We are seeing rapid growth in artificial intelligence (AI). Large Language tools like Chat GPT, CoPilot, Google AI, and Perplexity are being increasingly built into our online lives, and many people are using them unknowingly. But the computer power needed for these systems is making growing demands on our supplies of electricity and water. Some claim that this is unsustainable. Are they right?
Much more than large web searching
Traditional search engines search the web, indexing publicly accessible material. Asked a question, they find relevant sites and just provide a list. But Large Language models (LLMs), trained on huge quantities of online material and able to generate complex answers, raise more difficult issues: about intellectual property rights and plagiarism; about bias and inaccuracy; and future employment.
Increasingly LLMs are being built into websites. Google still searches for relevant websites, but it now routinely uses AI to summarise the material it has found, and organise it for the convenience of the enquirer. And any of these models can turn a query result into an extended essay with a single click.
Resources and the datacentres
There are many concerns about these developments, but here I want to look at the specific issue of sustainability. At the centre of this are the datacentres: huge banks of computers which need vast quantities of electricity, and either water or more electricity for cooling.
Datacentres provide a host of services, like cloud computing and streaming services. The development of autonomous vehicles and the internet of things will increase their importance. But currently, large language models account for two-thirds of their activity. Some of this is used to answer queries, but a very large amount is used in training the models themselves.
The UK is the sixth largest datacentre provider in the world, and the biggest in Europe. Our 400 large datacentres currently account for around 2.5% of UK electricity consumption, and at 148 Megawatts, the largest, in Cardiff, can consume more electricity than Norwich.
Cooling the computers
One striking claim from critics of AI is the assertion that a single Chat GPT query consumes half a litre of clean water, at a time when global water supplies are under threat.
The huge banks of computers which gather and process data generate heat and have to be cooled. Most do this with air driven by electric fans. But many use clean water, which is a globally scarce resource.
Big datacentre providers are experimenting with using water from less clean sources, and recycling the water that they do use, but there are challenges with corrosion and contamination. Nevertheless, four of the largest users, Amazon, Microsoft, Google and Meta, have all committed to going “water positive” by 2030.
Using electricity
Making datacentres themselves ‘water positive’ is a step forward. But most of the water-use is not in the datacentres themselves, but in the generation of the electricity, where that comes from fossil fuelled power stations. One study suggests that a typical Chat GPT query uses ten times as much electricity as a conventional Google search, and generates 300 times more CO2. Although the cost of a query is tiny, Chat GPT currently processes 200 million a day.
However, the proportion of electricity generated from renewables, which do not require cooling water, continues to rise, and last year in the UK, for the first time, the proportion passed 50%. That was a 6.5% rise in a year. Some datacentres, like the Cardiff one, already get most or all of their power from renewables.
A future?
The new study by the International Energy Agency (IEA) suggests that in the USA, which houses nearly half of all datacentres, AI may account for only 10% of the growth in electricity demand in the rest of the decade, exceeded by growth for industrial uses, air conditioning and electric vehicles. Nevertheless, in 2024, AI-focused datacentres consumed 1.5% of all global electricity, and its demand will double by 2030, with AI, in all its forms, the biggest element. They predict that half the overall growth will come from renewable sources (including small modular nuclear).
As search engines have matured, they have become more energy efficient, and IEA suggests that the same will happen with datacentres. Furthermore, the IEA report suggests that existing AI tools (not LLMs) can be deployed to increase the efficiency of energy use. They claim that, by improving the management of the grid, transport, industrial uses and heating and cooling buildings, they could offset all the expected growth in electricity demand up to 2030.
The Coalition for Sustainable AI is about to publish a mapping of sustainable AI initiatives. And the IEA’s new Observatory on Energy, AI and Data Centres is gathering worldwide data on AI’s electricity needs and tracking emerging AI applications. Their latest analysis paints an optimistic picture of the future, and suggests that some of the alarm about sustainability may be misplaced.
The way forward?
AI is not going to go away. We have got used to the convenience. Writing, editing and publishing this article has involved dozens of web searches and downloading multiple documents. And you have used water and electricity simply by reading it. Commercial pressures will drive more AI development and expand its uses. Realistically, some individuals refusing to use it will not make a significant difference.
So, we are going to need regulation to ensure that it is not used in inappropriate ways. There are proper concerns about theft of intellectual property, inventing and misrepresenting facts and stoking societal conflict. There are also issues about employment and taxation. Those are all going to be challenges, because the politics is complex and the technology companies creating AI systems are huge and beyond the control of any single government.
More fundamentally, water and energy systems in most countries are not well designed to manage the balance between competing resource needs. And, as always with technological developments, there is a question about how any benefits will be distributed.
We need to be careful and realistic about the resource issues. AI is only one of many ways in which we are depleting the resources of the planet. Optimists believe that it can actually help us limit, or reverse, those pressures.
There are many reasons to be cautious about AI developments, but perhaps the sustainability issue is not the most significant. The challenge is not whether or not we can create sustainable AI, but how to govern and manage it.
More from East Anglia Bylines
