The CO2 Emissions from AI

The CO2 emissions from AI consume about 0.5% of global energy^[1]In 2024 The Verge published: “How much electricity do AI generators consume?”. Artificial intelligence is based on programs called models that are trained by reading enormous amounts of data, e.g. the publicly available internet, in order to perform tasks. All data centres on the internet, for comparison, consume approximately 1 to 2% of global energy.

On the ChatGPT platform, the initial training is followed by deployment to customers, at which point the models answer questions on the data it saw (i.e. the internet) in fully expressive everyday language. The largest amount of AI energy consumption is accounted for by models which create images, e.g. current platforms DALL-E or Midjourney.

Unfortunate Business Models Drive CO2 Emissions from AI

Not only is the increase in CO2 emissions caused by AI dangerous, but it seems that the business models that have evolved in the world of Big Tech and free market capitalism have become locked into an ever-accelerating cycle of competition in which the corporations deploy more and more computer power, consume greater and greater amounts of energy and seek more and more training data to outcompete their rivals. The result is a classic example of the Jevons Paradox (or “Rebound Effect”) where cheaper hardware or energy or data will not result in cheaper prices, rather it will be eaten up to provide greater, faster, more human-seeming results.

So what can one do? Embrace AI and “unleash” it into the economy as the UK’s prime minister just announced^[2]UK Government: “Prime Minister sets out blueprint to turbocharge AI”, with an attempt to power it by small modular nuclear reactors, to avoid the fossil fuels emissions? The non-existence at the time of writing of small modular nuclear reactors as a deployable energy source might cause problems here.

One could deploy a carbon tax to make the energy used by AI more expensive and perhaps drive the transition to renewable energy power sources? The evidence to date on carbon taxes shows that they do have a limited but measurable impact compared to fossil fuel usage in their absence. However, carbon taxes are only designed to encourage the adoption of renewable alternatives, and it is unlikely that the tax impact here will have the desired effect in the hotly contested AI marketplace. Energy prices are not a significant limiting factor, because much of the costs of AI are funded by investment to capture market share.

Hence, Carbon Accounts. If corporations are obliged to fund fossil fuel energy purchases with carbon tokens derived ultimately from customers, and these are limited in supply, then corporations will be compelled to turn to other energy sources. Microsoft is already purchasing and re-commissioning old nuclear power stations. Corporations have shown their flexibility by relocating their data centres to locations with cheap, clean power, e.g. Iceland. Such actions would become standard practice, alongside all other kinds of innovation to clean up or reduce CO2 emissions from AI.

Our sister organisation EcoCounts has recommendations on personal actions one can take on the issue. Read on here for more of the advantages of the Carbon Accounts policy in combating society’s addiction to fossil fuels.

EcoCore is funded by the small regular subscriptions of individual donors like you. Please sign up via Patreon: