RIGHT NOW, AI IS CONSUMING
Every AI query you make costs fresh water. Training models costs millions of liters. And it's accelerating.
AI doesn't drink water — but the data centers that power it do.
AI models run on thousands of GPUs that generate extreme heat. Data centers use evaporative cooling systems where approximately 80% of withdrawn freshwater evaporates and cannot be recovered. The contaminated cooling water picks up dust, minerals, and chemicals, making it unsuitable for reuse.
EESI; UK Government Sustainable ICT Blog56% of electricity powering U.S. data centers comes from fossil fuels. Coal plants require ~19,185 gallons per MWh; natural gas ~2,800 gallons per MWh. This indirect water use totaled roughly 211 billion gallons in 2023 in the U.S. alone.
EESI — Data Centers and Water ConsumptionAn average chip fab consumes 10 million gallons of water daily. Producing ultrapure water requires ~1,500 gallons of piped water for every 1,000 gallons of ultrapure output. Each AI chip needs thousands of gallons before a single query is processed.
EESI — Data Centers and Water ConsumptionAI runs in data centers that rival small cities in water consumption. Here's what the numbers look like on the ground.
Data centers are being built in regions already facing severe water scarcity — competing with agriculture and drinking water supplies.
Experiencing 15 consecutive years of unprecedented drought. A proposed Google data center near Santiago would have required more water than consumed by nearby populations.
Undark, 2025The world's largest data center hub saw water consumption surge 63% in just 4 years (2019-2023), straining local water infrastructure and supply systems.
EESIGoogle's thirstiest data center drinks 2.7 million gallons daily in an agricultural state where groundwater reserves are being overpumped nationally.
Undark, 202568% of data centers are located near protected areas or Key Biodiversity Areas. Water demand is expected to exceed freshwater supply by 40% by decade's end.
UK Government Sustainable ICTHow many AI queries do you make per day? See what it costs in water.
Not all AI tasks are equal. Image generation and complex reasoning use significantly more water.
Estimates based on published research by Shaolei Ren (UC Riverside) and corporate sustainability reports. Includes both direct cooling water and indirect water from electricity generation. Actual usage varies by data center location, cooling technology, and energy source.
AI water demand is projected to rival the water withdrawal of entire countries by 2027.
Major AI companies are consuming billions of liters of water annually — and the numbers keep climbing.
Microsoft's 2025 ESR reports 6,399 ML water consumption for FY2024, down from 7.8B liters in 2023. New zero-water cooling datacenters launched August 2024 save 125M liters per facility annually. WUE improved 39% since 2021 to 0.30 L/kWh.
Microsoft 2025 Environmental Sustainability ReportGoogle's 2025 Environmental Report shows an 8% increase in water consumption driven by AI expansion. The Iowa facility alone consumed 1 billion gallons. Water stewardship projects replenished 4.5B gallons — 64% of freshwater consumption, targeting 120% by 2030.
Google 2025 Environmental Report; AA NewsBefore you ask your first question, millions of liters have already been consumed.
By 2027, global AI water withdrawal could exceed the total annual water withdrawal of multiple countries.
AI doesn't just consume water — it devours electricity. And generating that electricity consumes even more water.
Data center electricity projections include all workloads, not just AI. However, AI is the fastest-growing segment, with Goldman Sachs projecting AI alone could account for 200+ TWh by 2028. Each TWh of electricity from fossil fuels requires approximately 2-19 billion gallons of water for cooling at power plants.
Putting AI's daily global water consumption into perspective.
AI's estimated daily water use could provide this many showers (75L each)
Number of people who could have their daily drinking water needs met (2.5L/day)
Number of Olympic pools (2.5M liters each) that could be filled daily
Amount of rice that could be grown with the same water (2,500L per kg)
AI water consumption is not just growing — it's compounding.
Training consumed ~700,000 liters of water. AI water footprint begins gaining research attention.
Microsoft's water use surges 34%. Google's increases 20%. Researchers link the growth directly to AI infrastructure expansion.
200M+ weekly active users. The research paper "Making AI Less Thirsty" quantifies per-query water cost for the first time. Google water use hits 6.6B liters.
AI integrated into search, productivity, and consumer apps. IEA warns data center electricity demand may double to ~1,000 TWh by 2026. Northern Virginia data centers consume 2 billion gallons — up 63% from 2019. Google's Iowa facility hits 2.7 million gallons/day.
DeepSeek V3 shatters cost assumptions — trained for $5.6M with 1/10th the compute of Llama 3.1. xAI's Colossus in Memphis deploys 200,000 GPUs at 250 MW to train Grok 3. GPT-5 launches in August, consuming up to 40 Wh per response — 63x more than GPT-4o. Meta's Llama 4 uses MoE architecture to cut training energy 5x vs Llama 3. IEA reports global data center electricity hit 415 TWh.
OpenAI CEO Sam Altman defends AI water use at Energy Summit. Food and Water Watch publishes "The Urgent Case Against Data Centers" — one hyperscale center uses as much energy as 2M homes. IEA projects data centers at 650-1,050 TWh by year end. Microsoft launches zero-water cooling datacenters. Google and Meta report lower water figures, but researchers note inconsistent reporting methods and missing indirect water data.
AI water withdrawal projected at 4.2-6.6 billion m³ — exceeding the total annual water withdrawal of 4-6 Denmarks, or half the United Kingdom. Global freshwater supply gap widens.
Awareness is the first step. Here's what individuals, companies, and policymakers can do.
All data on this site is derived from peer-reviewed research, corporate sustainability reports, and international agency publications.
Ren, S., Li, P., et al. "Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models." University of California, Riverside & University of Texas, Arlington. arXiv:2304.03271, 2023.
arxiv.org/abs/2304.03271OECD.AI Policy Observatory. "How Much Water Does AI Consume?" Analysis of AI water withdrawal projections and per-query water intensity data.
oecd.ai/en/wonk/how-much-water-does-ai-consumeInternational Energy Agency. "Electricity 2024: Analysis and Forecast to 2026." Data center electricity consumption projections and efficiency analysis.
iea.org/reports/electricity-2024Microsoft Environmental Sustainability Report 2022, 2023, 2025. Data center water consumption figures, year-over-year increases.
Microsoft Sustainability Report 2025Google Environmental Report 2023, 2024. Data center water consumption, energy usage, and sustainability commitments.
Google 2024 Environmental ReportMeta Sustainability Report 2023. Data center water and energy consumption data.
Meta Sustainability PortalUK Government Sustainable ICT Blog. "AI's Thirst for Water." Data on data center placement risks, freshwater availability, and environmental impacts of AI infrastructure. September 2025.
sustainableict.blog.gov.ukEnvironmental and Energy Study Institute. "Data Centers and Water Consumption." Comprehensive analysis of U.S. data center water usage including daily consumption, regional impacts, semiconductor manufacturing, and indirect water from electricity generation.
eesi.orgUndark Magazine. "AI Data Centers and Water." Investigation into specific facility impacts including Google's Iowa data center (2.7M gallons/day), community water stress, and industry responses. December 2025.
undark.orgForbes. Cindy Gordon. "AI Is Accelerating the Loss of Our Scarcest Natural Resource: Water." Analysis of AI's impact on global water scarcity. February 2024.
forbes.comLincoln Institute of Land Policy. "Data Drain: The Land and Water Impacts of the AI Boom." Katharine Wroth, Land Lines Magazine.
lincolninst.eduThe Independent. "AI Artificial Intelligence Chatbot ChatGPT Data Water Use." UK-focused reporting on AI chatbot water consumption and data center impacts.
independent.co.ukCNBC. "OpenAI's Altman Defends AI Resource Usage." Coverage of Sam Altman's defense of AI water consumption at 2026 Energy Summit. February 2026.
cnbc.comTechTarget. "Data Center Heat Reuse: How to Make the Most of Excess Heat." Analysis of data center cooling methods, heat waste, and water consumption for thermal management.
techtarget.comEpoch AI. "How Much Energy Does ChatGPT Use?" Detailed analysis of per-query energy consumption across GPT-4o, o1, o3, and other models. 2025.
epoch.aiTom's Hardware. "ChatGPT 5 Power Consumption Could Be As Much As Eight Times Higher Than GPT-4." University of Rhode Island AI Lab estimates GPT-5 averages 18.9 Wh per medium response, peaks at 40 Wh. 2026.
tomshardware.comMIT News. "Explained: Generative AI's Environmental Impact." Overview of AI training energy costs, noting frontier models in 2025-2026 expected to exceed 100+ GWh per training run. January 2025.
news.mit.eduAll About AI. "AI Environment Statistics 2026." Comprehensive dataset: 415 TWh global data center electricity (2024), GPT-4 training at 50+ GWh, per-query energy data, company comparisons. 2026.
allaboutai.comDeepSeek. "DeepSeek-V3 Technical Report." Training details: 2.79M H800 GPU-hours, $5.6M cost, achieving competitive performance at ~1/10th the compute of Llama 3.1 405B. December 2024.
arxiv.org/html/2412.19437v1Meta. "Llama 4 Scout / Maverick Model Card." Training specifications: 7.38M H800 GPU-hours for Maverick, 5.17 GWh total energy. 2025.
github.com/meta-llamaData Center Dynamics; xAI. Coverage of xAI's Colossus supercomputer in Memphis: 200,000 H100 GPUs, 250 MW power draw, plans for 1M GPUs at 1-1.5 GW. 2024-2025.
datacenterdynamics.comThe Conversation. "Data Centers Consume Massive Amounts of Water — Companies Rarely Tell the Public Exactly How Much." Reporting on Google's 6B gallons (2024), Meta's 813M gallons, Microsoft's WUE data. August 2025.
theconversation.comFood and Water Watch. "The Urgent Case Against Data Centers." Report on AI data centers consuming outsized energy and water: one hyperscale center uses as much energy as 2M U.S. households. March 2026.
foodandwaterwatch.orgInternational Energy Agency. "Energy and AI" and "Electricity 2025." Updated projections: 415 TWh in 2024, 650-1,050 TWh by 2026, ~945 TWh by 2030 (base case). AI-specific servers 53-76 TWh in 2024.
iea.org/reports/energy-and-aiGoldman Sachs Research. "AI Is Poised to Drive 160% Increase in Data Center Power Demand." Analysis of AI-driven electricity demand growth and infrastructure investment requirements.
goldmansachs.comPatterson, D. et al. "Carbon Emissions and Large Neural Network Training." Google Research, 2021. Energy consumption data for training large language models including GPT-3 (1,287 MWh).
arxiv.org/abs/2104.10350Per-query estimates are based on the Ren et al. (2023) research which calculated that a ChatGPT conversation of 20-50 queries consumes approximately 500mL of water (direct evaporative cooling at Microsoft data centers). The real-time counter estimates global AI queries per second based on published user counts and average usage patterns, multiplied by per-query water intensity. Training water estimates for unreported models are extrapolated from GPT-3 data proportional to published compute requirements. All figures include both Scope 1 (on-site cooling) and Scope 2 (electricity generation) water consumption where data is available. Estimates marked with "~" are derived projections, not direct measurements.
In the spirit of transparency, here's what it cost to build this website using AI.
This entire website — research, HTML, CSS, JavaScript, data visualization, and content — was generated through conversations with Claude, an AI assistant by Anthropic. The irony is not lost on us.
These are estimates based on published per-query resource data. Actual figures depend on data center location, cooling technology, energy mix, and query complexity. The water and energy figures include only direct inference costs — they do not account for the water and energy used to train the Claude model itself, which would add substantially more.