Engineering sustainable AI

WIth all the hype surrounding AI and the huge changes it might bring to our day-to-day working lives, it’s easy to overlook the technology’s potential environmental impact. Working to address this, earlier this year the Royal Academy of Engineering – in partnership with the Institution of Engineering and Technology and BCS, the Chartered Institute of IT, under the National Engineering Policy Centre (NEPC) – produced a report calling for the extension of mandatory reporting on AI’s energy and water use, carbon emissions and e-waste recycling of data centres.

Speaking about the issues highlighted in the report, Dr Natasha McCarthy, Associate Director of Policy at the Royal Academy of Engineering emphasised: “AI is fundamentally a physical, engineered system. It's quite easy to think about AI as ephemeral; it’s digital; it doesn't exist in real form. But actually it's got this huge footprint, and we need to think about the way that we manage that – and we also need to think about the engineering and design choices that we have in order to build systems that work really well.”

So, how can the engineering community go about achieving environmentally safe, stable AI?  One of the Academy’s strongest calls is for better monitoring and standardisation across the AI value chain. 

“We are calling really clearly for better data, better monitoring, better standards,” McCarthy said. “There are frameworks already – for example, the UK’s Streamlined Energy and Carbon Reporting framework, and the EU’s Energy Efficiency Directive – but reporting is inconsistent. Only one in six data centres in Ireland are actually reporting, and that shows the scale of the challenge.”

AI’s global nature also makes consistent measurement difficult. “AI companies are transnational,” she explained. “The EU has a different system to the UK, and the US doesn’t have the same level of mandatory reporting. That makes it much less easy to get consistent, cross-national comparisons of different AI systems.”

However, McCarthy believes this inconsistency presents an opportunity for leadership. “We think the UK could be a real leader here,” she said. “If we strengthen our reporting mandates and call for standardisation, we can start to deliver AI where we’ve got a clear sense of how much energy and water it’s using, and how we can improve that.”

Data centres and resource use

A central concern is the sustainability of the data centres that form the industrial backbone of AI systems. “It’s well known that data centres have a huge appetite for water and energy,” she said. “We’d like to see more stringent sustainability requirements, particularly around the use of potable water.

“We can’t really afford to have data centres using potable water in cooling. Some are designed with closed-loop systems that recycle water more efficiently, but even those have an environmental footprint and can affect local ecosystems when wastewater is discharged.”

Locating data centres in areas with sufficient water and renewable energy resources could help, she argued, but these developments should also be treated as opportunities for local growth. 

“We’ve seen government talk about AI growth zones and the ambition to build more data centres in the UK. We’d love to see that become a broader opportunity to create innovation hubs, bring investment, and support local businesses and skills.”

Frugal AI

The Academy’s report highlights the growing concept of “frugal AI”,  which focuses on building systems that use less compute power and energy. McCarthy pointed to the example of the Chinese large language model DeepSeek, which surprised observers by achieving high performance with far less energy and less powerful chips. “Not all AI models are the same,” she emphasised. “Not all have to run on the most powerful chips.”

Supporting frugal AI, she argued, could also help smaller organisations and research institutions that struggle to access high-performance computing. “If there’s an opportunity to build AI differently so that it doesn’t use such intensive compute, then we can create wider opportunity as well as reach that environmental sustainability goal.”

Looking to the future, McCarthy described emerging computing architectures such as neuromorphic chips based on memristors, which mimic how the human brain processes information. “The human brain is the most powerful computational tool, and possibly the most energy efficient,” she said. “This technology could use up to a thousand times less energy than a classic computing chip.”

With UK researchers already working at the forefront of this field, there's an opportunity for the country to lead in the next generation of sustainable computing.

McCarthy also urged businesses to rethink how they approach data and algorithms. “We don’t always need to be using the largest language models for everything,” she said. Smaller, task-specific models that run on edge devices rather than in data centres can reduce energy use and bring wider benefits in privacy, data protection and fairness.

"If we can do that, we can create sustainability benefits as well as address some of the bias and copyright challenges that come from harvesting huge amounts of data from the internet.”

Sustainable AI

For McCarthy, the challenge of sustainable AI is ultimately one of systems thinking: viewing AI within the broader context of energy, infrastructure, and society.

 “We simply can’t afford for AI to make a disproportionate demand on our energy system, especially as we electrify transport and heat,” she said. “We need to think about AI as part of our overall transition towards a decarbonised grid.”

At the same time, AI can play a vital role in supporting sustainability. “AI has a huge contribution to decarbonisation,” she said, citing UK companies funded by the Academy’s Enterprise Hub, such as Q Flow, which uses AI to monitor sustainability across supply chains, and DonAAI, which applies AI to detect manufacturing faults and reduce waste.

“What we really need,” McCarthy concluded, “is a whole-systems approach that ensures AI has a net benefit rather than a disproportionate cost.” 

That means designing AI that is “proportionate to its appetite”, enabling applications that truly matter, from decarbonisation and healthcare to scientific discovery.

nepc.raeng.org.uk