AI training emissions: AI-enabled systems contribute to global carbon emissions

• Author:
• Author name

  Quantumrun Foresight
• May 3, 2022

Insight summary

The surge in artificial intelligence (AI) technology has brought with it an unexpected environmental challenge, as the power consumed during AI training leads to significant carbon emissions. Recognizing this issue, the industry is exploring solutions such as developing more energy-efficient AI models, partnering with renewable energy companies, and relocating data centers to minimize energy consumption. These efforts, along with potential regulatory measures, are shaping a future where technological advancement and environmental responsibility can coexist.

AI training emissions context

Artificial intelligence (AI)-driven systems are known to consume significant amounts of power during their training phases, leading to the emission of large quantities of carbon. This, in turn, contributes to climate change, creating an environmental concern that cannot be overlooked. As the AI industry continues to grow, with an increasing demand for larger and more complex models, the challenge becomes even more intricate. 

AI is playing an increasingly important role in the global economy and driving new developments in the healthcare, technology, and energy industries, to name but a few. However, amid the beneficial change being introduced by AI systems, studies have shown that high amounts of carbon are produced due to the power consumed by AI systems when they are being trained and when they perform a large number of calculations. According to research conducted in 2019 by the University of Massachusetts at Amherst, approximately 1,400 pounds of emissions are generated when training an off-the-shelf AI language processing system. In addition, depending on the power source, about 78,000 pounds of carbon are emitted when a deep learning AI system is built and trained from scratch.

In recognition of how the creation and training of AI systems contribute to climate change, the Green AI movement has emerged, which seeks to make AI-enabled processes cleaner and more environmentally friendly. The movement noted that some machine learning algorithms consume less power than other AI-based systems, while AI system training can be moved to remote locations and can use power from renewable sources. 

Disruptive impact

Companies that specialize in producing and training AI systems have the potential to make a positive impact on the environment by embracing renewable energy sources. Governments and regulatory bodies can encourage this shift by offering tax incentives and support to those who install renewable power systems to support their AI-based operations. Countries with strong renewable energy industries could become attractive destinations for these companies, providing the necessary infrastructure. 

The carbon emissions produced when training AI algorithms vary widely, depending on factors such as the source of electricity generation, the type of computer hardware used, and the algorithm design itself. Researchers, including those at Google, have found that it's possible to reduce these emissions significantly, sometimes by a factor between 10 and 100 times. By making thoughtful adjustments, such as leveraging renewable energy and utilizing different locations, the industry can make substantial strides in reducing its carbon footprint. 

Regulatory authorities have a role to play in ensuring that AI training projects adhere to environmental standards. If specific projects are identified as significant contributors to carbon emissions levels in their jurisdictions, authorities may enforce work stoppages until emissions are reduced. Taxes on AI centers that produce large amounts of carbon can be implemented as a deterrent, while AI companies can explore the latest developments in computational science to perform more calculations using less power.

Implications of AI training emissions 

Wider implications of AI training emissions may include:

• The prioritized development of new AI models that can more efficiently analyze data with minimal energy consumption, leading to a reduction in overall energy demands and a corresponding decrease in environmental impact.


• Companies invested in AI development partnering with renewable energy companies so that clean power infrastructure can be installed to support their operations, fostering collaboration between technology and energy sectors.


• Transferring the location of data centers to take advantage of tax incentives and avoid regulatory oversight, or relocating them to arctic locations to minimize energy spent on cooling servers, leading to new geographical hubs for technology and potential local economic growth.


• The creation of new educational programs focusing on sustainable AI development, leading to a workforce that is more skilled in balancing technological advancement with environmental responsibility.


• The emergence of international agreements and standards on AI carbon emissions, leading to a more unified global approach to managing the environmental impact of AI.


• A shift in consumer expectations towards environmentally responsible AI products and services, leading to changes in purchasing behavior and increased demand for transparency in AI energy consumption.


• The potential for job displacement in traditional energy sectors as AI companies increasingly turn to renewable energy sources, leading to labor market shifts and the need for retraining programs.


• The development of new political alliances and tensions based on renewable energy availability and AI industry needs, leading to changes in international relations and trade agreements.


• An increased focus on energy-efficient hardware design specifically tailored for AI applications, leading to technological advancements that prioritize sustainability alongside performance.


• The potential for rural areas with abundant renewable energy resources to become attractive locations for AI development, leading to demographic shifts and new opportunities for economic growth in previously underserved regions.

Questions to consider

• Do you think that regulations should be passed that stipulate that only renewable power be used when AI companies plan to train and develop deep learning AI systems? 


• Should environmentalists factor in the energy-saving benefits that result from AI system analysis (e.g., computing designs for new energy-saving materials, machinery, supply chain routing, etc.) to calculate the real/full environmental cost of AI systems?