How the Electrical Grid Is Being Rebuilt for AI

1.The Scale of AI Power Demand

A data center is a physical facility built to house computing hardware: servers, storage, and the networking and cooling equipment that keep them running. The concept is old, dating back to room-sized military computers in the 1940s, but the AI era has changed its scale. Training and running large models concentrates demand in a small number of enormous campuses, each drawing as much power as a mid-sized city.

The numbers are large and moving. Industry assessments have put global data centre electricity consumption in the mid-400s of terawatt-hours for 2024, with mainstream projections doubling that by 2030 under AI-driven growth. A terawatt-hour is a billion kilowatt-hours; total global electricity consumption is on the order of 30,000 terawatt-hours a year, so data centres are a modest share of the total but a very concentrated one, arriving in a handful of regions at once.

2.Why AI Load Is Different From Normal Grid Load

Electric power transmission is the bulk movement of energy from generating sites to substations; distribution then carries it to customers, and the combined network is the electrical grid. A traditional grid was built to serve demand that grows slowly and predictably, spreading new capacity over decades of incremental line and substation upgrades.

AI campuses break those assumptions three ways. First, the size of a single request: connecting hundreds of megawatts at one site requires dedicated substations and transmission spurs that would normally serve a whole industrial district. Second, the speed: developers want power in two or three years, not the decade a new transmission line can take to permit and build. Third, the location: fiber-rich rural sites with cheap land often sit at the weak end of the transmission network, exactly where spare capacity is scarcest.

3.The Interconnection Queue Problem

Before any large load or generator connects to the grid, it must pass an interconnection study proving the network can absorb it, and requests are processed largely in arrival order. As connection requests multiplied across renewables and now data centres, queues in key US regions stretched to years, with the volume of pending requests far exceeding what will actually be built, since many applicants hold places speculatively.

Reform efforts are under way, including first-ready-first-served cluster studies and deposit rules that penalize speculative applications, and grid operators have been clearing backlogs faster than they add them in recent cycles. Still, for a data center developer, queue position has become a strategic asset in its own right, which is why sites with existing power arrangements now command premiums.

4.Grid Hardware: Transformers, Lines, and Lead Times

The bottleneck is physical equipment as much as process. Large power transformers step voltage up and down between transmission levels and are enormous, custom-built machines; distribution transformers and high-voltage breakers are smaller but needed in far greater numbers. Surging demand across electrification, renewables, and data centres has pushed reported lead times for this hardware to multiple years, with industry surveys through 2024 and 2025 reporting roughly the ranges shown below.

Transmission lines face a parallel problem: permitting and siting can take longer than construction. The result is that even a well-funded project with land and chips can wait on a transformer the way a home builder waits on lumber, except the wait is measured in years and there is no substitute supplier at scale.

5.On-Site Generation and Behind-the-Meter Deals

Because the queue is slow, many developers now route around it. Behind-the-meter arrangements place generation on the customer's side of the utility meter, so the campus draws less from the public grid. The menu includes gas turbines for bridge power, fuel cells, battery storage to smooth peaks, and in a growing number of proposals, on-site nuclear, from small modular reactor designs to restarts of existing plants near retired facilities.

These deals are controversial. Critics argue self-generation lets large customers opt out of the shared system and its cost-sharing, while supporters note that on-site supply frees grid capacity for everyone else and can bring generation online faster than any transmission project. What is not disputed is the trend: the smart grid's promise of two-way communication and distributed resources has become an engineering necessity, not a research agenda, precisely because demand can no longer wait for the centralized build-out.

6.What Changes for Electricity Consumers

Ordinary ratepayers feel the shift in two ways. One is price pressure: when data centres compete for the same transformers, engineers, and generation capacity as public utilities, costs can spread across all customers, and regulators in several US states are already debating how much connection cost big customers should bear. The other is the opposite effect: an anchor tenant can finance transmission or generation upgrades that a region could not otherwise justify, which is why some consumer advocates and utilities have begun courting data centres as grid investment catalysts rather than merely fearing them.

Which force wins depends on regulatory design. Rules that assign costs to the load that causes them tend to leave ratepayers whole; rules that socialize everything across small customers do not. The next few years of utility commission decisions will matter more to household bills than any single data center announcement.

7.The Road Ahead

The grid is being rebuilt rather than merely expanded: high-voltage direct current lines to carry remote generation, smart controls that let data centres shift or curtail load second by second, storage sited where lines are weakest, and generation contracts signed directly between campuses and power plants. Bloomberg's Primer walks through this reconstruction in detail; its core framing, that the AI buildout is colliding with grid physics and permitting, matches what utilities themselves are telling regulators.

What to watch is whether equipment supply and interconnection reform can outpace demand growth. The projections in the first chart assume they mostly do; the lead times in the second chart describe how hard that will be. Both cannot relax, because the customers waiting on power are the fastest-moving enterprises in the economy, and they have shown they will build their own generation when the grid will not come to them.