02 Water · U.S. + Global
Water: U.S. residential outdoor use vs. AI data centers
American lawns drink 35 to 70 times what data centers use to cool themselves.
Compared to What? AI and Water, Electricity, & CO2
An interactive atlas of AI's environmental impact in scale, context, and comparison.
AI's environmental footprint is real, growing, and often described without enough scale. This atlas sets AI and data centers beside familiar systems — homes, roads, golf courses, cattle, aviation, Bitcoin, streaming, gaming, cement, steel, and the electric grid itself.
The question is not only “How large is AI's footprint?” but compared to what?
How to read
this atlas
this atlas
Most numbers are ranges, not single values.
Boundary choices change the answer.
Sources & methods are linked on every plate.
Last verified
May 16, 2026
Water: Data Centers vs. Golf Courses
Direct cooling alone makes data centers look tiny. Count the water used to generate their electricity and they rise beside U.S. golf — the boundary is most of the argument.
Global golf courses (site estimate) 800–1,500
U.S. golf courses, 2024 531
Global data centers — including electricity-generation water, 2025 500–700
U.S. data centers — including electricity-generation water, 2023 211
Global data centers — direct cooling only, 2025 45–90
U.S. data centers — direct cooling only, 2023 17
02,000 billion gallons / yr
500–700 billion gallons / yr
Global data centers, counting power plants
The atlas
The Atlas of Comparisons
Open a plate to see the numbers, assumptions, boundaries, uncertainty, and sources behind each comparison.
How to read these plates
- Point estimate
- Range, with central tick
- Higher uncertainty
- Projection
- Boundary-sensitive
Most figures are ranges because the sources differ and the boundary matters. A range is not a weakness in the data; it is usually the honest shape of the answer.
02 
03 CO₂ · U.S. + Global
CO₂: U.S. driving vs. global data centers
Every car, truck and bus in the U.S. out-emits the world's data centers about eightfold.
04 Electricity · U.S. + Global
Electricity: U.S. residential vs. global data centers
U.S. homes use more than three times the electricity of every data center on Earth.
05 CO₂ · Global
CO₂: Global commercial aviation vs. data centers
Global flights emit roughly five times what data centers do.
06 CO₂ · Global
CO₂: Global cement and steel vs. data centers
Making the world's cement and steel out-emits data centers more than twentyfold.
07 Electricity · Global
Electricity: Bitcoin mining vs. data centers
Bitcoin mining draws a third to a half of the whole data-center fleet's power.
08 Electricity · Global
Electricity: Global EV charging vs. data centers
EV charging already pulls about 40% of global data-center electricity, and climbing fast.
09 Electricity · Global
Electricity: Global video streaming vs. data centers
Streaming's annual electricity sits in the data-center range, but the estimate is shaky.
10 Electricity · Global
Electricity: Global video gaming vs. data centers
Global gaming may rival data centers, or sit at a quarter; the underlying data is weak.
11 CO₂e · Global
CO₂: Global cattle (beef and dairy) vs. data centers
The world's cattle out-emit data centers about twentyfold, mostly as methane.
12 Electricity · U.S. + Global
Electricity: U.S. residential air conditioning vs. global data centers
Cooling U.S. homes burns over half the world's data-center electricity each year.
13 CO₂ · U.S. + Global
Emissions: U.S. gas-powered lawn equipment vs. data centers
U.S. gas mowers and blowers emit about a sixth of the global data-center total, in far dirtier air.
14 Electricity · U.S. + Global
Electricity: U.S. holiday lighting vs. global data centers
A month of U.S. holiday lights is about 1.4% of the year-round data-center fleet.
15 CO₂ · U.S. + Global
CO₂ per hour: driving vs. AI, gaming & streaming
An hour of driving emits hundreds to thousands of times an hour of AI chat.
Supporting material
Field Notes on Scale
Short lessons on how to read the comparisons — the year, the water, the trajectory, the translations.
I
The Hour
An hour of driving outweighs an hour of AI chat hundreds to thousands of times over.
II
The Year
Over a year, AI is still the shortest bar beside homes, Bitcoin, and EV charging.
III
The Water
Direct cooling, or counting the power plants — the boundary changes the answer.
IV
The Trajectory
Flat through 2018, then bent upward as AI scaled toward 2030.
V
In Equivalents
What a terawatt-hour buys, translated into U.S. household-years.
VI
Training vs. Inference
One training run, once; a year of inference, a thousand times heavier.
Read the ledger: every figure has a derivation, a source trail, an uncertainty note, and a last-verified date.
Comparison
Water: AI data centers vs. U.S. golf courses
U.S. golf courses applied roughly 531 billion gallons of water in 2024. The global figure is harder to pin down — no clean primary source gives a total — but extrapolating from U.S. per-course averages and the R&A's count of about 38,860 courses worldwide puts it somewhere in the 800 to 1,500 billion gallon range. Every data center on the planet uses 45–90 billion gallons directly for cooling, and 500–700 billion when you also count the water used to generate their electricity. The U.S. data center share is about 17 billion direct and 211 billion indirect (LBNL, 2023). Which framing is fair depends on which boundary you draw — and the boundary is most of the argument.
Boundary note: the answer depends on where you draw the line — whether only direct use is counted, or the upstream water and power are included too.
Billion gallons per year
See this plate in the full archive →
Comparison
Water: U.S. residential outdoor use vs. AI data centers
American households spray roughly 3.2 trillion gallons of water on their lawns, gardens, and driveways every year — somewhere between 35 and 70 times what every data center on the planet uses directly for cooling, and roughly five times the broader figure that counts the water used to generate data-center electricity. The EPA estimates about half of that outdoor water is wasted to overwatering. Lawns are the largest irrigated crop in the United States by area; data centers are not yet close.
Boundary note: the answer depends on where you draw the line — whether only direct use is counted, or the upstream water and power are included too.
Billion gallons per year
See this plate in the full archive →
Comparison
CO₂: U.S. driving vs. global data centers
U.S. on-road transportation — every car, truck, and bus on American roads — emitted about 1,440 million metric tons of CO₂ in 2022. That's about eight times what every data center on the planet emitted in 2024 combined. The hourly hero on the home page already shows driving outpacing any digital activity per hour; the annual ledger says the same thing in different units. AI's footprint is large enough to argue about; the cars in just one country still dwarf it.
Million metric tons CO₂ per year
See this plate in the full archive →
Comparison
Electricity: U.S. residential vs. global data centers
American homes used roughly 1,550 terawatt-hours of electricity in 2024 — about 38 percent of total U.S. consumption, and more than three times what every data center on the planet drew that year. The world's data-center electricity converts to roughly 40 million U.S. households' worth: around 30 percent of American homes. Heating and cooling houses still moves more electricity than the entire global compute layer, even as the layer grows.
Terawatt-hours per year
See this plate in the full archive →
Comparison
CO₂: Global commercial aviation vs. data centers
Global commercial aviation emitted about 950 million metric tons of CO₂ in 2023 — more than five times what every data center on the planet emitted the following year. Aviation is one of the better-tracked comparisons here, and one of the harder ones to abate. Even with data-center electricity projected to roughly double by 2030, aviation will likely remain comparable or larger.
Million metric tons CO₂ per year
See this plate in the full archive →
Comparison
CO₂: Global cement and steel vs. data centers
Producing the cement and steel for the world's buildings, bridges, and roads emits roughly 4,200 million metric tons of CO₂ a year — more than twenty times the global data-center fleet's emissions. Cement alone, around 1,600 Mt, dwarfs data centers almost nine times over. The familiar industries — concrete, beams, chimneys — still set the climate scale. AI is becoming visible inside it; it has not yet rearranged it.
Million metric tons CO₂ per year
See this plate in the full archive →
Comparison
Electricity: Bitcoin mining vs. data centers
Estimates of Bitcoin's annual electricity use cluster between 140 and 200 terawatt-hours — somewhere in the range of 30 to 43 percent of the global data-center fleet. Cambridge's direct survey of mining operators sits at the low end; their hashrate-based CBECI index and market-based estimators sit higher. Bitcoin is a single-purpose load, hard to abate by design: the network's security depends on burning electricity. Whatever you make of the trade-off, it is in the same scale conversation as data centers without being part of them.
Terawatt-hours per year
See this plate in the full archive →
Comparison
Electricity: Global EV charging vs. data centers
The world's electric vehicles drew about 180 terawatt-hours of electricity in 2024, up nearly 60 percent in a single year — nearly 40 percent of the global data-center figure. The IEA expects EV charging to grow past 2,000 TWh by 2035 in its main scenario. EVs and data centers are now the same kind of question for grid planners: where the new gigawatts land, and what generates them.
Terawatt-hours per year
See this plate in the full archive →
Comparison
Electricity: Global video streaming vs. data centers
Estimates of global video streaming electricity range from roughly 100 to 300 terawatt-hours a year, depending heavily on whether end-user devices, home Wi-Fi, and last-mile networks are counted on streaming's side of the ledger or attributed elsewhere. Per hour, an HD stream emits about 36 to 56 grams of CO₂ — below an hour of gaming or driving and roughly comparable to ordinary AI chat. The annual number stays unsettled because most of the energy lives in components shared with everything else: routers, CDNs, screens. Treat any 'streaming uses as much as a small country' headline with care; the IEA has spent years correcting the high end.
Terawatt-hours per year
See this plate in the full archive →
Comparison
Electricity: Global video gaming vs. data centers
Estimates of global gaming electricity sit in a wide and contested band — somewhere between 75 and 285 terawatt-hours a year, depending on whose assumptions you accept. The high end approaches the global data-center figure; the low end is closer to a quarter of it. Almost no number on this site has weaker underlying data; it earns its place mostly because the comparison gets cited so often. Take it as a reminder that 'video games use as much electricity as Country X' is a claim resting on a long chain of rough multipliers.
Terawatt-hours per year
See this plate in the full archive →
Comparison
CO₂: Global cattle (beef and dairy) vs. data centers
The world's cattle — beef and dairy combined — generate about 3,800 million metric tons of CO₂-equivalent emissions a year, mostly from methane belched by ruminants and from clearing land for pasture. That is about twenty-one times the data-center figure, and the comparison is one of the few here where 'global vs. global' lines up cleanly. Cows have a head start of about ten thousand years; on this scale they are still ahead.
Million metric tons CO₂-equivalent per year
See this plate in the full archive →
Comparison
Electricity: U.S. residential air conditioning vs. global data centers
American homes used about 254 terawatt-hours of electricity for air conditioning in 2020 — about 55 percent of the entire global data-center electricity figure for 2024. AC use rises sharply in summer and is, in many U.S. regions, the single biggest driver of grid stress on hot afternoons. A seasonal residential load in one country sits in the same league as the year-round, planet-wide computing fleet.
Terawatt-hours per year
See this plate in the full archive →
Comparison
Emissions: U.S. gas-powered lawn equipment vs. data centers
Gas-powered lawn and garden equipment in the United States — mowers, blowers, trimmers — emitted about 30 million metric tons of CO₂ in 2020, plus large shares of the country's smog-forming pollutants. The CO₂ figure is around 17 percent of every data center on the planet combined. The bigger story is air quality: gas mowers and blowers account for nearly one in six U.S. volatile-organic-compound emissions and more than one in ten nitrogen-oxide emissions. AI's grid load gets headlines; gas mowers do quieter, dirtier work in the same neighborhoods.
Million metric tons CO₂ per year
See this plate in the full archive →
Comparison
Electricity: U.S. holiday lighting vs. global data centers
American holiday lighting consumes roughly 6.6 terawatt-hours of electricity each December — more than the entire annual electricity supply of El Salvador and several smaller nations. The figure dates to a 2008 Department of Energy estimate and is probably lower today as LED string lights have largely replaced incandescent bulbs. The U.S. still spends as much electricity on a month of yard ornaments as some countries spend running everything for a year. Compared with the year-round, planet-wide data-center fleet, it is a footnote — about 1.4 percent.
Terawatt-hours per year
See this plate in the full archive →
Comparison
CO₂ per hour: driving vs. AI, gaming & streaming
An hour of gasoline driving emits 14,000-18,000 grams of CO₂. An hour of ordinary AI chat emits 5-200 grams. Heavy AI workflows, high-end gaming, and HD streaming all cluster between those two -- and far closer to chat than to driving. The big numbers live at data-center scale: grid load, water draw, where new construction lands and what local utilities can carry. That's where the rest of the site looks.
Grams CO₂ per hour
See this plate in the full archive →
Field note I · The Hour
An hour of driving vs. an hour of AI
An hour of gasoline driving emits 14,000-18,000 grams of CO₂. An hour of ordinary AI chat emits 5-200 grams. Heavy AI workflows, high-end gaming, and HD streaming all cluster between those two -- and far closer to chat than to driving. The big numbers live at data-center scale: grid load, water draw, where new construction lands and what local utilities can carry. That's where the rest of the site looks.
Field note II · The Year
Annual electricity, in context
AI gets the loudest headlines. The data-center footprint it sits inside is already larger than Bitcoin or EV charging, and the whole compute layer, taken together, still draws about a quarter of what U.S. homes spend keeping the lights and appliances on for a year. The AI slice itself is one of the chart's softer numbers — there is no consensus yet on what counts as 'AI' inside a hyperscale facility — but even its high end leaves room above for everything else data centers do.
Field note III · The Water
Where you draw the boundary
U.S. golf courses use about 531 billion gallons of water a year. Globally — extrapolating per-course averages to the R&A's count of roughly 38,860 courses — golf probably uses 800 to 1,500 billion. The water that flows through data center cooling loops is much smaller: 45–90 billion globally, about 17 billion in the U.S. Add the water used to generate the electricity those data centers run on, and the global figure jumps to 500–700 billion (211 billion in the U.S.). Which framing makes AI look thirsty depends on which boundary you draw — and the boundary is most of the argument.
Field note IV · The Trajectory
Data centers, 2017 to 2030
The flat era ran from about 2017 to 2022 — efficiency gains roughly canceled growth, and global data centers held around 200 to 290 TWh per year. Then the line bent. By 2024 it was 460 TWh; in 2025 it was 485, a 17 percent jump in a single year. The IEA's base case puts 2030 at 950 TWh, with its lower scenario at 830 and Goldman Sachs' April 2026 update reaching 1,350 at the upper edge. The shaded band shows that spread; the central line is the IEA base case. Most of the new load is hyperscale build-out, and most of that is AI-adjacent.
Field note V · In Equivalents
Terawatt-hours, in household-years
A U.S. household burns about 11.9 megawatt-hours of electricity a year. That makes every terawatt-hour of grid load roughly 84,000 households' worth. The world's data centers in 2024 ran on 460 TWh — about 39 million American homes. By 2030 the IEA range puts the global figure between 830 and 1,350 TWh: 70 to 113 million households. The yardstick has limits. Data centers run flat all year; residences are peaky, summer-and-evening loads. And 11.9 MWh is a U.S. average applied to global numbers — a translation that measures scale better than it describes geography. When half a trillion kilowatt-hours starts to lose meaning, utility bills are a useful place to land.
Field note VI · Training vs. Inference
Where the carbon actually goes
Five published training emissions sit between 0.65 and 6,300 tCO₂ each. A single year of all AI inference sits between 8 and 27 million tCO₂ — roughly 1,300 to 4,000 times the largest training event on the chart. The 'but training is huge' framing stops working at 2024 scale.