The AI Power Crunch Is Now a Grid Capacity Problem, Not a Utility Bill

01 The Bill Stopped Being the Story

For most of computing history, electricity was an operating footnote. A company that doubled its compute fleet saw a noticeable utility invoice and adjusted. The AI buildout broke that relationship by changing the unit of scale. Data centers - defined in US law, in the Energy Independence and Security Act of 2007, as facilities for storing, managing, and disseminating data and housing IT infrastructure - were once measured in a few megawatts apiece. Announced AI campuses are now described in hundreds of megawatts and gigawatts, which is the language power systems engineers reserve for cities. Somewhere in that transition, the question for AI companies stopped being what power costs and started being whether power arrives at all.

This article treats the buildout the way a grid planner would: as an electrical-capacity story. The unit of analysis is megawatts and years, not FLOPs and dollars. Framed that way, the crunch looks less like an expense problem and more like a queueing problem at the boundary between private capital and public infrastructure.

02 Reading a Grid Like an Accountant

Plugging in a data center is not like plugging in a server rack. Connecting a large load requires generation capacity reserved for it, high-voltage transmission lines to carry it, transformers to step voltage down, and a substation to interconnect - each of which is engineered, permitted, and manufactured on timelines measured in years. Interconnection queues are the formal process where a new load or generator requests a place in that sequence, and utilities study the grid impact before approving equipment. Transformer and high-voltage equipment lead times are widely reported to have stretched badly during the current buildout, which converts money into time: a developer with full funding still waits on physical steel.

Grid accounting also distinguishes two numbers that press releases tend to blur. Connected capacity is what the substation can deliver at peak. Energy is what the facility consumes over time. An AI campus with a 500 MW interconnection does not draw 500 MW every hour, but the grid must be able to deliver it, and capacity reserved for that peak is capacity no one else can use. That distinction is the whole game in the sections that follow.

03 The LBNL Numbers: A Doubled Footprint

The best federal-commissioned baseline comes from Lawrence Berkeley National Laboratory, whose December 2024 report to Congress put US data center electricity use at 176 TWh in 2023, about 4.4 percent of total US electricity, and projected 325 to 580 TWh by 2028 - roughly 6.7 to 12 percent, depending on how aggressively AI demand grows. The spread matters: it is not a forecast so much as two scenarios, one where efficiency gains and workload mix temper growth, and one where accelerated AI buildout dominates. Both scenarios, notably, show growth; they disagree only about magnitude.

Label these carefully: the 2023 figure is a measured estimate from metered and modeled data; the 2028 pair are scenario projections. Even the low scenario implies adding on the order of 150 TWh of annual demand in five years - the kind of load that utilities normally plan across a decade or more, compressed into their next filing cycle.

04 From Rack to Campus: The Scale Ladder

It helps to climb the scale ladder one rung at a time. A single GPU rack draws roughly 100 to 130 kilowatts - around the combined load of 83 to 108 average US homes, using the common approximation of about 1.2 kilowatts average draw per home. A mid-size data center runs in the tens of megawatts, roughly 50 to 100 MW, which is on the order of 42,000 to 83,000 homes. The announced gigawatt-class AI campuses - Stargate-scale projects described in hundreds of megawatts to more than 1,000 MW - correspond to 250,000 to more than 833,000 homes worth of average draw. These are reported announcement figures, not metered loads, and the home conversions assume steady average draw rather than peak.

The ladder explains why AI facilities stress grids differently than earlier data centers. Each rung crosses a different procurement threshold: a rack is an IT purchase, a mid-size hall is a utility feed, and a gigawatt campus is effectively a new industrial customer asking a regional grid to grow around it.

05 Announced GW versus Energized GW

The most important discipline in this beat is refusing to equate the two headline numbers. Announced gigawatts are intentions: project reveals, land options, financing statements. Energized megawatts are what flows through a completed interconnection after transformers are installed and commissioned. Between the two sit environmental review, utility impact studies, equipment procurement, construction labor, and local grid upgrades. The gap between them is where multi-year delays live, and it is wide enough that two honest analysts can describe the same buildout as both a historic boom and a pipeline of paper.

The capacity-accounting lens also explains a subtle risk: because announced projects compete for the same equipment, the same engineering firms, and the same queue slots, some announcements function as place-holding - claims on scarce interconnection capacity filed early so rivals cannot take it. Utilities respond with deposit requirements and queue reforms, but the dynamic means announced GW systematically overstates near-term energized GW. Neither number is false; they measure different stages of the same physical pipeline.

06 Ratepayer Politics Enter the Chat

Once campuses reach hundreds of megawatts, the financial question stops being private. Grid capacity is socialized infrastructure: transmission upgrades and new generation are paid for through rates that residential customers share, under regulatory frameworks designed around gradual demand growth. A single AI campus that triggers a substation or transmission line invites a rate-case fight over who benefits and who pays. State regulators, consumer advocates, and utilities are now openly debating tariff structures that make large loads cover their own grid costs, and some jurisdictions have proposed special rate classes for data centers.

The residential equivalence is what makes this politically legible. When a project is described as drawing the average power of hundreds of thousands of homes, local audiences understand it immediately - both the scale of the investment it justifies and the scale of the intrusion. The politics do not stop the buildout, but they shape its geography: siting now favors places with available capacity, willing regulators, and existing generation, which is why the map of AI campuses follows the map of cheap power rather than the map of engineers.

07 What Actually Relieves the Constraint

If megawatts and years are the binding constraints, the relief valves are ones that save megawatts or years. On the megawatt side: hardware efficiency gains per unit of AI computation, which the LBNL scenarios implicitly assume; better cooling and power-usage discipline; and scheduling flexibility, where training workloads yield during regional peaks - a behavior grid operators call demand response. On the years side: expedited interconnection study processes, standardized transformer designs, pre-approved siting corridors, and on-site generation or storage that reduces the incremental grid upgrade a campus requires.

None of these eliminate the queue; they shorten it at the margins. That is why sophisticated builders now evaluate sites primarily on time-to-energization rather than on power price. A campus that energizes two years earlier on expensive power can beat one that waits on cheap power, because the constraint is the calendar, not the invoice.

08 Limits of the Capacity Frame

Honesty requires marking what this lens misses. Megawatt accounting says nothing about chip supply, model efficiency breakthroughs that could flatten demand, or the possibility that some announced projects are never built - the LBNL low scenario is partly a bet on exactly that. It also cannot capture geography: national TWh aggregates hide the fact that constraints bind locally, in specific queues and specific substations, long before they show up in national statistics. And the frame is US-centric; the global picture, as the IEA reporting emphasizes, includes regions with different generation mixes and different regulatory speed.

What would update this analysis: LBNL follow-up reports tracking actual consumption against the 325-to-580 TWh scenario band, utility filings showing energized interconnections by year, and tariff decisions that settle who pays for grid expansion. The core claim stands as stated: the AI buildout has crossed from a cost discussion into a capacity discussion, and in a capacity discussion the scarce resources are transformer lead times, queue positions, and years - commodities no amount of capital can order overnight.