In its annual report this year, the International Energy Agency (IEA) calculated the energy consumption associated with artificial intelligence.
By 2030, the electricity consumption of data centers around the world will double, while carbon emissions will increase by 70%. Three-quarters of this additional electricity demand will be generated by artificial intelligence technologies.
In numerical terms: Within four years, AI data centers will generate an additional 350 terawatt-hours of electricity each year – a figure that exceeds the total annual electricity consumption of the United Kingdom.
The reason why carbon emissions are increasing so rapidly is that the expansion of artificial intelligence still relies heavily on fossil fuels. Data centers are being built at a fast pace, but power grids and clean energy infrastructure are developing more slowly. As a result, gas-fired power plants are filling in the gap that arises in this process.
The report also states that since the conflicts in the Middle East have impacted the energy market, the uncertainty surrounding the global energy outlook has increased. Any fluctuation in electricity prices immediately affects the operating costs of data centers.
The news events of these past few days happen to be closely related to each other. Micron stated that memory supply will be even tighter in 2027 than it is this year; Anthropic’s prospectus mentions a commitment to provide 900.9 billion dollars’ worth of computing power; Google has ordered rockets in order to launch satellites dedicated to providing computing resources, and one of the reasons for doing so is that space offers almost uninterrupted access to solar energy.
The bottleneck in computing power is shifting from chips to energy supply.
If a chip is missing, one can still wait in line for a replacement; however, expanding the power grid often takes five to ten years. This is also the reason why large companies have been competing fiercely in acquiring nuclear power plants and land surrounding them in recent years. Whoever possesses stable and inexpensive electricity will have significantly lower training costs.
For ordinary users, this cost will not be directly reflected in their subscription fees at present. However, if electricity prices actually increase, the rate at which model prices decrease is likely to be slower than what everyone expects.