The public conversation about artificial intelligence has focused overwhelmingly on capability — what models can do, what they might replace, what risks they pose. Far less attention has been paid to a more mundane and more immediate constraint: the physical resources required to run them. Generative AI is not just software. It is an industrial activity with a real energy and water footprint, and that footprint is growing fast enough to strain electricity grids and water systems in the regions that host it.
The Scale of the Problem
Data centres have always consumed significant electricity, but the AI boom has changed the profile. Training a single large language model can consume thousands of megawatt-hours of electricity. Inference — running the model for users — is often the larger cumulative cost, because it happens billions of times.
The International Energy Agency estimated in its 2024 report Electricity 2024 that data centres, including AI, cryptocurrency, and traditional IT, consumed roughly 415 terawatt-hours of electricity in 2022, about 1.5% of global demand. The IEA projected that figure could more than double to around 945 TWh by 2026 in its base case — and its more aggressive scenarios see it approaching 1,000 TWh by 2030, comparable to the total electricity consumption of Japan.
In the United States, data centre demand is growing faster than the grid can accommodate. Several utility forecasts have been revised sharply upward. Dominion Energy in Virginia, home to the world’s densest concentration of data centres in “Data Center Alley,” has reported interconnection requests far exceeding its prior planning assumptions. In Ireland, data centres already consume more than 20% of the nation’s electricity, prompting restrictions on new connections.
Water: The Overlooked Cost
Electricity is only half the story. Data centres use water for cooling, both directly (evaporative cooling towers) and indirectly (the water consumed to generate the electricity they draw). A widely cited estimate from researchers at the University of California, Riverside, suggested that generating a single 100-word email with a large language model could consume roughly 500 millilitres of water once training, inference, and upstream electricity generation were accounted for.
Google’s environmental reports have disclosed total water consumption in the billions of litres annually, with data centres accounting for the majority. Microsoft’s sustainability disclosures revealed a sharp rise in water withdrawal as its AI infrastructure expanded. In water-stressed regions — Arizona, parts of Texas, Spain, Chile — the siting of large AI data centres has drawn local opposition.
The industry’s response has been to point to efficiency improvements and to the adoption of closed-loop liquid cooling, which recycles water rather than evaporating it. But efficiency gains have historically been outpaced by growth in demand — a classic rebound effect.
Carbon Emissions
Tech companies have made high-profile climate commitments. Google, Microsoft, and Meta all pledged to reach net-zero emissions, and Google committed to operating on 24/7 carbon-free energy by 2030. Yet their own sustainability reports show emissions rising, not falling. Google reported that its 2023 greenhouse gas emissions were roughly 48% higher than its 2019 baseline, attributing the increase primarily to data centre energy consumption. Microsoft reported a similar trajectory, with emissions roughly 29% above its 2020 baseline.
The reason is straightforward: AI demand is growing faster than clean energy can be added to the grid. When a data centre draws power from a grid that is still partly fossil-fuelled, its operational emissions rise even if it purchases renewable energy certificates. And the construction of that data centre — concrete, steel, and embodied carbon — adds further emissions.
The Grid Bottleneck
The constraint is increasingly not chips but power. Utilities and grid operators are being asked to supply gigawatts of new load on timelines that their planning cycles were never designed for. Interconnection queues for new generation stretch to five years or more. Gas turbine manufacturers are sold out for years. Nuclear small modular reactors, though promising, will not be deployed at scale until the 2030s.
Some operators are turning to unconventional sources. Microsoft signed a deal to restart Three Mile Island’s Unit 1 reactor. Google and Amazon have invested in small modular reactor companies. Others are experimenting with on-site fuel cells and geothermal.
Can Efficiency Save Us?
There is genuine progress in efficiency. Model compression, quantisation, speculative decoding, and better hardware all reduce the energy per inference. Nvidia’s newer accelerators deliver far more performance per watt than their predecessors. But historical patterns suggest that efficiency gains in computing tend to be consumed by increased usage rather than reducing total demand — a phenomenon economists call the Jevons paradox.
The more hopeful framing is that AI itself could accelerate the energy transition by optimising grids, materials, and processes. But that is a promise, not yet a demonstrated net benefit.
What Should Happen
Transparency is the starting point. Companies should disclose energy and water consumption per workload, not just aggregate figures. Regulators should require that large data centres account for their resource use, and grid planners should be given the tools to manage the surge. Investment in clean firm power — geothermal, advanced nuclear, long-duration storage — is essential, not optional.
For readers in Canada, the issue is especially live. The country’s relatively clean electricity mix is an asset, but it is also finite, and the same AI facilities that could be sited here to reduce global emissions will compete for power with homes, industry, and electrification goals.
The Chips and Their Embodied Carbon
The energy story of AI begins long before a model runs. Manufacturing advanced semiconductors is extraordinarily resource-intensive. A leading-edge fab consumes enormous amounts of electricity and ultra-pure water and uses potent greenhouse gases — notably fluorinated compounds such as nitrogen trifluoride and sulfur hexafluoride — that trap heat thousands of times more effectively than CO₂. Producing a single high-end GPU carries an embodied carbon footprint of hundreds of kilograms of CO₂ equivalent, and producing the thousands of GPUs in a large training cluster multiplies that into thousands of tonnes before the first token is generated.
Siting and the Water Battleground
Where data centres are built increasingly determines their environmental impact. In cool climates with clean grids — parts of Canada, Scandinavia, and the Pacific Northwest — the operational footprint is far lower than in hot, fossil-heavy regions. This has turned siting into a contested decision. Communities near proposed AI campuses worry about electricity rates rising, water tables falling, and noise from cooling systems. Several jurisdictions have imposed moratoria or required new data centres to meet efficiency standards. The industry’s response — building in Ireland, Virginia, and Arizona because of fibre, tax incentives, and proximity to customers — frequently conflicts with the goal of minimising environmental impact.
Measuring What Isn’t Measured
One of the most serious problems is the absence of standard, comparable disclosure. Companies report energy and water use in inconsistent ways: some count only direct consumption, others include purchased offsets; some report power usage effectiveness (PUE) for a campus average, hiding the load of specific AI clusters; almost none report emissions or water per query or per model. Without standardised metrics, regulators cannot set meaningful rules, investors cannot assess risk, and the public cannot hold the industry accountable. Efforts to create disclosure frameworks — analogous to financial reporting — are nascent but essential.
The Efficiency Argument, Examined
Industry optimists point to a long track record of efficiency gains. Data centres today deliver vastly more computation per kilowatt-hour than a decade ago, and modern accelerators are dramatically more efficient per operation than their predecessors. There is real truth here. But efficiency per operation and total consumption are different things. When a service becomes cheaper, demand grows — sometimes explosively. AI inference is on a trajectory to become one of the largest new loads on the world’s grids precisely because it is becoming useful and affordable. The Jevons paradox is not a hypothetical here; it is the observed pattern.
A Canadian Reckoning
Canada’s grid is unusually clean — roughly 80% non-emitting — which makes it an attractive location for AI infrastructure and a potential source of low-carbon compute for the world. But that advantage is finite. The same electricity that could power data centres is needed for electric vehicles, heat pumps, industrial electrification, and existing demand. Every large AI campus added to a province is a gigawatt-scale decision that competes with other climate priorities. The country faces a genuine trade-off: attract the industry and its jobs and investment, or preserve the clean-power surplus for other uses. Pretending there is no trade-off serves no one.
Offsets, Accounting, and the Honest Ledger
A final complication is accounting. When a company reports that it is “carbon neutral,” it often relies on renewable energy certificates, power purchase agreements, or carbon offsets rather than eliminating emissions. The methods of accounting vary widely, and the numbers can be gamed. A data centre might claim 100% renewable electricity while drawing from a grid where the marginal generation at its operating hours is gas. It might purchase offsets of uncertain quality. Standardised, rigorous accounting — where emissions are measured at the point of consumption, on an hourly basis, matched to actual generation — is technically feasible but not yet customary. Until the industry accounting matures, public claims about AI’s climate footprint should be read with appropriate scepticism.
Conclusion
Generative AI is a remarkable technology, and it may deliver enormous value. But it is not weightless. Every query is served by machines that draw power from a grid, cooled by water, built from materials, and maintained by infrastructure with a measurable carbon cost. The industry’s habit of treating those costs as externalities is becoming untenable. The next phase of AI’s growth will be shaped as much by energy and water as by algorithms.
