AI’s environmental cost could outweigh sustainability benefits

What are the environmental impacts of AI?

However, despite the potential sustainability benefits of AI, speakers were clear that the technology itself is having huge environmental impacts around the world, and that AI itself will come to be a major part of many organisations Scope 3 emissions.

Harvey Lewis – a partner in Ernst & Young’s AI division who was chairing the panel on how to leverage the tech for sustainability purposes – said that while there are a number of estimates out there, “broadly speaking, [training GPT] required around 25,000 Nvidia A100 GPUs running for about 100 days”. This is equal to the amount of energy required to power every single household in the UK for an hour, or 2,000 homes for 18 months.

“It’s a staggering amount of energy, and that’s just scratching the surface,” he said, adding that this does not take into account things such as the amount of energy it takes to make a single inference; the sheer and growing number of users such AI systems have; and the substantial increases happening in model sizes.

Pugh added that if the rate of AI usage continues on its current trajectory without any form of intervention, then half of the world’s total energy supply will be used on AI by 2040.

Pointing to the example of some AWS datacentres in Dublin having to turn off their AI services because Ireland’s grid could no longer meet the energy burden of the technology, Pugh said: “We’re already seeing that [AI] demand is outstripping [energy] supply.”

Eskelinen also pointed out that, at a time when billions of people are struggling with access to water, AI-providing companies are using huge amounts of water to cool their datacentres.

However, he added that while running AI systems is clearly extremely energy intensive, there is also the potential to build a degree of circularity into its operation by, for example, using the excess energy from datacentres to power surrounding buildings or infrastructure.

In another example, he said that datacentres in Finland are using sea water as a natural coolant, but further noted that it would likely require some sort of regulatory interventions from governments to make these kind of alternative approaches viable.

“There are ways to do this, but it will probably require regulatory intervention to tax water usage, which would create a financial incentive to be better at this,” he said.

Eskelinen added that it was also key for people in the tech sector to “internalise” thinking about the socio-economic and environmental impacts of AI, so that it is thought about from a much earlier stage in a system’s lifecycle.

On practical steps companies can take now to manage the environmental impact of their AI use, Pugh said the first step is working with IT teams and suppliers throughout the chain to understand where data is stored (for instance, to figure out if there are any renewable energy sources that can be used or if diesel generators are deployed), before then working on a plan of action for how to make the operation as clean as possible.

Part of this, he added, would also be assessing what applications are in use throughout an organisation, and whether any pre-trained algorithms can instead be used to minimise the emissions in model training.