Why are AI data centers taking so long to come online?

A data center can be fully built, chips installed and racks wired, and still sit dark for years because it's waiting on a power grid connection.

The reason is almost never the building itself.

Large power projects now take a median of more than four years to go from an interconnection request to commercial operation, according to Lawrence Berkeley National Laboratory's "Queued Up" research. Projects above 200 megawatts, the size many AI data centers need, are averaging over 55 months. As of the lab's most recent count, roughly 11,600 projects representing 2,600 gigawatts of combined generation and storage capacity were sitting in that queue nationwide.

That queue doesn't care how much a company spent on GPUs. It's a grid problem, and it sits upstream of every chip order.

What Virginia, Phoenix, Dallas and the Midwest have in common

Northern Virginia, Phoenix, Dallas and parts of the Midwest all host heavy AI data center construction, and all four regions are reporting that utilities can't meet large new power requests on the timeline developers want. The equipment and the queue are both regional constraints, which means a hyperscaler can't simply build its way around the problem by picking a different city. The same shortage shows up almost everywhere large loads are landing at once.

What's actually causing the shortage?

Two things: not enough transformers and not enough room in the interconnection queue to use the transformers that do exist.

Lead times for high-capacity transformers have stretched to roughly four years, according to PwC analysis. Demand for generator step-up transformers, the kind substations need to handle new large loads, rose 274% between 2019 and 2025, per Wood Mackenzie data. Prices for that equipment are up about 80% over the past five years. None of that is a chip shortage. It's a heavy-industrial manufacturing shortage that happens to be sitting directly under the AI boom.

The data center isn't one company's project. It's a supply chain.

Coverage of AI infrastructure usually credits one name, the hyperscaler on the press release, which flattens a long chain of separate companies into a single story.

Each of those companies solves one layer of the problem.

Decompose a single "new AI data center" announcement and the layers look roughly like this:

Layer What it solves Current bottleneck
Grid interconnection Connecting the site to the power grid 4+ year median queue wait (LBNL)
Power equipment Transformers, switchgear, backup generation ~4-year transformer lead times (PwC)
Thermal / cooling Keeping high-density AI chips from overheating Liquid cooling capacity still scaling up
Compute The chips themselves The layer that gets almost all the headlines

Big Tech rarely builds the first three layers in-house. It buys from specialists who already solved the problem, the same way it buys most of what goes into a finished product.