The rapid expansion of artificial intelligence and its reliance on data centres are placing unprecedented demands on the US electric power grid, according to recent reports and surveys. As AI technology continues to advance at a breakneck pace, the need for data processing power is surging, leading to a substantial increase in electricity consumption by data centres across the country.
The Electric Power Research Institute (EPRI) estimated in May that by 2030, data centres could use up to 9% of the entire electricity generated in the US. However, consultancy firm McKinsey projects an even higher figure, suggesting that data centres could account for 11% to 12% of the national electricity load, equating to around 80,000 megawatts (MW) of demand.
A survey conducted by EPRI between May and July, which included 26 utility companies, highlighted the growing burden of data centre demands. Sixty percent of these utilities reported requests for new data centre connections of 500 MW or more, while 48% had requests exceeding 1,000 MW. Alarmingly, nearly half of the utilities indicated that current requests for data connections outstrip 50% of their peak demand, although none currently handle data centre loads exceeding 500 MW.
One of the most impacted areas is Virginia, serviced by Dominion Energy, which acts as a global hub due to its high-speed fibre optic infrastructure. In 2022, data centres in Virginia utilised 2,800 MW, a figure that has almost doubled to nearly 5,000 MW, accounting for a quarter of the state's total power consumption. Moreover, Dominion Energy has 50,000 MW worth of data centre demand waiting for connections but has recently announced it will not establish new connections exceeding 100 MW for the next seven years due to capacity constraints.
Similarly, in Texas, the utility company Oncor is grappling with 59,000 MW worth of data centre connection requests. AEP Ohio, which currently serves 600 MW of data centres, has seen a surge in interconnection queries totalling over 40,000 MW.
The tech industry's advancements have been largely unpredictable, even to many insiders. AI's capabilities are evolving faster than anticipated. Historical milestones such as IBM’s Deep Blue defeating chess champion Gary Kasparov in 1997 were just the beginning. Further progress was seen in 2011 with Watson's victory in Jeopardy and in 2016, when AlphaGo, developed by DeepMind, defeated Korean Go champion Lee Sedol using strategies that appeared creative and unconventional. More recently, AI models have improved dramatically on academic tests, with ChatGPT 3.5's LSAT scores rising from the 40th to the 88th percentile in just one year, demonstrating the swift enhancement of AI capabilities.
The driving force behind these advancements is the increasing speed and computational capacity of modern chips, allowing machines to train faster and more efficiently. The next instalment in this series is set to delve into the specifics of these chips and the significant energy resources they require.
Source: Noah Wire Services