AI is already on a tight grid around the world, but according to a new report, we are just getting started.
According to the latest forecast from the International Energy Agency (IEA), by 2030, artificial intelligence data centers will almost swallow up electricity throughout Japan.
Today’s data centers have consumed 1.5% of the world’s electricity – about 415 Terawatt hours per year. The IEA expects that this will increase to nearly 950 TWH by 2030, accounting for nearly 3% of global electricity.
The dedicated hardware that runs AI systems is the real consumer. The power demand for these “accelerated servers” will increase by 30% annually in 2030, while regular servers will grow at less than 9% annually.
Some data centers that are already under construction will consume up to 2 million average home power, while others have announced future consumption of up to 5 million or more.
Very unbalanced distribution
By 2030, data centers in the U.S. will consume about 1,200 kilowatt-hours (kWh) per person — about 10% of the entire U.S. household use in a year and “an order of magnitude higher than any other part of the world.” Meanwhile, each person in Africa will reach almost 2 kWh.
On the area, some areas have felt the squeeze. In Ireland, data centers now consume 20% of the country’s incredible electricity. Six U.S. states use 10% of their power to data centers, with Virginia leading by 25%.
Can cleaning energy keep up?
Despite concerns that AI’s appetite may effectively undermine climate goals, the IEA believes these concerns are “exaggerated”.
By 2030, nearly half of the additional electricity required for data centers should come from renewable energy sources, although fossil fuels will still play a leading role.
The energy mixture varies with the region. In China, coal’s data centers today account for 70% of data centers. In the United States, the rate of natural gas is 40%, followed by renewable energy at 24%.
Looking ahead, small modular nuclear reactors (SMRs) are crucial to satisfying the power consumption of AI after 2030.
Technology companies such as OpenAI have already planned to fund SMR capacity over 20 GW – they are considering the mark of long-term energy security. Microsoft even wants to revive the resolved Three Mile Island nuclear power plant.
Efficiency and expansion
The IEA speculates several possible futures of AI energy consumption. Global data center power could exceed 1,700 TWH by 2035, nearly 45% higher than its base forecast, assuming the “lift” of AI adoption.
Additionally, their “high efficiency” schemes show that software, hardware and infrastructure improvements can reduce power demand by more than 15%, while delivering the same AI capacity and performance. However, if AI encounters problems, “headwind”, then energy consumption may be greatly reduced.
The main point of the IEA is that the next decade will test the careful balance of AI between power and energy efficiency.
Whether the tech industry can solve this problem will not only affect the future of AI, but also its role in solving the global climate crisis rather than worsening.
