The AI buildout is no longer only a race for chips, talent and capital. According to a Financial Times analysis cited in the report, the next major constraint for OpenAI, Microsoft and their peers may be far more physical: the US power grid.
Planned AI data centers are expected to require far more electricity than the grid is likely to deliver on the same timeline. That mismatch could turn energy access into one of the biggest limits on the next phase of AI growth.
The scale of the power gap
The report says new data centers could need an estimated 44 gigawatts (GW) of additional power by 2028. Over that same period, grid infrastructure bottlenecks mean only about 25 GW is expected to come online.
That leaves a 19 GW deficit, described as roughly 40 percent of total demand. For companies planning enormous AI infrastructure projects, the issue is not abstract. Without enough electricity, data centers cannot operate at the scale their owners are planning.
The largest hyperscalers, including Amazon, Google, Meta, and Microsoft, have outlined investment plans exceeding $400 billion, mainly for data centers. OpenAI alone has reportedly signed infrastructure contracts totaling over $1.4 trillion to secure roughly 28 GW of capacity over the next eight years.
CEO Sam Altman has characterized the energy shortage as an existential threat. The logic is direct: without enough compute power, the company cannot generate revenue or build models at the necessary scale.
Why the grid is slowing the AI buildout
The US power grid is facing a demand shift after two decades of stagnant growth. Electricity demand is now rising, and AI data centers account for more than half of the expected increase described in the report.
The physical infrastructure is part of the problem. Many poles and transformers date back to the 1960s and 1970s, according to the report. That aging equipment is being asked to support a wave of new demand from projects that require unusually large amounts of electricity.
Connection delays add another barrier. The average wait time from requesting a grid connection to commercial operation now exceeds eight years nationwide. For fast-moving AI companies, that timeline clashes with the pace at which they are trying to build.
The report also points to a planning problem caused by so-called phantom data centers. Developers may submit multiple applications to different utilities to find the best price. That can inflate queues and make it harder to distinguish firm projects from speculative ones.
Supply chains are under strain as well. Lead times for large transformers are three to four times longer than in 2020. Gas turbines, often used as a stopgap, now have delivery times of around four and a half years.
Companies are looking beyond the grid
To avoid long grid delays, AI companies are increasingly turning to behind-the-meter power generation. In practice, that means producing electricity directly for a facility rather than waiting for a conventional grid connection to be ready.
The report highlights xAI as a prominent example. According to the Southern Environmental Law Center (SELC), Elon Musk's company powered its controversial Colossus cluster in Memphis, Tennessee, for months using dozens of gas turbines without necessary environmental permits.
xAI received a permit for 15 backup turbines in July. The SELC claims up to 35 turbines were observed on-site.
OpenAI is also planning to use ten natural gas turbines with a 361-megawatt capacity for its Stargate project in Texas. Microsoft is taking a different route, with a deal to reactivate the Three Mile Island nuclear plant by 2027.
These examples show how energy procurement is becoming part of AI infrastructure strategy. Building data centers is not enough if the power cannot be delivered on time, and the report suggests companies are now trying to secure power through direct or alternative arrangements.
The China comparison raises the stakes
The tech industry is also framing the energy shortage as a national security issue. In an open letter to the US government, OpenAI warned that China is pulling ahead in infrastructure.
The comparison in the report is stark. China added roughly 429 GW of new power capacity in 2024, described as more than a third of the entire US grid. The US added just 51 GW in the same period.
That difference matters because AI companies view infrastructure as a foundation for model development and commercial growth. If one country can add power capacity much faster than another, the infrastructure race becomes part of the broader AI race.
What happens if the deficit remains
The US government is attempting to fast-track permits and delay coal plant retirements through emergency orders. Environmentalists warn that focusing on fossil fuels and blocking renewables is counterproductive, because solar and battery parks are much faster to build than gas plants.
The conflict is not simply about how much power AI needs. It is also about what kind of power gets built, how quickly it can connect, and whether the grid can support the projects already being planned.
For OpenAI, Microsoft and other hyperscalers, the issue now sits at the center of growth strategy. The report suggests the feared AI bubble may not burst because of weak demand for intelligence. It could run into a simpler barrier: a lack of electrons.