Big models, big bills

Today data centers around the world consume up to 1.5% of electricity, and by 2030 consumption will double.

The French antitrust authority has released a Report on competition issues related to the impact of AI on energy and the environment. According to the IEA, while a typical data center has a capacity of 10–25 MW, an AI-focused data center has a capacity of over 100 MW. This is comparable to the annual electricity consumption of 100,000 households.

It is evident that AI has an impact on the environment. Energy for data centers is often produced from fossil fuels. However, there is a trend toward investing in decarbonized energy sources, such as renewable energy and nuclear power, especially in areas with high electricity demand. For example, in 2024, Microsoft agreed with Brookfield to supply 10.5 GW of green power in the US and Europe. Tech giants Amazon, Google, and Oracle have announced the introduction of small modular reactors.

The EU AI Act (Regulation 2024/1689) already sets the task of developing codes of conduct to minimize the environmental impact of AI systems.

When it comes to competition, the French antitrust authority highlights three issues.

Difficulties represent t he first issue in accessing power grids and uncertainty about energy prices (energy costs account

of water consumption occurs during training and AI inference for 30–50% of a data center's operating costs).

The growth of data centers has led to an increase in the number of applications for technological connection to high-power grids. The risk of network overload in areas of high demand has increased: data centers are often organized as clusters with redundancy (i.e., the installation of several technical resources designed to replace each other in the event of a failure). There is also a risk that large players will “take over” profitable sites to the detriment of smaller ones.

In France, in particular, a number of measures have been taken. For example, an accelerated procedure has been introduced for connecting energyintensive industrial consumers (0.4–1 GW). The state selects several areas in advance where it is possible to connect large consumers and reserves capacity there for potential projects so that investors can then come in and quickly connect to the electricity grid. There is also the problem of some energy consumers having “predatory” strategies. For example, enormous capacity is reserved in the grid so that competitors cannot obtain this capacity, but in fact it is not used, which creates an artificial shortage. In this regard, a rule has been introduced: it is not possible to reserve capacity without confirming the rights to use the land on which the project is planned to be located. And if the requested capacity is not used, it can be reduced. The idea of “dynamic” capacity allocation is being discussed: capacity is reserved not for those who applied first, but for those who build and commission the facility faster.

The second issue is related to the emergence of the concept of “frugal AI”— the prioritization of solutions that minimize material and energy costs and environmental footprints to assess the “necessity” of AI use (i.e., AI solutions are only used when they are indispensable), and which are optimized throughout the entire chain (development, implementation, use) to minimize resources. In fact, the question has arisen of developing smaller AI models that help to “save” on computing power. For example, open-source helps: if a model has already been trained, it can be reused, saving resources, rather than having to train it from scratch every time.

In this regard, the state and companies are beginning to introduce “green” parameters into procurement and tenders. This creates the risk that companies may overstate the environmental benefits of their AI solutions and data centers due to the lack of scientifically sound calculation methods, gaining an undeserved advantage in procurement. Also, large operators may refuse to disclose data on their environmental footprint and frugality, reducing transparency.

The third issue is the need to develop uniform methods for assessing the environmental impact of AI. Currently, there are many different methods that are not always sufficiently scientifically sound. Alternatively, the standard may be developed by major players who lobby for their own interests without offering environmental improvements.

In Russia, the Ministry of Digital Development, Communications and Mass Media projects a 2.5-fold increase in energy consumption by Russian data centers by 2030, which may also drive up electricity prices amid energy shortages in Russia.

From the monitoring issue No. 12 (24), December 2025. Download the full issue (PDF) · issue page at the Gaidar Institute