Power becomes the defining constraint of the American AI boom as data centers strain the grid

The artificial intelligence revolution in the United States is increasingly running into a very physical obstacle. As technology companies race to build ever larger data centers to train and run their models, the availability of electricity is emerging as the single most important constraint on growth. What was once an abstract concern about computing power has become a concrete struggle over megawatts, transmission lines and generating capacity across the country.

A surge in electricity demand

The numbers behind this transformation are striking. According to research from the Lawrence Berkeley National Laboratory, data centers consumed 176 terawatt hours of electricity in 2023, representing about 4.4 per cent of total United States consumption. Projections for 2028 place that figure between 325 and 580 terawatt hours, or roughly 6.7 to 12.0 per cent of national electricity use, a remarkable jump in just a few years.

Other forecasts paint an equally dramatic picture. Goldman Sachs Research has projected that United States data center power demand will climb from 31 gigawatts in 2025 to 66 gigawatts by 2027. Looking further out, analysts at the Electric Power Research Institute estimate that data centers could account for anywhere between 9 and 17 per cent of national electricity consumption by 2030, up from roughly 4 to 5 per cent today.

The pace of this growth is what makes it so challenging. The International Energy Agency has noted that demand from accelerated servers, which include the graphics processing units and custom chips used for artificial intelligence, is expanding at around 30 per cent annually, compared with just 9 per cent for conventional servers. By some estimates, the electricity demand of AI data centers alone is set to triple between 2025 and 2030.

The scale of investment

Behind this energy appetite lies an extraordinary wave of capital spending. In 2024, the largest technology companies dramatically increased their investments in infrastructure. Amazon spent 85.8 billion dollars, a 78 per cent increase over the previous year, while Microsoft spent 44.5 billion, Google 52.5 billion and Meta 39.2 billion, bringing the combined total to more than 200 billion dollars in a single year.

The trend has only accelerated since then. For 2025, Amazon was projected to exceed 100 billion dollars in capital expenditure, while both Microsoft and Google were each expected to surpass 80 billion dollars. This concentration of spending reflects a belief among the major players that building out computing and energy infrastructure now is essential to maintaining their competitive position in the years ahead.

Regional pressure points

The strain is felt most acutely in a handful of regions. Northern Virginia, often described as the data center capital of the world, already has more than 4,900 megawatts of operating capacity, with an additional 1,000 megawatts under construction. It is estimated that about 70 per cent of global internet traffic passes through the region on a daily basis, underscoring its central role in the digital economy.

The intensity of this concentration carries real risks. In July 2024, a voltage fluctuation in Virginia caused the simultaneous disconnection of 60 data centers, which suddenly created a power surplus of around 1,500 megawatts on the grid. Such incidents highlight how the clustering of enormous electricity loads in a single area can introduce new forms of instability into the power system.

Texas faces its own version of the challenge. The grid operator known as ERCOT recorded a peak demand of 85 gigawatts in 2024, but projections suggest that figure could rise to as much as 145 gigawatts by 2031. Data centers are expected to contribute more than 32 gigawatts of that new demand, forcing the state to confront difficult questions about how to expand generation and transmission quickly enough.

The grid struggles to keep up

The core problem is that building new power infrastructure takes far longer than building a data center. Grid interconnection delays have become the primary obstacle to deployment, with waiting times that can range from four to ten years in some cases. The interconnection queue across the country now holds more than 2,600 gigawatts of proposed projects, with an average wait of around five years to connect.

This mismatch is already reshaping the economics of electricity. Property firm CBRE has found that power shortages are now the leading cause of construction delays for data centers. Goldman Sachs has estimated a United States power shortfall of 9.3 gigawatts in 2026, a gap that could widen to as much as 45 gigawatts by 2028 if new supply fails to arrive in time to meet the surging demand.

The technical demands of AI hardware compound the difficulty. Modern AI optimized server racks can draw anywhere from 30 kilowatts to more than 100 kilowatts each, compared with just 5 to 15 kilowatts for traditional racks. Individual AI data center sites now commonly require between 100 and 750 megawatts of power, a scale that can overwhelm the capacity of local grids that were never designed for such loads.

Turning to nuclear power

Faced with these constraints, the largest technology companies are increasingly turning to nuclear energy to secure reliable, carbon free power. Since 2023, hyperscalers have collectively committed to nearly 10 gigawatts of nuclear and small modular reactor capacity, a strategy designed in part to bypass the lengthy interconnection queues that slow conventional grid connections.

The specific commitments are substantial. Microsoft has secured 835 megawatts through the revival of the Three Mile Island site, now called the Crane Clean Energy Center, targeted for the second half of 2027. Amazon has committed to 2 gigawatts through an extension at the Susquehanna facility, while Meta has arranged for 6.6 gigawatts across multiple partners, with deliveries expected in the early 2030s.

A global race for power

The United States is not alone in this struggle. The International Energy Agency projected in April 2026 that global data center electricity consumption would rise from 485 terawatt hours in 2025 to around 950 terawatt hours by 2030. Between 2025 and 2030, the United States is expected to add roughly 240 terawatt hours of new demand, while China adds around 175 terawatt hours and Europe about 45 terawatt hours.

Taken together, these trends mark a genuine watershed moment for the relationship between technology and energy. The future of artificial intelligence in the United States will depend not only on advances in chips and algorithms, but also on the far more mundane question of whether the country can generate and deliver enough electricity. For now, power has become the defining constraint of the American AI boom, and resolving it will shape the industry for years to come.