Every time you ask a chatbot a question, a machine somewhere in a windowless building spins up, burns a tiny bit of electricity, and gives off heat that has to go somewhere. Multiply that by billions of requests a day and you get one of the fastest-growing power loads on the planet. So how does ai use water and energy in practice? The short version: electricity runs the chips that do the math, and water helps carry away the heat those chips create. The long version involves power plants, cooling towers, chip design, and choices that companies and communities are making right now.
This guide breaks the whole system down in plain language. You will learn where the electricity actually goes inside a data center, why some facilities drink millions of gallons of water while others barely sip, how training a giant model differs from answering your one-line prompt, and what the real numbers look like per query. You will also see the myths that get repeated online, real case studies from towns hosting these facilities, practical steps that cut consumption, and where the technology is heading. By the end, you will be able to judge headlines about AI’s footprint for yourself instead of guessing.
What AI Actually Consumes Behind the Screen
Artificial intelligence is not magic living in a cloud. It is thousands of physical computer chips packed into racks inside data centers, all crunching numbers at once. AI uses energy because its specialized chips perform trillions of mathematical calculations per second, and it uses water because those chips turn nearly all that electricity into heat that cooling systems must remove, often by evaporating water in cooling towers or by pulling water through chilled loops. Energy comes first; water follows as a consequence of energy.
Think of an AI server rack like a very large space heater that happens to do math. A single Nvidia H100 accelerator draws up to 700 watts. Newer rack-scale systems pull 120 kilowatts or more per rack, which is roughly the power of 80 to 100 average American homes squeezed into a footprint the size of a refrigerator. That heat has to leave the room within seconds, or the chips throttle themselves or fail.
There are also two kinds of water use that people mix up constantly. Direct water is what a data center pulls from a utility or well for its own cooling. Indirect water is what power plants consume to generate the electricity the data center buys. Coal, gas, and nuclear plants boil water to spin turbines and then use more water to condense the steam. In many regions, indirect water use is actually larger than the water used on site.
- Direct energy: electricity for chips, memory, storage, networking, fans, pumps, and chillers.
- Direct water: evaporative cooling, humidity control, and occasional system flushing.
- Indirect energy: the fuel burned or resources used at the power plant.
- Indirect water: water evaporated or withdrawn during electricity generation.
- Embodied impact: the energy and ultrapure water used to manufacture the chips themselves in semiconductor fabs.
That last point surprises people. Chip fabrication plants are enormous water users. A single advanced fab can use millions of gallons of ultrapure water per day to rinse silicon wafers. So part of AI’s water story happens years before a server ever powers on.
Inside a Data Center: Where the Electricity Really Goes
Once electricity enters a data center, it splits into a few predictable buckets. The biggest share goes to the IT equipment itself, and the second biggest goes to cooling. The rest disappears into power conversion losses, lighting, and backup systems that idle constantly so they can take over during an outage.
Engineers measure this split with a number called Power Usage Effectiveness, or PUE. You get PUE by dividing total facility power by the power delivered to the computers. A PUE of 2.0 means half the electricity does useful work and half goes to overhead. A PUE of 1.1 means only 10 percent is overhead. The global average sits around 1.5 to 1.6, while the best hyperscale campuses run between 1.08 and 1.2.
The typical energy breakdown
| System | Share of total power | What it does |
|---|---|---|
| Servers and AI accelerators | 50-65% | Runs the actual model math |
| Cooling (chillers, fans, pumps) | 20-40% | Removes heat from chips and rooms |
| Power delivery losses | 5-12% | Transformers, UPS units, conversion |
| Storage and networking | 5-10% | Moves and holds data |
| Lighting, security, other | 1-3% | Building operations |
Here is a practical way to picture the scale. The International Energy Agency estimated that data centers worldwide consumed roughly 415 terawatt-hours of electricity in 2024, about 1.5 percent of global electricity, and projected that number could climb toward 945 terawatt-hours by 2030. In the United States, Lawrence Berkeley National Laboratory reported data centers used about 176 terawatt-hours in 2023, around 4.4 percent of national electricity, with projections reaching 6.7 to 12 percent by 2028. AI servers drive most of that growth because they are far denser than traditional web servers.
Density is the key word. A standard rack of web servers might draw 5 to 10 kilowatts. An AI training rack can draw 100 kilowatts or more. Air alone struggles to cool that, which pushes operators toward liquid cooling and changes the water equation completely.
Why AI Needs Water, and How Cooling Systems Differ
Water shows up in data centers because evaporation is an incredibly cheap way to move heat. When water evaporates, it absorbs a lot of energy and carries it into the air. A cooling tower can dump megawatts of heat using far less electricity than a mechanical chiller running the same job. The tradeoff is simple and stark: you either spend water or you spend electricity.
Operators track this with Water Usage Effectiveness, or WUE, measured in liters per kilowatt-hour of IT energy. The average US data center runs near 1.8 liters per kilowatt-hour. Google has reported a fleet-wide average close to 1.1, and some facilities using air cooling or closed loops report numbers below 0.2.
The main cooling approaches
- Evaporative or open-loop cooling towers: Water evaporates to shed heat. Very energy efficient, very water hungry. Losses come from evaporation plus “blowdown,” the mineral-heavy water drained off to keep pipes clean.
- Closed-loop water systems: The same water circulates over and over inside sealed pipes. Almost no water leaves the loop, but the system needs more electricity because chillers do the heavy lifting.
- Air-cooled chillers and free cooling: Cold outside air handles the load in cool climates. Near zero water use, higher power use in warm months.
- Direct-to-chip liquid cooling: Cold plates sit right on the processors and carry heat away in a closed loop. Highly effective for dense AI racks and increasingly standard.
- Immersion cooling: Servers sit fully submerged in a non-conductive fluid. Excellent heat transfer and near zero water use, but it requires new hardware designs.
Which method uses what
| Cooling method | Typical water use | Typical energy overhead | Best fit |
|---|---|---|---|
| Open evaporative towers | High (1.5-2.5 L/kWh) | Low | Hot, water-rich regions |
| Closed-loop chilled water | Very low | Moderate to high | Water-stressed areas |
| Air-cooled with free cooling | Near zero | Low in cold climates | Northern latitudes |
| Direct-to-chip liquid | Low to moderate | Low | Dense AI clusters |
| Immersion | Near zero | Very low | New builds, high density |
Location matters as much as technology. A facility in Oregon can run on cool outside air most of the year. The same design in Phoenix would need constant mechanical cooling or heavy evaporation. That is why the same company can report wildly different water numbers at two different campuses, and why comparing single sites without context leads people to wrong conclusions.
Training Versus Inference: Two Very Different Bills
AI energy use splits into two phases, and they behave nothing alike. Training is the one-time (or occasional) process of building a model by feeding it enormous amounts of data. Inference is what happens every time someone actually uses that model. People often assume training dominates, but at large scale, inference usually wins over time because it never stops.
Training: a short, brutal sprint
Training a frontier model means running thousands of accelerators nonstop for weeks or months. Researchers estimated that training GPT-3 consumed roughly 1,287 megawatt-hours of electricity and produced around 552 metric tons of carbon dioxide. A widely cited study estimated training that same model in US data centers could evaporate about 700,000 liters of fresh water on site, with several million liters more counted across the full supply chain. Later models are far larger; independent estimates for GPT-4 class training runs land in the tens of gigawatt-hours.
Inference: a marathon that never ends
One chat response is tiny. Google published a measurement showing a median text prompt to its Gemini assistant used about 0.24 watt-hours of electricity and roughly 0.26 milliliters of water, which is about five drops. OpenAI’s leadership shared a similar figure of about 0.34 watt-hours per average query. For comparison, 0.24 watt-hours is close to running a microwave for one second, or watching a streaming video for about nine seconds.
Now do the math on volume. If a service handles a billion prompts per day at 0.3 watt-hours each, that is 300 megawatt-hours daily, or roughly 110 gigawatt-hours per year, from text queries alone. Image and video generation cost far more per request, sometimes 10 to 100 times the energy of a text reply, because the model runs many more computation steps.
| Task | Rough energy per request | Everyday comparison |
|---|---|---|
| Traditional web search | 0.1-0.3 Wh | A few seconds of a light bulb |
| Short AI text reply | 0.2-0.5 Wh | One second of a microwave |
| Long reasoning answer | 2-20 Wh | Charging a phone a few percent |
| AI image generation | 2-5 Wh | Running a laptop for a minute |
| Short AI video clip | 50-500 Wh | Running a laptop for an hour or more |
One more wrinkle: reasoning models that “think” before answering generate long chains of hidden text. That extra thinking multiplies the compute per answer. So the trend toward smarter answers can push energy per query up even as the chips themselves become more efficient.
Myths and Misunderstandings About AI’s Footprint
This topic attracts bad numbers. Some come from outdated studies, some from mixing units, and some from comparing a worst-case facility to a global average. Sorting the real concerns from the noise makes the conversation far more useful.
The most famous example is the claim that a single AI chat uses a 16-ounce bottle of water. That figure came from a research estimate covering roughly 20 to 50 exchanges with an older model in a specific hot-climate data center, and it counted both on-site and power-plant water. Measured figures from newer, efficient systems land closer to a fraction of a milliliter per prompt. Both numbers can be technically defensible under their own assumptions, which is exactly why context matters.
- Myth: The water is destroyed. Evaporated water returns to the water cycle as rain. The real problem is local and timing based: a watershed under drought loses supply now, even if the vapor falls somewhere else later.
- Myth: All data centers are water hogs. Many run closed-loop or air-cooled designs and consume almost no water on site. Water use varies by an order of magnitude between facilities.
- Myth: AI is the main driver of rising electricity demand. AI is a fast-growing slice, but electrification of vehicles, heating, and industry drives large shares of load growth too. AI’s share is significant and rising, not total.
- Myth: Your personal chatbot use is the problem. Individual prompts are small. The aggregate matters, and infrastructure decisions by operators and utilities matter far more than personal restraint.
- Myth: Efficiency gains will fix everything. Chips get more efficient every generation, but demand grows faster. Economists call this the rebound effect, and it has held true so far.
There is also a common unit mix-up. Water “withdrawal” means water pulled from a source, most of which may return to the river. Water “consumption” means water that evaporates and does not come back locally. Headlines that treat withdrawal as consumption can overstate impact by a factor of five or more. When you read a statistic, check which word it uses.
Real Places, Real Impacts: What Communities Are Seeing
Abstract terawatt-hours become concrete when a data center lands next to a town. Several patterns repeat around the world, and they show both the strain and the solutions.
Water-stressed desert regions
In parts of Arizona, data center campuses sit in areas where groundwater levels have dropped for decades. Some operators responded by switching to air-cooled or closed-loop designs that trade extra electricity for near-zero water use. Others signed agreements to use reclaimed wastewater instead of drinking water, which sidesteps the direct competition with households.
Grid pressure in concentrated hubs
Northern Virginia hosts one of the densest data center clusters in the world, and utilities there have warned about transmission constraints and multi-year interconnection queues. In Ireland, data centers grew to consume more than a fifth of national metered electricity, prompting the grid operator to pause new connections in the Dublin area. These cases show the bottleneck is often not generation but wires, substations, and transformers.
Heat reuse success stories
Northern Europe offers the most encouraging counterexample. Several facilities in Denmark, Finland, and Sweden pipe waste heat into district heating networks that warm thousands of homes. Instead of dumping heat into the air or evaporating water to get rid of it, they sell it. That turns a liability into a revenue stream and cuts community heating fuel at the same time.
A useful scenario to picture: imagine a 100-megawatt AI campus running at full load. Over a year it uses roughly 876,000 megawatt-hours of electricity. At a WUE of 1.8 liters per kilowatt-hour, on-site water consumption approaches 1.5 billion liters per year, comparable to the household use of a small city. Switch that campus to a closed-loop design at 0.1 liters per kilowatt-hour and on-site water drops below 90 million liters, though electricity use may rise 5 to 15 percent. Same building, radically different footprint, based on one engineering decision.
Practical Ways to Cut AI’s Water and Energy Use
Plenty of levers exist, and they operate at different levels. Some belong to data center operators, some to model developers, and a few belong to everyday users and businesses buying AI services.
What operators can do
- Deploy direct-to-chip or immersion cooling to handle dense racks without evaporative towers.
- Raise allowed server inlet temperatures so free cooling works more hours per year.
- Use reclaimed, recycled, or non-potable water instead of drinking water supplies.
- Site new campuses in cool climates with abundant clean power and grid capacity.
- Sign clean energy contracts matched hour by hour, not just annually averaged.
- Capture and sell waste heat to district heating, greenhouses, or industrial neighbors.
- Publish PUE, WUE, and carbon data per facility so communities can verify claims.
What model builders can do
- Right-size the model. A small, task-specific model often matches a giant general model on narrow jobs while using a fraction of the compute.
- Distill and quantize. Compressing a large model into a smaller one, or running it at lower numerical precision, can cut inference energy by 50 to 90 percent.
- Cache repeated answers. Many queries repeat. Serving a stored answer costs almost nothing.
- Batch requests. Processing many prompts together uses accelerators far more efficiently than one at a time.
- Schedule training flexibly. Shifting jobs to hours and regions with surplus clean power lowers both carbon and grid strain.
- Route smartly. Send easy questions to a cheap small model and only escalate hard ones to the big model.
What users and businesses can do
If your company runs AI features, ask your vendor for energy and water disclosures, choose regions with low-carbon grids when you pick a cloud location, and avoid defaulting every task to the largest available model. For individuals, the honest advice is that your prompts are small, but thoughtful use still adds up across an organization. Skipping unnecessary image or video generation saves far more than trimming text queries.
Helpful resources exist for anyone who wants to dig deeper: the IEA’s energy and AI reporting, Lawrence Berkeley National Laboratory’s US data center reports, the Uptime Institute’s annual surveys, cloud carbon dashboards from major providers, and open tools like CodeCarbon and ML CO2 Impact that estimate the footprint of a specific training run.
What Is Changing: The Next Wave of Efficient AI
The trajectory is not fixed. Several forces are pushing in opposite directions, and how they balance will decide whether AI’s footprint keeps climbing steeply or flattens out.
On the efficiency side, each chip generation delivers far more performance per watt. Liquid cooling is moving from exotic to standard for AI racks, which cuts both cooling energy and, in closed-loop form, water use. Model architectures are getting leaner too, with mixture-of-experts designs activating only a slice of the network per request instead of the whole thing. Smaller models running directly on phones and laptops shift some work off data centers entirely.
On the growth side, demand keeps climbing. Video generation, agents that run for minutes or hours on a task, and reasoning models that produce long internal thought all raise compute per user. Rack power densities are heading toward 250 kilowatts and beyond, forcing entirely new building designs.
- Behind-the-meter generation: Campuses building their own gas turbines, solar farms, fuel cells, or contracting nuclear power to avoid grid queues.
- Small modular reactors: Several operators have signed agreements for future nuclear capacity, though most units are years away.
- Mandatory disclosure: Regulations in the EU and proposals elsewhere require reporting on data center energy, water, and heat reuse.
- Grid flexibility deals: Data centers agreeing to reduce load during peak hours in exchange for faster connections.
- Waterless designs by default: New builds in dry regions increasingly commit to closed-loop cooling from day one.
- Heat as a product: More projects treating waste heat as something to sell rather than something to discard.
The most likely outcome is a split world. Well-sited, well-designed campuses will run on clean power with minimal water, while poorly sited ones will keep drawing criticism for straining local grids and watersheds. Transparency will separate the two, which is why disclosure rules matter more than any single technology.
Frequently Asked Questions About AI, Water, and Energy
Does using a chatbot really waste water?
A single text prompt uses a tiny amount, on the order of a fraction of a milliliter to a few milliliters at most, depending on the facility and cooling method. The concern is not one prompt but billions of them concentrated in specific watersheds. Location determines whether that matters locally.
Is AI worse for the environment than streaming video or gaming?
Per hour of use, streaming video and gaming still account for far more total household-linked energy globally because so many people do them for hours daily. AI’s per-request cost is small, but its growth rate is much steeper, and it concentrates demand in a few dozen regions rather than spreading it out.
Can AI actually save energy too?
Yes, and this is a real part of the ledger. AI systems optimize building HVAC, forecast wind and solar output, reduce grid losses, cut fuel use in shipping routes, and speed up materials research for batteries and solar cells. Google reported using AI to cut data center cooling energy substantially. Whether the savings outweigh the consumption depends on how widely those applications spread.
Which uses more water, training or everyday use?
Training a single model consumes a large amount in a short burst. Inference consumes less per event but runs forever across millions of users. For popular deployed models, cumulative inference typically passes training within months.
Do data centers use drinking water?
Many do use municipal potable supplies because that is what the pipes deliver. A growing number switch to reclaimed wastewater, harvested rainwater, or industrial water instead. Some publish the split between potable and non-potable sources.
How can I check a specific provider’s footprint?
Look for annual environmental or sustainability reports that list PUE, WUE, carbon-free energy percentage, and water source breakdowns by region. Cloud providers also publish region-level carbon intensity data you can use when choosing where to run workloads.
AI runs on physical infrastructure, and that infrastructure runs on electricity and, in many cases, water. Chips do the math and turn nearly all the power they draw into heat, cooling systems carry that heat away, and the choice between evaporating water or spending more electricity shapes each facility’s footprint. Training builds the model in an intense burst, inference serves users forever, and per-query costs are small while totals are large. The numbers vary enormously by location, cooling design, and grid mix, so any single statistic without context tells you very little.
Understanding this matters because the decisions being made now, about where campuses get built, how they get cooled, and how honestly operators report their numbers, will lock in impacts for decades. The encouraging part is that the fixes are not science fiction. Closed-loop cooling, reclaimed water, cool-climate siting, heat reuse, smaller models, and smarter routing all work today, and many operators already use them. Keep asking for real data, favor providers who publish it, and you will help push the industry toward the version of AI that delivers its benefits without draining the places it calls home.
