Power supply is again drawing attention as a key bottleneck that could stand in the way of developing more powerful AI. As AI models become more advanced, the data centres and computing resources they need increase sharply. Power generation and transmission infrastructure is difficult to expand at the same pace.

On Aug. 20, IT outlet TechRadar revisited past remarks by OpenAI CEO Sam Altman (샘 알트먼) warning that breakthroughs in the energy sector are needed for AI progress. Altman made the remarks at the World Economic Forum in Davos, Switzerland, in January 2024.

Altman said at the time that future AI would consume far more energy than people expect and that there was no way to reach the goal without a breakthrough. He cited nuclear fusion as well as solar power and energy storage with dramatically lower costs as possible solutions. He also suggested that nuclear fission should be used more actively.

Two years later, the rise in power demand is showing up in more concrete numbers. The International Energy Agency forecast that global data centre power consumption will more than double to about 945 TWh in 2030 from about 415 TWh in 2024. That is just under 3 percent of global electricity consumption in 2030. AI is expected to be the biggest driver of the increase, with power consumption by accelerator servers mainly used for AI projected to rise about 30 percent a year on average.

That does not mean the world will soon run out of power overall. The IEA forecasts that renewables, natural gas and nuclear power will meet growing data centre demand. The issue is speed and location. Data centres can be built within 2 to 3 years, but power infrastructure such as power plants and transmission networks takes much longer from planning to completion. In some regions, delays in connecting to the power grid could limit the expansion of AI data centres.

The bottleneck is not limited to power. The outlet noted that as data centre expansion accelerates, supplies of network connection devices, optical components and memory are also emerging as factors constraining the expansion of AI infrastructure.

Ultimately, AI competition is entering a phase in which it is difficult for model performance alone to determine the outcome. How quickly power and transmission networks, semiconductors and networks can be secured to support greater computing capacity is becoming another competitive edge that will determine the pace of next-generation AI development.