AI’s environmental problem is now well known, with the heavy energy use of large language models as the main culprit. Less talked about is AI’s potential to go green in countries with high AI investment per capita, like the US and Singapore. Angel Melguizo writes that most countries are on the wrong side of the curve and suggests four actions that governments can take on the road to green AI.
Beyond the astonishing news we read every week on AI advances lies a key challenge: AI’s voracious appetite for energy and its growing carbon footprint. Training large language models (LLMs) and running ever-more complex algorithms require vast data centres, which in turn draw on national power grids at unprecedented rates.
To give a flavour of the size of the environmental challenge, electricity consumption by Alphabet, Amazon and Microsoft surpasses 100 terawatt-hours, roughly equivalent to the combined annual consumption of Colombia and the Dominican Republic together, and their greenhouse gas emissions are growing double digits. These figures may understate the true impact, as many leading AI developers do not disclose emissions from model training, which is particularly energy-intensive.
Optimists argue it is a matter of (little) time for AI to turn green, based on the existence of a “green AI curve”: when AI is first introduced and used more widely, both energy use and carbon emissions go up. This is because running AI systems requires a lot of electricity, and most of that electricity still comes from fossil fuels. But after a country (or a company) invests a lot in both AI and clean energy, the curve turns around—emissions start to drop, and the share of renewable energy grows.
AI spending and carbon emissions
Our recent working paper at Oxford University’s TIDE centre analyses the dynamics of AI spending, energy consumption, CO2 emissions and renewable energy adoption in 23 middle- and high-income countries—nearly 80 per cent of the global economy and all leading AI-developers and users. The analysis covers the period from 2019 to 2023 and brings good and bad news on green AI prospects.
Starting with the good news. We estimated a model where energy consumption, carbon emissions and renewable energy adoption are explained by country characteristics (per capita income, education, government spending, industry structure and trade) and AI spending, and plotted the results of AI spending against carbon emissions and renewable energy adoption.
We saw that relations are not linear. If AI spending is plotted versus the share of renewables in the energy matrix, the relationship follows an “inverted-U” shape. When a threshold of AI spending per capita is reached, the share of renewables grows and CO2 emissions decline.
In other words, AI can turn green. But only Singapore and the United States have reached the levels of AI investment per capita ($220–$580) that allow the environmental benefits—reduced emissions and greater reliance on renewables—to begin to outweigh the costs.
However, there is bad news too. While AI holds potential to accelerate the green transition, most countries are currently experiencing the opposite. AI is significantly augmenting energy consumption, and almost all nations will remain stuck on the “bad” side of the AI-emissions curve—where digitalisation drives up both energy use and emissions. And this does not consider the use of critical minerals, rare earths and water for semiconductors and machinery.
Figure 1. CO2 emissions and AI spending per capita
Notes: The left panel shows GDP per CO2 emissions as a function of AI spending per capita. The right panel shows CO2 emissions per capita as a function of AI spending per capita.
A further cause for concern is that in most countries, the energy matrix powering AI remains dominated by fossil fuels. The share of renewables in national energy mixes has not kept pace with the surge in AI-driven demand. In fact, the study finds that at lower levels of AI adoption, the share of renewables declines, as efficiency gains in fossil-heavy grids are prioritised over clean energy expansion.
Again, it is only after surpassing high levels of AI spending per capita that the positive relationship between AI and renewables emerges. Two main hypotheses might be driving this movement: fossil fuels are depleted or become too expensive, and AI allows power grids to go green. Hopefully, with more data we will be able to test these hypotheses in a near future.
For the vast majority of nations, Europe included, this inflection point is years—if not decades—away, at current rates of investment.
Figure 2. AI spending and renewable energies adoption
Note: The figures shows the share of renewables in the power grid as a function of AI spending per capita.
If no action is taken, the AI revolution will worsen environmental outcomes for most countries in the short and medium term. So now´s the time to move thresholds closer (and not as high as Singaporean or US spending levels) and to shift emission curves down (lightening the energy bill all along the way).
Four steps to green AI
Here are four main areas for action.
Make renewable energy more affordable and accessible
As long as AI is powered predominantly by fossil fuels, its environmental costs will continue to mount. Clean energy subsidies, smart(er) grids and cross-border energy trading and regional grid integration are urgent.
Mandate energy efficiency and emissions disclosure
Transparency and efficiency must become the norm in AI development and deployment. If the stick of mandate on public reporting of energy use and emissions is not feasible, let´s push for carrots, linking public procurement, tax incentives, and R&D funding to demonstrated improvements in energy efficiency and emissions reductions. These measures should come with international standards for measuring AI’s energy use and emissions across the lifecycle (development, training, deployment), so we compare apples with apples, and amazons with amazons.
Foster energy-saving AI applications
AI itself can be a powerful tool for energy savings—if deployed wisely in grid optimisation and manufacturing, transport and agriculture applications.
Innovate in financing green AI
A way to do it would be to issue “digital development bonds” specifically targeted at financing AI infrastructure powered by renewables, with returns linked to emissions reductions and energy savings and leverage blended finance to crowd in private capital for green AI projects, especially in emerging markets.
The time to bend the curve toward a greener AI future is now. Which side are you on?
Sign up for our weekly newsletter here.
- This blog post is based on Can AI grow green? Evidence of an inverted-U curve between AI, energy use and emissions, by Angel Melguizo, Raul Katz and Juan Jung, a working paper of the TIDE centre (Technology and Industrialisation for Development) at Oxford University.
- The post represents the views of its author(s), not the position of LSE Business Review or the London School of Economics and Political Science.
- Featured image provided by Shutterstock
- When you leave a comment, you’re agreeing to our Comment Policy.
