Is AI Sustainable?

^{Source: Luccioni, Sasha, Yacine Jernite, and Emma Strubell. "Power hungry processing: Watts driving the cost of AI deployment?." In The 2024 ACM Conference on Fairness, Accountability, and Transparency, pp. 85-99. 2024.}

Academic evidence suggests that, at present, training and inference account for roughly equal proportions of AI’s lifecycle carbon emissions, with the embodied carbon of AI hardware accounting for the rest.⁶

^{Source: Wu, Carole-Jean, Bilge Acun, Ramya Raghavendra, and Kim Hazelwood. "Beyond Efficiency: Scaling AI Sustainably." IEEE Micro (2024).}

^{Source: US Census Bureau.}

Second, researchers may exaggerate the likely future energy demand from more sophisticated AI. In the past we have seen catastrophic predictions on future data centre power use, which have proven to be illusory.¹⁵ Things have turned out better than expected in part because data centres have become more efficient at converting energy to brainpower. Over the past decade, internet traffic has increased 25-fold, yet power usage from data centres has risen only slightly.¹⁶ Improvements in efficiency have probably sped up during the AI era. If this continues, AI may not need as much extra power as some fear.

The third issue is the most tricky. This is the question of where AI power comes from. It matters a great deal whether AI is powered by fossil fuels or renewable energy. In theory, cloud companies could build vast numbers of renewable-energy power plants to power their AI data centres, meeting additional energy demands without increasing carbon emissions. But things are more complicated than that.

Unfortunately, it is usually too risky to try to power data centres solely with on-site renewables. Renewables tend not to be reliable enough to provide power at all times; the wind, after all, does not always blow. As a recent Bernstein paper argues, the 'intermittency problem' does not go away even with large-scale energy storage (such as batteries) on site.¹⁷

As a result, data centres usually need to be hooked up to the grid. This could massively increase demand on grids which are already overworked. More capacity will be required. And yet, just like with data centres, there are often large regulatory and political barriers to the construction of additional power plants. “If you’re talking about building large new power plants or large build-outs,” says Zuckerberg, “and then building transmission lines that cross other private or public land, that’s just a heavily regulated thing.” Power-plant construction in the US remains surprisingly weak, according to official data.

You also have to connect the data centres to the grid. It sounds trivial, but in fact these interconnections need regulatory approval, and possibly the construction of dedicated substations, power lines and backup power systems. This whole process can be slow. For years there have been lengthy interconnection queues -- and last year in the US they grew by 27%.¹⁸ We know that last year, production by US companies involved in the transmission, control and distribution of electric power actually declined, a surprising statistic given the excitement over AI.¹⁹ It is not much of an exaggeration to say that the main constraint for AI is not energy supply per se, but the grid. As someone put it to us: “renewables are cheap, copper is expensive.” (You need copper to build and upgrade grid infrastructure.)

This leads to a worrying possibility: that to meet the demands of generative AI without relying on a constrained grid, companies will turn to fossil fuels. Fossil fuels could provide a steady, 'behind-the-meter' (i.e., off-grid) supply of energy for data centres. Experts we have spoken to suggest that to meet the extra energy demand, we might have to keep old gas- or coal-power plants running longer, or even build new ones. Already there are anecdotes of this happening. That is deeply concerning. Just as the world needs to make progress on cutting carbon emissions, we risk going backwards.

The world therefore needs to think about workarounds, to avoid a situation in which fossil fuels power the growth of AI. One solution is to do a large amount of training and inference a long way from busy population centres. Here, supplies of renewable energy tend to be more reliable. Wind farms, for instance, are typically located in rural or coastal areas where wind speeds are higher and more consistent. Rather than carrying energy across large distances to power AI models in cities, you could instead carry the data from the AI models. This approach is fine when it comes to training models. It is trickier when it comes to inference, however. If data centres were located far away from demand, there would be more 'latency' -- a delay between when you ask a question of an AI model and when you get an answer. For many AI applications this small delay might not be a big problem. But for others, latency would be far from ideal.