Why a data center can’t just plug into the grid

A large data center can be announced years before it receives full power. That sounds strange. The United States already has thousands of power plants and an enormous electric grid. Why can’t a new customer simply connect?

Because the grid is not a lake of electricity that anyone can dip into from any shoreline. It is a delivery system. Every line, transformer, and substation has a job, a location, and a limit.

A generator can produce enough energy somewhere in the region while the exact data center parcel still lacks a reliable path to receive it. This difference—between producing electricity and delivering firm power to one location—is one of the most important parts of the AI infrastructure story.

The grid is closer to a road network than a wall outlet

Power plants inject electricity into a network. High-voltage transmission lines carry large amounts over long distances. Substations route the flow and use transformers to change voltage. Distribution equipment or dedicated high-voltage facilities make the final handoff to customers.

The U.S. Energy Information Administration describes this as an interconnected system of power plants, substations, transformers, and lines. The connections improve reliability because power can often take multiple paths. But every path still has a thermal, voltage, and stability limit.

Think about a city with plenty of food in regional warehouses but only one narrow bridge into a neighborhood. Adding another warehouse does not clear the bridge. The bottleneck is delivery.

For a large new load, engineers study more than whether nearby wires exist. They model normal operation and contingencies: what happens if a transmission line trips, a transformer fails, or a power plant goes offline? The system must remain within safe limits after credible failures, not only on a perfect day.

A connection request is also a promise about the future

Utilities plan years ahead. A data center developer may request service for an initial building, later phases, and eventually an entire campus. The requested capacity can be much larger than the site’s first-day demand.

This creates a difficult planning problem. Build too little and the utility may strand the project or overload equipment. Build too much for a project that is delayed, downsized, or canceled and other customers can be left carrying infrastructure built for demand that never arrived.

That is why large-load agreements increasingly address financial security, construction contributions, minimum charges, ramp schedules, and who pays if the customer does not materialize as promised. Colorado Springs Utilities, for example, publicly states that qualifying large data centers must pay the full cost of infrastructure upgrades needed to serve them, along with additional system-support charges. The exact policy differs by utility and regulator, but the underlying issue is widespread: risk has to sit somewhere.

“Who pays?” is not a side debate. It is part of the engineering decision because transmission lines, substations, and generation are financed assets with long lives.

Transformers can become the schedule

A parcel beside a transmission corridor can still be years from service. New substations need land, permits, breakers, protection systems, control equipment, and very large transformers. Some upgrades require new rights of way. Others depend on work elsewhere in the network before power can safely flow to the site.

The transformer is basically a giant electrical adapter. It changes voltage so power can travel efficiently on high-voltage lines, then be used at levels appropriate for the facility. Unlike the adapter in your laptop bag, a grid-scale transformer can be custom equipment that is expensive to manufacture, transport, install, test, and protect.

Inside the data center, the voltage may step down again through multiple layers before reaching the chips. Switchgear isolates faults. Uninterruptible power systems bridge brief outages. Generators provide longer-duration backup. The utility connection is the start of a chain, not the end of one.

Power and energy are not the same measurement

Data-center conversations often mix megawatts and megawatt-hours. A megawatt measures power—the rate of electricity use at a moment. A megawatt-hour measures energy—one megawatt sustained for one hour.

The distinction matters because utilities must prepare for the highest credible rate of demand while also procuring enough energy over time. A 100 MW facility that operates near that level around the clock is a very different grid customer from a 100 MW industrial process that runs occasionally or can pause when the system is stressed.

Berkeley Lab’s 2025 update estimates U.S. data centers could use 521 to 843 terawatt-hours of electricity in 2030, with a reference case of 649 TWh. Those are annual energy estimates, not a single instantaneous power requirement, and the wide range reflects uncertainty in equipment shipments, utilization, efficiency, and other assumptions.

Forecast uncertainty does not mean the problem can be ignored. It means utilities have to build plans that work across multiple plausible futures.

Flexibility can be infrastructure too

One possible tool is load flexibility: shifting some computing work, charging batteries at different times, or temporarily reducing demand when the grid is constrained. Not every AI workload can pause, and not every facility will offer the same flexibility. Reliability commitments, customer contracts, hardware utilization, and software architecture all matter.

But even a portion of flexible demand can have value. It may help a utility manage peaks or emergencies while longer-term upgrades are built. Berkeley Lab is studying these strategies because the fastest megawatt to connect may be the one that does not require every part of the grid to be expanded for the same hour.

On-site generation and batteries can also support a facility, but they do not magically erase the grid connection. Fuel supply, emissions permits, operating limits, maintenance, and the duration of an outage all shape what those systems can do. A generator designed for emergencies is not automatically a year-round power plant.

The important question is not “Does the region have enough electricity?” It is “Can the system deliver the required power to this site, at the right time, through a failure, without shifting unreasonable cost or risk to everyone else?”

The connection is the project

Land can be acquired quickly. Server orders can be announced loudly. The grid moves at the speed of studies, permits, factories, rights of way, construction, and commissioning.

That mismatch explains why some data-center proposals include their own substations, why utilities are creating special tariffs for large loads, and why communities are suddenly debating transmission corridors and transformer supply chains alongside AI.

The chips may be the product. The connection is what makes the product possible.

When I look at a proposed AI campus, the first question is no longer just how many GPUs will go inside. I want to know the path: which lines, which substation, which upgrades, which backup systems, which payment protections, and which date each piece can actually be ready.

That path is the physical infrastructure behind the promise.