A llama spits on a zookeepr at the London Zoo.
Joe Maher/Getty Images
Meta just released a big new AI model called Llama 2.
Training this model took a lot of water, according to researcher Shaolei Ren.
Meta and other tech giants are planning new data centers in arid and hot Arizona.
The Rocky Mountain Llama and Alpaca Association says llamas don’t need to drink much water.
Meta’s AI version, however, is damn thirsty.
Earlier this week, Meta released a big new AI model called Llama 2. People in the tech industry are excited. Except for maybe Shaolei Ren.
He’s an associate professor at UC Riverside who researches how much water is used when data centers are fired up to train these giant new AI models. He took a look at Llama 2 and is worried.
The total water footprint for training Llama 2 is 10.9 million liters, or 2.8 million liters if hydropower is excluded, Ren estimated. (For context, you’re supposed to drink about 3 liters of water per day).
Meta didn’t disclose how much water it used to train this AI model. However, the company did disclose power consumption. Ren took that and looked at how efficient Meta data centers are when it comes to using energy and water.
That yielded his estimate, which he said is almost double the water footprint of Meta’s previous big AI model, Llama 1, which came out earlier this year.
Data centers already used a lot of energy. The generative AI boom is expected to increase this consumption massively.
“If the power usage increases, everything else like carbon and water footprints will also increase,” Ren told Insider. A Meta spokesperson didn’t respond to a request for comment on Ren’s estimates.
Meta’s data centers used just over 5 million cubic meters of water in 2021. That’s about 1.33 billion gallons. The company has yet to share 2022 data.
By the end of this year, or maybe next year, Meta will have finished building and expanding a data center in Arizona. The state is running out of water.