Data centers for training and operating AI include millions of chips, cost hundreds of millions of dollars, and if current trends apply, they will need the equivalent of a big city power grid.
This is according to new research from researchers from Georgetown, Epoch AI and Rand, which examined the growth trajectories of AI data centers around the world from 2019 to this year. The co-authors have analyzed datasets from over 500 AI data center projects in a collective analysis and found that data centers’ computational performance is more than doubled each year, but so are power requirements and capital expenditures.
The findings illustrate the challenges involved in building the infrastructure needed to support the development of AI technology over the next decade.
Recently, around 10% of the world’s population has said they use the CHATGPT platform, but they have partnered with SoftBank and others to raise up to $500 billion to establish a network of AI data centers in the US (and possibly elsewhere). Other tech giants, including Microsoft, Google and AWS, have collectively pledged to spend hundreds of millions of dollars this year alone to expand their data center footprint.
According to a Georgetown, Epoch and Rand research, hardware costs for AI data centers like Xai’s Colossus, with a price tag of around $7 billion, increased 1.9 times each year between 2019 and 2025, while Power doubled each year. (Colossus has acquired an estimated 300 megawatts of electricity and has earned 250,000 households.)
The study also found that data centers have been much more energy efficient over the past five years, with one key metric (computational performance per watt) increasing by 1.34 times each year from 2019 to 2025. By June 2030, major AI data centers have 2 million AI chips, costing $200 billion and needing 9 GW of electricity.
It’s not a new revelation that the power demand for AI data centers is at a pace that puts a great burden on the power grid. A recent Wells Fargo analysis shows that data center energy intake is projected to increase by 20% by 2030. This could push renewable power sources that are dependent on a variety of weather. This spurs ramp up at non-renewable and environmentally damaged power sources like fossil fuels.
AI data centers pose other environmental threats, such as high water consumption, erode valuable real estate and state tax bases. A study by Good Jobs First, a nonprofit based in Washington, D.C., estimates that at least 10 states are losing more than $100 million per year in annual tax revenue to their data centers. This is the result of an overly generous incentive.
Of course, these projections may not pass. It is also possible that the time scale is off-quilter. Some hyperscalers like AWS and Microsoft have returned to data center projects in the past few weeks. In a note to investors in mid-April, Cowen analysts observed that there was “cooling” in the data center market in early 2025, demonstrating the industry’s fear of unsustainable expansion.