AI is not one thing. Neither is its environmental impact.
We need to keep our eyes, our debates, and our policymaking focused on the problems we’re trying to solve. On the outcomes we’re trying to achieve and the society we want to build.
AI can be machine learning, Grok, or basic (mislabelled) data analytics and simple process automation software.
Its impact on sustainability can of course be environmental – its emissions, energy use, water consumption, effect on the local environment, or through the physical materials and supply chains that build datacentres and all manner of other infrastructure.
AI’s impact can also be social and economic – on employment, education, policing, for creatives, or on local community cohesion.
When considering governance, often the third pillar of sustainability (and the acronym ESG) AI opens up new questions of accountability from self-driving cars to corporate management.
I’m not going to rank these impacts. Or pretend that in one blog post I can fully account for their breadth and complexity. But as a starting point for our collective future conversations, here are some numbers, questions, and counterarguments to consider…
The whole technology industry, including chip manufacturing and electronics supply chains, is 2-3% of global emissions, based on an analysis combining sources like Our World In Data and the World Economic Forum. This is, it’s worth noting, a variable estimate depending on how you define “the sector” as is the case with all carbon accounting. The International Energy Agency (IEA) reports that AI is far less than the 0.5% of global emissions generated by all current datacentres, which provide storage and processing for our digital lives far beyond AI. Now that footprint will grow with the rapidly increasing number of datacentres before it comes back down as electricity is decarbonised. And given emissions saved now are especially vital in arresting global warming (vs emissions saved years and decades down the line) the construction of new gas and coal to fuel datacentres should be stopped. In the long run the energy transition is building critical mass: over $2 trillion will be spent this year on clean energy vs $1 trillion in fossil energy. Agriculture, steel, cement, and other sectors are more worrying as far as a successful transition to net-zero is concerned. But all this said, we must address both the now and long term.
Regarding electricity, it’s a major headache for planners and the grid. Numbers across NESO, DESNZ, and National Grid project datacentres as 2-3% of UK electricity demand in 2026. A recent London Assembly report cited that will grow to between 30-71TWh by 2050 (roughly between 1 and 3 Hinkley Point C nuclear plants) – an immensely variable number and likely supplied by heavily decarbonized electricity.
The impact of the physical infrastructure follows the concerns of the broader tech and infrastructure sector: security of critical mineral supply, refining ability, and human rights abuses in those opaque supply chains. Something we desperately need to address in legislation but not uniquely to AI.