Meeting AI’s demand for power

The National Grid could struggle to meet the energy needs of the AI-based technology we now use. According to one study AI systems might use 33 times more energy than machines designed for specific tasks. This is because AI queries are handled in data centres which are energy hungry. By 2026 they could consume 1,000TWh of electricity globally – which is about the same as all of Japan’s energy use. 

 

The UK’s current electrical infrastructure will face a huge increase in demand. It’s not just due to data centres supporting AI but also electric vehicle (EV) charging and other technology. 

 

AI is becoming embedded in everything from digital assistants to mapping and navigation tools. Data centres use many interconnected servers and the services you use ‘in the cloud’ are really rooted on the ground. They use a lot of electricity which is putting increasing pressure on energy supply. 

 

With renewable energy systems contributing more and more, the National Grid can keep pace for now. However, the future is less certain. 

 

Each time an AI query is made, the whole Large Language Model (LLM) is activated. In terms of processing and energy use, that’s very inefficient. 

 

Many generative AI (GenAI) systems were ‘trained’ using massive stores of written information. That’s why they can provide responses to almost any question. However, to do this they need to access the whole model. 

 

The measurement of data centre efficiency is usually power usage effectiveness or PUE. The lower the number, the more efficient they are. The latest data centres have a PUE of around 1.1 but they still generate a lot of waste heat. This could be useful to heat swimming pools, for example, as long as they’re close enough. However, demand will soon outgrow the efficiency benefits we can achieve. 

 

One option is to build new data centres with their own on-site generators or renewable energy sources. Others are looking at ways to measure energy ratings for AI to help develop more efficient models. It’s even been suggested that AI use is scheduled according to energy demand. For example, shorter tasks could run overnight while planning larger projects for cooler months in places where air conditioning is widespread. 

 

On the other hand, AI can help analyse and predict wider energy use to help everyone become more energy efficient. It could help optimise energy storage, improve carbon capture and predict weather and climate changes.