Weighing up AI’s climate costs

LONDON – Historically, technological revolutions have brought higher carbon dioxide emissions – with the first Industrial Revolution being powered by coal, and the second heavily by oil. Will artificial intelligence – the general-purpose technology of our time – do the same? The early signs are concerning. Microsoft’s CO2 emissions jumped some 30% since 2020 as the company invested in AI infrastructure, and Google’s are up almost 50% over the past five years.

But there are two countervailing forces to consider: demand and efficiency. While demand has grown, efficiency has improved. Chips from companies like Nvidia are getting better, and the next generation is expected to be five times faster than the current one. Equally, OpenAI and other industry leaders are making their models more efficient to train and run.

Still, given the surging demand for AI, energy consumption worldwide could still grow, even as models become more efficient. What really matters is emissions, and to project those, we need to know how the electricity to power AI data centers will be generated, and how AI will impact carbon-intensive industries.

According to the International Energy Agency, data centers accounted for roughly 1-1.5% of electricity use worldwide in 2023, and this share is sure to grow in the short term. Microsoft, Google, and Meta nearly doubled their cumulative electricity consumption between 2020 and 2022, and that was before the arrival of ChatGPT. Since then, they have only strengthened their commitments to expanding this infrastructure.

While data centers represented roughly 1% of energy-related CO2 emissions in 2023, the electricity systems that power them are rapidly decarbonizing. In the United States, 41% of electricity was produced from zero-carbon sources in 2023 – marking a one-quarter increase over the past decade – and in Europe, the proportion is closer to 60%. In the US, Europe, the United Kingdom, and China, renewables are the fastest-growing means of producing electricity.

At the same time, Goldman Sachs expects data-center energy demand to grow 15% per year until 2030, with AI accounting for one-fifth of that growth. Even if two-fifths of US data centers’ energy needs are met by renewable energy, AI infrastructure would emit roughly 26 million tons of additional CO2 annually.

But while that is a huge amount in absolute terms, it needs to be put in context. The additional emissions from AI would represent 0.4% of current emissions, and less than the “scope 1” (direct) emissions of any of the three biggest US airlines. Notwithstanding the breathless headlines about AI’s carbon footprint, America’s energy system is so large that the direct impact of AI represents more of a perturbation than a systematic change.

Moreover, there is compelling evidence that AI can reduce emissions across a variety of hard-to-decarbonize sectors. Since aircraft contrails alone are responsible for about 35% of aviation emissions, Google and American Airlines are exploring how machine learning can be used to minimize contrail formation. The early results show that roughly one-sixth of aviation emissions worldwide could be avoided (more than all the current output from AI data centers in the US combined).

Similarly, food waste (which accounts for 6% of global emissions) can be reduced by using AI to forecast demand, manage production levels, and optimize schedules across the supply chain. AI is already being used to reduce emissions from industrial processes (currently 30% of the global total), such as by aiding in the development of biologically inspired materials that are less reliant on fossil fuels (but still meet industries’ mechanical standards), and by reducing the cost and increasing the efficiency of material recycling. And AI will aid climate adaptation as well, by improving weather forecasting and early-warning systems. Timely preparation has the potential to save lives and reduce economic losses.

Thus, while AI will most likely increase global energy consumption in the short term, its potential to drive down emissions across a wide range of industries is immense. We must remember that technological progress can decouple economic growth from emissions. The UK, for example, has boosted per capita GDP by almost 50% since 1990 while slashing domestic emissions by half. AI could be the key to accelerating this trend globally.

Realizing AI’s full potential as a tool for decarbonization will, however, require stronger climate policies. Putting a price on carbon and doubling down on support for clean energy would create powerful incentives for businesses to invest in AI solutions that minimize emissions and hasten the transition to a sustainable future. If we play our cards right, AI may well prove to be our ace in the hole in the fight against climate change.