AI and energy use

‍Good morning from New Economy Brief.

‍

We recently explored the government's plans to accelerate the adoption of Artificial Intelligence (AI) across the UK economy and what this could imply for productivity, growth and jobs.  

‍

This week’s New Economy Brief asks what AI’s environmental implications could be, particularly after a new innovation by a Chinese company (DeepSeek) wiped a trillion dollars off the US stock market by showing how AI could be trained much more efficiently. 

‍

--

Turbocharging AI in the UK.

Governments across the world are investing huge amounts of money in infrastructure and R&D to accelerate the development of AI technology and its use in their economies. Last month New Economy Brief explored the UK government’s plans to ‘turbocharge’ AI and accelerate its rollout across the economy to increase productivity. (Read for more on the potential economic implications of an AI-led disruption to jobs, and whether the UK’s regulatory system is prepared for this.) 

‍

What is compute? The government’s latest plans included a goal to increase the UK’s AI compute capacity twentyfold by 2030, through AI sector growth zones, new data centres and other digital infrastructure. Compute is the processing power needed to train a Large Language Model (such as ChatGPT) on vast quantities of data. It is a mix of hardware, software and other infrastructure – most of the calculations are done on GPU chips in data centres, but these depend in turn on everything from servers, miles of cabling and special cooling equipment to carefully optimised algorithms. Multiplying the UK's compute capacity by 20 is likely to have environmental consequences due to the amount of resources needed to manufacture, train and use AI systems.

‍

--

Projections of AI’s resource use.

A recent report from the National Engineering Policy Centre (NEPC) warns that “growing demand for AI and data centres could have far-reaching consequences, such as competition for renewable energy or drinking water sources.” 

‍

Energy. In some jurisdictions, electricity consumption by AI systems is projected to “outstrip renewable generation capacity” within the next 10 years. Last year the head of the National Grid warned that total energy demand from the UK’s data centres could grow sixfold over this time, which will mean the UK needs to generate more renewable energy to meet the government's Clean Power 2030 target. A recent study found that “If additional renewable capacity cannot be deployed quickly enough, the UK might face a scenario where AI-driven electricity demand increases overall emissions.”

‍

Water. Data centres withdraw and consume potable water from nearby sources, putting local businesses and communities at risk of water scarcity. Even without the extra demand from new data centres, seven regions in England are already predicted to be severely water stressed by 2030. This includes London, which already has a high concentration of data centres (more than three quarters of the UK total). Water is also consumed to make components needed for compute: an average chip manufacturing facility uses as much water per day as 33,000 US households. Local communities in Chile and Uruguay are resisting new Microsoft and Google data centres being built by Microsoft and Google due to fears of exacerbating water shortages.

‍

Critical materials. Hardware components needed for compute capacity, such as computer chips, need a variety of rare earth metals. Mining these causes many environmental harms such as habitat and biodiversity loss, as well as leading to conflict, exploitation and devastating impacts on local communities, often in the Global South. Demand for these materials is rising with the complexity of AI technology development, and the UK is particularly bad at recycling – a 2022 study of how much e-waste a group of countries produce ranked it as the second-worst offender. (The UK government is expected to release a Critical Minerals Strategy this year, detailing its plans to shore up supply chains.)

‍

--

Is AI likely to become less resource intensive?

Governments have thrown significant financial resources into winning the race to develop better AI models. Decisions like the Trump administration’s $500bn ‘Stargate’ investment are partly bets that demand for compute power will continue to grow exponentially. Expensive computer chips have been the main physical bottleneck for AI development so far, which is why governments are investing so much in digital infrastructure like new data centres – to stay in the competition to be global leaders in an industry they think will define the 21st century. 

‍

Will ‘DeepSeek’ make AI more efficient? In January the Chinese company DeepSeek launched a new AI large language model and demonstrated how to train AIs that perform like OpenAI’s ChatGPT4 in much more energy- and resource-efficient ways than was previously thought possible. Professor Ben Ansell from the University of Oxford estimates they achieved similar results at “around a thirtieth of the cost… roughly the same as the decline in the cost of light in the UK between 1925 and 2005, as we moved from coal and gas lamps to LED bulbs. Except this happened over a few months.” Before this, Nvidia, a manufacturer of computer chips, had become the world's largest company at one point last year, as investors believed the exponential growth of compute capacity (and thus demand for its computer chips) would need to continue to service growing demand for AI. But DeepSeek's launch suggested that future AI models might need much less of Nvidia’s hardware than had been assumed, wiping $600bn (18%) off its market capitalisation in one day, though it has since recovered some of these losses.

‍

What about Jevons Paradox? History shows that technological developments which make resource use more efficient tend to go one of two ways. If a new breakthrough lets us do something more cheaply and efficiently, we sometimes do the same amount of it as before while enjoying the fact that it now costs less. But the lower price can also encourage people to consume more. For example, as cars get more fuel-efficient, drivers tend to drive more so their overall petrol consumption goes up – what's known as the Jevons Paradox. Training AI models has previously had very high barriers to production, as only huge tech firms that can afford expensive compute capacity could develop them. On the one hand, after DeepSeek, much lower costs of production could make AI less resource intensive and increase competition in the market from smaller players. But on the other, it could allow more powerful AI models to be developed and applied more widely. This is why the Big Tech moguls are arguing that the huge investments in AI infrastructure were not a mistake. In short, as Macrodose’s James Meadway explains, DeepSeek is unlikely to be the “best environmental news we have had in a long time”. Tech firms are competing to build and disseminate the most powerful AI model, and they will likely continue to need more and more compute capacity to stay in the race.

‍

--

What does this mean for the UK government?

When announcing plans to accelerate the rollout of AI systems across the UK economy, Prime Minister Keir Starmer talked about the UK becoming a world leader in AI development and championed using the technology to make public services more productive. Aside from the regulatory issues explored in our previous digest, recent technological developments may have significantly reduced the cost of deploying AI in public services. As Ben Ansell notes, “the worst thing that could happen to this plan is for Labour to get continually snookered by high-paid IT consultants recommending extremely expensive proprietary AI systems, run by Big Tech…If the government can instead use a basically as good set of AI models for, let’s say, three percent of the cost, well why wouldn’t they?”. For a Labour government desperate for rapid productivity growth and under significant spending pressures from multiple directions, AI’s potential benefits could now look even more attractive. 

‍

Managing AI’s resource intensity. The government is developing an AI Energy Council to discuss balancing supply and demand for energy and compute. But as the NEPC report notes, the sustainability challenge for AI systems includes critical materials and water too. It recommends forcing data centres to report on energy, water use and e-waste recycling to improve our understanding of their environmental impact. Strong regulation will be needed to ensure growing compute capacity and demand for AI doesn’t exacerbate water shortages or come at the expense of either the government’s clean energy mission.