ChatGPT uses half a liter of water per conversation

Popular artificial language models like ChatGPT or google’s Bard require massive amounts of for their training Energy. This training takes place in data centers that need to be cooled. The cooling in turn requires vast amounts of water, with mostly Drinking water is used.

4 Researchers from the University of Colorado Riverside and the University of Texas Arlington calculated this water consumption. According to her Study needed the training for GPT-3 alone 700,000 liters of water. That is so much that a cooling tower of a nuclear power plant could be filled with it. In addition, consumption is country-specific: Should data centers in Asia used for training, the water consumption would have tripled due to the warmer climate and poorer efficiency.

Cooling even during operation

The scientists are based on estimates, since the company OpenAI does not make public how long the language model was trained. However, OpenAI partner Microsoft provides insight and states that their new Supercomputer for AI development out of 10,000 graphics cards and 285,000 processor cores would exist. According to researchers, the water consumption corresponds to that for the production of 320 Teslas is needed.

But not only the training needs water, the hardware also has to be cooled during operation. The experts assume that a Conversation with 25 to 50 questions to ChatGPT around 500 milliliters of water needed.

GPT-4 needs even more water

While that doesn’t sound like much, the sheer volume of users makes ChatGPT a “thirsty” service. Water is also a precious commodity in arid regions. In addition, the researchers assume that the training of the currently most up-to-date model GPT-4 more water needed.

“The water consumption of AI models can no longer be ignored,” the researchers state in their study. “Water consumption needs to be addressed as part of efforts against the global water problems be treated as a priority.”

In their calculations, the researchers distinguished between water withdrawalin which water is taken from a river or lake and returned again and water consumption. During consumption, which is the main focus of the study, the water evaporates in cooling towers into the atmosphere.

All tech giants affected

However, the high water consumption does not only affect the data centers of Microsoft and OpenAI, other tech giants are also affected. In 2019, Google data centers in 3 US states required 8.7 billion liters of water for cooling. Drinking water is usually taken to Corrosion and bacteria prevent in the plants.

Googles Chatbot Bard and their model LaMDa should consume significantly more water than ChatGPT, according to the researchers. LaMDa alone could require millions of gallons of water since some of Google’s data centers are located in hot states like Texas. However, the researchers emphasize that these are estimates.

Global warming is affecting data centers

The cooling of data centers is becoming more and more of a challenge in times of global warming. Only last year did one Heatwave in UK Google and Oracle data centers closed, which caused some servers to shut down. Older data centers in particular are only dependent on outside temperatures 35 Grad designed.

However, it is not only the water consumption that is problematic, but also the power consumption the computer cluster. In Ireland, Microsoft and Amazon were not allowed to build their planned data centers because the state network operator Blackout feared.

The power grid has reached its limits, 11 percent of all Irish electricity needs are already going to data centers. If the expansion continued at this rate, the share would increase to a third by 2030.