IEA Announces AI Data Center Capacity Doubled in One Year

Analysis of global data centers by the International Energy Agency (IEA) using satellite tracking technology revealed that the installed capacity of dedicated AI data centers doubled in 12 months. This aligns with the trend of major tech companies allocating the majority of their 2025 capital expenditure to data center construction, with the scale expected to exceed the total cost of NASA's Apollo program (approximately $300 billion).

Power Demand and Technological Changes

Power demand per server rack is surging. Power consumption per rack, which was approximately 13kW in 2020, is projected to increase to 600kW with chipsets launching within the next year. This level of consumption is equivalent to that of 65 households for equipment the size of a refrigerator. Hyperscale facilities now consume power equivalent to that used by approximately 100,000 households, and are estimated to account for about 3% of global electricity usage by 2030.

Regional Impact and Grid Challenges

In the United States, nearly half of the increase in power demand by 2030 is expected to come from data centers, with data center power consumption surpassing the combined total of the aluminum, steel, cement, and chemical industries. Data centers tend to concentrate near urban areas rather than near power plants, potentially consuming up to 30% of local power supply in some regions. Additionally, AI training processes exhibit extreme load volatility, with power consumption fluctuating by 400% in just 0.25 seconds, posing significant challenges to grid operations.

Efficiency and Energy Saving Potential

Text-based AI queries consume power similar to a few seconds of TV watching, but video generation and agent workloads consume hundreds to thousands of times more energy. Conversely, the IEA forecasts that utilizing small specialized models could achieve energy savings of 13.5 EJ (equivalent to Indonesia's total consumption) by 2035.