AI results in a large increase of data center power demand, and does have a large effect on natural resources.
In 2018, OpenAI concluded that the computing power required to train a large AI model had doubled every 3.5 months from 2012 onwards. The accuracy of results and time efficiency that can be achieved by harnessing the computing power of a vast number of computers in data centres necessitates a considerable amount of electricity. A significant proportion of data centres globally continue to rely, to some extent, on fossil fuels, resulting in a notable surge in CO₂ emissions.
In 2020, researchers at the University of Massachusetts conducted an analysis of several natural language processing (NLP) models and determined that the energy expenditure associated with training a single model resulted in CO2 emissions of approximately 300,000 kg on average (equivalent to 125 return flights from New York to Beijing). The training of ChatGPT-3 has been found to require the consumption of 1.3 gigawatt hours of electricity, resulting in the generation of 550,000 kg of CO2. It is estimated by Bloomberg that the energy consumption necessary for training is only 40% of that required for operational purposes. Moreover, the training process necessitates the consumption of approximately 700,000 litres of water for the purpose of computer cooling. This quantity of water is equivalent to that which would be required by a nuclear power plant cooling tower.
In 2023, data centres operated by Google extracted a total of 24 billion litres of water from the environment. This represents a 14% increase compared to figures recorded in the previous year. In 2022, 20 billion litres of water were employed for the purpose of cooling. Two-thirds of this quantity was comprised of potable water.
Furthermore, data from Microsoft’s facilities indicate a 34% increase in cooling water consumption during the same period. In 2024, Microsoft’s CO₂ emissions were 30% higher than in 2020, while Google’s emissions increased by 48% over the past five years.
See this blogpost by GoldmanSachs
For literature used to compose this post, see here.
Publishd in a LinkedIn-post, november 2024.
