AI is now part of how software is built and used, and it will not be legislated away. Open models are widely available, the tools improve faster than laws can be drafted, and governments see AI as a matter of competitiveness and security. The useful question for policy is how the energy it uses, and the energy of the software around it, can be kept in proportion to the value it delivers.
The scale is real. The International Energy Agency estimates that data centres used around 415 TWh in 2024, about 1.5% of the world's electricity, and expects that to more than double to around 945 TWh by 2030, with AI the most important driver. Locally the share can be far higher: in Ireland, data centres accounted for 23% of metered electricity consumption in 2025, according to the Central Statistics Office.
The IEA also finds that wider use of existing AI applications could cut emissions, but warns that AI "is not a silver bullet" and that the net effect will depend on how AI is rolled out, the incentives around it and how regulation responds. Luccioni, Strubell and Crawford make the same point from the other direction: efficiency gains can increase total consumption, the rebound effect known as Jevons' paradox, so efficiency alone does not guarantee a lower footprint. Both conclusions point to policy and practice, not prohibition.
The first saving is not to use AI where a simpler method will do. Public funders and procurers can ask projects to show that they have assessed the need, and to publish that assessment where they can.
Where AI is justified, model choice matters. A systematic comparison of inference costs found that multi-purpose generative models are orders of magnitude more expensive than task-specific systems for the same tasks. Classification, extraction and forecasting are often well served by classical machine learning or small specialist models. Guidance and procurement criteria that ask for the smallest adequate model would move behaviour quickly.
How a model is used matters too. Energy use rises with input length, output length and model size, so short, well-structured prompts that ask for what is needed and no more are cheaper to run. That is a skill organisations can be taught. Our pages on green-optimised generative AI and green AI and energy-aware computing describe how GreenCode approaches the same problem in its own pipeline.
Training runs, batch inference and large data processing jobs can often wait a few hours. Grid carbon intensity varies through the day, and forecasts are public: Great Britain's Carbon Intensity API predicts regional intensity more than 96 hours ahead. Public compute programmes and cloud framework agreements can require flexible workloads to be scheduled against these signals, and energy policy can reward data centres that shift load rather than treating them as fixed demand. Our green hosting page shows how we estimate and report the footprint of this website.
Organisations using AI services are not usually told how much energy their requests consume. That can change. Google's published methodology shows that a provider can measure the energy, emissions and water of a median text prompt, and its authors argue that comprehensive measurement is needed to compare models and reward efficiency.
Where governments build or buy centralised AI services, they can require inference energy to be logged and reported back to each user organisation. People change behaviour when they can see what it costs. Reporting of this kind would also fit the direction of EU rules for data centres, which already require operators above 500 kW of IT power to report energy and sustainability indicators and are moving towards a common rating scheme.
France already has a model. Its general policy framework for the ecodesign of digital services (RGESN), published by Arcep and Arcom with ADEME in 2024, sets out 78 criteria covering websites, applications and AI tools, with the aims of extending device life, reducing resource use and improving transparency. Governments expanding AI and data centre capacity can build requirements like these into their plans and into public procurement from the start.
A simple baseline would help smaller suppliers. The UK's Cyber Essentials scheme shows how it can work: a government-backed minimum standard with a small number of controls and a certification that buyers recognise. A sustainability equivalent, paired with tax credits for measured improvements, would give suppliers a clear first step and a reason to take it.
Every new device carries the emissions of its manufacture, so the longer each one stays in use, the fewer need to be made. Policy can ask when an IT refresh is really needed, as opposed to recommended by the manufacturer, and treat forced obsolescence as a problem to be addressed. The EU's rules for smartphones and tablets now require operating system updates for at least five years after the last unit of a model is sold. Efficient software helps too: code that needs less memory and processing runs well on older hardware for longer.
Some common cybersecurity practices work against decarbonisation. Retaining and analysing large volumes of log and backup data costs storage and energy. Keeping workloads apart on dedicated hardware, instead of sharing resources, lowers utilisation. And when an operating system reaches end of support, as Windows 10 did in October 2025, working machines that cannot run the successor are often retired for security reasons alone. None of these is wrong in itself, but security guidance and sustainability policy are written separately. They should be reviewed together, so that retention periods, isolation requirements and support lifetimes are set with both risks in view.
The EU, its member states and the UK are all expanding AI capacity under similar energy constraints and similar climate targets, and their policies do not always align. Sharing what works, openly, avoids duplicated effort and wasted energy. Our climate case sets out the potential of software optimisation, and our further resources page collects the standards and tools referred to here.
If you work on digital, energy or AI policy and would like to discuss any of this, please get in touch.
