The scale of the artificial intelligence industry's energy appetite is becoming increasingly tangible. According to my analysis of the latest data from Global Energy Monitor (GEM), the announced capacity of gas-fired power plants aimed at directly powering data centers in the U.S. has surged from 97 GW at the end of 2025 to an impressive 189 GW by mid-2026. For context: at the beginning of 2024, this figure barely reached 4 GW. Thus, over two and a half years, the volume of relevant projects has grown more than 40-fold, and in just the last six months, it has nearly doubled.
My sample included announced initiatives, projects at the pre-construction stage, and plants already under construction. It is important to emphasize: a significant portion of these gigawatts may remain on paper, but the trend itself demands close attention.
Autonomous Generation: A New Standard for Data Centers
The key growth driver is the behind-the-meter model. By locating power plants directly at their campuses, operators can bypass overloaded power grids, connections to which in some regions stretch over years. This trend is no longer just noticeable—it has become systemic. The expansion of gas generation in the U.S. is now inextricably linked to data center construction, and discussing these two phenomena separately is simply impossible.
As early as January, my analysis showed: the total portfolio of gas projects in the U.S. reached 252 GW, of which more than a third (97 GW) was intended directly for data centers. Implementing the entire January portfolio would increase the country's installed gas generation capacity by nearly 50%, requiring capital expenditures exceeding $416 billion. New data confirms: it is the segment tied to data centers that continues to grow exponentially.
Political Consensus and Market Risks
This shift coincided with a change in White House policy. In March, leading market players—Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI—committed to independently financing new energy capacity and grid infrastructure. The goal is to avoid shifting the costs of the AI boom onto ordinary consumers. By July, more than 200 additional organizations, including utility companies and developers, had joined this initiative.
However, 189 GW should not be viewed as a guaranteed forecast. Project implementation faces hurdles in financing, permitting procedures, local community positions, and, critically, a shortage of gas turbines. Order backlogs at the largest manufacturers are already booked through 2030, and for two-thirds of projects in the GEM global database, the equipment supplier has not even been identified. Even partial implementation of these plans will have long-term consequences, as gas plants are designed for decades of operation.
Geopolitical Divide: Gas vs. Green Energy
Notably, thanks to this boom, the U.S. has overtaken China in the total capacity of gas projects under development (252 GW vs. 153 GW). But the approaches differ radically. Chinese data centers are strategically located in regions with surplus solar and hydro energy, reducing dependence on fossil fuels. The U.S., meanwhile, bets on speed and stability, achieving them at the cost of additional emissions and the risk of creating excess infrastructure if AI demand forecasts fail to materialize.
The epicenter of the gas boom has become Texas: 80.6 GW under development, of which about 40 GW is for data centers. Also telling is Amazon's purchase of a site in Pecos County for the GW Ranch campus with its own 7.65 GW plant and 35 turbines. The company plans to use autonomous generation until it connects to the main grid.
My verdict: We are witnessing the formation of a new energy paradigm where AI becomes the primary driver of demand. However, betting on gas is a double-edged sword. A quick solution to today's connection problems could create long-term environmental and economic obligations that will become a drag on the industry in the next decade. Investors and operators should factor into their strategies not only current gains but also potential regulatory and reputational costs.