According to The Washington Post, a new study from the University of California, Riverside in United States reveals the staggering environmental costs of generative AI, particularly its water consumption. Studies have found that AI needs a lot of water to cool down servers, even just by generating text. This, combined with the enormous pressure it places on the power grid, further highlights the environmental challenges of AI development.
The study notes that water usage for AI varies by state and data center location, but overall, water consumption is inversely proportional to electricity bills. IT House notes that Texas has the lowest water consumption, requiring only 235 milliliters of water to generate a 100-word email, while Washington has a whopping 1,408 milliliters, which is equivalent to three 16.9-ounce bottles of mineral water.
In addition, data centers themselves are large consumers of water and electricity, which also drives up the utility bills for residents in their areas. For example, meta trained its LLaMA-3 model using 22 million liters of water, equivalent to the amount of water needed to grow 4,439 pounds (about 2,013.5 kilograms) of rice, or the annual water consumption of 164 United States. This is by no means a small number, especially considering that this is only a very light use case for GPT-4's target audience.
The cost of electricity for GPT-4 is also very high, and if one in 10 United States office workers used GPT-4 once a week, over the course of a year, their electricity demand would be equivalent to the electricity consumption of the entire Washington, D.C. (about 671,803 people) for 20 days.
Representatives of OpenAI, meta, Google, and Microsoft responded to this, mostly reaffirming their commitment to reducing environmental requirements without giving specific courses of action. Microsoft representative Craig Cincotta said the company is committed to developing a cooling method that does not consume water at all, but did not provide specific details.