Every AI query draws on something physical. Electricity to run the chips, water to cool the buildings that house them, metals and minerals to make the hardware. The cost is real, and the numbers are large enough to take seriously.
They are also smaller, in global terms, than the alarm suggests. Data centres of all kinds, not only those running AI, accounted for around 1.5% of the world's electricity consumption in 2024, or 415 terawatt-hours, according to the International Energy Agency. The honest position sits between panic and dismissal.
This article sets out what AI actually costs the environment, how those costs compare to the benefits, and what a small business should do about it. We work with AI every day, and our view is that the right response is care, not avoidance.
What "AI's environmental impact" actually means
The environmental impact of AI is the energy, water, and raw materials consumed across the life of an AI system: building the hardware, training the model, and answering each query after it ships. Two of those costs dominate the public debate, electricity and water, because data centres need a great deal of both.
It helps to separate two phases. Training is the one-off process of building a model, which is energy-intensive and concentrated. Inference is what happens every time you use it, a small cost per query multiplied by billions of queries. Training makes the headlines. Inference is where the steady, cumulative demand lives, and where the International Energy Agency sees most of the growth coming from.
How much energy does AI use?
Data centres, the warehouses of servers that run AI and most of the modern internet, are the unit to measure. The International Energy Agency puts their 2024 demand at 415 TWh, about 1.5% of world electricity, and projects it will "more than double to around 945 TWh by 2030," a figure the agency describes as "slightly more than Japan's total electricity consumption today." AI is the main driver of that rise.
The increase is not spread evenly. The United States "accounts for by far the largest share of this projected increase, followed by China," and in the US alone data centres are set to make up nearly half of all electricity demand growth this decade. That concentration is the real strain. A grid adds load fastest where the data centres cluster, which is why AI's energy use shows up as a local infrastructure problem before a global one.
Keep the global share in view, though. Even on the agency's own projection, data centres reach roughly 3% of world electricity by 2030. That is a meaningful slice and worth managing well. It is not the dominant cause of the world's emissions.
The picture at the level of a single use is smaller still, and moving. Per-query energy has fallen sharply as models and chips have become more efficient, even as total demand climbs because usage is growing faster. For a small business sending a few hundred prompts a week, the direct energy footprint of using AI is modest. The figures that matter are the aggregate ones, set by how the whole industry is built and powered.
How much water does AI use?
Water is the cost most people miss, because it is invisible at the keyboard. Data centres run hot, and many are cooled by evaporating clean freshwater, the same grade we drink. That water leaves the local supply as vapour.
The clearest measurement comes from a peer-reviewed study by Li, Yang, Islam and Ren, later published in Communications of the ACM. They calculated that "training the GPT-3 language model in Microsoft's state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater". That is one training run of one model, and it covers only on-site cooling, not the water used to generate the electricity in the first place.
Scaled up, the projection is large. The same researchers estimate "the global AI demand is projected to account for 4.2-6.6 billion cubic meters of water withdrawal in 2027", which they note is "more than the total annual water withdrawal of 4-6 Denmark or half of the United Kingdom." Data centre water usage matters most where it competes with farming and households for a stressed supply, which is exactly where some of the largest facilities sit.
The fix here is partly engineering. Air cooling, recycled or non-potable water, cooler climates, and timing heavy work for cooler hours all cut the freshwater drawn. The fix is partly siting, building where water and clean power are abundant rather than scarce. Neither is solved yet, which is why water belongs in any honest account of AI's footprint.
Is AI bad for the environment? The case on the other side
AI carries a real carbon footprint. It can also cut one, and the size of the possible cut is large enough to change the verdict.
A 2025 study from the Grantham Research Institute at the London School of Economics and Systemiq, published in Nature's npj Climate Action, found AI could reduce global emissions by "3.2 to 5.4 billion tonnes of carbon-dioxide-equivalent annually by 2035." The savings come from three sectors, power, transport, and food, which together cause roughly half of all greenhouse gas emissions. AI helps by making grids, logistics, and farming more efficient, and by speeding up the science behind cleaner materials.
The crucial line is what those savings are measured against. The authors conclude the reductions "would outweigh increases from global power consumption of data centres and AI," and that this holds across "all of AI's activities, not just those related to decarbonisation." On that analysis, the net effect of AI on emissions is positive, provided the technology is pointed at problems worth solving.
This is the part the loudest takes leave out. So "is AI bad for the environment?" has no clean yes or no. AI used carelessly, on trivial tasks, powered by coal, is a cost with little return. The same technology used to decarbonise a power grid can pay for its own footprint many times over. The verdict depends on use, not on the technology in the abstract.
The Australian view
Australia shows the global pattern at national scale, and with more room to get it right. The Climate Council reports that "in 2024-25, data centres used around four terawatt-hours (TWh), or 2%, of the electricity in Australia's main grid". The market operator AEMO "expects data centre energy demand in the NEM to triple to nearly 12 TWh by 2030," equal to about 6% of the grid. The growth is steep, the base is still small, and the timing lines up with Australia's shift to renewables.
Water is the more reassuring number here. The industry "currently uses less than 0.1% of Australia's total water," though that demand is projected to "more than triple from 5.5 GL to 17 GL over the next five years," again per the Climate Council. On a dry continent, where new data centres land relative to water and clean power will decide whether that growth is a problem or a non-event.
For an Australian small business, the practical reading is simple. Your own AI use sits far below these grid-level figures. The lever you actually hold is choosing efficient tools and using them where they earn their keep. The larger questions of siting and clean supply belong to operators, regulators, and the grid, and they are being decided now.
The Enki Approach
We treat AI as a tool with a real cost, to be used where it returns more than it consumes. In practice that means matching the model to the task, a small efficient model for routine work rather than the largest one out of habit, and being honest with clients when AI is not the right answer at all. The aim is a build that creates genuine value, not AI for its own sake. That connects to a wider point we make in The Misuse of AI: waste is its own cost, financial and environmental.
Responsible use is part of how we work, and it sits alongside how we give. Enki Digital commits 10% of profits to effective charities, including environmental and AI safety causes, so the upside of the work we do is shared beyond our clients. AI is going to be built and used at scale either way. The question worth answering, for a small business and for the industry, is whether it is used well. Using it where it creates value, choosing efficient implementations, and pushing for cleaner infrastructure is how the benefits the LSE study describes actually arrive.