plants that supply its data centres. And as a former water sustainability manager put it, carbon is a global problem, but water is more localised.

2 sources

• Grist, 29 August 2024: "the company's water stewardship goal doesn't include the water consumed by the power plants that supply its data centers" · case · 2026-10-11
• SourceMaterial, 9 April 2025, quoting a former water sustainability manager at a cloud company: "Carbon is a global problem—water is more localised" · case · 2026-10-11

Leaving out the chips and the buildings

Embodied emissions can exceed 40% of a data centre's total greenhouse gas emissions, so a footprint that leaves them out can miss a large part.

1 source

• ICEF Sustainable Data Centers Roadmap, October 2025: "embodied emissions can exceed 40% of a data center’s total greenhouse gas emissions and may dominate the lifetime greenhouse gas emissions footprint" · case · 2026-10-11

Efficiency, not the total

Researchers warn that because of rebound effects, efficiency alone won't ensure lower total harm. US data-centre electricity use climbed from 58 TWh in 2014 to 176 TWh in 2023.

2 sources

• Luccioni, Strubell and Crawford, arXiv January 2025: "Rebound effects undermine the assumption that improved technical efficiency alone will ensure net reductions in environmental harm." · case · 2026-10-11
• Lawrence Berkeley National Laboratory, 15 January 2025: "The total data center electricity usage climbed from 58 TWh in 2014 to 176 TWh in 2023" · case · 2026-10-11

We changed our numbers

When better information comes out, we update. Every change is listed here, newest first. Numbers are low / middle / high.

1. 2026-10-25 · version 11

   • Sources updated; no numbers changed.

2. 2026-10-12 · version 10

   • CO2 per kWh generated from coal (United States, 2023): 2.31 / 2.31 / 2.31 (first version)
   • CO2 per kWh generated from natural gas (United States, 2023): 0.96 / 0.96 / 0.96 (first version)
   • CO2 per kWh generated from petroleum (United States, 2023): 2.46 / 2.46 / 2.46 (first version)

3. 2026-10-03 · version 9

   • Water in an Olympic-size swimming pool: 2.5 million / 2.5 million / 2.5 million (first version)

4. 2026-10-03 · version 8

   • Data-centre electricity in 2035, IEA High Efficiency case: 970 / 970 / 970 (first version)

5. 2026-10-03 · version 7

   • Sources updated; no numbers changed.

6. 2026-10-03 · version 6

   • Growth in 2025 of electricity used by AI-focused data centres: 0.5 / 0.5 / 0.5 (first version)

7. 2026-10-03 · version 5

   • Electricity per short AI video (about 5 seconds): 30.3 / 944 / 944 → 25.3 / 90 / 944
   • CO2 per kWh of electricity: 125 / 473 / 473 → 125 / 473 / 560

8. 2026-10-02 · version 4

   • AI's share of data-centre electricity (United States, 2024): 0.265 / 0.32 / 0.38 (first version)

9. 2026-10-01 · version 3

   • Electricity per AI picture: 0.63 / 1.22 / 2.91 → 0.63 / 1.22 / 2.9
   • Water used at the AI building (cooling) per kWh: 0.12 / 1.08 / 1.8 → 0.12 / 0.36 / 1.08
   • Water used at power plants to make each kWh: 0.49 / 2.18 / 3.94 → 0 / 2.18 / 4.4
   • People in the world: 8.16 billion / 8.16 billion / 8.16 billion → 8.2 billion / 8.2 billion / 8.2 billion
   • The (v1) world-health scale was re-derived.

10. 2026-10-01 · version 2

    • Sources updated; no numbers changed.

11. 2026-10-01 · version 1

    • Electricity per text answer: 0.24 / 0.34 / 1.86 (first version)
    • Electricity per AI picture: 0.63 / 1.22 / 2.91 (first version)
    • Electricity per short AI video (about 5 seconds): 30.3 / 944 / 944 (first version)
    • Water used at the AI building (cooling) per kWh: 0.12 / 1.08 / 1.8 (first version)
    • Water used at power plants to make each kWh: 0.49 / 2.18 / 3.94 (first version)
    • CO2 per kWh of electricity: 125 / 473 / 473 (first version)
    • Electricity used by all data centres worldwide per year: 415 / 415 / 415 (first version)
    • People in the world: 8.16 billion / 8.16 billion / 8.16 billion (first version)
    • The (v1) world-health scale was re-derived.

Your privacy

• No accounts, no cookies, no ads, and no tracking across sites.
• We count visits with Cloudflare Web Analytics, run by Cloudflare, which already hosts this site. It sets no cookies and doesn't fingerprint you; it records the page, where you came from, your country and your kind of device, and only as totals. It is the one thing that loads from outside this site.
• The calculator runs in your browser. We never see what you enter.
• Hum Daily keeps your answers and streak on your device only (in your browser's storage). Clear your site data and they're gone; nothing is sent to us.
• Hum's first version saved a "world" on your device; the current site removes it the next time you visit.