Environment 11 min read

AI’s Energy Ledger

Abstract navy and teal waves threaded with cream and amber ribbons.

The cost of an AI answer runs through power plants, water systems, chip factories, and balance sheets. The benefits need to be measured against that entire chain.

TL;DR

Data centre electricity demand is rising, but data centre totals are not AI totals. The impact of a new facility depends on its local grid, cooling system, water source, and construction costs. Claims that AI offsets its footprint need measured outcomes and a credible baseline.

An AI answer appears on a screen in seconds. Behind it sits a chain of physical decisions: chips designed and manufactured, servers assembled, a site connected to a grid, heat removed, water drawn or avoided, and money committed years before a customer sends a prompt. The visible response is only the last step.

The public debate often compresses this chain into a single claim about energy per prompt. That number can be useful, but it cannot describe the load from a growing service, the cost of building capacity, or the effect of one large campus on a local water system. The right unit changes with the decision. A user may care about a task. A utility must plan for megawatts. A community may care about water during a dry season. A business analyst must ask whether the service earns a return after electricity, hardware, and infrastructure are paid for.

The scale, with the right denominator

The International Energy Agency estimated that all data centres consumed about 415 terawatt hours in 2024, roughly 1.5% of global electricity use. Its 2025 base case put the 2030 total near 945 terawatt hours. A 2026 update estimated about 485 terawatt hours in 2025 and projected about 950 in 2030. These are totals for data centres, not measurements of AI alone. The comparison shows the pace of infrastructure growth and the limits of a popular shortcut that assigns every server’s electricity to AI.

Global data centre electricity use

2024 estimate415 TWh
2030 base case945 TWh
IEA Energy and AI (2025). The 2030 value is a projection. Both values cover all data centres, not AI alone. The IEA report is linked in References.

AI is nevertheless a major driver of new demand. The IEA’s 2026 update estimated that electricity use from AI focused data centres rose about 50% in 2025, while data centre electricity overall rose 17%. At the same time, energy per simple AI task has fallen rapidly as software and chips have improved. These facts can both be true: a unit becomes cheaper while the total grows because more people use it, more tasks are automated, and some new tasks consume far more power than short text queries.

Video generation, lengthy reasoning, and agents that run many tool calls should not be priced or assessed as if they were one short response. Training demand and inference demand also have different timing. Training is concentrated; a popular deployed service repeats inference millions of times. Any meaningful comparison needs the task type, model, output length, hardware, utilisation, and date of measurement.

A global percentage can hide a local bottleneck

One and a half percent of global electricity can sound modest. A single computing campus is connected to one regional grid, at one point in time. It may require new transmission, transformers, substations, and generation. The resulting cost can fall on the operator, a utility, customers, or a mix of all three, depending on local rules. A national average does not show whether the nearby grid has spare capacity on a hot evening.

The IEA reports that data centre developers are encountering shortages of power equipment and slow grid connections. It also finds that some US projects are considering on site gas generation. That may help a project reach an operating date, but it changes the emissions and fuel risk. On site power is not automatically cheaper or cleaner. The IEA estimates that reliable gas based supply for variable data centre load may require 30% to 70% overbuilding of generation relative to demand.

Electricity contracts need the same care. Annual renewable matching means a company has procured enough renewable electricity over a year to match its annual consumption in accounting terms. It does not mean every hour of operation uses carbon free power on the local grid. The difference matters when a cluster runs through nights, cloudy periods, or grid stress. For a community assessing a proposed site, hourly load and local generation are more useful than a global annual percentage.

Heat, water, and the geography of harm

Nearly all electricity consumed by computing equipment becomes heat. Cooling moves that heat out of the servers. Evaporative systems may consume water directly at the site. Electricity generation can consume water indirectly. Semiconductor fabrication uses water and energy before a server is ever installed. These are different parts of the lifecycle; combining them into one unsourced “water per prompt” figure creates false precision.

Withdrawal means water taken from a source. Consumption is the part not returned to that source, often because it evaporates. The distinction is important for local planning. A facility with a small direct consumption number may still depend on a water intensive power supply. A design that uses less water at the site may require more electricity for cooling. Reclaimed water can reduce competition for drinking water, but its availability and treatment costs vary by location.

The global crisis perspective is about distribution. Fresh water is not interchangeable across watersheds, seasons, and communities. A litre consumed where water is plentiful has a different consequence from a litre consumed during a drought or near an overdrawn aquifer. The 2024 Lawrence Berkeley National Laboratory report models large variation in US data centre water demand under different cooling and expansion scenarios. The useful questions for a proposed site are therefore local: source of water, seasonal availability, cooling design, discharge, backup operations, and the alternatives considered.

For a resident, the issue is tangible. Additional demand can meet households, farms, and industry in the same supply system. For a business analyst, the same constraint appears as a permitting delay, a higher tariff, an operating restriction, or a reputational cost. A site with cheaper land but uncertain water and grid access may be more expensive over its lifetime than a site with higher initial rent and reliable infrastructure.

Google as a case study in measurement boundaries

Google’s 2026 environmental report, covering 2025, provides an unusually useful view because it publishes demand, emissions, and water accounting together. The company reported that its electricity demand grew 37% year over year. It also reported a 2% fall in operational emissions, while supply chain emissions grew 25%. These statements can coexist. Cleaner electricity and efficiency can improve the emissions from operating a facility while the construction and manufacture needed for expansion increase emissions elsewhere in the value chain. None of these company wide figures measures AI alone; Google’s infrastructure serves many products.

Google says it matched 100% of its annual electricity consumption with renewable energy purchases in 2025. Its reported global average hourly carbon free energy for data centres was 65%. The gap is the point. Annual purchasing and hourly supply on the grids where data centres operate answer different questions. A reader should not treat a 100% annual match as proof that every prompt ran on carbon free electricity. The hourly figure is itself a global average: Google’s report shows much lower matches in some grid regions and much higher ones in others.

Google researchers also measured the serving stack for a median Gemini Apps text prompt and estimated 0.24 watt hours of energy and 0.26 millilitres of water. Their method included active accelerators, host systems, idle capacity, and facility overhead. The result is valuable because the boundary is stated. It is not a universal figure for every model, video request, training run, or hardware lifecycle. A median can also hide the long tail of more expensive requests.

The detailed 2026 report lists 10.87 billion gallons of operational water consumption in 2025, including 10.52 billion gallons at data centres. Separately, Google says its projects replenished about 7.7 billion gallons, equal to 78% of the freshwater consumption used in its replenishment calculation. The two denominators differ because the replenishment metric concerns freshwater, while operational consumption includes other water sources. Replenishment may help watersheds, but an annual global ratio does not erase a local shortage at a particular facility in a dry season. These disclosures are more useful together than as one reassuring headline.

The financial ledger behind the energy ledger

The IEA reported that capital spending by five large technology companies exceeded US$400 billion in 2025 and projected another 75% rise in 2026. That spending covers more than electricity. It funds land, buildings, accelerators, networking, grid connections, cooling, and long lived power agreements. Some assets may depreciate rapidly as chips improve; transmission and water infrastructure can take years to permit and build. The business risk is a mismatch between fast changing AI demand and slow physical construction.

A simple calculation makes the exposure visible. Consider a hypothetical 100 megawatt IT load running continuously with power usage effectiveness of 1.2. The whole facility would draw about 120 megawatts, or 1.05 terawatt hours in a year. At an illustrative electricity price of US$0.10 per kilowatt hour, the annual electricity bill would be about US$105 million. This excludes chips, buildings, financing, network charges, cooling water, and staff. It is an example, not a claim about any named company or site.

The investment case then depends on utilisation and revenue per unit of compute. A half empty campus still carries financing and some fixed operating costs. High utilisation improves asset productivity but can increase local load and cooling demand. Efficiency improvements lower cost per task, yet may also stimulate more demand. A responsible forecast therefore tests several scenarios: slower adoption, higher power prices, hardware replacement, grid delay, and a shift from simple text toward video or autonomous agents.

The cost cannot be evaluated solely on a company balance sheet. Utilities may fund grid upgrades, municipalities may offer tax incentives, and communities may bear water or land use effects. A transparent project assessment should disclose the likely allocation of costs and benefits. If public infrastructure is required, projected local jobs and tax revenue should be compared with the full lifetime obligations, not only with the ribbon cutting announcement.

When would AI repay its footprint?

AI can forecast renewable output, detect faults on power networks, improve industrial controls, and optimise heating and cooling in buildings. These are credible routes to savings. The IEA estimated that broad use of existing AI applications could reduce energy related emissions by about 1.4 billion tonnes of carbon dioxide in 2035 in its widespread adoption case. The agency explicitly warned that adoption is not on a path that guarantees those savings. Data access, skills, infrastructure, incentives, and security all affect whether a model changes real operations.

Google estimated that nine of its products enabled about 41 million tonnes of carbon dioxide equivalent reductions in 2025 through uses including thermostats, solar planning, and fuel efficient routing. This is a modelled estimate of changes enabled outside Google’s own operations, based on product specific methods. The same report estimates that hardware, software, and clean power initiatives avoided more than 58 million tonnes within its operations and supply chain compared with its chosen counterfactuals. These are different claims with different baselines. Neither is a directly measured net cancellation of the company’s reported emissions.

A claimed benefit needs a counterfactual. If a building uses less energy after an AI controller is installed, compare it with weather adjusted use under a conventional controller, not with an unusually wasteful baseline. If an AI system identifies a methane leak, count the reduction after the leak is repaired, not after an alert is generated. If a factory runs more efficiently, include the energy to run the model and check whether output growth offsets the saving.

Business analysis has to distinguish a technical possibility from realised cash flow and realised emissions reduction. The first may justify a pilot. The latter requires metering, a baseline, uncertainty, and a decision owner capable of acting on the recommendation. Some applications may produce benefits in one region while creating power or water burdens in another. Net impact needs a stated geographic and lifecycle boundary.

A decision ready scorecard

An engineering team can make this debate much clearer by publishing a few consistent fields: electricity by workload and site, hourly power source, direct and indirect water consumption, hardware lifecycle assumptions, task volume, and the measured outcome of any claimed saving. Projections should be labelled as projections, and data centre totals should never be silently relabelled as AI totals.

For an AI product, the functional unit might be one completed customer task rather than one prompt. For a data centre, the useful unit may be annual megawatt hours at a named site. For a city, it may be peak demand and water use during a constrained season. Each is valid for a different decision. The mistake is expecting one number to settle every question.

AI can help solve expensive physical problems. It also creates new physical demand. The credible position is neither automatic optimism nor automatic rejection. It is a ledger with the boundary drawn clearly enough that a customer, utility, investor, and community can see the same costs and test the same promised benefits.

Acknowledgements

The IEA, Lawrence Berkeley National Laboratory, and Google published the data and methods examined here. The 100 megawatt cost example is illustrative. This article reports no original site measurement.

References

  1. IEA, Energy and AI, global demand outlook (2025)
  2. IEA, Key Questions on Energy and AI, 2026 update
  3. IEA, AI and climate change scenarios (2025)
  4. Google, 2026 Environmental Report and detailed metrics (2025 data)
  5. Google, explanation of 2025 energy, emissions, water, and enabled savings
  6. Google, measuring the impact of Gemini Apps serving (2025)
  7. Lawrence Berkeley National Laboratory, US Data Center Energy Usage Report (2024)