An AI system can lower the electricity needed to cool a building and increase the electricity needed to run the data centre that optimises it. Both statements can be true. The climate result depends on the size, timing and carbon intensity of each effect—and on what people do with the money or capacity the efficiency gain releases.[1]
This last response is the rebound effect. When a service becomes cheaper or more productive, demand may expand. A model that uses half as much energy per answer does not cut total energy use if the number of answers triples. A routing system that reduces fuel per delivery can still increase total travel if lower costs make more deliveries profitable.
AI’s climate balance has three moving parts
| Channel | Potential climate benefit | Potential offset |
|---|---|---|
| Direct computing footprint | More efficient chips, models, cooling and workload scheduling | Rapid growth in training, inference, agents and generated media raises total demand. |
| Application in other sectors | Optimised grids, buildings, industrial processes and transport can avoid emissions | Adoption barriers, poor data and operational constraints limit realised savings. |
| Economy-wide rebound | Higher productivity can free resources for clean investment | Lower costs and higher output can increase energy, materials and consumption elsewhere. |
Two current reference points
| Measure | Value | Context |
|---|---|---|
| Data-centre emissions today | Around 150 million tonnes of CO2 | Latest IEA estimate for emissions associated with data-centre electricity use. |
| Central 2035 projection | Around 350 million tonnes | Latest IEA central scenario, approximately 1.5% of global energy-related emissions. |
| Water-stress exposure | Nearly 40% | Share of identified AI-dedicated data-centre capacity in high or extremely high baseline water-stress areas. |
| Older application potential | Reductions equivalent to about 5% of 2035 energy-related emissions | IEA estimate for broad adoption of existing AI applications, before barriers and rebound are considered. |
Efficiency per task is not the same as an emissions cap
Technical efficiency is improving quickly because computing is expensive. Models use lower precision, route only part of a network, cache repeated work and run on better accelerators. Data centres improve cooling and power utilisation. Those gains reduce the cost of supplying an AI task. They can also make new uses viable: continuous agents, real-time video generation, personalised simulations and larger populations of automated workers.
The climate effect therefore depends on an elasticity: how strongly use expands when the effective cost falls. If a tenfold efficiency improvement leads to a hundredfold increase in tasks, total computing demand rises. This is not an argument against efficiency. It is an argument for measuring the system boundary rather than celebrating one denominator.
The electricity at the margin matters
A data centre can buy annual renewable-energy certificates while still increasing demand during hours when the local grid relies on gas or coal. New generation, transmission queues, backup equipment and the timing of flexible workloads shape the real emissions effect. The same computing load can have a different footprint in two regions or two hours.[1][2]
Water and local affordability add another layer. The latest IEA review covers policies in 34 jurisdictions representing 90% of global data-centre capacity. It finds governments trying to reconcile rapid investment with grid security, prices, emissions and water use. A global percentage can hide a severe local constraint where one cluster competes with households and industry for power or water.[2]
AI can still be a powerful climate tool
The IEA identifies substantial potential in energy systems and other sectors, and the BIS paper describes benefits for forecasting, adaptation, resource optimisation and climate innovation. The problem is not that these applications are imaginary. It is that technical potential must survive deployment cost, institutional adoption, behavioural response and the expansion of activity that success itself can trigger.[1][3]
AI’s net climate impact will not be decided by whether the technology is labelled efficient or wasteful. It will be decided by whether useful applications displace more emissions than the computing and additional activity they induce—and whether policy treats rebound as a measurable design variable rather than an inconvenient surprise.
Sources & further reading
- Artificial intelligence and climate change: balancing innovation and sustainability Bank for International Settlements · 8 Oct 2026 · Accessed 16 Oct 2026
- Powering Artificial Intelligence: Executive summary International Energy Agency · 14 Oct 2026 · Accessed 16 Oct 2026
- Energy and AI: Executive summary International Energy Agency · Accessed 16 Oct 2026