Meta builds Hyperion data center: 10 gas power plants for one AI project

Meta has announced that its new AI supercomputer data center codenamed Hyperion will be powered by ten new gas-fired power plants. The volume of energy consumed is unprecedented in the industry: the combined capacity of these facilities is comparable to the total electrical consumption of South Dakota — approximately one and a half million residents. Over the past two years, the world's largest technology companies have faced a common and fundamental problem: training and inference of large language models require such an enormous amount of electrical energy that traditional power supply schemes have become insufficient.

According to the International Energy Agency, by 2026, the combined electricity consumption of US data centers could exceed that of several small European countries. GPT-4, Gemini, Llama — each of these products conceals gigawatts of continuously consumed power behind its interface. Training one large model is comparable in energy consumption to the yearly electricity use of an entire residential district.

Hyperion is one of Meta's most ambitious responses to this challenge. The company has already announced investments exceeding $60 billion in 2025 solely for AI infrastructure development. The Hyperion data center will be one of the largest facilities of its kind in the world and the first for which the company directly orders the construction of new generating capacity — rather than simply purchasing energy from utility companies.

This is a fundamental shift: Meta takes control over the entire energy supply chain and ceases to depend on the capacity of regional networks. The choice of natural gas as the primary energy source has provoked criticism from environmental organizations and some investors. Just a few years ago, Meta consistently promoted the narrative of transitioning to renewable sources and annually reported achieving carbon neutrality in its operations.

Direct contracts for the construction of ten new gas-fired power plants send an unambiguous signal to the market: climate declarations are taking a back seat to the demands of deployment speed. The reason is pragmatic and understandable. Wind and solar capacity cannot provide stable load 24 hours a day, 7 days a week without large-scale industrial energy storage systems — a technology that is not yet ready for deployment at the required scale.

AI data centers operate under constant and uniform computational load — unlike, for example, office infrastructure with peaks during business hours. They require predictable, continuous, and powerful power supply. Natural gas, despite all its climate drawbacks, addresses this task here and now.

Competitors are moving along similar but somewhat different paths. Microsoft has reached an agreement to restart the Three Mile Island nuclear reactor in Pennsylvania. Google is actively investing in next-generation small modular nuclear reactors.

Amazon is building its own nuclear facilities near existing reactors. Meta is betting on gas — probably because it is faster and easier in terms of timeline for commissioning capacity. A nuclear program takes 10-15 years; gas turbines can be brought online in 2-3 years.

The Hyperion case exposes a fundamental contradiction within the technology industry in the AI era. Companies publicly proclaim ambitious climate goals, sign agreements on transitioning to green energy, and promise carbon neutrality by 2030. Yet the physics of supercomputers leaves no room for maneuver: energy is needed here and now, in volumes that existing renewable infrastructure simply cannot handle.

According to forecasts by Goldman Sachs analysts, by 2030, demand for electricity from data centers will grow four-fold relative to 2023 levels. In this race, the victor will be the one who can secure a reliable, scalable, and accessible energy source. For now, Meta is making the largest and most direct single bet in this game.