Data centers to train and run AI may soon have millions of chips, spend hundreds of billions of dollars, and if the current trends are made, a large city power grid is required equal to electricity.
According to A new study Georgetown, Epoch AI and Rand researchers who saw the development trajectory of AI data centers around the world from 2019 to this year. Co-writers compiled and analyzed a dataset of over 500 AI data center projects and found that, while the computational performance of data centers is more than doubled annually, so there are power requirements and capital expenses.
The conclusions describe the challenge in the creation of the infrastructure required to support the development of AI technologies in the coming decade.
Openai, who recently said that About 10% of the world’s population Using its CHATGPT platform, a partnership with SoftBank and others to increase a network of AI data centers in the US to increase to $ 500 billion (and possibly else). Other technical giants including Microsoft, GoogleAnd AWS has collectively promised to spend millions of dollars alone this year, which is expanding the footprints of their data center alone.
According to Georgetown, APOCH and Rand Study, the cost of hardware for AI data centers such as XE’s colossus, which costs around $ 7 billion, increased by 1.9x each year between 2019 and 2025, while electricity requirements rise 2x annually in the same period. (Colossus draws an estimated 300 MW power, as much as 250,000 homes.)
The study also found that the data centers have become very energy efficient in the last five years, with a major metric – per watt per computational performance – growing by 1.34x each year from 2019 to 2025. Nevertheless, these improvements will not be enough for increasing power needs. By June 2030, the major AI data center can have 2 million AI chips, the cost is $ 200 billion, and 9 GW power is required – broadly produced nine atomic reactors.
This is not a new revelation that AI data centers are demanding electricity To put the power grid under great stressData center energy intake is estimated to increase by 20% by 2030, According to a recent Wales Fargo analysisIt can push the renewable sources of power, which are dependent on the weather, to their extent – A ramp-up In non-renewable, environmentally harmful power sources such as fossil fuels.
AI data centers also give other poses Environmental threatsSuch as high water consumption, and valuable immovable property, as well as separating state tax targets. Good Jobs First, a Washington, DC-based non-profit, a study, Estimate Data centers in at least 10 state tax revenue lose more than $ 100 million in tax revenue, the result of extremely liberal incentives.
It is possible that these estimates may not be passed, or that the scales of time are off-cilters. Some hypersscalers, such as AWS and MicrosoftData centers have been pulled back on projects in the last several weeks. In the middle of April, investors in a note, Analyst in Cowen It was observed that in early 2025 the data center is “cooling” in the market, which reflects the fear of uncertain expansion of the industry.
