{
"metadata": {
"title": "AI Consumption in Context",
"subtitle": "AI energy and water use compared with everyday activities, with measurement boundaries labelled",
"lastUpdated": "2026-09-02",
"sources": "IEA, LBNL, EPA, USGS, USDA, ASCE, GCSAA/USGA, NGF, OpenAI, Google, Mistral, Epoch AI, Meta, DeepSeek, AI2, Hugging Face AI Energy Score, ML.Energy, Pew Research Center, PJM/Monitoring Analytics, Morgan Stanley, Cornell (Nature Sustainability), UT Austin Bureau of Economic Geology, Microsoft, The Guardian/Cleanview, MIT Technology Review, Harvard/UCLA Study"
},
"waterConsumption": {
"unit": "litres/year",
"unitDisplay": "litres per year",
"scaleNote": "Bar lengths are drawn to scale (linear). The bottom four bars are shown at the minimum visible width - at true scale each would be thinner than a pixel, and ChatGPT's cooling-water bar would be less than 1/400,000th of the top bar.",
"items": [
{
"id": "global_network_losses",
"label": "Global Water Network Losses",
"sublabel": "Modelled estimate, not a census",
"value": 126054153000000,
"displayValue": "~126 trillion",
"color": "#dc2626",
"icon": "pipe",
"source": "Liemberger & Wyatt 2019 (IWA), cited by ASCE",
"sourceUrl": "https://infrastructurereportcard.org/cat-item/drinking-water-infrastructure/",
"year": 2025,
"notes": "Extrapolated from utility-reported non-revenue water rates (~346 million cubic metres/day) - no one meters every country's losses. Frequently misattributed to the US alone; the US share is roughly 8.3 trillion litres - a reminder that a number's boundary matters as much as its size"
},
{
"id": "corn_irrigation",
"label": "US Corn Irrigation",
"sublabel": "Irrigation water only, not rainfall",
"value": 20440000000000,
"displayValue": "20.4 trillion",
"color": "#eab308",
"icon": "corn",
"source": "USDA NASS Irrigation Survey 2023",
"sourceUrl": "https://www.nass.usda.gov/Newsroom/2024/10-31-2024.php",
"year": 2023,
"notes": "~35-40% of US corn becomes ethanol; each litre of ethanol carries a ~1,200-1,500 litre water footprint (mostly evapotranspiration)"
},
{
"id": "us_infrastructure_leaks",
"label": "US Treated Water Lost to Leaks",
"sublabel": "~6 billion gallons lost every day",
"value": 8290000000000,
"displayValue": "8.3 trillion",
"color": "#ef4444",
"icon": "pipe",
"source": "EPA / ASCE",
"sourceUrl": "https://www.epa.gov/watersense/statistics-and-facts",
"year": 2025,
"notes": "~240,000 water main breaks per year; nearly 20% of US water mains have exceeded their useful life"
},
{
"id": "residential_leaks",
"label": "US Household Leaks",
"sublabel": "Dripping taps, toilets, pipes",
"value": 3406869000000,
"displayValue": "3.41 trillion",
"color": "#f97316",
"icon": "faucet",
"source": "EPA WaterSense",
"sourceUrl": "https://www.epa.gov/watersense/statistics-and-facts",
"year": 2026,
"notes": "Equivalent to the water use of 11 million homes; the average family wastes 9,400 gallons (35,600 litres) a year"
},
{
"id": "golf_courses",
"label": "US Golf Courses",
"sublabel": "~16,000 courses nationwide",
"value": 2010000000000,
"displayValue": "2.01 trillion",
"color": "#22c55e",
"icon": "golf",
"source": "GCSAA/USGA National Water Survey (Dec 2025)",
"sourceUrl": "https://gcsaa.org/who-we-are/media/news-release/2025-news-releases/2025/12/30/golf-courses-reduce-water-usage-by-31-percent-according-to-national-survey",
"year": 2024,
"notes": "Down 31% since 2005; 29.1 million on-course golfers"
},
{
"id": "ai_global_all",
"label": "All Global AI Data Centres",
"sublabel": "Incl. water behind their electricity (Scope 1+2)",
"value": 540000000000,
"displayValue": "312-765 billion",
"color": "#a855f7",
"icon": "ai",
"source": "de Vries (VU Amsterdam) / UC Riverside via IEEE Spectrum",
"sourceUrl": "https://spectrum.ieee.org/ai-water-usage",
"year": 2025,
"notes": "Wide uncertainty band; counts on-site cooling plus water consumed generating the electricity. Morgan Stanley projects ~1,068 billion L/yr by 2028 on this boundary"
},
{
"id": "us_datacenters_onsite",
"label": "All US Data Centres, On-site Cooling",
"sublabel": "Scope 1 only; every workload, not just AI",
"value": 66000000000,
"displayValue": "66 billion",
"color": "#c084fc",
"icon": "datacenter",
"source": "LBNL 2024 US Data Center Energy Usage Report",
"sourceUrl": "https://escholarship.org/uc/item/32d6m0d1",
"year": 2023,
"notes": "Under 1% of US water consumption. Same on-site boundary as the ChatGPT bar; excludes the water behind electricity generation"
},
{
"id": "google_datacenters",
"label": "Google Data Centres + Offices (Global)",
"sublabel": "All Google operations, consumption",
"value": 30660000000,
"displayValue": "30.7 billion",
"color": "#4285f4",
"icon": "datacenter",
"source": "Google Environmental Report",
"sourceUrl": "https://sustainability.google/google-2026-environmental-report/",
"year": 2024,
"notes": "8.1 billion gallons consumed in 2024; ~9.9 billion gallons in 2025 (derived from Google's replenishment disclosure)"
},
{
"id": "microsoft_datacenters",
"label": "Microsoft (Global, Consumption)",
"sublabel": "FY2024; withdrawal was 10.7 billion L",
"value": 5810000000,
"displayValue": "5.81 billion",
"color": "#0ea5e9",
"icon": "datacenter",
"source": "Microsoft Environmental Data Fact Sheet 2025",
"sourceUrl": "https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2025-Microsoft-Environmental-Data-Fact-Sheet-PDF.pdf",
"year": 2024,
"notes": "Consumption (evaporated), not withdrawal - the two are often conflated in coverage"
},
{
"id": "chatgpt_annual",
"label": "ChatGPT Cooling Water (Global, Annual)",
"sublabel": "On-site cooling only, all queries",
"value": 292000000,
"displayValue": "292 million",
"color": "#8b5cf6",
"icon": "ai",
"source": "OpenAI (Sam Altman, reaffirmed Feb 2026)",
"sourceUrl": "https://www.datacenterdynamics.com/en/news/sam-altman-chatgpt-queries-consume-034-watt-hours-of-electricity-and-0000085-gallons-of-water/",
"year": 2026,
"notes": "0.32 mL/query x 2.5B+ queries/day x 365. Excludes power-generation and chip-fab water; no methodology published. Corrects an arithmetic error in earlier versions of this page (111.2B)"
}
],
"perQuery": {
"chatgpt": {
"value": 0.32,
"unit": "mL",
"source": "OpenAI (Sam Altman, Jun 2025; reaffirmed Feb 2026)",
"notes": "On-site cooling only; no methodology published"
},
"gemini": {
"value": 0.26,
"unit": "mL",
"source": "Google (August 2025, still latest official disclosure)",
"notes": "About 5 drops of water; median text prompt, full-stack methodology"
},
"mistral": {
"value": 45,
"unit": "mL",
"source": "Mistral Large 2 lifecycle analysis (July 2025)",
"notes": "Looks 170x higher than Google's figure because Mistral also counts water consumed generating the electricity (Scope 2) - the gap is methodology, not thirst"
},
"comparison": "A single dripping tap wastes 11,000+ litres/year - equivalent to 35 million ChatGPT queries"
}
},
"energyConsumption": {
"unit": "kWh",
"scaleNote": "Bar lengths are drawn to scale (linear). Even so, the sub-1 Wh query bars are only a few pixels wide - at true scale a ChatGPT query is about 1/350th of an hour of big-screen TV. AI video generation is left off this chart because it would flatten everything else: a 10-second Sora 2 clip is an estimated ~1,000 Wh (1 kWh), roughly 2,900x a text query - see 'Energy Varies by AI Task Type' below.",
"items": [
{
"id": "gemini_query",
"label": "Google Gemini Query",
"value": 0.00024,
"displayValue": "0.24 Wh",
"color": "#4285f4",
"source": "Google (Official)",
"sourceUrl": "https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference",
"year": 2025,
"notes": "Median text prompt, full-stack methodology; still the latest official disclosure as of mid-2026"
},
{
"id": "chatgpt_query",
"label": "ChatGPT Query (average)",
"value": 0.00034,
"displayValue": "0.34 Wh",
"color": "#8b5cf6",
"source": "OpenAI (Sam Altman, reaffirmed Feb 2026)",
"sourceUrl": "https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use",
"year": 2026,
"notes": "No methodology published; independently corroborated by Epoch AI's ~0.3 Wh estimate for GPT-4o"
},
{
"id": "gemini_query_2024",
"label": "Google Gemini Query (2024)",
"value": 0.008,
"displayValue": "~8 Wh",
"color": "#ea4335",
"source": "Google (Official)",
"year": 2024,
"notes": "Same query type, 12 months earlier - 33x less efficient (derived: 0.24 Wh x 33)"
},
{
"id": "google_search",
"label": "Google Search",
"value": 0.0003,
"displayValue": "0.3 Wh",
"color": "#34a853",
"source": "Google (2009 figure - stale baseline, no modern disclosure)",
"year": 2009
},
{
"id": "reasoning_query",
"label": "Reasoning Query (complex)",
"value": 0.0039,
"displayValue": "~3.9 Wh",
"color": "#ec4899",
"source": "Epoch AI (o3-style estimate)",
"sourceUrl": "https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use",
"year": 2025,
"notes": "Epoch AI point estimate (~11x a standard query); measured open-model averages run higher at 25-30x (HF AI Energy Score v2, ML.Energy v3)"
},
{
"id": "image_generation",
"label": "AI Image Generation",
"value": 0.002,
"displayValue": "0.1-4 Wh",
"color": "#f97316",
"source": "Measured, 17 models (arXiv 2506.17016)",
"sourceUrl": "https://arxiv.org/abs/2506.17016",
"year": 2025,
"notes": "46x range across models; higher resolution adds 1.3-4.7x"
},
{
"id": "netflix_hour",
"label": "Netflix Streaming (1 hour HD)",
"value": 0.077,
"displayValue": "77 Wh",
"color": "#e50914",
"source": "IEA / Carbon Trust (DIMPACT)",
"sourceUrl": "https://www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines",
"year": 2025,
"notes": "~72% of this is your TV/device, not the data centre"
},
{
"id": "tv_oled_65_hour",
"label": "165 cm OLED TV (1 hour)",
"value": 0.12,
"displayValue": "120 Wh",
"color": "#0e7490",
"source": "RTINGS / Solar Tech Online",
"sourceUrl": "https://www.rtings.com/tv/learn/led-oled-power-consumption-and-electricity-cost",
"year": 2025,
"notes": "2025 flagship OLEDs range ~110-165 W"
}
],
"comparisons": [
{
"description": "1 hour of 165 cm TV watching",
"equivalentQueries": 279,
"calculation": "95 Wh ÷ 0.34 Wh = 279 queries"
},
{
"description": "1 hour of 191 cm OLED TV",
"equivalentQueries": 544,
"calculation": "185 Wh ÷ 0.34 Wh = 544 queries"
},
{
"description": "1 hour of Netflix streaming",
"equivalentQueries": 227,
"calculation": "77 Wh ÷ 0.34 Wh = 227 queries"
},
{
"description": "One 10-second AI video generation",
"equivalentQueries": 2941,
"calculation": "1,000 Wh ÷ 0.34 Wh = ~2,900 queries"
}
],
"householdEquivalents": {
"geminiQuery": {
"value": 0.24,
"unit": "Wh",
"equivalents": [
{ "activity": "Microwave", "duration": "1 second" },
{ "activity": "Television", "duration": "9 seconds" },
{ "activity": "Refrigerator", "duration": "6 seconds" },
{ "activity": "Laptop", "duration": "17 seconds" },
{ "activity": "Phone charging", "amount": "2% charge" }
],
"source": "Sustainability by Numbers"
}
},
"uncertaintyNote": "Energy per query ranges from 0.1-4 Wh for standard text depending on model, token count, and complexity. Reasoning modes average 25-30x more (measured); extreme cases reach 150-700x."
},
"reasoningModels": {
"title": "Reasoning Models Changed the Maths",
"averageMultiplier": "25-30x",
"extremeMultiplier": "150-700x",
"tokenMultiplier": "~10x more output tokens",
"sources": [
{
"name": "Hugging Face AI Energy Score v2",
"finding": "Reasoning models average 30x more energy; same model with reasoning on vs off measured at 150-700x",
"url": "https://huggingface.co/blog/sasha/ai-energy-score-v2",
"date": "2025-12-04"
},
{
"name": "ML.Energy Leaderboard v3",
"finding": "Problem-solving (reasoning) responses used 25x more energy than conversation, driven by ~10x more output tokens at lower GPU utilisation",
"url": "https://ml.energy/blog/measurement/energy/diagnosing-inference-energy-consumption-with-the-mlenergy-leaderboard-v30/",
"date": "2025"
}
],
"disclosureGap": "Only Google (Gemini: 0.24 Wh, 0.26 mL, 0.03 g CO2e per median prompt) and Mistral (Large 2: 1.14 g CO2e, 45 mL per 400-token reply, Scope 2 included) publish audited per-model figures. OpenAI offers an aggregate CEO figure with no methodology; Anthropic, xAI and Meta publish nothing per-query as of mid-2026. Anthropic declined environmental disclosure in Stanford's Dec 2025 transparency index and has no sustainability report, though it joined the Frontier carbon-removal coalition (Jun 2026) and is hiring emissions-reporting staff.",
"newerNotGreener": "Newer models are not automatically more efficient: 8 of 14 comparable models used equal or more energy than their early-2025 predecessors (HF AI Energy Score v2). MoE architectures help - active-parameter count drives energy, with a measured 3.6x lower energy/token vs a similar-size dense model.",
"tokenEfficiency": {
"fixedCapability": "Reaching the same ~27% accuracy on FrontierMath took ~43 million output tokens with o4-mini (high effort, April 2025) and ~5 million with GPT-5.2 (low effort, December 2025) - about 9x fewer tokens, roughly 3x cheaper once per-token prices are counted, in eight months. Epoch AI puts the trend at roughly 5-10x cost reduction per year to reach a fixed capability level.",
"fixedCapabilitySource": "Epoch AI, Gradient Updates: How persistent is the inference cost burden?",
"fixedCapabilityUrl": "https://epoch.ai/gradient-updates/how-persistent-is-the-inference-cost-burden",
"effortControls": "Every major lab now ships a thinking dial - Anthropic's effort parameter with adaptive thinking, OpenAI's reasoning_effort and verbosity, Google's thinking_level, DeepSeek and Qwen thinking/fast modes - so reasoning token spend is increasingly a per-request choice rather than a model property. Anthropic claims Opus 4.5 at medium effort matched Sonnet 4.5's SWE-bench score with 76% fewer output tokens (vendor claim, not independently verified)."
}
},
"boundaries": {
"title": "Why the Water Numbers Disagree",
"explainer": "Resource figures depend on where the measurement boundary sits. Small numbers count only on-site cooling at query time; large numbers add the water consumed generating the electricity (Scope 2), chip fabrication (Scope 3), and projections of growth. Both can be arithmetically honest.",
"bookends": {
"small": {
"figure": "0.32 mL per query",
"source": "OpenAI (Sam Altman)",
"boundary": "On-site cooling only, per query, today"
},
"large": {
"figure": "~1,068 billion litres/year by 2028",
"source": "Morgan Stanley Research (Sep 2025); range 637-1,485B",
"boundary": "All AI data centres, cooling + power-generation + chip-fab water, projected 2028 (~11x 2024)"
},
"peerReviewed": {
"figure": "731-1,125 billion litres/year by 2030 (US only)",
"source": "You et al., Cornell, Nature Sustainability (2025)",
"sourceUrl": "https://www.nature.com/articles/s41893-025-01681-y",
"boundary": "US AI servers, on-site cooling + power-generation water (Scope 1+2, no chip fab), depending on buildout scale. Upper bound is roughly New York City's annual drinking water supply"
}
},
"withdrawalVsConsumption": "US thermoelectric power plants account for 41% of total water withdrawals (34% of freshwater withdrawals, USGS) - but only ~3-4% of that is consumed (evaporated); the rest returns to the source. Counting withdrawals as 'use' inflates any electricity-linked water figure ~25x.",
"locality": {
"fact": "~2/3 of new US data centres built since 2022 are in high water-stress areas",
"source": "Bloomberg / S&P Global",
"plannedBuilds": "517 of 809 planned US data centres are on land that was in drought throughout the preceding year - developers favour arid regions for cheap land, tax breaks and less equipment corrosion. A large site can draw up to 5 million gallons (~19 million litres) a day, the water use of a town of up to 50,000 people",
"plannedBuildsSource": "Guardian analysis of Cleanview and US Drought Monitor data, June 2026",
"plannedBuildsUrl": "https://www.theguardian.com/us-news/2026/jun/08/datacenter-ai-drought-water",
"sitingSwing": "Best vs worst distribution of US AI servers across states moves the 2030 water footprint by -52% to +354% (You et al., Nature Sustainability 2025)",
"texas": "Data centres drew ~0.75% of Texas water withdrawals in 2025 (43.5 billion gallons incl. power generation); projected 3-5% by 2030-2040 at 40-60 GW or 5-9% at 70-110 GW. Withdrawals, not consumption",
"texasSource": "UT Austin Bureau of Economic Geology, Water Requirements for Data Centers white paper (2025)",
"texasSourceUrl": "https://compass.beg.utexas.edu/files/publications/Water_Requirements_for_DC_White_Paper.pdf",
"note": "Water is local: the same facility can be benign in a wet region and a real problem in a dry one. Quantity matters less than where it is drawn.",
"operatorResponse": "Sam Altman said in Feb 2026 that OpenAI's data centres have largely moved off evaporative cooling (CNBC); Project Rainier, the 2.2 GW AWS campus that trains Claude, is air-cooled ~98% of the year at an estimated PUE ~1.15, though AWS won't quantify operational water (MeasuredAI, from permits and utility filings); all new Microsoft designs since Aug 2024 use closed-loop chip-level cooling that evaporates no water (pilots in Phoenix and Wisconsin in 2026; fleet WUE 0.30 L/kWh in FY2024 vs 0.49 in 2021); Google aims to replenish 120% of the freshwater its data centres consume by 2030; facility-level water data remains largely undisclosed (AGU Advances, 2026)"
}
},
"videoVsText": {
"title": "Video Generation: The Real Energy Consumer",
"comparison": {
"textQuery": {
"label": "Text Query (Gemini/ChatGPT)",
"energy": 0.34,
"unit": "Wh"
},
"imageGeneration": {
"label": "Image Generation (measured)",
"energyMin": 0.1,
"energyMax": 4,
"unit": "Wh"
},
"videoGeneration": {
"label": "10-second Video (Sora 2, estimate)",
"energy": 1000,
"unit": "Wh"
}
},
"ratio": "~2,900x more energy for a 10-second video vs a text query",
"insight": "Concerns about AI energy should distinguish between text (minimal), reasoning modes (25-30x text), and video generation (thousands of times text)"
},
"efficiencyImprovements": {
"title": "AI Efficiency is Improving - With Caveats",
"google": {
"model": "Gemini",
"period": "12 months (May 2024 - May 2025)",
"energyReduction": 33,
"emissionsReduction": 44,
"before": {
"energy": 8,
"emissions": 1.32,
"units": { "energy": "Wh", "emissions": "g CO₂e" }
},
"after": {
"energy": 0.24,
"emissions": 0.03,
"units": { "energy": "Wh", "emissions": "g CO₂e" }
},
"source": "Google (self-reported, no independent verification as of mid-2026)",
"sourceUrl": "https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference"
},
"costPerToken": {
"finding": "LLM inference price per token fell ~50x/year (median across benchmarks; ~200x/year after Jan 2024), expected to decelerate to 3-5x/year through 2027",
"source": "Epoch AI",
"sourceUrl": "https://epoch.ai/data-insights/llm-inference-price-trends"
},
"hardware": {
"finding": "NVIDIA markets Blackwell as 10-50x more efficient than Hopper; independent teardown puts the real per-watt silicon gain at ~47%, with the rest from FP4 quantisation and architecture changes",
"source": "NVIDIA claims vs independent analysis (adrianco)"
},
"caveat": "Efficiency per token improves rapidly, but reasoning models spend ~10x more tokens per query - so energy per QUERY can rise even as energy per token falls. 8 of 14 comparable models used equal or more energy than their early-2025 predecessors (HF AI Energy Score v2).",
"potentialOptimisation": {
"trainingCarbonReduction": "Up to 75%",
"source": "University of Michigan (2023 study, still cited)",
"method": "Tuning GPU power limits, batch size, and location selection"
},
"moe": {
"description": "Mixture-of-experts models: active-parameter count drives energy, not total - measured 3.6x lower energy/token vs similar-size dense model (ML.Energy v3)",
"source": "ML.Energy Leaderboard"
}
},
"dataCentreIndustry": {
"title": "The Bigger Picture: Data Centre Industry",
"usEnergyShare": {
"current": 4.4,
"previous": 1.9,
"previousYear": 2018,
"currentYear": 2023,
"unit": "% of US electricity (2023, latest full-economy data)",
"source": "LBNL 2024 US Data Center Energy Usage Report",
"sourceUrl": "https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf"
},
"global2025": {
"actual": 485,
"unit": "TWh",
"growth": "+17% year on year; AI-focused data centres +50%",
"source": "IEA Electricity 2026 (Dec 2025)",
"sourceUrl": "https://www.iea.org/reports/electricity-2026"
},
"usEmissions": {
"value": 105,
"unit": "million metric tons CO₂",
"year": 2023,
"percentOfNational": 2.18,
"comparison": "Comparable to domestic aviation (~2.5%)",
"source": "Harvard/UCLA Study (2023 data; hyperscale subset independently reproduced June 2026)"
},
"carbonIntensity": {
"dataCentreAverage": 548,
"usAverage": 369,
"premium": 48,
"unit": "g CO₂e/kWh",
"reason": "95% of US data centres in locations with dirtier-than-average electricity due to 24/7 power needs",
"source": "Harvard/UCLA Study; independently reproduced (~545 g) by June 2026 hyperscale study"
},
"regionalVariation": {
"bestCase": { "location": "Quebec, Canada", "intensity": 32 },
"worstCase": { "location": "Central US (coal regions)", "intensity": 1000 },
"variationFactor": "16x difference between best and worst locations",
"unit": "g CO₂e/kWh"
},
"electricityPrices": {
"title": "The 2026 story: electricity prices",
"pjmRecord": "$333.44/MW-day record capacity price (2027/28 auction, third consecutive record); $16.4B total market cost",
"attribution": "PJM's market monitor attributes 63% of the 2025/26 auction price increase ($9.3B recovered from customers) to data centres",
"billImpact": "Residential bills in PJM states rising 1.5-5% from June 2026 (~$16-18/month in the worst-hit areas); NRDC projects up to $163B in PJM bill increases through 2033 (model-dependent scenario)",
"sources": "PJM, Monitoring Analytics, Canary Media, IEEFA (Dec 2025 - 2026)"
},
"projections": {
"iea2030": {
"globalDemand": 950,
"unit": "TWh",
"percentOfGlobalElectricity": "~3%",
"source": "IEA Electricity 2026; central case unchanged, higher scenarios now judged less likely due to supply-chain bottlenecks"
},
"iea2035": {
"globalDemand": 1200,
"unit": "TWh",
"source": "IEA Electricity 2026 (Base Case)"
},
"lawrenceBerkeley2028": {
"usShare": "6.7-12% of US electricity",
"source": "Lawrence Berkeley National Laboratory / DOE"
},
"fossilFuelShare": {
"value": 60,
"unit": "% of new demand met by fossil fuels by 2030",
"source": "Goldman Sachs; consistent with 2026 gas-turbine order books (~100 GW vs 60-70 GW/yr manufacturing capacity)"
}
},
"overbuildRisk": "The buildout may not materialise as promised: Microsoft froze ~1.5 GW of self-build capacity and Satya Nadella conceded 'there will be an overbuild'; the Bank for International Settlements lists an AI capex bust among its top systemic risks (2026), noting ~$1T/year hyperscaler capex against ~$50-60B AI revenue",
"context": "Data centres are ~8% of projected global electricity demand GROWTH to 2030 - less than EVs or air conditioning (Carbon Brief). Current scale is modest; the trajectory and its local grid/water impacts are the legitimate concerns."
},
"fuelConsumption": {
"unit": "litres/year",
"items": [
{
"id": "personal_car_commute",
"label": "Average US Car Commute",
"sublabel": "12,282 km/year @ 8.65 L/100km",
"value": 1066,
"displayValue": "1,066 litres",
"color": "#ef4444",
"source": "EPA Automotive Trends (MY2024) / US Census",
"sourceUrl": "https://www.epa.gov/automotive-trends",
"year": 2026,
"notes": "69.2% of Americans drive alone to work (ACS 2024)"
},
{
"id": "pickup_commute",
"label": "Full-Size Pickup Commute",
"sublabel": "12,282 km/year @ 13.07 L/100km",
"value": 1605,
"displayValue": "1,605 litres",
"color": "#dc2626",
"source": "US Gas Price News",
"sourceUrl": "https://news.usgasprice.com/how-much-fuel-does-the-average-american-commute-use-each-year/",
"year": 2024
},
{
"id": "hybrid_commute",
"label": "Hybrid Car Commute",
"sublabel": "12,282 km/year @ 4.70 L/100km",
"value": 579,
"displayValue": "579 litres",
"color": "#22c55e",
"source": "US Gas Price News",
"year": 2024
}
],
"context": {
"totalNational": "$90+ billion spent annually on commuting fuel",
"idlingWaste": "57-76 litres/year wasted from idling alone",
"averageCost": "$2,043/year average commuting cost"
}
},
"aiTraining": {
"models": [
{
"id": "gpt3",
"label": "GPT-3 Training",
"parameters": "175 billion",
"energyGWh": 1.287,
"energyDisplay": "1.287 GWh (1,287 MWh)",
"carbonTonnes": 552,
"perUserWh": {
"users300M": 4.29,
"users900M": 1.43
},
"equivalents": [
"120 US homes for 1 year"
],
"source": "MIT Technology Review / Various research",
"year": 2020
},
{
"id": "gpt4",
"label": "GPT-4 Training",
"parameters": "~200-280 billion (estimated)",
"energyGWh": 50,
"energyDisplay": "~50 GWh",
"carbonTonnes": "12,456-14,994",
"perUserWh": {
"users300M": 166.7,
"users900M": 55.6
},
"equivalents": [
"San Francisco for 3 days",
"~300 round-trip flights NYC to San Francisco"
],
"vsGpt3": "40-48x more energy than GPT-3",
"source": "MIT Technology Review (leaked data, unverified)",
"sourceUrl": "https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/",
"year": 2023,
"notes": "Training cost over $100 million; ~25,000 A100 GPUs for 90-100 days"
},
{
"id": "gemini_ultra",
"label": "Gemini 1.0 Ultra Training",
"energyGWh": 70,
"energyDisplay": "~70 GWh",
"perUserWh": {
"users300M": 233.3,
"users900M": 77.8
},
"source": "Derived from Epoch AI models database (facility power x duration, average-draw basis)",
"sourceUrl": "https://epoch.ai/data/notable-ai-models",
"year": 2023,
"notes": "The only Gemini model with any published training-energy estimate - Google discloses nothing. Long superseded by Gemini 2 and 3, for which no credible estimate exists"
},
{
"id": "llama31_405b",
"label": "Llama 3.1 405B Training",
"parameters": "405 billion",
"energyGWh": 21.6,
"energyDisplay": "~21.6 GWh",
"carbonTonnes": 8930,
"perUserWh": {
"users300M": 72,
"users900M": 24
},
"source": "Derived from Meta's model card: 30.84M H100 GPU-hours x 700 W (GPU-only)",
"sourceUrl": "https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md",
"year": 2024,
"notes": "Best-disclosed large run: Meta publishes GPU-hours and 8,930 tCO2e (location-based; offset to zero market-based). GPU-only energy - facility-level runs higher (Epoch estimates ~35-48 GWh)"
},
{
"id": "deepseek_v3",
"label": "DeepSeek-V3 Training",
"parameters": "671 billion (37B active)",
"energyGWh": 1.95,
"energyDisplay": "~1.95 GWh",
"perUserWh": {
"users300M": 6.5,
"users900M": 2.2
},
"source": "Derived from DeepSeek's technical report: 2.788M H800 GPU-hours x 700 W (GPU-only)",
"sourceUrl": "https://arxiv.org/abs/2412.19437",
"year": 2024,
"notes": "Roughly 25x less than GPT-4's estimated training energy - the efficiency outlier among frontier-class models. GPU-only; facility energy ~1.1-1.3x higher"
},
{
"id": "llama4_maverick",
"label": "Llama 4 Maverick Training",
"parameters": "400 billion (17B active, 128 experts)",
"energyGWh": 1.67,
"energyDisplay": "~1.67 GWh",
"carbonTonnes": 645,
"perUserWh": {
"users300M": 5.6,
"users900M": 1.9
},
"source": "Derived from Meta's model card: 2.38M H100 GPU-hours x 700 W (GPU-only)",
"sourceUrl": "https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md",
"year": 2025,
"notes": "An eighth of Llama 3.1 405B's training energy a year earlier - mixture-of-experts efficiency. Scout used ~3.5 GWh (5.0M GPU-hours). GPU-only; 645 tCO2e location-based, offset to zero market-based"
},
{
"id": "grok4",
"label": "Grok 4 Training",
"energyGWh": 310,
"energyDisplay": "~310 GWh",
"carbonTonnes": 154000,
"waterLitres": 754000000,
"perUserWh": {
"users300M": 1033.3,
"users900M": 344.4
},
"source": "Epoch AI (significant uncertainty flagged)",
"sourceUrl": "https://epoch.ai/data-insights/grok-4-training-resources",
"year": 2025,
"notes": "Largest credibly-estimated training run: ~6x GPT-4; 246M H100-hours, ~$490M compute"
},
{
"id": "gpt5",
"label": "GPT-5 Training",
"compute": "~5x10^25 FLOP (median estimate)",
"source": "Epoch AI",
"sourceUrl": "https://epoch.ai/gradient-updates/why-gpt5-used-less-training-compute-than-gpt45-but-gpt6-probably-wont",
"year": 2025,
"notes": "Used LESS training compute than GPT-4.5 - OpenAI scaled post-training on a smaller base model. Frontier training is not monotonically growing; no credible GWh estimate exists"
}
],
"trend": "Frontier training runs now draw >100 MW; training power demand grows ~2.2x/year while hardware efficiency improves ~40%/year (Epoch AI). GW-scale announcements (Stargate, Colossus 2) are build-out CAPACITY targets, not measured consumption.",
"context": "Training is a one-time cost amortised across all users. Over a deployed model's lifetime, INFERENCE dominates: estimates range 60-90% of total footprint depending on boundary.",
"chartNote": "Model vintages are forced by disclosure, not choice: after 2024 the frontier went dark. Grok 4 is the newest frontier-scale estimate anywhere; no credible training-energy figure exists for GPT-4.5, GPT-5, Gemini 2 or 3, or any Claude model. Figures also mix measurement boundaries: Grok 4 and Gemini 1.0 Ultra are Epoch AI facility-level estimates, while the Llama and DeepSeek figures are GPU-only, derived from developer-disclosed GPU-hours (facility energy runs roughly 1.3-2x higher)."
},
"chatgptStats": {
"dailyQueries": {
"value": 2500000000,
"display": "2.5+ billion",
"source": "OpenAI (July 2025 - latest officially confirmed volume)",
"year": 2025
},
"annualEnergy": {
"value": 310,
"unit": "GWh",
"equivalent": "Powering ~30,000 US homes annually",
"source": "Derived: 0.34 Wh x 2.5B/day x 365; MIT Technology Review independently estimated ~300 GWh"
}
},
"chatgptUsers": {
"global": {
"weeklyActive": [
{ "date": "2024-12", "users": 300000000, "display": "300 million" },
{ "date": "2025-08", "users": 700000000, "display": "700 million" },
{ "date": "2025-10", "users": 800000000, "display": "800 million" },
{ "date": "2026-Q1", "users": 900000000, "display": "900+ million" }
],
"monthlyActive": {
"value": 1000000000,
"display": "~1 billion",
"date": "2026-06",
"source": "Reuters / Sensor Tower estimate (MAU, a different metric to WAU)"
},
"source": "OpenAI / TechCrunch",
"sourceUrl": "https://techcrunch.com/2026/02/27/chatgpt-reaches-900m-weekly-active-users/"
},
"usa": {
"trafficShare": "~17-19% of global usage (US); ~19.8% for North America (OpenAI)",
"source": "OpenAI / Similarweb-based estimates",
"year": 2026
},
"adoption": {
"usedChatbots": 49,
"usedChatbots2024": 33,
"dailyUse": 24,
"platformShare": "ChatGPT 44%, Gemini 24%, Copilot 17%, Meta AI 14%, Grok 8%, Claude 6%",
"unit": "% of US adults",
"source": "Pew Research Center (survey Feb 2026, n=5,119, published Jun 2026)",
"sourceUrl": "https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/"
}
},
"individualImpact": {
"title": "Personal AI Usage in Context",
"dailyUsage": [
{
"queries": 10,
"energyWh": 3.4,
"percentOfDailyUse": 0.012,
"context": "Less than running a fridge for 2 minutes"
},
{
"queries": 50,
"energyWh": 17,
"percentOfDailyUse": 0.06,
"context": "About 10 seconds of microwave use"
},
{
"queries": 100,
"energyWh": 34,
"percentOfDailyUse": 0.12,
"context": "About 20 minutes of TV watching"
}
],
"usHouseholdDaily": 28800,
"unit": "Wh",
"source": "Derived from EIA average household consumption (~10,500 kWh/yr)",
"note": "Even 100 standard queries/day is ~0.12% of typical household electricity. Corrects the 0.3% previously shown. Reasoning-heavy use runs higher (~25-30x per query)"
},
"dailyHabitsComparison": {
"title": "Typical Daily Usage Patterns",
"description": "How energy consumption compares for average daily use",
"tvStreaming": {
"hoursPerDay": 6.75,
"hoursDisplay": "6 hours 45 minutes",
"breakdown": {
"traditionalTV": 2.48,
"streaming": 3.83
},
"energyPerHour": {
"streaming": 77,
"tv65inch": 95,
"average": 77
},
"dailyEnergyWh": {
"low": 520,
"high": 675,
"typical": 520
},
"sources": ["eMarketer US Time Spent With Media 2025", "Nielsen"],
"year": 2025,
"notes": "2025 reported as the peak year; traditional TV component now declining"
},
"chatgpt": {
"queriesPerDay": {
"typicalWeeklyUser": 2.8,
"dailyActiveUser": 22,
"range": "2-3"
},
"energyPerQuery": 0.34,
"dailyEnergyWh": {
"typicalUser": 0.95,
"dailyActiveUser": 7.5
},
"sources": ["OpenAI usage data"],
"year": 2026,
"methodology": "2.5B+ daily prompts / 900M weekly active users ≈ 2.8 prompts/day"
},
"comparison": {
"ratioTypicalUser": 550,
"ratioDailyActiveUser": 70,
"description": "TV/streaming uses ~550x more energy daily than typical ChatGPT usage"
}
},
"perspectives": [
{
"title": "Household Leaks",
"fact": "A single dripping tap wastes 11,000+ litres per year",
"comparison": "Equivalent to 35 million ChatGPT queries",
"context": "EPA estimates 10% of US homes have leaks wasting 340+ litres daily"
},
{
"title": "Daily Media Consumption",
"fact": "Average American watches 6 hours 45 minutes of TV/streaming per day",
"comparison": "Uses 520 Wh daily vs ~1 Wh for typical AI use (2-3 queries)",
"context": "Sources: eMarketer, Nielsen (2025)"
},
{
"title": "Streaming Energy",
"fact": "1 hour of Netflix streaming uses 77 Wh",
"comparison": "Equivalent to 227 ChatGPT queries - and ~72% of it is your TV, not the data centre",
"context": "Video streaming accounts for ~60% of global internet traffic"
},
{
"title": "Golf Course Water Use",
"fact": "US golf courses use 2.01 trillion litres of water per year",
"comparison": "Roughly 3-6x ALL global AI data centres combined - including the water behind their electricity",
"context": "~16,000 courses serving 29.1 million golfers vs 900+ million ChatGPT users"
},
{
"title": "Corn and Ethanol",
"fact": "US corn irrigation draws ~20.4 trillion litres per year - and ~35-40% of the crop is burned as car fuel",
"comparison": "26-65x the water footprint of all global AI data centres",
"context": "Each litre of ethanol carries a ~1,200-1,500 litre water footprint (USDA/Water Resources Research). Different water kinds: irrigation is largely non-potable; data centre cooling is often municipal"
},
{
"title": "Same Industry, 170x Apart",
"fact": "Mistral discloses 45 mL of water per reply; Google discloses 0.26 mL per prompt",
"comparison": "The 170x gap is methodology, not thirst",
"context": "Mistral counts water consumed generating the electricity (Scope 2); Google counts on-site cooling. Never compare figures without checking the boundary"
},
{
"title": "Water is Local",
"fact": "~2/3 of new US data centres since 2022 are in high water-stress areas",
"comparison": "The same facility can be benign in a wet region and a problem in a dry one",
"context": "Bloomberg / S&P Global; a Jun 2026 Guardian analysis found 517 of 809 planned US facilities on drought-hit land. Facility-level water data remains largely undisclosed"
},
{
"title": "Reasoning Costs Tokens",
"fact": "Reasoning modes use 25-30x more energy than standard queries on average",
"comparison": "Extreme cases measured at 150-700x, driven by ~10x more output tokens",
"context": "Measured: Hugging Face AI Energy Score v2, ML.Energy v3 (Dec 2025). One visible query can be many hidden ones"
},
{
"title": "Electricity Prices",
"fact": "Data centres drove 63% of the record PJM capacity price increase - $9.3B recovered from customers",
"comparison": "Residential bills in PJM states rising 1.5-5% from June 2026",
"context": "PJM market monitor / Canary Media. The strongest evidence yet of AI's cost to ordinary bill-payers"
},
{
"title": "Commuting Emissions",
"fact": "One year of average US car commuting uses 1,066 litres of fuel",
"comparison": "Produces approximately 2.5 tonnes of CO₂",
"context": "69% of Americans drive alone to work"
},
{
"title": "Claude's Undisclosed Footprint",
"fact": "Anthropic holds ~40% of enterprise LLM API spend but publishes no query volume or per-query figures",
"comparison": "Only 6% consumer reach, yet the largest enterprise share - an API-shaped footprint",
"context": "Menlo Ventures (year-end 2025); Pew (Feb 2026). With no disclosed volume or per-query data, Claude's total footprint can't be independently estimated"
}
],
"keyInsights": [
{
"stat": "3-6x",
"description": "More water used by US golf courses than ALL global AI data centres",
"tone": "good"
},
{
"stat": "~11,700x",
"description": "More water lost to US household leaks (dripping taps, toilets) than ChatGPT's global cooling water",
"tone": "good"
},
{
"stat": "~35x",
"description": "More energy for 1 hour of big-screen TV than a 10-prompt ChatGPT conversation",
"tone": "good"
},
{
"stat": "~2,900x",
"description": "More energy for a 10-second AI video than a text query (estimate) - the real outlier",
"tone": "caution"
},
{
"stat": "2/3",
"description": "New US data centres built since 2022 sit in high water-stress areas - where matters more than how much",
"tone": "caution"
},
{
"stat": "33x",
"description": "Gemini efficiency gain in 12 months (Google-reported)",
"tone": "good"
}
],
"mythBusters": [
{
"myth": "AI is destroying the climate",
"reality": "US data centres are 2.18% of national emissions (2023 data) - comparable to aviation. The concern is the trajectory: global data centre electricity grew 17% in 2025, with AI-focused facilities up ~50%.",
"source": "Harvard/UCLA Study; IEA Electricity 2026"
},
{
"myth": "Every AI query wastes massive resources",
"reality": "A typical text query uses 0.24-0.34 Wh and ~5 drops of water - like microwaving for 1 second. Reasoning modes (25-30x) and video generation (~2,900x) are the real outliers.",
"source": "Google, OpenAI, HF AI Energy Score"
},
{
"myth": "Each query uses a bottle of water",
"reality": "Official on-site figures are 0.26-0.32 mL - about 5 drops. The viral ~500 mL figure counted the water behind electricity generation for older models. Boundary, not dishonesty, explains the 1,000x gap.",
"source": "Google, OpenAI, UC Riverside"
},
{
"myth": "AI efficiency isn't improving",
"reality": "Inference cost per token fell ~50x/year (Epoch AI) and Google reports a 33x per-query gain. But reasoning models spend ~10x more tokens, so energy per query can still rise.",
"source": "Epoch AI, Google, ML.Energy"
},
{
"myth": "Data centres run on clean energy",
"reality": "95% of US data centres sit on dirtier-than-average grids (48% above US average carbon intensity, independently reproduced in 2026), ~2/3 of new builds are in water-stressed areas, and the near-term buildout is gas-heavy.",
"source": "Harvard/UCLA, S&P Global, RBC"
},
{
"myth": "Training is the main energy cost",
"reality": "Over a model's lifetime, inference dominates (60-90% of footprint). And frontier training isn't monotonically growing: GPT-5 used less training compute than GPT-4.5.",
"source": "Epoch AI, UNU-INWEH"
},
{
"myth": "The water numbers must be lies - they differ by 1,000x",
"reality": "Altman's 0.32 mL counts on-site cooling per query; Morgan Stanley's ~1 trillion litres by 2028 adds power-generation and chip-fab water, projected forward. Both can be arithmetically honest. Always ask where the boundary sits.",
"source": "OpenAI, Morgan Stanley, USGS"
}
],
"quotes": [
{
"text": "Individual usage of ChatGPT and other LLMs for most people is a small part of their carbon and energy footprint.",
"source": "Sustainability by Numbers",
"context": "Analysis of AI environmental impact"
},
{
"text": "We should stop trying to reverse-engineer numbers based on hearsay and put more pressure on these companies to actually share the real ones.",
"author": "Sasha Luccioni",
"role": "Creator of AI Energy Score",
"source": "MIT Technology Review"
},
{
"text": "It suggests that energy demand for data centres and AI will still be pretty small for the next five years at least. I read it as them saying: 'Everyone just needs to chill out a bit.'",
"author": "Hannah Ritchie",
"role": "Data scientist, Our World in Data (on the IEA outlook)",
"source": "Sustainability by Numbers"
},
{
"text": "Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust.",
"author": "Bank for International Settlements",
"role": "Annual Economic Report 2026",
"source": "BIS"
}
],
"credits": {
"framing": "The measurement-boundary framing of the water section was prompted by Hank Green's video 'Why is Everyone So Wrong About AI Water Use?' (June 2026). Figures were independently verified against primary sources; where his numbers didn't hold (corn total, the 'training is 50%' claim), the verified figures are used instead.",
"url": "https://youtu.be/H_c6MWk7PQc"
}
}