How sustainable is Artificial Intelligence?

Artificial intelligence (AI) is poised to play a critical role in the transition to a lower-carbon economy, notably in overall energy optimization as the US modernizes its power grid, as well as in energy management of the built environment.

The potential uses for AI stretch across nearly every aspect of modern life, leading to intense public scrutiny of its potential benefits and challenges, including those related to climate. Training AI models requires a tremendous amount of computational power and energy, and models continue to consume a vast amount of energy once the technology is deployed.

So, what are the tradeoffs between the benefits of integrating AI into existing workflows and the emissions produced with its development and utilization? How can companies deploy AI in the context of their decarbonization strategies?

Optimizing operational energy use

The US energy grid has already begun a massive transformation in order to adequately meet the scale-up of renewables and future electricity demand. Electrification is critical on the path to decarbonization, but the increase in electricity demand is outstripping the aging grid. At present “nearly 70% of the nation’s grid is more than 25 years old [and] a recent study estimates electricity demand could grow by 20-38% by 2050 under certain scenarios” (USDOE).

As public and private entities work together to prepare for a decarbonized future, the country must manage intermittent and distributed energy sources from wind, solar, batteries, green hydrogen, small nuclear (modular) reactors, and electric vehicles. AI can play a crucial role in accurately forecasting demand, detecting outages, and balancing load based on real-time data. Virtual power plants will continue to help connect decentralized energy sources and employ AI to ensure they can work cohesively to optimize energy production and consumption.

On a smaller scale, AI can be deployed in buildings to optimize energy consumption. For example, AI can detect inefficiencies from equipment and identify when it may need repair or replacement. It can also use data from smart sensors and predictive capabilities to respond to changes in room occupancy and external factors, such as weather.

The commercial building sector accounts for 18% of total US energy consumption (EIA); therefore, improved efficiency in this space could yield benefits in emissions reduction efforts.

Powering the AI transition

There is a flipside to the sustainability benefits of utilizing AI, as significant emissions are generated from developing AI models from both the operational carbon (the electricity needed to power the systems) and the embodied carbon (the emissions associated with manufacturing hardware for the systems).

In one attempt to quantify the energy and emissions output of such endeavors, researchers at the University of Massachusetts conducted lifecycle assessments which found that training AI models “can emit more than 313 tons of carbon dioxide equivalents – nearly five times the lifetime emissions of the average American car, [which] includes the manufacturing of the car itself” (MIT Technology Review).

Newer studies further contextualize the energy requirements needed to create and operationalize AI models. A 2021 study estimated that the training of GPT-3 (the language model that powers ChatGPT) generated 502 tons of carbon emissions (The Journal of Machine Learning Research¹). This does not include the emissions associated with inference (the actual use of the model that generates predictions and solutions). Some data shows that that 40% of total [AI] energy consumption can be attributed to training, while the other 60% comes from inference (The Journal of Machine Learning Research²).

Emissions generated from training, retraining, and deploying AI models will increase as they become more complex and more common. Employing AI within business operations has the potential to unlock value and reduce both Scope 1 and Scope 2 emissions, but it can also have negative implications on Scope 3 emissions.

AI can reduce Scope 1 and Scope 2 emissions by supporting process optimization and improving energy efficiencies. However, when organizations create their own models or integrate their data with a third party, they must utilize data centers and computing services to store and process information, leading to increased emissions.

At present, many of the largest data storage and cloud computing providers are working toward 100% renewable energy use. However, as demand for AI increases, more energy will be required to meet it. Renewable energy production will continue to grow, but competition for resources will be a key challenge, especially as organizations across all sectors strive to decarbonize and reach reduction targets. Data centers need to be hyper-focused on their own emission reduction efforts as demand for their services grows.

Theoretically, as the grid transitions and integrates AI to deploy renewable energy, renewables will make up a growing portion of energy production, allowing clean energy to power the training and use of AI models. But currently, in practice, renewable energy only generates 20% of all US electricity (USDOE). While renewables’ share will grow as the race to net-zero captures investment, it is unlikely to keep pace with the rapid increase in energy demand in the immediate future. Thus, AI's potential to drive sustainability hinges on aligning its energy consumption with renewable sources.

Moving forward

Currently, the development, deployment, and use of AI in business operations for the purpose of decarbonization poses a sustainability conundrum. But AI is not alone in this conversation about reducing carbon emissions and transitioning energy production to renewable and more sustainable options. The potential for real gains exists, but companies must be cognizant across the entire lifecycle and understand both the benefits and limitations. With all this in mind, organizations should:

• Develop tools to effectively track and measure changes in emissions across all scopes
• Assess how AI and data service providers are powering operations
• Work with AI partners that have robust emissions reduction strategies (especially those aligned with SBTi)
• Conduct cost-benefit analysis to assess the impact of AI systems on emissions
• Develop a portfolio of investment options to achieve energy and climate targets.

^1. Citing Patterson, et al., Carbon emissions and large neural network training (2021).

^2. Citing Patterson, et al., The carbon footprint of machine learning training will plateau, then shrink (2022).