ChatGPT can use a pint of water for five tasks – E24

In 2022, Microsoft used 6.4 billion liters of water, more than a third more than the previous year. The big talk this year – artificial intelligence – is behind this increase.

The costs behind artificial intelligence (AI) tools like ChatGPT can be difficult to measure.

But the only thing Microsoft-backed OpenAI needed was massive amounts of water. They got it from the Raccoon and Des Moines Rivers in the central part of the US state of Iowa.

The water was needed to cool a powerful supercomputer that taught artificial intelligence systems how to imitate human language. Tech giants such as Microsoft, OpenAI and Google have admitted that the huge increase in demand for so-called generative AI has brought with it huge costs. This applies to everything from expensive semiconductors to significant increases in water consumption.

Developing such generative AI – a language model that can be used to create new content – ​​therefore requires analyzing patterns in massive amounts of text. It is a very complex computational process that requires large amounts of electricity and generates large amounts of heat. On hot days, these data centers must be cooled by a large amount of water pumped into cooling towers.

0.1 liter can be used for one question

Microsoft reveals its latest releases Environmental report The group’s global water consumption rose by 34 percent from 2021 to 2022. In total, consumption stands at 6.4 billion liters – the equivalent of nearly 2,500 50-meter swimming pools.

– It’s fair to say that most of the increase is due to artificial intelligence, including significant investments in generative AI and cooperation with OpenAI, says researcher Shaoli Ren.

Later this year, he and other researchers will present a report that estimates that ChatGPT swallows a pint of water in between five and 50 user questions — depending on server location and time of year. This would mean a water consumption of 0.1 liters per question.

Google was thirsty too

The estimate includes indirect water consumption that companies themselves do not measure – for example to cool power plants that supply electricity to data centres.

– Most people are not aware of the resource usage behind ChatGPT. If people aren’t aware of the problem, we won’t be able to try to provide resources as well, Ren says.

Google Water consumption increased by 20 percent in the same period. This is also largely due to the expansion of artificial intelligence, according to Ren. There were also significant local differences here: Google previously received a lot of attention for its water consumption in Oregon, but here no significant increase was recorded. On the other hand, consumption doubled in the desert city of Las Vegas. And in Iowa, where Microsoft also has large data centers, Google was at its most intense.

Promises of recovery

Resource utilization in data centers is not an unknown issue. Many people are familiar with the calls to delete old emails so they are not left in the cloud. China cracked down on cryptocurrency mining in 2021 after commercial players were found to be abusing schools, community centers and other public institutions that receive subsidized electricity.

However, with generative AI growing rapidly and being able to create new content on its own, this takes the problem to a new level. Technology companies themselves are aware of the problem.

Microsoft told the AP that it is investing in research that measures the climate and energy footprint of artificial intelligence. In addition, the giant says they are working to make systems more energy efficient and are investing in renewable energy.

Collaboration partner OpenAI also acknowledges that training AI systems can lead to a significant rise in water and energy consumption, and the company is working to simplify this.

Cold conditions are better

Microsoft and OpenAI have opened a large number of their data centers in West Des Moines, Iowa, a flat Midwestern state known for its endless cornfields. And in many ways, it’s a relatively efficient place to build these AI systems. By comparison, Microsoft’s data centers in desert Arizona use much more water.

– If you train AI models at Microsoft, you must add them to Iowa, not Arizona. It doesn’t make a difference in training, but it does make a big difference in water and energy consumption, Ren says.

For most of the year, data centers only need air for cooling in Iowa. Only when the temperature exceeds 29 degrees Celsius does Microsoft have to draw additional water, according to the company itself.

Emphasizes in

However, in the summer, it can get very hot in Iowa. In July 2022, there was still one month left until OpenAI finished training GPT-4, which would later become publicly available ChatGPT. Microsoft then pumped 43.5 million liters of water to its data centers in Iowa, according to numbers from waterworks in West Des Moines.

This alone represented 6 percent of the water stations’ consumption that month, and in the same year the water stations and the municipality announced restrictions.

They explained in one document that they would not consider permitting new data center projects unless Microsoft adopted technology that significantly reduces water consumption from current peak levels.