AI water usage compared to other industries tells a very different story than the headlines suggest. Yes, a large data centre can consume up to 5 million gallons of water per day — roughly equivalent to the daily needs of a town of 50,000 people, according to the Environmental and Energy Study Institute. That sounds alarming. But the global cattle industry alone uses approximately 4,555 billion litres of water annually, compared to roughly 18.2 billion litres attributed to AI systems like ChatGPT — nearly 250 times more for a single agricultural sector. The real picture requires comparison, not isolation. Put AI's water use next to steel, paper, agriculture, and semiconductor manufacturing, and a more nuanced truth emerges: AI isn't the water villain it's been painted as.
How much water does AI actually use?
The water demand of AI comes almost entirely from one source: data centre cooling. Servers generate enormous heat. To stop them from failing, facilities pump water through cooling towers, chillers, and increasingly, direct-to-chip systems. When water evaporates in those towers, it's gone — that's the consumption that ends up in the headlines.
The figure most cited is from research by Shaolei Ren at the University of California, Riverside, who estimated that training GPT-3 consumed around 700,000 litres of freshwater. A single conversation with ChatGPT — roughly 20 to 50 questions — uses approximately 500 millilitres, about the volume of a standard water bottle. Scaled to millions of daily users, that adds up fast.
Globally, data centres are estimated to account for around 0.2% of total freshwater withdrawals worldwide. That's not nothing. But it's a useful number to hold in your head before we start comparing.
| AI / Data Centre Activity | Estimated Water Use | Context |
|---|---|---|
| Training GPT-3 (one training run) | ~700,000 litres | UC Riverside estimate |
| One ChatGPT conversation (~25 queries) | ~500 ml | Equivalent to one water bottle |
| Large data centre (daily) | Up to 19 million litres | Per EESI; serves 10,000–50,000 people equivalent |
| Global data centres (annual) | ~18.2 billion litres (AI systems) | Bryant Research estimate, 2025 |
The numbers are real. But they only mean something when stacked against what else we consume water for.
How does AI compare to agriculture and food production?
Agriculture is the single largest consumer of freshwater on the planet — responsible for roughly 70% of all global freshwater withdrawals, according to the UN Food and Agriculture Organization. That's not a close race. It's a different category entirely.
Dairy production alone consumes around 4,555 billion litres of water annually. Beef is even thirstier — producing one kilogram of beef requires approximately 15,400 litres of water when you account for feed crops, drinking water, and processing. A single cotton T-shirt takes around 2,700 litres to produce.
| Industry / Activity | Annual Water Use | Ratio vs AI Systems |
|---|---|---|
| AI systems (ChatGPT etc.) | ~18.2 billion litres | Baseline (1×) |
| Global dairy production | ~4,555 billion litres | ~250× more |
| Global beef production | Trillions of litres | Far exceeds dairy |
| Global cotton farming | ~200 billion litres (est.) | ~11× more |
| Global rice production | ~1,000 billion litres | ~55× more |
The comparison isn't meant to dismiss AI's footprint. It's meant to calibrate the conversation. If the goal is water conservation, redirecting even 5% of global beef consumption would save multiples of everything AI data centres use in a year. That's not a political statement — it's arithmetic.
None of this means AI gets a free pass. But framing AI as a primary water threat while ignoring agriculture misdirects both public concern and policy energy.
Where does heavy industry actually stand?
Before AI entered the discourse, industries like steel, paper, and semiconductor manufacturing were the quiet giants of industrial water use. They still are.
Steel production requires enormous amounts of water for cooling, descaling, and processing. Producing one tonne of steel can require anywhere from 25,000 to 45,000 litres depending on the process and the facility's recycling efficiency. Global steel output exceeds 1.9 billion tonnes per year — the maths there is staggering.
The paper and pulp industry is similarly water-intensive. Producing one tonne of paper can require 10,000 litres or more. Semiconductor fabrication — the industry that makes the chips AI runs on — uses ultrapure water in huge volumes. A single chip fabrication plant can consume millions of litres per day in the purification and etching processes alone.
| Industry | Water Use per Unit of Output | Global Scale Note |
|---|---|---|
| Steel production | 25,000–45,000 litres per tonne | 1.9 billion tonnes/year globally |
| Paper & pulp | ~10,000 litres per tonne | 400+ million tonnes/year globally |
| Semiconductor fabs | Millions of litres per day per plant | Ultrapure water required for chip etching |
| Thermal power generation | ~1,500 litres per MWh (cooling) | Powers most data centres indirectly |
| AI data centres | ~500 ml per 25-query session | Growing, but currently minor at macro scale |
Thermal power generation deserves special mention. Every time a data centre draws electricity from a coal or gas plant, that power station uses water too — often more than the data centre itself. The indirect water cost of AI is real, and it's tied directly to how clean the electricity grid powering those servers actually is. Countries running data centres on hydropower or renewables have a fundamentally different footprint from those still reliant on fossil-fuelled generation.
Is AI's water footprint actually growing?
Honestly? Yes. And that's the part that deserves honest scrutiny.
The AI boom is driving a rapid expansion of data centre infrastructure globally. Microsoft, Google, Amazon, and Meta have all announced multi-billion dollar data centre buildouts in 2024 and 2025. The International Energy Agency projects that data centre electricity consumption could more than double by 2030, and water use scales alongside power demand.
New AI workloads — particularly the large generative models used in tools like ChatGPT, Gemini, and Claude — are significantly more compute-intensive than traditional search or streaming. Inference (running the model to answer your question) is less demanding than training, but it happens billions of times per day. The aggregate matters.
| Factor | Current Status | Trend |
|---|---|---|
| Data centre water intensity | Improving with newer cooling tech | Efficiency rising, but scale growing faster |
| Total data centre count | 8,000+ globally (est.) | Expanding rapidly through 2030 |
| AI model size (parameters) | Hundreds of billions in frontier models | Continuing to grow |
| Cooling innovation (immersion/direct-chip) | Early adoption phase | Could cut water use 90%+ vs air cooling |
| Grid carbon intensity | Varies widely by region | Slowly decarbonising in most markets |
The technology to dramatically reduce data centre water use already exists. Direct-to-chip liquid cooling and full immersion cooling — where servers sit in tanks of non-conductive fluid — can cut water consumption by up to 90% compared to traditional evaporative cooling towers. Some newer hyperscale facilities are targeting a Water Usage Effectiveness (WUE) rating of near zero for direct water consumption, relying entirely on closed-loop systems.
The trajectory matters as much as the current number. AI's water footprint is growing — but so is the industry's ability to shrink the water cost per query.
What does context-adjusted responsibility actually look like?
The goal of this comparison isn't to let AI off the hook — it's to assign responsibility proportionally.
Water stress is a genuine global crisis. Roughly 2 billion people currently live in water-stressed regions, according to the United Nations. When a new data centre opens near a community already facing aquifer depletion, that's a real conflict worth scrutinising. Local impact and global percentage share are different things. A data centre drawing from an already-strained river basin is a problem regardless of what the aggregate global numbers show.
But the policy response should match the scale of the problem. If a government is debating water regulation and focuses primarily on tech campuses while ignoring irrigated monoculture farms using orders of magnitude more water — that's a misallocation of regulatory attention.
| Sector | Share of Global Freshwater Withdrawals (est.) | Primary Use |
|---|---|---|
| Agriculture | ~70% | Irrigation for crops and livestock feed |
| Industry (steel, paper, chemicals) | ~20% | Cooling, processing, manufacturing |
| Municipal / domestic use | ~10% | Drinking, sanitation, household |
| Data centres (all, not just AI) | ~0.2% | Server cooling |
The honest answer is that AI's water footprint is a legitimate and growing concern — but it's currently a minor contributor within a much larger industrial water system. The industries consuming 99.8% of the rest deserve proportionally more scrutiny, more regulation, and more investment in efficiency.
Holding AI accountable while ignoring the rest is bad policy dressed up as environmentalism.
AI's water use is real, measurable, and growing. That deserves honest scrutiny — especially when new data centres are built near already-stressed water supplies. But perspective is not the same as dismissal. Agriculture uses 70% of global freshwater. Steel and paper dwarf AI's consumption. The semiconductor fabs that build AI chips use more water per facility than most data centres. The conversation about AI and water needs better numbers, not just louder headlines. Context is how we get to solutions.
Originally published on SnackIQ
Top comments (0)