Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteSome links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
A 2025 analysis estimated that artificial intelligence systems could account for 32.6 million to 79.7 million metric tons of carbon dioxide and 312.5 billion to 764.6 billion liters of water over the year. The study’s upper carbon estimate is comparable to New York City’s annual emissions; its water range is on the scale of annual global bottled-water consumption. These are modeled estimates, not a direct count of every AI system worldwide—and the claim that AI definitively used more water than the bottled-water market is too strong.
What the 2025 study says—and what it does not
The figures come from a peer-reviewed analysis by Alex de Vries-Gao published in Joule. It estimated AI’s 2025 footprint at 32.6–79.7 million metric tons of CO₂ and 312.5–764.6 billion liters of water. The author compared the high end of the carbon range with New York City’s annual emissions and described the water estimate as comparable to the annual global bottled-water market. Read the study.
That makes the headline directionally grounded, but two qualifications matter. First, the study estimated a possible range; it did not measure every AI query, data center, or provider. Second, the water estimate spans a wide interval: only the upper end clearly exceeds commonly cited bottled-water volumes. “On a similar scale” is more accurate than saying AI definitively exceeded global bottled-water demand.
Free tools Windows power users keep installed
One-click scans. No signup required.
The analysis is best understood as a warning about the potential scale of a fast-growing infrastructure load, not as an audited global environmental ledger. Its totals depend on assumptions about how much data-center activity belongs to AI and how much energy and water that activity requires.
#1 Best Overall
- Easy to use - simply fill to the top, snap to close and hang in the toilet
- For use with older toilets that are not considered high efficiency, low-flow toilets
- Every flush saves up to 0.8 gallons per flush
- Constructed of non-corrosive materials
- No mess - the anti-evaporation snap-lock means the bag never needs refilling
| Measure | 2025 estimate | How to read it |
|---|---|---|
| Carbon emissions | 32.6–79.7 million metric tons of CO₂ | The upper estimate is compared with New York City’s annual emissions; it is not a direct city-to-AI measurement using necessarily identical accounting boundaries. |
| Water consumption | 312.5–764.6 billion liters | A global annual volume compared with bottled-water market volumes, which vary by year and definition. |
| Electricity demand | A substantial global data-center load; some reporting places the upper estimate near 23 gigawatts | This is an estimate, not a separately metered total for all AI workloads. |
Why the estimate is a range
Companies usually report environmental information for entire data-center operations, not a clean split between AI and non-AI work. A facility may run AI models alongside search, video streaming, cloud storage, enterprise software, databases, social-media services, and ordinary web hosting. Researchers therefore have to infer AI’s share from broader data-center figures, public company disclosures, and estimates of AI’s share of electricity demand.
Other inputs also vary. The result depends on the mix of training and inference, hardware utilization, server locations, local electricity generation, cooling systems, climate, and whether water used to generate electricity is counted. Limited and inconsistent disclosure makes it difficult to verify the inputs or compare providers on equal terms. The U.S. Government Accountability Office has likewise highlighted gaps in information about generative AI’s environmental effects.
A wide range is not automatically a flaw: it can make uncertainty visible rather than hide it behind a single precise-looking number. The risk comes when coverage reports only the high end or turns a modeled possibility into a measured fact.
What counts as AI’s carbon footprint?
At a minimum, operational electricity includes more than the processors doing calculations. It can include GPUs and other accelerators, CPUs, memory, storage, networking, cooling equipment, and the data center’s power infrastructure. The emissions associated with that electricity depend on where and when it is used. A kilowatt-hour supplied by a coal-heavy grid generally has a different carbon impact from one supplied by nuclear, hydropower, wind, solar, or a mix of sources.
Rank #2
- Find how much water you are using | Made in the USA
- Detect silent leaks
- Saves costly plumbing repairs
- flow gauge measuring bag
- Ways to Save water in and around your home
Renewable-energy contracts can change how companies account for electricity emissions, but they do not by themselves prove that a workload has no environmental impact. The accounting method, timing, location, and relationship between purchased power and actual grid supply all matter.
A full lifecycle footprint could also include making semiconductors and servers, mining and processing materials, constructing data centers, transporting equipment, and replacing or disposing of hardware. The de Vries-Gao estimate is primarily an operational energy and resource estimate based on data-center activity. It should not be described as a complete lifecycle total unless the accounting boundary is explicitly established.
“Water use” is not one simple number
Water accounting distinguishes withdrawal—water taken from a source—from consumption, the portion not promptly returned to the same usable water system, often because it evaporates. Data centers can use water directly, particularly for cooling. There can also be indirect water consumption at power plants generating the electricity the facility uses.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The study’s estimate includes modeled indirect effects as well as assumptions about data-center operations. It does not mean every liter was pumped into a server hall, was drinking-quality freshwater, or was permanently removed from the planet. Some facilities may use reclaimed water or other sources; their impacts differ from a facility drawing freshwater from a stressed local supply.
Water’s location matters as much as the global total. A liter consumed in a water-abundant basin is not equivalent, in practical impact, to a liter consumed where households, farms, and ecosystems already face shortages. A global annual estimate signals scale but cannot by itself show which communities or watersheds bear the pressure.
How far does the bottled-water comparison go?
The comparison puts two annual volumes side by side; it does not suggest that data centers literally use water in the same way bottled-water companies do. Estimates of global bottled-water consumption or production vary by year and by what is counted—sales, product volume, production, or the water footprint of the supply chain. Historical research has put worldwide consumption in the hundreds of billions of liters annually, but that is not a single fixed denominator. See the published work on bottled-water consumption and production and lifecycle impacts.
So the useful takeaway is scale, not equivalence: the study’s AI water range is of the same broad order as a large global consumer market’s annual product volume. The low estimate need not exceed that volume, and neither comparison automatically counts the same kinds of water use.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →AI is not the whole data-center story
Data centers support many digital services, and their total electricity use is not synonymous with AI use. The study attempts to estimate AI’s contribution within that larger system, but companies’ limited workload-level reporting prevents a definitive public split. For context, the GAO reported that U.S. data centers used about 4% of the country’s electricity in 2022 and could reach about 6% in 2026, citing International Energy Agency estimates. Those figures concern data centers overall, not AI alone.
Rank #4
- 🚰 BUILD A FLOATING WATER TAP → FOR AGES 8-12: Assemble a stunning water tap that creates a non-stop stream of water! It pumps water like a normal tap but appears to float in mid-air – a magical illusion that amazes friends and family.
- 💧 LEARN THE SCIENCE OF WATER CIRCULATION: A motor inside the module operates a small water pump, drawing water from the container and pushing it through a hidden transparent tube. The falling water hides the clear centre pipe, making it look like the tap is floating!
- 🚰COMPLETE ASSEMBLY KIT: Includes water container, motor module, water tap, transparent tube, T tube, template, sticker, terminal caps, screws, and detailed instructions. Requires 2 AAA batteries and a small crosshead screwdriver (not included)
- 🔬 STEM LEARNING & ENVIRONMENTAL AWARENESS: This Green Science kit is designed to raise awareness about water conservation . Use the included templates to create a rocket launch illusion or decorate for World Water Day
- 🎁 PERFECT GIFT FOR CURIOUS MINDS: Ideal for science fairs, classroom projects, or home experiments. Builds problem-solving skills and fine motor coordination. Adult supervision recommended for assembly . Hours of magical, educational fun!
Other research underscores why no single global estimate settles the question. A 2025 Nature Sustainability study modeled U.S. AI-server pathways and found that location, electricity mix, cooling, and water-management choices materially affect outcomes. Provider- or model-specific measurements, including Google-related workload measurement research, can offer more detail for a defined workload, but should not be generalized to every provider or data center. The resource demands of inference also vary with model, hardware, infrastructure, and request type.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why AI’s footprint could keep growing
The total depends on both the resource cost of each task and how many tasks are run. Larger models, longer context windows, image and video generation, audio processing, and reasoning-intensive workloads can require more computation. Agentic systems may make multiple model calls to complete one user-facing task. Training and repeated fine-tuning consume resources too, while inference—the repeated serving of models—can become significant at enormous scale.
Demand can rise even as each request becomes more efficient. Faster, cheaper inference may encourage wider use, more capable services, and larger deployments. That rebound effect means lower energy or water per request does not guarantee a lower total footprint.
What can reduce the impact?
- Use the right-sized model. Smaller or task-specific models may handle routine work with fewer resources than a large general model, though they may not suit complex tasks.
- Improve inference efficiency. Quantization, batching, caching, and better utilization can reduce redundant computation or energy per task.
- Choose infrastructure deliberately. Low-carbon electricity can reduce operational emissions; efficient cooling, reclaimed water, and siting outside water-stressed areas can reduce local water pressure. Trade-offs matter: air cooling can reduce on-site water use while raising electricity demand in hot conditions.
- Measure the whole deployment. Organizations should track training and inference separately, along with electricity, peak demand, cooling, water source, location, and hardware life cycle where possible.
- Require comparable disclosure. Public reporting should separate AI from non-AI workloads and provide facility-level water withdrawal and consumption, power demand, grid emissions factors, workload geography, embodied hardware impacts, and the method used to account for renewable electricity.
For an organization evaluating an AI footprint number, ask what it includes: training, inference, cooling, electricity generation, and hardware manufacturing; whether it isolates AI; which locations and time period it covers; whether emissions are location-based or market-based; whether “water” means withdrawal or consumption; how utilization is treated; and whether the uncertainty and any possible double counting are disclosed. Without those details, figures may not be meaningfully comparable.
What readers should conclude
The study supports a serious but qualified conclusion: AI’s 2025 environmental footprint could be large enough to compare with a major city’s annual emissions and with the volume of the global bottled-water market. It does not establish that AI definitively emitted New York City’s exact amount of carbon, exceeded all bottled-water demand, or used a fixed quantity of water for every prompt.
For individual users, a single interaction has no universal carbon or water cost; the answer varies with the model, hardware, workload, electricity supply, cooling, and accounting boundary. System-wide impacts are shaped more directly by infrastructure, deployment choices, demand, and company disclosure. Better efficiency helps, but transparent, comparable reporting is the necessary first step to determine where reductions are real and where impacts are being shifted.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
