Do you know the environmental cost of GenAI and how to mitigate against it? 

1. How does the amount of CO2 generated in training a single AI model compare to the carbon emissions produced by the average passenger car over its lifetime?

A. It is approximately one-fifth the amount.

B. It is roughly equivalent.

C. It is five times the amount.

Answer: C. According to one recent US study, getting one GenAI model up and running for a business generates an estimated 284,000kg (626,155lbs) of CO2 emissions – almost five times the emissions produced by the average US car over its entire lifetime.

2. Every ChatGPT query consumes energy in the form of electricity. How does this compare to a simple web search?

A. It is approximately one-fifth the amount.

B. It is roughly equivalent.

C. It is five times the amount.

Answer: C. Significant energy consumption of GenAI models does not end following training. According to a recent MIT report, a ChatGPT query – which could be a basic request to summarize an email – consumes about five times more electricity than a simple web search.

3. Most large-scale AI deployments are housed in data centres, with computers that consume vast quantities of raw materials. How many kilos of raw material go into building a 2kg computer?

A. 10 times the amount of the finished product.

B. 100 times the amount of the finished product.

C. 400 times the amount of the finished product.

Answer: C. According to a UNEP report, making a 2kg computer uses a “staggering” 800kg of raw materials. Added to the environmental cost is the microchips that power AI hardware, which need rare earth elements that are often mined in environmentally destructive ways.

Ways to mitigate your GenAI energy consumption

• Rather than build your own models from scratch, try to make more use of foundational large language models (LLMs). This will require far less data, and therefore far less energy, reducing emissions accordingly.
• You should also try to use smaller models where appropriate. Not every LLM is better because it is bigger – smaller models trained on carefully curated data may deliver better results while operating far more efficiently.
• Pay attention to location. Edge computing enables businesses to lower their energy use by processing data in more places that are perhaps closer to the business, reducing travel time, or in a data center that has access to renewable energy.
• Put the right infrastructure in place. AI models running on processors specifically developed for the purpose will use much less energy.
• As a CEO, you may not be best equipped to oversee these interventions, but you can challenge your operations team on what they are doing in these areas, and task your CTO with developing metrics to gauge the results.