Data centers are gobbling up electricity like never before. The rise of AI workloads means power demand is soaring. But the electric grid wasn’t built for this kind of rapid growth.

In the U.S., data center power use jumped from 23 gigawatts in 2023 to 42 gigawatts in 2026. That’s nearly double in just three years. The new GPUs powering AI eat up far more energy per rack — reaching 130 kilowatts compared to 10-15 kilowatts a decade ago. This puts huge strain on local grids, especially in hotspots like Northern Virginia.

The problem is supply. Building new power plants and transmission lines takes years, usually 4 to 8 years or more. Meanwhile, data centers can be built and operational in 12 to 24 months. This mismatch creates a bottleneck that slows down AI infrastructure growth. Some projects sit idle, waiting for power connections that won’t arrive for years.

Flexibility Is the Key to Unlocking Power Bottlenecks

Here’s the twist: data centers don’t have to suck up power 24/7 at full throttle. Modern software can throttle power use during grid stress without hurting AI tasks that matter most. This is called demand flexibility or power flex.

For example, in Phoenix during a hot May afternoon in 2025, a data center cut its power use by 25% for three hours during peak demand. It did this with no hardware changes and no service interruptions. The trick was smart software that adjusted workloads and shifted tasks in real time.

This kind of flexibility can reduce the need for new power plants. It also helps avoid blackouts and keeps prices from spiking. Utilities have long used similar programs to manage air conditioners and water heaters, but data centers have mostly been exempt. Now that’s changing.

Policies Are Starting to Catch Up

Some states and utilities are pushing for rules that require large data centers to accept interruptible power service. This means the center agrees to reduce power use during rare grid emergencies. In return, they get faster and cheaper grid connections.

Federal regulators introduced new tariffs that reward data centers willing to be flexible. States like Texas already mandate remote disconnect switches to quickly curb power use when needed. Montana offers lower rates to data centers that allow curtailment up to 200 hours a year. These changes create a win-win: data centers get power faster and cheaper, while the grid stays stable and ratepayers save money.

But voluntary programs aren’t enough. Most data centers still connect as firm customers with no curtailment duty. Without mandatory rules, utilities tend to build more infrastructure, which means higher bills for everyone. The best path is clear: tie grid access to flexibility requirements.

The Backlash and the Road Ahead

Local communities and regulators are pushing back against data center growth. They worry about rising electricity costs, noise, pollution, and the strain on the grid. Some cities have passed permanent bans on new data centers. Others require operators to build their own clean power or pay for grid upgrades.

Restricting new projects only delays the problem. The grid needs to evolve to handle big, flexible loads. Data centers that can adjust power use quickly and automatically will face fewer hurdles. This approach also helps cut the carbon footprint by reducing reliance on fossil-fuel plants during peak times.

Meanwhile, hyperscalers like Microsoft and Google are investing heavily in off-grid power solutions, nuclear deals, and battery storage. These strategies help meet demand but require massive capital and long lead times.

The power challenge is now the biggest factor shaping AI data center growth. It will determine who builds where and how fast. The companies that master flexibility and secure reliable clean power will lead the race for AI dominance. Everyone else may face delays and higher costs.

The grid won’t keep up unless data centers become smarter customers. Power flex is no longer optional—it’s the future of AI infrastructure.