By Betty Huang –
Behind every online search, video stream, or use of AI is massive data centers, quietly consuming huge amounts of electricity and water to keep the digital world running. Data center energy consumption is expected to nearly double from 448 terawatt hours (TWh) in 2025 to 980 TWh by 2030, driven almost entirely by the rapid industrialization of AI. The adoption of AI optimized servers is projected to represent 44% of data center power consumption by 2030.
The level of growth has put serious strain on both energy and water supplies, in ways that barely existed just five years ago, before the rise in AI. A mid-sized data center is estimated to use as much water as a small town, while the largest facilities consume up to 5 million gallons per day, comparable to a city of about 50,000 people. Most of this demand comes from cooling systems required to manage on-site power generation equipment.
Geographically, we see major clusters of data centers concentrated in Northern Virginia, Texas, California, Western Europe, and parts of East Asia. Larger populations do not always drive data center development, instead, location decisions come down to:
- Access to cheap, reliable electricity grids
- Water availability for cooling
- Proximity to fiber networks and subsea cables
- Favorable zoning laws and tax incentives
Northern Virginia alone handles an estimated 70% of the world’s internet traffic. This is largely due to cheap power, dense fiber infrastructure, and proximity to major federal and enterprise clients. The Pentagon, the National Security Agency, the Central Intelligence Agency, and dozens of civilian agencies are headquartered in the DC metro area, with their cloud and data contracts flowing directly to facilities just across the border in Virginia.
Much of the recent demand surge can be attributed to an increase in AI training. This process requires thousands of high-performance chips to run continuously for weeks or months, consuming a substantial amount of electricity. A review from MIT Technology Review found that AI inference makes up as much as 80-90% of computing power used for AI. AI inference is a process where a trained model acts on new data to generate a response. A standard Google search uses roughly 0.3 watt-hours of electricity. A query to a large language model (LLM) like those from OpenAI, Google, or Anthropic uses about 2.9 watt-hours, nearly 10 times more.
The same companies driving this demand, including Google, Microsoft, Amazon, and Meta, made public commitments to carbon neutrality between 2019 and 2020. Since then, total emissions have only gone up, with Google’s up nearly 50%, Meta’s up more than 60%, Amazon’s up 33%, and Microsoft’s up more than 23%, according to a recent Fortune article.
Map: National Renewable Energy Laboratory (NREL)
Source: Gartner, Pew Research Center, Brookings, MIT Technology Review, Visual Capitalist, Fortune