AI Environmental Impact: Water and Carbon Use

A new study published in the journal “Patterns” has brought to light the substantial environmental consequences of artificial intelligence (AI), suggesting that the training and operation of systems like ChatGPT may result in more annual carbon emissions than New York City and more water consumption than all the bottled water consumed globally. This research highlights the escalating power usage of AI and its respective carbon and water footprints, which are reportedly higher than previously estimated.

According to the study’s author, Alex de Vries-Gao, a PhD candidate at Vrije Universiteit Amsterdam, the numbers are staggering. The report indicates that major AI developers, including OpenAI and Google, provide little information regarding the energy consumption and environmental impacts associated with their technologies. This lack of transparency has prompted researchers to rely on fragmented information to develop estimates.

Previous efforts to understand AI’s environmental impact have focused on the general energy consumed by data centers. However, de Vries-Gao distinguishes AI’s specific energy use and its environmental repercussions from the broader activities of data centers, which also support operations like email routing and document processing.

De Vries-Gao’s findings build on his previous research in June published in the journal “Joule,” which estimated the power consumption of AI by analyzing the energy usage of specialized AI chips by industry leaders like Nvidia and AMD. By multiplying the energy each chip uses by the estimated production figures, he determined that AI systems required between 5.3 and 9.4 gigawatts of energy to operate by the end of last year. Predictions suggest that this need could grow to 23 gigawatts by the end of the current year.

This rising energy demand translates to a staggering consumption rate of 201.5 terawatt-hours annually, positioning AI’s energy use comparable to that of entire countries, ranking it 25th worldwide, just behind Egypt.

The energy demands of AI systems inevitably have environmental repercussions, primarily due to the reliance on fossil fuels for electricity generation. Such practices contribute significantly to greenhouse gas emissions. Additionally, the cooling requirements for the data processing of AI systems mean further water evaporation, with the water not being reused in the local ecosystem.

In assessing the disclosures provided by major tech companies like Google, Meta, Amazon, and Apple, de Vries-Gao found inconsistencies and gaps. For instance, while Apple provides details on its data center energy consumption, it lacks transparency regarding how much carbon is emitted from the plants supplying that power. Conversely, Meta discloses power plant emissions but is reticent about water used therein. Amazon, on the other hand, offers limited information on the environmental impacts of its operations.

Using available data from these companies, de Vries-Gao averaged their disclosed figures to estimate carbon emissions and water usage per watt of electricity consumed. His calculations indicated that American tech companies generated between 0.32 and 0.35 tons of carbon dioxide equivalents per megawatt-hour, with companies in China being significantly more carbon-intensive.

Overall, examining the publicly available data corroborated estimates from the International Energy Agency, which indicated global data centers contribute roughly 0.4 tons of carbon emissions per megawatt-hour. De Vries-Gao’s extensive analysis concluded that AI activities generate between 32.6 million and 79.7 million tons of carbon dioxide equivalents annually—surpassing the total emissions produced by New York City.

De Vries-Gao also scrutinized water consumption associated with AI’s electrical demand. He estimated AI systems consume between 312.5 billion and 764.6 billion liters of water each year, significantly exceeding the estimated 446 billion liters of bottled water consumed globally every year. This suggests that previous forecasts underestimated the water demands of AI technology.

Ultimately, the critical question surrounding AI’s impact remains whether the benefits derived from the technology justify its significant environmental costs. While both de Vries-Gao and fellow researcher Shaolei Ren agree that the current lack of comprehensive data hampers this assessment, they highlight an urgent need for improved transparency from tech entities. Until such information is adequately available, making informed decisions regarding the societal implications of AI remains elusive.