Output beats environment: One in four companies checks AI's energy consumption

27 percent of companies check the power consumption of an AI model before making a selection. For more than half of companies, performance is more important than measuring the environmental impact of artificial intelligence. This is the result of a study by IT consulting firm Capgemini. At the same time, 48 percent of managers believe that AI has increased their company's greenhouse gas emissions. Companies that measure their environmental impact expect an increase of 85 percent over the next two years.

Environmental impact only secondary

When choosing AI models, companies focus primarily on performance. More than three quarters rank it among their top five selection criteria, just ahead of scalability. Cybersecurity, costs, and efficiency are also important to companies, with more than half of them ranking them in the top 5. In contrast, only 20 percent ranked ecological impact. In contrast, 38 percent of the managers surveyed are aware of the environmental impact of using AI.

So far, 42 percent of decision-makers have had to reconsider the environmental goals in their companies. Nevertheless, only around one in ten companies is currently examining the environmental impact of using artificial intelligence. However, more than 80 percent of managers state that they intend to carry out such investigations in the future. Around half of them plan to do so in the next 12 months, while the other half expect it to be 24 months before the tests are introduced.

Three out of four respondents point to the lack of transparency on the part of AI manufacturers with regard to the ecological consequences. However, two thirds also see themselves as responsible and recognize a lack of awareness at management level. They also point out that it is difficult to measure the environmental impact. An emissions calculator for AI models could help them with this. Nevertheless, a third of companies are already planning to integrate sustainability measures into the AI life cycle. Half are consciously focusing on smaller language models and renewable energies or are planning to do so in the coming year.

AI will double the power consumption of data centers

Compared to a conventional Google search, the power consumption of queries in ChatGPT 4 is ten times higher. According to the Capgemini study, OpenAI required between 51 and 62 gigawatt hours of electrical energy to train the language model. In the case of Google's machine learning program, the authors put the ratio at 60 percent consumption for inference and 40 percent for training the AI.

With reference to figures from the International Energy Agency, the researchers expect the global electricity demand for operating data centers to more than double from 460 terawatt hours in 2022 to around 1,000 terawatt hours by next year. Water consumption for IT infrastructure in the Data Center Valley in the US state of Virginia increased by 69 percent between 2019 and 2023. Around one liter of water is needed for cooling for every 40 to 100 requests.

For the study, Capgemini surveyed 2,000 managers from companies in 15 countries with an annual turnover of more than one billion US dollars that use artificial intelligence. The authors selected executives at director level and above who work in technology, innovation or in corporate functions such as finance, sales, or marketing.

(sfe)