Artificial intelligence is already an environmental problem

The era of generative artificial intelligence (AI) is changing the world, both figuratively and literally. The energy and water consumption of large technology companies, the main developers of this technology, as well as their carbon emissions, have skyrocketed in recent years. And projections show that the trend will not change. Although no company officially says that this increase is due to the emergence of AI, the numbers show a significant jump in 2022, the year in which OpenAI launched ChatGPT and inaugurated the generative AI race.

Seven of the 10 largest companies in the world by market capitalization are technological, which gives an idea of the importance of the sector. Large industries have large resource needs. All in all, the data shows an important jump. Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, Riverside and a specialist in AI sustainability, believes it is safe to infer that AI is responsible for this escalation in pollution and resource consumption. The increase in the last two years, he maintains, has been very large and coincides in time with a strong investment in generative AI and other services related to AI.

More energy

The latest figures available from Google and Microsoft, the main developers of this technology, reflect large growth for the second consecutive year in the three key magnitudes. Google, responsible for the Gemini model, has reported a 16.2% increase in energy consumption in 2023 compared to the previous year. Microsoft, owner of Copilot and which has lent its infrastructure to OpenAI to develop all versions of ChatGPT and the Dall-E image generator, has recorded a growth of 28.7%, as reflected in its annual sustainability report. The company founded by Bill Gates has doubled its energy needs between 2020 and 2023, going from 11.2 million megawatt-hours (MWh) to 24 million MWh. Almost the same thing has happened at Google, with an increase of 67% in this period.

GPU processors, those used in training AI models, are much more powerful than CPUs, until now predominant in data centers, and therefore consume more energy (up to 10 times more). Training large language models requires tens of thousands of GPUs operating day and night for weeks or months. The most advanced models are periodically retrained to incorporate updated data and every time a user types a prompt on their mobile device or computer, the response is computed in a data center. All this activity has stretched energy demand, to the point that some companies, aware that the trend will continue to rise for some time, are studying developing small nuclear power plants to ensure a sufficient and stable supply.

More water

The data centers in which AI (and all digital activity) is operated are large industrial warehouses populated with rows and rows of racks, processors arranged in the shape of a cabinet or refrigerator. All those processors and servers, which host our data and run online programs, run day and night. That activity emits a lot of heat; if the temperature is not controlled, the equipment can break down.

Water is used to cool data centers and is sprayed to cool the environment. The consumption of this resource has also recorded increases of 13.8% and 21% in 2023, respectively, figures similar to those of the previous year. Microsoft, for example, has reported using almost 13 billion liters of water. More than half of that volume (about 8 billion liters) was evaporated or consumed, so it could not be reused. Google, for its part, needed less water, about 8.6 billion liters, but only returned 26.6% of that amount to the system.

These figures, however, do not provide a complete picture of the real consumption of AI developers. The companies only provide data on the water they use to cool the data centers, but do not include in their reports either the water used to generate the electricity they consume or the water used in the supply chain of the products (mainly in the manufacturing of chips and other hardware), as is the case, for example, with carbon emissions.

“Companies intentionally hide this information,” says Ren. “That's why it's very telling that Apple accidentally said in its latest environmental performance report that its indirect water consumption due to the supply chain represents 99% of its total water footprint.” Based on Apple's direct water consumption data, Ren concludes, that would imply that Apple's actual consumption in 2023 was at least 300 billion liters. “That volume of water is enough to irrigate 0.1% of the wheat harvested annually around the world,” he illustrates.

More emissions

Regarding carbon emissions, Google's have grown by 13% and Microsoft's have grown by 3.8% in the last year. The increase is 67% and 40%, respectively, if the last four years are observed.

According to Ren, most of the pollution emitted by these companies has to do with their supply chain. “The main driver of the increase in global carbon emissions is that associated with the manufacturing of AI chips and the construction of data centers,” he explains.

While the energy efficiency of the hardware used to develop and run AI has increased in recent years and will continue to do so in the coming years, the researcher notes, “it is very unlikely that embodied carbon (the amount of carbon emissions associated with the extraction, production, transportation and manufacturing stages of a product’s life) will decrease to short term due to increased demand for AI hardware.

The race for AI

From Google to Microsoft, Meta or Amazon (which have not yet published their environmental reports for this year) and Apple, all the big technology companies are immersed in programs to improve their carbon emissions records and reduce the amount of water they use. The goal for many of them is to reach 2030 with a very low environmental footprint.

In an article that has just been accepted in the journal Communications of the ACM, a reference in the computing sector, Ren and his colleagues presented projections based on current consumption and trends in the sector. Global demand from AI will be responsible for the use of between 4.2 and 6.6 trillion liters by 2027, the equivalent of half the water used each year in the United Kingdom. That same year, the energy demand of AI will be between 85 and 134 TWh. In comparative terms, global battery production in 2023 stood at around one terawatt hour (1 TWh).

“If we only look at the emissions derived from their direct energy and water consumption, they can achieve no emissions or not use more water than they contribute by 2030, perhaps even sooner,” concludes Ren. “But if we look at their real footprint, it is quite unlikely that they will achieve neutrality by 2030.”

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition