US Data Center Power Demand Is Outrunning the Grid

The number that should worry anyone planning AI infrastructure is the gap. According to a new report from insurance broker Howden, US data center power demand is growing about ten times faster than the grid can expand. That mismatch is pushing developers to look well beyond the utility connection, and the alternatives they're choosing each carry costs of their own.

The numbers behind the bottleneck

Howden drew on data from the US Energy Information Administration and S&P Global going back to 2000. Its headline finding is that data center power demand has grown about 19% year on year, while the electricity grid has expanded roughly 2% annually. Data centers' share of grid capacity reached 5% in 2025, up from 0.1% in 2000.

Those growth rates compound quickly. A load rising almost ten times faster than the system serving it doesn't create a gentle squeeze. It forces a choice between waiting for transmission and distribution upgrades that take years, or finding power somewhere else. Howden estimates that keeping the current pace of data center expansion without adding strain to the grid would require up to 58% of demand to come from alternative sources.

The comparison to the dot-com era is apt, though the bottleneck has moved. Then it was fibre cabling. Now it's electrons. Chips and memory have been tight, too, and new supply chains and manufacturing capacity are expected to ease those constraints by late 2027 or early 2028. Power doesn't have a comparable timeline. Grid buildouts run on permitting, equipment lead times and interconnection queues that move at utility speed.

Who ends up paying for the grid

Grid expansion is typically funded through ratepayers, which means the cost of building new lines and connections for data centers gets spread across household and small business bills. That's the part of this story that tends to get lost in the gigawatt headlines. Experts are predicting that power costs in some states could rise by as much as 50% by 2030.

This dynamic is one reason the arrangement is politically sensitive. When a utility spends heavily to serve a handful of large loads, the people with the smallest margins often absorb a share of the bill. Operators that bring their own generation sidestep some of that exposure, but they introduce different trade-offs around air quality, noise and fuel supply.

Gas turbines: fast to deploy, heavily exposed to fuel

Onsite generation is the quickest route to power that doesn't depend on the grid. Anthropic recently disclosed a $3.4 billion investment in onsite gas turbine generators to bring new sites online. My read is that this is less a sign of confidence in gas than a sign of how little time developers have.

The catch is fuel. By relying on natural gas, Anthropic's plan ties data center operations to gas supply and pricing. Gas also has competing demand from other industries, and the source notes that Middle East tensions have pushed other sectors toward more natural gas consumption, lifting prices across the board. A campus designed around turbines is effectively making a long-term bet on gas markets, and that bet is harder to hedge than it looks on a spreadsheet.

Local impact is the other constraint. Some hyperscale campuses now run ranks of gas turbines that produce noise comparable to an international airport. Residents near these sites have reported symptoms they link to a persistent low hum, and the source describes unexplained illnesses in local residents. The evidence base here is still developing, but community opposition tends to harden quickly when people feel they've been given no say, and that can slow permitting for the next project as well.

Nuclear: the long game

Nuclear is attracting attention for the opposite reason. It's slow to build but offers steady, carbon-free output once running. Three projects are central to the current push: the Palisades Nuclear Plant in Michigan, the Crane Clean Energy Center in Pennsylvania and the Duane Arnold Energy Center in Iowa. Together they're set to receive $4.4 billion in grants to be recommissioned, with Microsoft and Google first in line for their output.

Amazon has taken a different route. It plans to buy more than 960 megawatts from 12 small modular reactors to be built in Iowa, at a cost that will likely exceed $1 billion. Those sites haven't been approved yet, and they may not run until the mid-2030s. That timeline is the core problem: the reactors that matter for the next few years are the ones already in the ground, and there aren't many of those. Progress on new designs is real, though. Plans for the first data center-specific small modular reactor have just been approved for a site in Tennessee, with each unit providing up to 300 MW. If that design performs as planned, it could give developers a template that doesn't need to be rebuilt from scratch every time.

Still, nuclear carries its own price tag. New reactors are expensive. Restarting old ones requires substantial investment to extend operating life and to meet current safety requirements. Those costs don't disappear because the demand is urgent.

What to watch

The Gartner report released in June warned that the current growth of data center power demand could halt new construction by 2030 if supply doesn't keep up. That's the scenario the industry is racing to avoid, and it's why the mix of power sources matters as much as the total amount. Here's the thing: none of the options on the table is fast, cheap and clean at the same time. Gas is quick but fuel-exposed. Nuclear is clean but slow and capital-heavy. Grid upgrades are necessary but depend on the same ratepayers who would fund them.

Watch three things over the next two years: whether turbine and reactor projects clear permitting on schedule, whether gas prices stay volatile, and whether utilities can bring large transmission projects online faster than they have been. The answers will determine whether the AI buildout keeps its current pace or gets rationed by electricity.