A preliminary US Energy Department report found that at least half of the growth in power demand from US datacenters since 2016 has come from AI.
The country's datacenters are expected to consume about 375 TWh in 2024, compared to less than 100 TWh in 2016, according to yet-to-be-published data shared with industry members on April 18. AI computing alone in 2023 used more than 100 TWh, according to the data.
The 375-TWh figure equates to about 9% of total annual electricity consumption in the US, according to the US Energy Information Administration's projection for 2024 in its April 9 "Short-Term Energy Outlook."
"The primary growth is coming from the growth in AI," said Arman Shehabi, a staff scientist at the DOE's Lawrence Berkeley National Laboratory and co-author of the report. "We don't know exactly how much AI is going to increase electricity use, but we do know it's going to increase and it's going to increase pretty quickly."
The US Energy Act of 2020 directed the DOE to update a 2016 study on the energy usage of datacenters. Though Berkeley Lab is still collecting data for the final version, the report's co-authors gave attendees of a datacenter conference in Washington, DC, a preview of their findings.
According to those early estimates, US datacenters consumed about 275 TWh in 2022 and could consume as much as 500 TWh in 2026. The annual totals represent the combined demand of all datacenter servers, storage, infrastructure and networks in the US.
However, "there's a lot of uncertainty in there," Shehabi said. The scientist said any projection past 2022 could be off by 100 TWh or more. A recent International Energy Agency report projected much lower levels of datacenter power consumption in the US, from 200 TWh in 2022 to almost 260 TWh in 2026.
One source of uncertainty is the speed of AI deployment, including the time and resources needed to build new AI processors and the speed of AI adoption.
"What's the value of AI?" Shehabi asked. "How much is this going to get picked up? How well will it be monetized? Who's going to be using it? Is there going to be that growth? We don't know that yet."
Preparedness of the power grid
The surge in computing is one of many forces expected to drive grid load in the coming years, along with electric vehicle adoption, building electrification and new manufacturing technologies. Meanwhile, consumers and state lawmakers are pressing US utilities to retire their fossil fuel generation.
Despite those added constraints, the researchers cast doubt on the fear that datacenters will overload the power grid.
In regions such as the San Francisco Bay Area, datacenter developers may face long waits for utility interconnection, said Sarah Smith, a research scientist at Berkeley Lab and fellow report co-author.
"But on the national scale, it would surprise me if that power really could not be found," Smith added.
The difference between datacenters and other drivers of power demand is that "it's going to happen first," Shehabi said.
However, innovation could manage that demand or allow datacenters to cut down on energy usage. In the early 2010s, the industry shifted to cloud computing and flattened its electricity consumption, the researchers found.
"What will be the big technology shift or industry shift in the '20s? We're not sure yet," Shehabi said. "But there's a lot of opportunities here in terms of how that electricity use could be met."