AI Energy Use Is a Real Problem. But It Is Not the Whole Question
Spend a few minutes in almost any discussion about artificial intelligence, and the same concern appears quickly: energy. Data centres already consume substantial amounts of electricity, and AI is accelerating their growth. The International Energy Agency’s April 2026 update projects that global data-centre electricity consumption will rise from 485 terawatt-hours in 2025 to about 950 TWh in 2030, reaching approximately 3% of global electricity demand. Consumption by AI-focused data centres is expected to triple over the same period.
Those figures cover data centres as a whole, including cloud services, storage, conventional computing and other digital activity. They are not a direct measurement of AI alone. AI is the largest source of current growth, but it is not the entire load. The distinction matters because otherwise the debate begins by assigning every server, cooling system and stored file to artificial intelligence.
The physical constraint is nevertheless real. Training large models requires substantial computation, but deployment matters too. Inference—the repeated use of models after training—expands as AI enters search, office software, customer service, coding tools, media production, research and industrial systems. AI is becoming infrastructure, and infrastructure cannot be evaluated only by asking whether an individual use feels impressive or frivolous.
The serious question is not merely how much electricity AI consumes. It is where that demand appears, what physically supplies it, who pays for the surrounding infrastructure and what the resulting computation actually enables.
The Number Is Real, but the Percentage Misleads
The rate of growth deserves attention. According to the IEA’s 2026 update, electricity consumption by data centres increased by 17% in 2025, while consumption by AI-focused facilities rose by approximately 50%. Efficiency per task improved rapidly, but adoption and the growth of more energy-intensive applications outpaced those gains.
Global context still matters. The IEA estimated that data centres consumed about 415 TWh in 2024, or roughly 1.5% of worldwide electricity. Their electricity consumption was associated with approximately 180 million tonnes of indirect CO₂ emissions, about 0.5% of global fuel-combustion emissions. These figures include all data-centre workloads, of which AI is only a subset.
That is neither negligible nor evidence that data centres currently dominate global energy use. The pressure comes from their growth rate, their concentration and the speed with which projects can appear compared with the slower development of power stations, substations, transmission lines and manufacturing supply chains.
A global share can therefore reassure and mislead at the same time. Three per cent of worldwide electricity demand may sound manageable in aggregate, while a cluster of facilities can transform the planning problem faced by a particular utility or region.
A Data Centre Has an Address
A data centre does not connect to an abstract global grid. It connects to a particular substation, transmission system and generation mix. Its electricity demand arrives in a place with existing consumers, bottlenecks, land constraints, water pressures, permitting disputes and infrastructure plans that may already be years behind schedule.
The IEA’s analysis of the geographical concentration of data centres illustrates the difference between global and local significance. Data centres consume about 20% of metered electricity in Ireland. Six US states already devote more than 10% of their electricity supply to them, with Virginia at approximately 25%.
A new facility may require a larger substation, additional transmission capacity, new generation, backup systems or grid capacity reserved years before the site reaches full operation. The cost may be paid by the operator, spread across other electricity customers or supported through public investment. Those arrangements determine whether the surrounding community receives an investment or a bill.
This is why the issue cannot be reduced to watt-hours per model response. Electricity is a system of generation, storage, transmission, balancing and demand. As the Journal’s essay on why the energy transition is a system-design problem argues, producing enough electricity over a year does not guarantee that it is available in the right place or at the right hour.
Training, Inference and the Rebound Problem
Public discussion often concentrates on training because a major training run is visible, dramatic and relatively easy to describe. Deployment is more diffuse. Once a model is integrated into millions of searches, office interactions, generated images, coding sessions, automated agents or industrial decisions, the accumulated electricity used for inference can become substantial.
There is no universal rule that training or inference must dominate every workload. The balance depends on model size, hardware, optimisation, frequency of use and how long the system remains in service. A model trained once and used rarely has a different profile from a smaller model invoked billions of times.
The larger problem is opacity. Companies rarely publish sufficiently comparable information to separate training, routine inference, storage, cooling and non-AI cloud services. This allows reassurance and alarm to coexist without confronting one another. A company can point to lower electricity consumption per query while a critic points to rapidly increasing total data-centre demand. Both claims can be true.
Efficiency measures the resources required for one unit of computation. Total consumption depends on how many units the system performs and which activities acquire an AI layer that previously had none. Cheaper computation may reduce the cost of an existing task, but it can also encourage new uses, larger models and more frequent calls. Efficiency remains essential; it is simply not self-executing.
What Physically Powers the Load
A facility supported by newly built low-emissions generation presents a different environmental problem from one increasing demand on a constrained, fossil-heavy grid. The phrase “powered by renewable energy” can nevertheless describe several arrangements. A company may finance a new wind or solar project, buy certificates associated with generation elsewhere or match its annual demand with renewable output produced at different hours.
The distinction between the contractual and physical electricity mix is important. The IEA’s analysis of energy supply for data centres considers the generation actually feeding facilities through their local grids and onsite systems, rather than relying on corporate procurement claims. On that basis, renewables are expected to provide nearly half of the additional electricity required through 2030. Natural gas and coal together are expected to supply more than 40%.
Clean investment and fossil growth can therefore happen at the same time. New renewable projects may be built while existing gas and coal plants operate more frequently or new fossil capacity is added to satisfy demand that arrives faster than low-emissions infrastructure can be permitted and connected.
The relevant tests are additionality, timing and location. Was new capacity built because of the facility? Does it generate when the facility consumes electricity? Can the network carry the power to the site? Does the project relieve a local constraint or deepen it? Similar questions apply to cooling water, land use and backup generation. A clean annual contract does not by itself answer them.
A Load, but Also a Lever
AI is not merely another electricity-consuming activity. It can also be applied to the systems whose energy use and emissions we are trying to reduce. The IEA identifies existing applications in weather forecasting, transmission monitoring, predictive maintenance, methane detection, industrial process control, demand response and the search for new energy materials.
In its Widespread Adoption scenario, the IEA estimates that AI-assisted operation could unlock up to 175 gigawatts of additional capacity on existing transmission lines. It also models energy savings in sectors such as light industry. These are estimates of what wider adoption of known applications could achieve if barriers involving data, skills, regulation and digital infrastructure were overcome. They are not automatic credits against the electricity consumed by data centres.
A model used to improve grid forecasting and a model used to generate disposable advertising copy may rely on related infrastructure, but they do not create the same value. Nor can a company justify every workload by pointing to the most socially valuable application someone else may eventually develop.
The possibility of scientific leverage still matters. AI may help search larger design spaces, identify patterns humans would miss or accelerate parts of experimental work, connecting this debate to the Journal’s broader examination of whether AI offers science more paths through an existing map or helps draw a new one. Useful outcomes can justify substantial costs. They do not make those costs disappear.
Govern Infrastructure, Not Individual Prompts
The temptation to judge AI energy use one prompt at a time leads quickly to absurdity. A medical-research query appears worthy; a generated joke appears frivolous. Yet governments and utilities cannot realistically assign an energy entitlement to every model interaction, and the social value of many uses will remain contested or impossible to measure in advance.
Personal restraint is legitimate. People can decline uses that seem pointless, and aggregate behaviour can influence product design and demand. It is still the wrong scale at which to govern infrastructure. Data centres are shaped more decisively by siting rules, connection agreements, electricity prices, procurement systems, utility regulation and capital markets than by one person deciding not to generate an image.
The practical policy questions concern effects that can be observed:
- how much electricity a facility requires and how rapidly that requirement may grow;
- where and when the electricity is consumed;
- which generation sources physically supply it;
- whether new generation and grid capacity are added;
- how infrastructure costs are divided between operators and other customers;
- whether workloads, backup systems and storage can respond during grid stress;
- how much water and land the facility requires;
- what emissions and local environmental effects result;
- whether comparable public reporting allows those claims to be tested.
Flexibility deserves particular attention, but it should not be romanticised. The IEA’s 2026 update notes that AI server loads can swing by more than 50% of rated capacity within a second. Storage is becoming necessary simply to manage those fluctuations reliably, and the agency estimates that 20–25 GW of battery capacity could be installed in data centres by 2030.
Those batteries and spare computing capacity could make facilities useful to the grid if incentives and operating agreements are designed accordingly. Some workloads might move across hours or locations, allowing greater use of abundant low-emissions electricity or reduced demand during emergencies. Other tasks require immediate processing, and expensive AI equipment gives operators a strong reason to keep utilisation high. Flexibility must be contracted and engineered; it will not emerge because a company describes itself as grid-friendly.
Public procurement and research funding can favour applications with demonstrable public value, particularly in health, science and energy. Infrastructure rules should nevertheless remain capable of governing a facility even when its promised breakthrough never arrives.
Make the Footprint Visible
The debate around AI and energy is often trapped between two incomplete positions. AI is not environmentally harmless because it may enable useful discoveries. It is not environmentally indefensible merely because it consumes electricity. Electricity use is a cost to be measured and managed, not a moral verdict that settles the value of the technology in advance.
Better hardware, cooling, software and model design remain essential. Without efficiency gains, the same level of use would require more electricity and more infrastructure. But efficiency per task cannot substitute for reporting total facility consumption, peak demand, emissions, water use and grid requirements.
Transparency is therefore not a secondary concern. Without comparable facility-level information, companies can highlight selected efficiency improvements and contractual clean-energy purchases while critics build arguments from estimates and worst-case projections. Neither side can be tested properly when the underlying physical system remains hidden.
The footprint is real, and so is the possibility of leverage. Those facts do not cancel one another; they create a governance problem. The debate becomes useful only when “AI energy use” is translated into a facility, a grid connection, a generation mix, a cooling system and a bill—and when it becomes clear who receives the capability and who is expected to carry its cost.
Comments
Post a Comment