• AI could use nearly half of data centre power by late 2025.
• AI chip demand and training needs are straining power grids and climate goals.

AI systems may soon be responsible for nearly half of all data centre electricity use. That’s the view of Alex de Vries-Gao, founder of the tech sustainability website Digiconomist, whose latest research estimates AI could take up around 49% of global data centre energy demand by the end of 2025.

As reported by The Guardian, his analysis, set to be published in the journal Joule, builds on figures from the International Energy Agency (IEA), which last year reported that data centres consumed 415 terawatt hours of electricity—excluding energy used for crypto mining. That’s already more than the annual power usage of some countries.

De Vries-Gao believes AI systems may have already been using around 20% of that total in 2023. His estimates are based on the energy draw of AI-focused chips made by Nvidia, AMD, and others like Broadcom. The calculations also consider how much electricity is needed to cool servers and run the rest of the infrastructure.

By the end of this year, AI’s power demand could hit 23 gigawatts. That’s about twice the electricity use of the Netherlands. These figures come as the IEA forecasts that AI alone could consume as much energy by the end of the decade as Japan does today.

Data centres are at the heart of AI technology. Their energy needs have raised growing concerns about sustainability, especially as more businesses and governments increase their AI investments. The growing use of tools like ChatGPT and other language models means more servers, more chips, and more power.

Still, de Vries-Gao points out that future demand may not keep rising at the same pace. For one, public interest in some AI applications could decline. Other risks include limits on the supply of chips due to geopolitical tensions, such as export restrictions that have made it harder for China to access high-end AI hardware.

In some cases, these restrictions have pushed developers to find workarounds. De Vries-Gao points to the recent release of the DeepSeek R1 model in China, which reportedly used fewer chips. Efforts like these, along with better chip design and more efficient code, may help reduce the energy cost of AI systems.

“New designs can lower both the compute and energy needs of AI,” said de Vries-Gao. But he added that better efficiency might also lead to broader adoption, pushing demand back up.

The growing trend of “sovereign AI” — where countries develop their own models instead of relying on outside vendors — may also fuel further demand for hardware. Meanwhile, some firms are turning to fossil fuels to meet their power needs. De Vries-Gao noted that Crusoe Energy, a US-based data centre startup, recently secured 4.5 gigawatts of gas-fired power for its operations. One of its possible partners is OpenAI, through a joint venture called Stargate.

“There are early signs that projects like Stargate could increase reliance on fossil fuels,” de Vries-Gao wrote.

OpenAI has since announced that a new Stargate site will be built in the United Arab Emirates — its first location outside the United States.

Both Microsoft and Google have acknowledged that the growth of AI is making it harder to stick to their environmental goals. And according to de Vries-Gao, getting clear data on AI’s energy use is becoming more difficult. He described it as an “opaque industry,” where much of the key information remains hidden.

The EU AI Act does require companies to report how much energy is used to train models, but not to operate them once deployed.

Professor Adam Sobey from the UK’s Alan Turing Institute said more transparency is needed — not just to track AI’s environmental impact, but also to measure whether AI can help cut emissions in other sectors.

“We probably don’t need a lot of successful AI use cases to balance out the energy used up front,” Sobey said. “But we need to be able to measure it properly.”