AI's Real Limit Isn't Chips. It's Electricity.

The Real Bottleneck Is Power

Research firm Sightline Climate estimates that 30 to 50 percent of the AI data centers scheduled to open in 2026 will be delayed or cancelled. Out of about 16 gigawatts of capacity that was announced, only around 5 gigawatts is actually under construction. The reason is not a shortage of money or chips. It is the physical electrical equipment.

The big transformers that connect a data center to the grid now take three to five years to deliver, and a key component called switchgear is sold out through 2028. The tech giants plan to spend more than 650 billion dollars on AI infrastructure this year, but no amount of money can rush the physics of building out a power grid.

Why AI Eats So Much Electricity

A single AI request can use up to roughly 1,000 times more electricity than a traditional web search. Multiply that by hundreds of millions of people using AI every day, and the demand becomes staggering. Global data center electricity use is projected to pass 1,000 terawatt-hours in 2026, which is on the order of the yearly electricity use of an entire large country.

Interestingly, the biggest power draw is no longer training the AI models. It is the everyday running of them for users, what the industry calls inference. And despite all the green pledges, more than 60 percent of the electricity feeding US data centers still comes from fossil fuels, because clean power simply is not being built fast enough to keep up.

It Is Reaching Your Power Bill

In some regions, people are already feeling it. One Virginia homeowner reported his monthly electric bill jumping from around 100 dollars to 281 dollars in a single month. More broadly, retail electricity prices are up about 42 percent since 2019, outpacing general inflation over the same period.

It is worth being fair here, because experts disagree on how much data centers are really to blame. Some point more to the price of natural gas and to the way grids buy backup power, rather than to AI alone. But the strain is real: PJM, the largest US grid operator, which serves more than 65 million people, has warned it could fall about 6 gigawatts short of the power it needs to stay reliable by 2027.

Communities Are Pushing Back

Local opposition has become a serious force. By one count, around 98 billion dollars of data center projects were blocked or delayed in just a few months of 2025, and cancellations roughly quadrupled over the year. Maine became the first state to pass a statewide pause on large data center construction, and about a dozen more states are considering similar limits.

In response, tech companies are chasing power wherever they can find it. They are signing deals for nuclear energy and small modular reactors, building their own on-site generation, and looking overseas. Microsoft, for example, committed 15.2 billion dollars to power-rich projects in the UAE. The race is no longer just for chips. It is for electrons.