How the figures were calculated
Every figure on this site has a section here. Each one names the display the figure appears on, restates what the figure shows, walks through how it was derived, names the assumptions behind the math, and says where the wide error bars come from. The sources are cited at the bottom of each section, so this page can be read on its own without flipping between tabs.
Home page hero
Per-hour CO₂ comparison
The hourly hero shows five activities arranged by grams of CO₂ equivalent emitted in one hour: gasoline driving, heavy AI workflow, high-end PC gaming, ordinary AI chat, and HD video streaming.
Derivation
Gasoline driving (14,000–18,000 g/hr). EPA's typical passenger vehicle emits about 400 g CO₂ per mile and 8,887 g per gallon of gasoline burned. At a roughly 30 mph average speed (mixed urban and highway driving), one hour produces 12 kg of tailpipe CO₂. Adding a 15–30% upstream factor for fuel production, refining, and distribution gives 14–18 kg of CO₂-equivalent per hour.
Heavy AI workflow (100–500 g/hr). Heavier sessions — long-context queries, reasoning models, agentic workflows, document uploads — can consume 0.1–0.5+ kWh per active hour at the data-center level. Multiplying by a typical grid intensity around 400 g CO₂ per kWh, with a data-center overhead factor of roughly 1.2–1.5×, lands the range at about 100–500 g. This figure is the most uncertain of the five because per-query disclosures are sparse and the workload category itself is fuzzy.
High-end PC gaming (100–300 g/hr). Lawrence Berkeley's gaming-computer study estimates a typical high-end setup at about 1,400 kWh/year including display; intensive segments can use double that. A rig drawing 300–700 W for an hour consumes 0.3–0.7 kWh; at typical grid intensities around 400 g CO₂ per kWh, the per-hour CO₂ range lands around 100–300 g.
Ordinary AI chat (5–200 g/hr). Epoch AI estimates a typical GPT-4o-style text query at about 0.3 Wh. An hour of ordinary back-and-forth chat might involve 20–200 turns, or 6–60 Wh of compute before data-center overhead and grid emissions. The wide 5–200 g range absorbs three sources of variation: turns per hour, the still-uncertain per-query energy at the data-center level, and grid intensity at the location of the inference.
HD video streaming (36–56 g/hr). Carbon Trust's 2020 European estimate for one hour of video-on-demand streaming is about 55–56 g CO₂-equivalent, with energy use around 188 Wh/hr and the viewing device responsible for the largest share. IEA's earlier central estimate for one hour of streaming video in 2019 was 36 g. The site uses the IEA lower bound and the Carbon Trust upper bound as a defensible 36–56 g range.
Assumptions
Three assumptions sit underneath all five figures. Driving uses a 30 mph mixed-driving speed; faster or slower averages would shift the hourly figure proportionally. Both AI categories assume typical grid intensities around 400 g CO₂ per kWh; cleaner grids (Quebec, Iceland) push the figure down meaningfully, and dirtier grids (West Virginia, India) push it up. The streaming and gaming figures are device-centric; they exclude upstream content delivery, server-side video transcoding, and game-server compute, which can be substantial for live-service titles.
Uncertainty
The figures span very different confidence levels. The driving figure (medium) is anchored to a single well-documented EPA conversion factor with a small upstream-factor band. The streaming figure (medium) is anchored to two published per-hour estimates that span a narrow range. The gaming figure (medium-low) translates aggregate annual estimates into a per-hour reading, with hardware tier and play intensity adding spread. The two AI figures (low-medium) carry the widest bands because per-query energy is still poorly disclosed and the workload categories themselves are fuzzy. The relative ordering — driving an order of magnitude or two above everything else, the digital activities clustered close together — is robust across reasonable assumption shifts.
Home page primer
Why every figure here is a range
The primer takes one quantity — AI workloads' electricity use globally in 2024 — and shows it twice: first as a single dot, then as a striped band with a tick at the dot's value. The arithmetic of the two framings is identical; only the editorial choice differs.
Derivation
The single-point framing (65 TWh). Most central estimates of AI's 2024 share of global data-center electricity cluster around 65 TWh; the IEA's Energy and AI figure, repeated in the April 2026 update, is the most prominent single value. If the site quoted one number for AI's 2024 footprint, it would land here.
The range framing (30–80 TWh). Across the wider literature for the same year and the same metric, the lower bound sits near 30 TWh and the upper bound near 80. The spread comes from definitional variation — what counts as an "AI workload" inside a hyperscale facility — and from how analysts construct the number, which is bottom-up from accelerator shipments and utilization assumptions rather than from facility-level metering. Different definitional choices and different shipments-to-utilization heuristics each move the answer by tens of TWh; stacked together they produce the 30–80 band.
The licensed exception. The site does use point estimates where a primary source publishes one and the underlying measurement is well-bounded. EPA's 400 grams of CO₂ per mile of typical passenger-vehicle driving is the canonical example: a single figure derived from regulatory fleet averages, used directly in the hourly hero with only a small upstream-factor band layered on top. Single numbers appear on the site only on those terms.
Assumptions
The primer assumes the reader has not yet scrolled past the home-page hero, so it teaches the rule using the figure most likely to come up next. The same 30–80 TWh band reappears in the annual snapshot, the AI-share trajectory, and the household-equivalents chart. Demonstrating the rule on a number that recurs is the closest thing to a useful "key" the site has.
Uncertainty
The demo figure is deliberately the wobbliest number on the site. AI's 2024 share of data-center electricity carries a low-medium confidence label everywhere it appears, for the reasons above. The primer is the case that pushed the site to range-not-point in the first place; its job is to make that convention visible to a reader on their first scroll.
Home page infographic
How big is AI, really?
The annual-TWh chart shows six categories of global electricity use stacked vertically. The U.S. residential figure sits at the top as a scaffolding bar; five compute-adjacent loads (data centers, gaming, Bitcoin, EV charging, and AI) sit below in display order. The bars use ranges where the underlying data spreads across published estimates; single bars where a primary tracking source publishes one number.
Derivation
U.S. residential electricity (1,550 TWh). EIA's Electric Power Annual reports U.S. residential electricity consumption from utility filings — the 2024 figure rounds to 1,550 TWh. The bar carries a reference role on the chart, styled in faint accent at low opacity so it reads as scaffolding for the comparison set.
Global data centers (460 TWh). The IEA's Energy and AI report is the central source; the April 2026 Key Questions update raised the 2024 figure from 415 TWh in the original report to 460, which the site now reflects everywhere. The IEA's methodology is bottom-up — hyperscale facilities tracked at the operator level where disclosures permit, enterprise data centers estimated from server-shipment data, and crypto categorized separately. The figure excludes Bitcoin mining, which appears as its own bar.
Global video gaming (75–285 TWh). Gaming has the widest band on the chart and the weakest underlying data. Lawrence Berkeley's gaming-computer study estimates U.S. gaming-computer electricity at around 32 TWh; extrapolating to a global figure depends on platform mix (PC vs. console vs. mobile), how peripherals are counted, and whether streaming and download infrastructure roll up to gaming or to data centers. The 75-TWh lower bound counts gaming hardware alone; the 285-TWh upper bound includes game distribution, streaming, and the ad-tech that wraps around them. Both endpoints are defensible; neither is settled.
Bitcoin mining (140–200 TWh). Cambridge's CBECI maintains two estimators in parallel. Their direct miner survey lands at the lower end; their hashrate-based index, which models the network's installed mining hardware, lands at the upper end. The range here brackets that spread for 2024. Bitcoin is single-purpose load — the network's security model depends on burning electricity, and the protocol adjusts difficulty to match available hashrate, so efficiency improvements lower energy per hash but not energy per unit of security.
Global EV charging (180 TWh). IEA's Global EV Outlook 2025 estimates 2024 global EV electricity consumption at around 180 TWh, up nearly 60% from 2023. The figure is computed from the global EV fleet times an average per-vehicle annual kWh consumption, derived from country-level driving and efficiency data. The high confidence comes from comparatively good data on both inputs — fleet registration in major markets and per-vehicle consumption from regulatory filings.
AI workloads within data centers (30–80 TWh). Same range and same reasoning as the primer's section above. The IEA's central estimate is around 65 TWh; the wider literature brackets the same year at 30 to 80 depending on what counts as an AI workload at hyperscale. The bar is plotted as a range because no single number inside the band is more defensible than the others.
Assumptions
The chart compares figures with different definitional boundaries. Data-center electricity is metered at the facility level and rolls up through utility filings; gaming and Bitcoin are mostly modeled bottom-up from device populations and per-device energy assumptions; EV charging is modeled from fleet times average use. Putting them on one axis gives the right scale picture but flattens the methodological differences. The U.S. residential reference bar sits at the top as scaffolding — a calibration line for how big the others are relative to one country's homes — styled at low opacity to keep it visually subordinate to the comparison set.
Uncertainty
Two of the bars carry the lowest confidence labels: gaming (low) and AI (low-medium). Gaming is the widest because the underlying data is mostly LBNL's U.S. gaming-computer study extrapolated globally with assumption-heavy multipliers. AI is low-medium for the definitional reasons walked through above. Bitcoin sits at medium because Cambridge's two estimators agree on order of magnitude but diverge by roughly a third on the exact figure. The remaining three (residential, data centers, EV charging) come from primary tracking sources with comparatively narrow bands. The chart's bottom-line ordering — data centers clearly above the gaming/Bitcoin/EV cluster, AI clearly below it — holds across the uncertainty bands.
Home page infographic
Are AI data centers thirstier than golf courses?
The water-bracket chart sets four data-center bars (two boundaries times two scopes) against two golf bars (U.S. and global), all on one axis in billions of gallons per year. The boundary question — whether to count just the water flowing through cooling loops or also the water consumed at the power plants generating the data center's electricity — is the editorial pivot of the whole infographic. Moving the line shifts the answer by an order of magnitude.
Derivation
U.S. golf courses (531 Bgal). GCSAA's Golf Course Environmental Profile (Phase 4) Water Use & Management Practices survey reports 2024 U.S. golf-course irrigation at roughly 531 billion gallons. The figure is built from operator survey responses scaled to the national course count. Confidence is medium-high: the methodology is mature, response rates are decent, and recent updates have moved the figure only modestly.
Global golf courses (800–1,500 Bgal — site estimate). No primary source publishes a global golf water figure. The 800–1,500 range is derived in-house: U.S. per-course averages from GCSAA, multiplied across the R&A's count of approximately 38,860 courses globally, with bracketing assumptions about how non-U.S. courses' water use compares to the U.S. baseline (cooler and wetter regions lower the average; arid regions raise it). "Site estimate" on the chart is literal — this is the only figure on the site computed in-house. If a primary source surfaces, the band tightens or shifts.
Data centers — direct cooling water. This is the water flowing through cooling loops at the facilities themselves. Lawrence Berkeley's 2024 U.S. Data Center Energy Usage Report estimates U.S. data centers at about 17 billion gallons in 2023, with high confidence — the figure rolls up from facility-level disclosures and engineering estimates of cooling-load water demand. Globally, the 45–90 billion gallon range extrapolates U.S. per-MW water-use intensity to the global data-center electricity footprint, with the band reflecting variation in cooling architecture (water-cooled, air-cooled, and evaporative) and climate.
Data centers — including electricity-generation water. This boundary adds the water consumed at the power plants generating the data center's electricity, mostly thermoelectric plants (coal, gas, nuclear) that use evaporative cooling on their condensers. LBNL's U.S. 2023 figure is about 211 billion gallons, again with high confidence. Globally, the 500–700 billion gallon range applies the same boundary to the global data-center electricity figure, weighted by regional grid mixes — coal- and nuclear-heavy grids consume more water per kWh than wind- and solar-heavy grids, which is why the band is wide. The boundary shift takes the global figure from 45–90 to 500–700, and the U.S. figure from 17 to 211. That is the order-of-magnitude jump.
Assumptions
Three modeling assumptions move the bars. The global golf range (800–1,500) depends on a multiplier for non-U.S. per-course water use relative to the U.S. baseline; the band brackets a range of climate and irrigation-practice assumptions across Europe, Asia, and Latin America. The global data-center direct-cooling range (45–90) extrapolates U.S. per-MW water intensity globally; the band reflects variation in cooling architecture and climate. The global inclusive range (500–700) further depends on regional grid mixes — coal- and nuclear-heavy grids consume more water per kWh than wind- and solar-heavy grids — so any shift in the global generation mix shifts this number.
A separate note about reading the chart: golf irrigation and data-center water differ in environmental terms. Golf irrigation is mostly applied to permeable ground; some returns to local aquifers, some evaporates, some runs off. Cooling-tower evaporation and thermoelectric consumptive use are largely lost to atmosphere. The chart shows annual volumes; environmental impact is a different question.
Uncertainty
The two single-bar U.S. figures (17 and 211 Bgal, direct and inclusive) carry high confidence — LBNL's report is the canonical primary source and its facility-level methodology is well-documented. The U.S. golf figure (531) is medium-high; survey-based but mature. The two global data-center ranges (45–90 direct, 500–700 inclusive) are medium-low and low-medium respectively, with most of the spread in the inclusive case coming from regional grid-mix variation. The global golf figure (800–1,500) is the lowest-confidence number on the chart; it is the only figure on the site without a primary source. The chart's editorial point — that the boundary choice swings the comparison by an order of magnitude — holds across the uncertainty bands.
Home page infographic
Data-center electricity, 2017 to 2030
The trajectory chart plots global data-center electricity in terawatt-hours per year, with five historical anchor points from 2017 to 2025 and a projection band fanning out from 2025 to 2030. The boundary between actual and projected sits at 2025; the right side of the chart is scenario rather than data.
Derivation
The flat era (2017–2022). The IEA's tracking pegs global data-center electricity at roughly 200 TWh in 2017 and 2018, climbing to about 290 TWh by 2022. Through that five-year stretch, efficiency gains — Moore's law, improvements in cooling and PUE, the consolidation of enterprise workloads into hyperscale facilities — roughly canceled the underlying growth in compute demand. The line runs nearly flat from 2017 to 2018 and rises gently to 2022.
The bend (2024–2025). Two more recent anchors sit on the chart: 460 TWh in 2024 and 485 TWh in 2025. The 2024 figure is the IEA's central estimate for global data-center electricity, revised from 415 to 460 in the April 2026 Key Questions update — the same revision that propagates through every site display showing a 2024 number. The 2025 figure comes from the IEA's April 2026 commentary on the surge year, which reports a 17 percent jump from 2024. The visual effect is a clear bend away from the 2017–2022 baseline.
Lower bound — IEA Headwinds (830 TWh by 2030). The bottom edge of the projection band runs from 485 in 2025 to 830 in 2030. The IEA publishes its Headwinds scenario as a constrained case in which permitting delays, transformer shortages, water constraints in arid build sites, and higher capital costs slow construction. 830 TWh in 2030 still represents a near-doubling from 2024, but well short of the base case.
Central — IEA base case (950 TWh by 2030). The dashed central line runs from 485 in 2025 to 950 in 2030. This is the IEA Energy and AI base-case scenario as updated in April 2026. It assumes continued hyperscale build-out at announced rates, ordinary efficiency gains in accelerator silicon and cooling, and no major grid-side bottleneck stalling deployments.
Upper bound — Goldman Sachs (1,350 TWh by 2030). The top edge of the band uses Goldman Sachs's April 2026 forecast for global data-center power demand, which lands at roughly 1,350 TWh by 2030 — the highest among the major industry trackers' published figures. Pairing it with IEA Headwinds at the bottom gives the chart the widest defensible band.
Assumptions
Three modeling choices set the shape of the chart. The historical line uses the IEA's published anchor years rather than interpolating annual values; 2019, 2020, 2021, and 2023 are not plotted because the IEA did not publish definitive figures for those years, and the slopes between anchors are therefore artifacts of irregular spacing rather than annual change rates. The projection band is bracketed by IEA Headwinds at the bottom and Goldman Sachs at the top because pairing them yields the widest defensible span; using only IEA scenarios would narrow the band and arguably overstate confidence. The 2024 figure of 460 TWh reflects the IEA's April 2026 revision and replaces the 415 TWh figure from the original Energy and AI report.
A separate framing note: most of the new load shown by the projection is hyperscale build-out, and most of that hyperscale build-out is AI-adjacent — meaning the facilities themselves run a mix of training, inference, and ordinary cloud workloads, and the AI-specific share within them is its own question. The next chart on the home page (AI's slice) addresses that share directly.
Uncertainty
Confidence climbs across the historical line. The 2017 and 2018 figures are medium; the IEA's methodology improved between the original Digitalisation and Energy report and the Energy and AI series, and the early figures carry more spread than the later ones. The 2022 anchor is medium-high. The 2024 and 2025 figures are both high — facility-level disclosures and hyperscaler earnings filings have made the recent past well-bounded. The projection band is medium overall; the spread itself is the confidence statement, with most of the disagreement between IEA and Goldman coming from differing assumptions about how fast accelerator efficiency will improve and how aggressively hyperscalers will buy ahead of demand.
Home page infographic
Data-center electricity, in household-years
This chart converts terawatt-hours into a unit most readers have a feel for: U.S. household electricity. A horizontal hairline runs between the 2024 row and the 2030 row, with actual values above and projection below. Six bars in descending priority — U.S. residential reference at the top, then five compute-adjacent loads, ending with the 2030 data-center projection.
Derivation
The conversion factor (11.9 MWh per household per year). The EIA's Electric Power Annual divides total U.S. residential electricity (1,550 TWh in 2024) by the country's roughly 130 million households, which yields about 11.9 MWh per household per year. Every TWh therefore stands in for roughly 84,000 American homes' annual electricity. The chart applies this same factor to every bar.
U.S. residential reference (130 million households). The reference bar at the top is U.S. total residential electricity — 1,550 TWh in 2024 — converted at the chart's own factor, which by construction yields 130 million households. The bar exists as scaffolding: a calibration line for how big the others are relative to one country's homes. It's styled in faint accent at low opacity so it reads as the chart's background scale rather than as a comparison item.
AI workloads, 2024 (2.5–6.7 million households). The 30–80 TWh AI range, divided by 11.9 MWh per household, yields 2.5 to 6.7 million U.S. households' worth of electricity. Two and a half to seven million homes is roughly the size of a mid-Atlantic state — Pennsylvania at the lower end, somewhere near New Jersey at the upper.
Bitcoin mining, 2024 (12–17 million households). The 140–200 TWh Bitcoin range converts to 12 to 17 million households — roughly the size of Texas, give or take. Bitcoin is bigger than AI by this yardstick today, by a factor of roughly two to five.
Global video gaming (6.3–24 million households). The 75–285 TWh gaming range yields 6.3 to 24 million households. The wide spread reflects the underlying figure's weakness, walked through in the annual-TWh section above. The bar is plotted as a range in the same width and styling as AI's, since both are genuine ranges rather than central estimates.
Global data centers, 2024 (39 million households). The 460-TWh global data-center figure yields about 39 million households — roughly 30 percent of all American homes. This is the chart's anchor row: the figure other bars are sized against.
Global data centers, 2030 (70–113 million households, central 80). The 830–1,350 TWh IEA-Headwinds-to-Goldman band, drawn from the trajectory chart above, converts to 70 to 113 million households at the same 11.9 MWh factor. The IEA base case at 950 TWh maps to 80 million — the central tick on the bar. The 2030 row sits below the actual/projected hairline; its visual treatment matches the other range bars but the hairline cues the reader that this row is scenario, not measurement.
Assumptions
Two assumptions deserve explicit framing. The first is the yardstick itself: 11.9 MWh is a U.S. average applied to global figures, which measures scale rather than geography. A bar saying "39 million households" does not mean 39 million homes in any specific country; it means an amount of electricity that, if drawn by U.S.-average households, would supply that many. The second is shape: data centers run flat all year, while residential load is peaky — heavy on summer afternoons in the South, on winter evenings in the North. Equating their annual totals is a useful translation for scale and a misleading one for grid impact at any specific hour.
A separate framing note about the projection bar: the 2030 range is bracketed by IEA Headwinds at the low end and Goldman Sachs at the high end, the same bracketing used by the data-center trajectory chart. Using IEA scenarios alone would yield a narrower band; pairing IEA Headwinds with Goldman gives the widest defensible spread.
Uncertainty
Each row inherits the confidence label of its underlying TWh figure, walked through in the annual-TWh and trajectory sections above. Bitcoin (medium) and global data centers (medium-high) are well-bounded. AI (low-medium) and gaming (low) carry the widest fractional uncertainty. The 2030 data-center projection (medium) is bracketed by the spread between IEA Headwinds and Goldman, which is the chart's most explicit confidence statement. The 11.9 MWh conversion factor itself is high-confidence — EIA reports both the numerator and denominator on a regular schedule, and the factor moves slowly from year to year.
Home page infographic
The watt-scale primer
The site's headline figures run from a fraction of a TWh to a few thousand TWh. Most readers have intuition at the bottom of that range — a phone charge, a microwave running — but not at the top. The watt-scale primer is a small ladder for crossing the gap. Each rung is a thousand times the one below, anchored in an everyday reference; the site's actual figures appear along the same axis so the reader can place them.
Derivation
The ladder has five rungs, each a single decade in scientific notation: 1 kWh, 1 MWh, 1 GWh, 1 TWh, 1,000 TWh. The middle three rungs convert through the same per-household factor used elsewhere on the site — 11.9 MWh per U.S. residential customer per year, the EIA 2023 RECS midpoint.
1 kWh — a microwave running for an hour. Most domestic microwaves are rated 700–1,200 W; a 1,000-W unit running for one hour consumes 1 kWh by definition. The same rung is also roughly 30–45 minutes of average U.S. residential draw.
1 MWh — a U.S. home for about a month. 11.9 MWh per year ÷ 12 ≈ 0.99 MWh per month, rounded.
1 GWh — 85 U.S. homes for a year. 1,000 MWh ÷ 11.9 MWh per home = 84.0, rounded up.
1 TWh — a city of 85,000 homes for a year. 1,000 × 84.0 = 84,034 households; rounded to 85,000. Cities in roughly that residential range, depending on measurement year: Charleston, SC; Topeka, KS; Bellingham, WA.
1,000 TWh — about two-thirds of U.S. residential electricity in a year. EIA puts the residential total around 1,550 TWh; 1,000 ÷ 1,550 ≈ 64.5%.
The site's three anchor figures appear along the same axis: AI in 2024 (30–80 TWh), global data centers in 2024 (460 TWh, IEA April 2026), and U.S. residential electricity (1,550 TWh).
Assumptions
11.9 MWh per U.S. household per year is the conversion factor used throughout the site (EIA 2023 RECS midpoint), the same factor anchoring the household-equivalents bar chart.
"A microwave for an hour" approximates a 1,000-W unit running continuously. Real domestic microwaves are rated 700–1,200 W; the round number is the closest readable anchor at this scale.
The "city of 85,000 homes" reference is residential-only. A real city's full electricity footprint includes commercial and industrial loads not counted on this rung.
Uncertainty
The ladder is pedagogical, not measurement. Each rung is approximate; the named references — a microwave, a home, a city — are illustrative. A reader shouldn't infer that every microwave draws exactly 1 kWh per hour or every city consumes exactly 1 TWh of residential electricity. The point is to give the eye a stepwise feel for what a thousand-fold jump actually looks like.
Home page infographic
Where AI's carbon actually goes
Public arguments about AI emissions often fixate on training. A model's training run is large, dramatic, and reported as a single tCO₂ figure that sticks in memory — GPT-3 at 552 tonnes, Llama 3 at 6,300. These are real numbers, but they describe a single event in a model's life. Once a model is deployed, every query and every API call adds incremental inference emissions, day after day, for as long as the model is in service. By 2024, the aggregate inference burn dwarfs every published training run combined.
Derivation
The chart shows five published training-event emissions, ordered by ascending tCO₂:
BERT-base (2019) — 0.65 tCO₂. From Strubell et al.'s "Energy and Policy Considerations for Deep Learning in NLP," which estimated 1,438 lbs CO₂ per training run. The same paper is the source of the famous "five cars" headline, which referred to a neural-architecture-search run rather than to BERT itself; the BERT figure is the more representative number for a single language-model training pass at the scale of that era.
BLOOM (2022) — 25 tCO₂. From the BigScience BLOOM impact paper. The unusually low figure reflects training on the Jean Zay supercomputer in France, which runs largely on nuclear electricity (~57 g CO₂/kWh) rather than the global ~400 g/kWh average.
Llama 2 (2023) — 539 tCO₂. Disclosed by Meta in the Llama 2 paper.
GPT-3 (2020) — 552 tCO₂. From Patterson et al.'s 2022 paper, which back-computed from OpenAI's published 314 zettaFLOPs of training compute, with assumptions about the V100/A100 hardware mix and U.S. grid intensity at training time.
Llama 3 (2024) — 6,300 tCO₂. Disclosed by Meta in the Llama 3 paper, summing the 8B, 70B, and 405B training runs across multiple Meta data centers.
The chart shows annual AI inference for 2024 as a range:
All AI inference, 2024 — 8.4 to 27.2 MtCO₂. Derived from the site's existing AI 2024 electricity range (30–80 TWh, IEA April 2026), assuming inference accounts for 70–85% of that total. Patterson et al. (2022) estimated inference at 60–90% of lifetime emissions for widely-deployed models in Google's stack; the chart uses 70–85% as a slightly tighter band consistent with that range. At an average grid intensity of 400 g CO₂/kWh: 30 TWh × 70% × 400 g/kWh ≈ 8.4 MtCO₂; 80 TWh × 85% × 400 g/kWh ≈ 27.2 MtCO₂.
The largest single training event on the chart (Llama 3, 6,300 tCO₂) is roughly 1,300 times smaller than the lower bound of the 2024 inference range. Per-event training and per-year inference are different units — one is a one-time emission, the other repeats every year — but the chart shows them on the same log axis because the editorial point is that the gap is roughly four orders of magnitude in absolute terms.
Assumptions
The biggest assumption is the inference share. Patterson et al. estimated 60–90% inference for widely-deployed Google models in the 2019–2022 timeframe. The chart uses 70–85% as a wider band, which is conservative on the low side and consistent with industry estimates for the 2024 LLM mix. The actual share depends on how much of the year's compute went to one-off training runs versus heavily-used products; the order-of-magnitude conclusion holds across the plausible range.
The grid-intensity assumption (400 g CO₂/kWh) is the global average used elsewhere on the site. Real per-data-center intensities run from ~50 g/kWh on nuclear- or hydro-rich grids up to ~700 g/kWh on coal-heavy grids. The chosen figure is moderately conservative for U.S.-based hyperscalers, where most LLM inference runs.
The 2024 inference range inherits the AI 2024 electricity range of 30–80 TWh, which is the site's lowest-confidence electricity figure overall.
Uncertainty
The five training figures are well-sourced. Three are first-party disclosures from BigScience and Meta; the other two come from peer-reviewed papers. Each is accurate within roughly ±20%.
The inference figure is the loose end. The 8.4–27.2 MtCO₂ range stacks two uncertainties: the underlying TWh range (30–80, "low-medium" confidence) and the inference-share assumption (70–85%). The true 2024 figure is plausibly anywhere in 5–35 MtCO₂. The order-of-magnitude claim — that annual inference is hundreds to thousands of times larger than the largest single training run on the chart — is robust to the full range of plausible inputs.
The chart does not aggregate training emissions across all known runs in a year. Rolling that up would tighten the comparison somewhat: a few dozen frontier-model trainings at low-thousands of tCO₂ each total roughly 100,000 tCO₂ per year, still two orders of magnitude smaller than the inference band's lower bound.
Comparison card
Water: U.S. residential outdoor use vs. AI data centers
This card sets U.S. residential outdoor water use — lawns, gardens, driveways — against the global data-center fleet, in billions of gallons per year. Three bars: U.S. lawn-and-garden water at roughly 3,200 billion gallons (3.2 trillion), global data-center direct cooling at 45–90 billion, and global data centers including electricity-generation water at 500–700 billion.
The editorial point parallels the water-bracket infographic but inverts the geographic scope. The water bracket compares U.S. golf to global data centers; this card compares U.S. residential outdoor use to global data centers. The figure that comes out is even more lopsided than the golf comparison, because U.S. residential outdoor water dwarfs U.S. golf — roughly six times larger. Lawns are the largest irrigated crop in the United States by area, and the EPA estimates about half of all outdoor water is wasted to overwatering.
Derivation
U.S. residential outdoor water, recent (2,900–3,300 Bgal). From the EPA WaterSense outdoors statistics page, which aggregates data from the Water Research Foundation's residential end-use surveys and U.S. Geological Survey water-withdrawal accounting. The 2,900–3,300 range brackets recent estimates that vary depending on whether the figure is computed from per-household surveys multiplied by household counts, or from utility-side billed-water accounting with an outdoor allocation factor. EPA's headline figure of "about half of outdoor water is wasted" comes from the same underlying surveys. Confidence is medium; the input methodologies are sound but reach different totals depending on which lever the analyst pulls.
Global data centers — direct cooling only, 2025 (45–90 Bgal). Same LBNL-extrapolation derivation as the water-bracket infographic. Direct cooling means water flowing through cooling towers and evaporative chillers at the facilities themselves. Confidence is medium-low; the global figure extrapolates U.S. per-MW intensity to the global data-center electricity footprint.
Global data centers — including electricity-generation water, 2025 (500–700 Bgal). Same source and same derivation as the water-bracket infographic's inclusive bar. The boundary adds water consumed at the power plants generating data-center electricity — mostly thermoelectric plants that use evaporative cooling on their condensers. Confidence is low-medium; the band reflects regional grid mix variation.
Assumptions
The U.S. residential outdoor figure includes irrigation, pool fill and top-off, vehicle washing, and other outdoor uses; the dominant share is lawn-and-garden irrigation. The global data-center figures use the same U.S.-to-global extrapolation as the water-bracket infographic; the boundary question (direct cooling only vs. inclusive of electricity-generation water) is the editorial pivot of both displays. A separate framing note: lawn irrigation and data-center water differ in environmental terms. Lawn water is mostly applied to permeable ground; some returns to local aquifers, some evaporates, some runs off. Cooling-tower evaporation and thermoelectric consumptive use are largely lost to atmosphere. The chart shows annual volumes; environmental impact is a different question.
Uncertainty
The lawn figure (medium) is bracketed by methodology variation as described above; both endpoints are defensible. The data-center bars carry the same uncertainty profile as the water-bracket infographic: medium-low for direct, low-medium for inclusive. The chart's headline ratios — 35–70x for lawns vs. direct cooling, roughly 5x for lawns vs. inclusive — hold across reasonable adjustments to any of the three figures.
Comparison card
Water: AI data centers vs. U.S. golf courses
This card sets U.S. golf-course irrigation against the water that runs through global and U.S. data centers, in billions of gallons per year. Six bars: U.S. and global golf, data centers under a "direct cooling" boundary, and data centers under a wider boundary that adds the water used to generate their electricity.
The comparison is widely cited and rarely complete. "Data centers use less water than golf courses" is true under one boundary and false under another, and the boundary is most of the argument. The card pairs the same data-center fleet against the same golf figure twice, once narrow and once wide, so the rhetorical move is visible rather than implicit. The water-bracket infographic on the home page makes the same point with the same numbers; this card is the prose-card form of it, sitting alongside the other comparisons rather than standing alone as a teaching visual.
Derivation
All six figures are derived in §Are AI data centers thirstier than golf courses? — golf U.S. and global, data centers direct and inclusive, U.S. and global pairs within each.
Comparison card
Electricity: U.S. residential vs. global data centers
This card sets U.S. residential electricity consumption against every data center on the planet, in terawatt-hours per year. Two bars: U.S. homes at 1,550 TWh in 2024, global data centers at 460.
The card exists to keep residential scale visible while the data-center conversation gets louder. American homes — heating, cooling, lighting, refrigerators, water heaters, screens — still move more electricity than every data center in the world combined, by roughly three to one. The comparison frames AI's grid presence less as a coming flood and more as a new, growing slice inside an energy economy whose biggest consumer remains, at the U.S. scale, just people running their houses.
Derivation
Both figures sit in §How big is AI, really? — U.S. residential as the reference scaffolding bar at 1,550 TWh, and global data centers at 460 TWh, with the IEA April 2026 reconciliation (415 → 460) noted there.
Comparison card
Electricity: U.S. residential air conditioning vs. global data centers
This card sets U.S. residential air conditioning against the global data-center fleet, both in terawatt-hours per year. U.S. AC alone runs about 254 TWh annually — around 55 percent of every data center on the planet combined.
The editorial point is one of seasonal scale. AC use rises sharply in summer; on hot afternoons it is the single biggest driver of grid stress in most U.S. regions, and it dwarfs everything else households do with electricity during those hours. A peaky, seasonal residential load in one country sits in the same scale league as the year-round, planet-wide computing fleet — which says more about how big residential cooling has become than about how small data centers are.
Derivation
U.S. residential air conditioning, 2020 (254 TWh). From EIA's Residential Energy Consumption Survey (RECS), the 2020 vintage being the most recent published wave. RECS combines a national household survey with utility-bill data and engineering simulations to estimate end-use electricity consumption — heating, cooling, water heating, refrigeration, and so on — for the U.S. residential sector. The 254 TWh AC figure includes central air conditioning, room AC units, evaporative coolers, and heat pumps in cooling mode. Confidence is high; RECS is the canonical source and updates on a roughly four-year cadence.
Global data centers, 2024 (460 TWh). Same IEA Energy and AI source as the rest of the site. April 2026 revision raised this from 415 to 460. Confidence is medium-high.
Assumptions
The two figures are reported on different vintages: 2020 for residential AC, 2024 for data centers. AC use has likely grown modestly since 2020 with continued cooling-degree-day increases and population growth in hot states; data-center electricity has grown sharply over the same period (from roughly 290 TWh in 2022 to 460 in 2024 to 485 in 2025). The 55 percent ratio quoted on the card therefore probably overstates the current AC-to-data-center ratio; the true number today is likely closer to 50 percent or below, though no fresh RECS data is yet available to confirm. A separate framing note: the AC figure is U.S.-only residential. Adding U.S. commercial air conditioning would roughly double it; adding global air conditioning would multiply it several times over.
Uncertainty
The AC figure (high) carries the smallest band of any electricity comparison on the site — RECS is well-instrumented and the 2020 wave is a definitive source. The data-center figure (medium-high) is well-bounded but moves with new IEA reports. The chart's qualitative point — one country's seasonal residential cooling sits in the same league as the world's compute layer — holds regardless of which year's figures are used.
Comparison card
Electricity: U.S. holiday lighting vs. global data centers
This card sets U.S. December holiday lighting against the global data-center fleet, both in terawatt-hours per year. The holiday-lighting figure — 6.6 TWh — is small in absolute terms, around 1.4 percent of global data centers. It earns its place on the site because the comparison cuts the other way too: 6.6 TWh is more than the entire annual electricity supply of El Salvador and several other smaller nations.
The card serves as a calibration exercise for "X uses as much electricity as Country Y" headlines. Aesthetic energy use can rival national grids without being structurally significant against the major loads. Holiday lighting is genuinely large at the country-scale and genuinely small at the global-electricity-system scale; both framings are correct, and the card lets readers feel both.
Derivation
U.S. holiday lighting, 2008 (6.6 TWh). Originally estimated by the U.S. Department of Energy in 2008 and propagated through the Center for Global Development and several news outlets in subsequent years. The estimate combines household-survey data on holiday-lighting penetration, average bulb counts and wattages, and average days-and-hours-per-day of operation. Confidence is low-medium; the 2008 vintage is the latest publicly available defensible figure, and the actual current number is almost certainly lower because LED string lights have largely displaced incandescent bulbs over the intervening years. An LED string draws roughly 80–90 percent less than an incandescent string of comparable length, so a back-of-envelope adjustment for current LED penetration would put the figure closer to 1–2 TWh today. The chart uses the 2008 number because it's the last one with a clear methodology trail.
Global data centers, 2024 (460 TWh). Same IEA Energy and AI source. Confidence is medium-high.
Assumptions
The holiday-lighting figure is U.S.-only; the data-center figure is global. The card uses the 2008 estimate without adjustment because the LED-corrected figure has no published authority behind it — applying a back-of-envelope efficiency factor would substitute one set of guesses for another. The "more than El Salvador" framing comes from the original Center for Global Development analysis and uses 2008-era national electricity totals; both numbers have moved since, but the comparison still holds at order of magnitude.
Uncertainty
Holiday lighting (low-medium) is the lower-confidence figure on the chart by some distance; the underlying methodology is sound but the 2008 vintage and the LED transition since make the current actual figure highly uncertain. Data centers (medium-high) is well-bounded. The ratio quoted on the card (around 1.4 percent of global data centers) is probably a high estimate; today's true figure is likely closer to 0.3 percent. The country-comparison framing is more durable than the data-center ratio because both holiday lighting and small-country electricity demand have moved together over the years.
Comparison card
Electricity: Global video gaming vs. data centers
This card sets the global video-gaming ecosystem — consoles, PCs, handhelds, and the streaming infrastructure that supports them — against every data center on the planet, in terawatt-hours per year. Two bars: gaming somewhere between 75 and 285 TWh, global data centers at 460.
The pairing is included because the claim "video games use as much electricity as Country X" gets repeated often, and the card lets a reader see what the underlying spread looks like. The high end of the gaming range approaches the global data-center figure; the low end is closer to a quarter of it. The point is less to settle the comparison than to make its softness legible: gaming is the weakest electricity figure on the site, and the headline "as much as a country" rests on which end of a four-fold range you pick.
Derivation
Both figures are derived in §How big is AI, really? — gaming as the 75–285 TWh range and global data centers at 460 TWh.
Comparison card
Electricity: Global video streaming vs. data centers
This card sets the global video-streaming layer — Netflix, YouTube, Disney+, Twitch and the rest — against every data center on the planet, in terawatt-hours per year. Two bars: streaming somewhere between 100 and 300 TWh, global data centers at 460.
The card pairs streaming and data centers because the "streaming uses as much electricity as a small country" headline circulates in much the same shape as the gaming and AI ones. The IEA has spent years correcting the upper end of the streaming range, and the range remains wide: most of the energy lives in components shared with everything else — routers, content-delivery networks, screens, home Wi-Fi — and analysts disagree on which side of the ledger to put each one. The hourly hero on the home page renders the same activity per hour rather than per year (36–56 grams of CO₂ for an HD stream on a TV); the per-hour figure is steadier than the annual one because it sidesteps most of the boundary fight.
Derivation
Global video streaming, recent (100–300 TWh). A composite range from the IEA streaming fact-check (which argues for the lower end after correcting widely-cited 2019 figures), the Carbon Trust's video-streaming carbon study, and the IEA Energy and AI rollup. Whether end-user devices, home Wi-Fi, and last-mile networks count on streaming's side of the ledger or get attributed elsewhere is the dominant source of spread; choices on those three accounting questions move the figure two- or three-fold. Confidence is low.
Global data centers, 2024 (460 TWh). Same IEA Energy and AI source as the rest of the electricity comparisons, derived in §How big is AI, really? with the IEA April 2026 reconciliation (415 → 460) noted there.
For the per-hour version of this comparison — what an hour of HD streaming looks like alongside an hour of driving, gaming, or AI chat — see §Per-hour CO₂ comparison.
Comparison card
Electricity: Bitcoin mining vs. data centers
This card sets Bitcoin's mining electricity against every data center on the planet, in terawatt-hours per year. Two bars: Bitcoin somewhere between 140 and 200 TWh, global data centers at 460.
Bitcoin earns its place because it is one of the only other pure-compute, planet-scale electrical loads on the chart, and because it answers an obvious question: is AI bigger than Bitcoin yet? Today the answer is no — AI's 30–80 TWh slice fits inside Bitcoin's range. But Bitcoin's load is structurally fixed by its consensus mechanism (the network's security depends on burning electricity) while AI's load grows with usage. Whatever you make of the trade-offs of either, they sit in the same scale conversation without being part of each other: most Bitcoin mining lives outside hyperscale data centers and is excluded from the IEA's data-center figures.
Derivation
Both figures are derived in §How big is AI, really? — Bitcoin as the 140–200 TWh range (Cambridge CBECI low end; market-based estimators higher) and global data centers at 460 TWh.
Comparison card
Electricity: Global EV charging vs. data centers
This card sets electricity drawn by the world's electric vehicles against every data center on the planet, in terawatt-hours per year. Two bars: EV charging at 180 TWh in 2024, global data centers at 460.
The two are growing on similar curves and for similar reasons: both are new gigawatt-scale loads landing in the same regional grids that are also trying to retire coal. EVs grew nearly 60 percent in a single year and the IEA expects them past 2,000 TWh by 2035 in its main scenario; data-center load is on its own steep climb in the same window. For grid planners the two questions converge — same kilowatt-hours, same substations, often the same utilities deciding what to commission next.
Derivation
Both figures sit in §How big is AI, really? — EV charging at 180 TWh (IEA Global EV Outlook 2025) and global data centers at 460 TWh (IEA Energy and AI, April 2026 reconciliation applied).
Comparison card
CO₂: U.S. driving vs. global data centers
This card sets U.S. on-road transportation against the global data-center fleet, both in millions of metric tons of CO₂ per year. The home page's hourly hero already shows driving outpacing any digital activity per hour; this card moves the same point to the annual ledger and to a national-vs-global scope.
The pairing is among the most legible benchmarks in the AI-footprint debate. "Cars vs. AI" is a comparison readers reach for instinctively, and the numbers reward the instinct: even restricted to one country's road fleet, transportation runs roughly eight times the global data-center total. Pairing U.S.-only driving against worldwide data centers sets a deliberately tough test — the smaller geographic scope still wins by an order of magnitude.
Derivation
U.S. on-road transportation, 2022 (1,440 Mt CO₂). The figure comes from the EPA's Inventory of U.S. Greenhouse Gas Emissions and Sinks, the canonical U.S. GHG accounting document, which reports on-road transportation separately from off-road equipment, aviation, and rail. The 2022 value covers every passenger car, light truck, heavy-duty vehicle, and bus on U.S. roads, with emissions calculated from national fuel-consumption volumes (DOT and EIA data) and the EPA's per-gallon CO₂ factor — the same 8,887 g/gallon constant the hourly hero uses for its per-mile derivation. Confidence is high; the inputs are well-instrumented and the methodology is mature.
Global data centers, 2024 (180 Mt CO₂). Pulled from the IEA's Energy and AI report. The figure is computed bottom-up from global data-center electricity (460 TWh, the same anchor used across the site) times the world's average grid carbon intensity. The April 2026 Key Questions update revised this figure downward — the original report quoted around 220 Mt; the April 2026 version's consistent figure is 180. The revision reflects a lower assumed grid intensity rather than a change in electricity use. Confidence is medium-high; the underlying TWh figure is well-bounded, but global average grid intensity shifts year over year as more renewables come online.
Assumptions
Two scope choices set the comparison. The driving figure is U.S.-only by deliberate constraint: U.S. transportation data is the cleanest available, and pairing one country's road fleet against the world's data centers makes the comparison harder for driving to win, not easier. The data-center figure is global because the IEA does not publish a clean U.S.-only figure on the same recent timetable, and the AI buildout the comparison is meant to address is itself global.
A separate framing note: 180 Mt is operational CO₂ from electricity generation only. Embedded emissions from data-center construction — concrete and steel for new builds, GPU manufacturing — are not in this figure. The cement-and-steel comparison card flags those embedded emissions in a different bar; readers thinking about the full lifecycle of AI should look at both cards.
Uncertainty
The driving figure (high) is anchored to EPA's regulatory accounting, which has a small annual revision history and tracks closely with national fuel-sales data. The data-center figure (medium-high) is bracketed mostly by uncertainty in the global average grid intensity used to convert TWh to Mt — a 10 percent shift in grid intensity moves the data-center figure by roughly 18 Mt. That spread is much smaller than the 8x ratio at the heart of the comparison, so the editorial point — driving comfortably exceeds the global data-center fleet even when the geographic scopes are stacked against driving — holds across reasonable assumption shifts.
Comparison card
CO₂: Global commercial aviation vs. data centers
This card sets global commercial aviation against the global data-center fleet, both in millions of metric tons of CO₂ per year. Both figures are global; both are operational (in-flight fuel burn for aviation, electricity-driven for data centers).
The pairing is unusual on this site in that it's apples-to-apples on geographic scope — most of the CO₂ comparisons here pit U.S. activities against a global data-center figure. Aviation also happens to be one of the most-tracked emissions categories outside electricity, with annual disclosures from major carriers and IATA-level rollups, which makes it a tighter comparison than most.
Derivation
Global commercial aviation, 2023 (950 Mt CO₂). From the IEA's Aviation tracker, which compiles emissions from passenger and freight flights worldwide using fuel-burn data reported by airlines and national aviation authorities. The figure includes domestic and international commercial flights but excludes military, general aviation, and contrail/non-CO₂ radiative forcing. The 2023 number reflects a near-full recovery from the COVID-era trough, climbing back to within a few percent of 2019 levels. Confidence is medium-high; aviation reports more granularly than most sectors, and the methodology is mature.
Global data centers, 2024 (180 Mt CO₂). Same IEA Energy and AI source as the driving and cement comparisons. The April 2026 Key Questions revision lowered this figure from approximately 220 Mt by adopting a lower global grid carbon intensity. Confidence is medium-high.
Assumptions
The aviation figure is operational only — fuel-burn CO₂ during flight. It excludes ground operations at airports, aircraft manufacturing, fuel refining, and (most importantly for the climate debate) non-CO₂ radiative effects. Cirrus clouds from contrails and high-altitude NOx emissions roughly double aviation's effective warming impact in many studies; this card uses CO₂ alone for parity with the data-center figure, which is also CO₂-only. Readers thinking about full warming impact should mentally double the aviation bar and not the data-center one.
Uncertainty
Aviation (medium-high) and data centers (medium-high) carry comparable confidence labels, which is why this comparison reads tighter than most on the site. The 5x ratio holds across most plausible adjustments to either figure. The harder forward-looking question is what happens by 2030: data-center electricity is projected to roughly double in the IEA central case, while aviation is projected to grow modestly, so the ratio narrows but does not invert. Even at the IEA's higher 2030 data-center scenarios paired with central grid-intensity assumptions, aviation likely remains comparable or larger.
Comparison card
CO₂: Global cement and steel vs. data centers
This card sets two of the world's largest industrial emitters — cement and iron-and-steel — against the global data-center fleet, all in millions of metric tons of CO₂ per year. The chart shows three bars: cement at roughly 1,600 Mt, iron and steel at roughly 2,600 Mt, and global data centers at 180 Mt.
The editorial point is one of scale calibration. AI's footprint is real and growing, but the materials industries that built and maintain modern infrastructure run roughly twenty times larger combined. Anyone arguing about whether AI will "swamp the climate" is asking a question whose framing is misleading: the climate ledger is dominated by the things that pour foundations and forge beams, not by the things that draw electricity to compute on those foundations.
Derivation
Global cement, 2023 (1,600 Mt CO₂). From the IEA's Cement tracker, which separates process emissions (the calcination of limestone, which releases CO₂ chemically) from energy emissions (kiln fuel). Process emissions account for roughly 60 percent of cement's footprint; the rest is fuel. The 2023 figure reflects continued growth in cement production, particularly in China and Southeast Asia. Confidence is high; cement emissions are well-instrumented through both production volumes and stack monitoring.
Global iron and steel, recent (2,600 Mt CO₂). From the IEA's Iron and Steel tracker. Steel is the harder of the two industries to decarbonize because most production still uses blast-furnace technology that consumes coking coal as both fuel and chemical reductant. The 2,600 Mt figure includes both blast-furnace primary steel and electric-arc-furnace recycled steel, weighted to current production shares. Confidence is medium-high.
Global data centers, 2024 (180 Mt CO₂). Same IEA Energy and AI source as the other CO₂ comparisons. The April 2026 revision lowered this from around 220 Mt to 180 by adopting a lower global grid carbon intensity. Confidence is medium-high.
Assumptions
All three figures are operational CO₂ — emissions from making the product or running the facilities, not from upstream extraction or downstream use. The cement and steel figures are global; the data-center figure is also global, so the geographic scope lines up cleanly. A scope mismatch worth flagging: data-center construction itself uses cement and steel (concrete pads, structural steel), so a portion of the cement and steel bars is properly attributable to AI's lifecycle footprint. The chart treats the bars as separate categories; readers thinking about lifecycle accounting should know the categories overlap at the margins.
Uncertainty
Cement (high) and steel (medium-high) are among the better-tracked industrial categories. The data-center figure (medium-high) carries the most movement-room of the three, but the 23x combined ratio is robust to reasonable adjustments. Even doubling the data-center bar (a rough proxy for the IEA's 2030 central case in CO₂ terms, holding grid intensity flat) leaves cement and steel still about ten times larger combined.
Comparison card
CO₂: Global cattle (beef and dairy) vs. data centers
This card sets the world's cattle herd — beef and dairy combined — against the global data-center fleet, both in millions of metric tons of CO₂-equivalent per year. The cattle figure includes methane belched by ruminants and CO₂ from clearing forest and grassland for pasture; the data-center figure is operational electricity-driven CO₂ only.
The comparison is one of the few on the site where global lines up against global. It also surfaces a non-energy axis of climate impact — methane-driven warming, which moves much faster than CO₂ over the next 20 years. The pairing's editorial point: the things humans do on the surface of the planet to feed themselves are still climate-larger than the things humans do in concrete buildings to compute.
Derivation
Global cattle (meat + milk), recent (3,800 Mt CO₂-eq). From the UN FAO's "Tackling Climate Change Through Livestock" report, which converts methane and nitrous oxide emissions to CO₂-equivalent using GWP100 (the IPCC's 100-year global warming potential standard). Roughly 60 percent of the cattle figure is enteric methane (belching from rumination); the rest is feed production, manure management, and land-use change. Confidence is medium-high; FAO methodology is mature, but the GWP100 framing systematically understates methane's near-term warming impact compared to GWP20.
Global data centers, 2024 (180 Mt CO₂). Same IEA Energy and AI source. April 2026 revision applies. Confidence is medium-high. The cattle figure is CO₂-equivalent (CH₄ converted via GWP100) while the data-center figure is direct CO₂ only — the units match by convention even though the underlying gases differ.
Assumptions
The cattle figure uses GWP100 for methane conversion, which is the standard for national greenhouse-gas inventories. Switching to GWP20 (which weights short-term warming more heavily) would roughly double the methane share of the cattle bar, taking the total closer to 5,500 Mt. The chart's 21x ratio is therefore conservative for near-term warming impact. A separate framing note: cattle emissions also include land-use change in countries where forest is cleared for new pasture, primarily in Brazil. The FAO figure attempts to capture this; some independent estimates run higher.
Uncertainty
Cattle (medium-high) is bracketed mostly by methodological choices around GWP framing and land-use accounting. Data centers (medium-high) is bracketed by grid-intensity assumptions. The 21x ratio holds across both reasonable adjustments. Even if the data-center figure doubled by 2030 in the IEA central case, cattle would remain about ten times larger.
Comparison card
Emissions: U.S. gas-powered lawn equipment vs. data centers
This card sets U.S. gas-powered lawn and garden equipment — mowers, blowers, trimmers, edgers — against the global data-center fleet, both in millions of metric tons of CO₂ per year. The CO₂ comparison alone is striking: thirty million tons from one country's small-engine garden tools is around 17 percent of the global data-center figure.
The editorial point of the card extends past CO₂. The bigger story is air quality. Gas-powered lawn equipment uses two-stroke and small four-stroke engines with no meaningful emissions controls, so per unit of work delivered they emit orders of magnitude more nitrogen oxides, volatile organic compounds, and fine particulates than passenger vehicles. AI's grid load gets headlines; gas mowers and blowers do quieter, dirtier work in the same neighborhoods.
Derivation
U.S. gas-powered lawn equipment, 2020 (30 Mt CO₂). From the EPA's Banks report on national emissions from lawn and garden equipment, the canonical U.S. inventory for this sector. The 30 Mt figure aggregates residential and commercial small-engine equipment — push mowers, riding mowers, leaf blowers, string trimmers, hedge trimmers, chainsaws — using fuel-consumption estimates from population surveys and operating-hour assumptions. Confidence is medium; the inventory methodology is mature, but the inputs (how many hours per year a typical leaf blower runs) carry meaningful spread.
The same Banks dataset also reports criteria air pollutants from this equipment. Gas-powered lawn and garden equipment accounts for roughly one in six U.S. volatile-organic-compound emissions and more than one in ten nitrogen-oxide emissions — disproportionate shares for a sector that consumes a small fraction of total fuel.
Global data centers, 2024 (180 Mt CO₂). Same IEA Energy and AI source. April 2026 revision applies. Confidence is medium-high.
Assumptions
The lawn-equipment figure is U.S.-only and the data-center figure is global, so the 17 percent ratio understates how thoroughly small-engine yard tools punch above their weight per unit of energy delivered. Restricting both figures to the same country would narrow the ratio considerably — U.S. data-center CO₂ is a small fraction of the global figure, so U.S. lawn equipment runs much closer to parity against U.S.-only data centers than the global ratio suggests. The card uses the global denominator because the rest of the site does.
Uncertainty
Lawn equipment (medium) carries spread mostly from operating-hour assumptions and population estimates for residential equipment. Data centers (medium-high) carry the usual grid-intensity uncertainty. The 17 percent CO₂ ratio is approximate; the air-quality story (one in six VOCs, one in ten NOx) is firmer because EPA's criteria-pollutant accounting has tighter inputs than its CO₂ accounting for this sector.
Back matter
Consolidated sources
Every source cited anywhere on the site, gathered in one place and de-duplicated. Each per-figure section above carries its own short source line; this is the full ledger, alphabetized. 30 sources in all.
- BigScience — Estimating the Carbon Footprint of BLOOM
- Cambridge Centre for Alternative Finance — Bitcoin Electricity Consumption Index (CBECI)
- Carbon Trust - Carbon impact of video streaming
- Digiconomist — Bitcoin Energy Consumption Index
- EIA — Residential Energy Consumption Survey (RECS), 2020
- EIA — Use of Electricity / Electric Power Annual
- EPA — Inventory of U.S. Greenhouse Gas Emissions and Sinks
- EPA — National Emissions from Lawn and Garden Equipment (Banks)
- EPA WaterSense — Outdoors / Statistics & Facts
- Epoch AI - How much energy does ChatGPT use?
- GCSAA Golf Course Environmental Profile (Phase 4) — Water Use & Management Practices
- Goldman Sachs — AI to drive 165% increase in data center power demand by 2030
- IEA — Aviation
- IEA — Cement
- IEA — Data centre electricity use surged in 2025 (April 2026)
- IEA — Energy and AI: Energy Demand from AI
- IEA — Global EV Outlook 2025: Outlook for Energy Demand
- IEA — Iron and Steel
- IEA — Key Questions on Energy and AI (April 2026 update)
- IEA — The carbon footprint of streaming video: fact-checking the headlines
- Lawrence Berkeley National Laboratory - Taming the Energy Use of Gaming Computers
- Lawrence Berkeley National Laboratory — 2024 United States Data Center Energy Usage Report
- Meta — Llama 2: Open Foundation and Fine-Tuned Chat Models
- Meta — The Llama 3 Herd of Models
- Patterson et al. 2022 — Carbon Emissions and Large Neural Network Training
- R&A — Global Golf Participation Report 2024
- Strubell et al. 2019 — Energy and Policy Considerations for Deep Learning in NLP
- U.S. Department of Energy / EIA — 2008 holiday lighting estimate (cited via Center for Global Development)
- U.S. EPA - Greenhouse Gas Emissions from a Typical Passenger Vehicle
- UN FAO — Tackling Climate Change Through Livestock