Northern Virginia is already a customer the old grid was never built to serve. Harvard Kennedy School’s Belfer Center, in AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment, puts more than 4, 900 megawatts of data centers operating there, with another 1, 000 megawatts under construction. The load is large, nearly constant, and still able to jump when the machines or the jobs get harder.
A round capital headline will not get the next campus interconnected. Spending on data centers can lift growth this year. Interconnection speed, financing, and the question of who pays for the upgrade decide whether that campus gets built. The constraint on AI for business is increasingly physical: power, permits, trades, and a fight over the bill when the forecast misses.
A local rate case is now a national constraint
Northern Virginia did not stumble into the role. Dense fiber, relatively low power costs, reliability, incentives, and early permit acceleration in some counties made the region the default. A local tax exemption and a utility rate case are now national politics for anyone planning AI infrastructure.
Dominion’s 2024 resource plan calls for nearly 27 gigawatts of new generation by 2039. In the Belfer Center’s summary, that includes about 21 gigawatts covering solar, wind, and nuclear small modular reactors, plus 5.9 gigawatts of gas. The gigawatts add up. The word “renewables” does not. Nuclear is not renewable, and a reader who treats that 21 gigawatts as a renewable build will misread what the utility is proposing.
In February 2025, Dominion proposed its first base-rate increase since 1992: about $8.51 a month in 2026 and $2.00 a month in 2027 for a typical household. Belfer places that proposal in a discussion of data-center demand and rising Virginia rates. It does not isolate how many of those dollars data centers alone caused. Households facing the first base-rate case in a generation do not need a perfect split to object. The political heat is already a contracting risk.
Lawmakers have pushed without stopping the cranes. On Belfer’s account, efforts to scale back data-center tax exemptions stalled. Several 2025 bills on efficiency-linked incentives, a Northern Virginia pause, and uniform standards did not advance. A statewide standards bill reached the governor and was vetoed. Construction continued. The next bills will be written by people who watched this round fail.
The argument that matters more is who pays. More contract-based financing moves projects off rate-base recovery (putting a utility asset into the charges all customers pay) and toward special tariffs and power purchase agreements. A special tariff is a rate written for a large customer or class, not the ordinary household schedule.
Those deals are less transparent, and Belfer notes they may shift power costs onto other consumers. The brief leaves the allocation question open: who pays for grid upgrades spurred in part by data centers, who captures the benefit, and how communities are protected from higher bills and heavier resource use.
If special tariffs socialize the upgrade and privatize the AI upside, community pushback becomes a constraint on the campus. A communications plan will not clear it.
You can bond a campus faster than you can permit a line
The same brief splits the buildout by who can pay for speed. Data-center shells and the computing load inside them are financed mainly from parent balance sheets, corporate bonds, and public incentives. Project finance, Belfer says, is occasional. Green bonds are a supplement. Transmission lines, substations, and generation sit with utilities, under tighter financial and regulatory constraints. Speed dies on that side of the ledger.
Large buyers are answering with collateralized third-party power contracts, long-term power purchase agreements, availability payments, and upfront capital. Availability payments are fees for capacity whether or not you draw the power. Some are also talking about putting generation on the same site as the campus. Co-location sounds tidy until a siting rule, an ownership question, or a regulator says no. Belfer’s judgment is that the spending exposes both firms and utilities to stranded assets and higher financing costs. The buildout holds only if the financing still works when the demand does not fully arrive.
The shape of the load makes old planning habits look quaint. Data centers run large and nearly flat. They have limited ability to ramp down, and usage can still jump when the equipment or the job gets more complex. A mild Tuesday at 3 a.m. does not give the grid a break. PJM, the regional grid operator that covers this load, was built around gradual, weather-sensitive demand. It is now reviewing how it defines firm service (power the operator must deliver, not power it offers when it can) and the reliability obligations that go with it. For a procurement team, the definition of a guaranteed customer is in motion.
The forecast is a range, and the range is the warning
History is firmer than the outlook. The 2024 United States Data Center Energy Usage Report, as summarized by Belfer, describes stagnation from 2014 to 2016, then growth from 2017. From 2018 to 2023, data-center electricity use rose from roughly 76 terawatt-hours, 1.9 percent of U.S. annual electricity, to 176 terawatt-hours, 4.4 percent.
Then the bands open. That same report puts 2028 use between 325 and 580 terawatt-hours, or 6.7 to 12.0 percent of projected national electricity use that year. Belfer stresses that the range is still uncertain. The World Resources Institute, cited in the brief, finds institutional projections for 2030 running from about 200 terawatt-hours to more than 1, 000. A 2022 review by Mytton and Ashtine looked at 258 data-center energy estimates and found systematic methodological defects, especially poor data availability and weak transparency.
Anyone who picks one terawatt-hour figure and calls it the forecast is overclaiming. Site plans are opaque. Efficiency gains are real, and they are also easy to double-count. Utility planners are working in the same fog as the energy studies. An AI automation roadmap that assumes a single official number is assuming a confidence the specialists have already refused.
Data-center construction is carrying growth. That is not a plan.
A Washington Examiner account of the mid-2025 expansion puts this construction near the center of U.S. growth. It reported real GDP growing at a 3.8 percent annualized rate in the second quarter of 2025, and pointed to third-quarter expectations of at least 2.5 percent, nearly 4 percent on the Atlanta Fed’s GDPNow tracker at the time. Treat both as a snapshot. Early GDP prints are revised, and a nowcast is replaced the next time the model runs.
The piece credits the rebound, despite tariffs, to high-end consumer spending and, more importantly, to AI and the physical work behind it: data centers, substations, high-voltage lines, and the skilled trades that build them. It never says how many points of that 3.8 percent belonged to AI. Slowing a line or a crew can still be a growth risk. We cannot put a number on it.
Questions worth asking before the next commitment
Is the buildout already big enough to matter for growth?
In one account, yes. The Washington Examiner ties a strong mid-2025 expansion, including 3.8 percent annualized real GDP growth in the second quarter, in important part to AI and data-center construction. It does not quantify AI’s share of that print, and early GDP figures get revised.
How fast has data-center electricity use actually grown?
From 2018 to 2023, the 2024 U.S. data-center energy report, as summarized by Belfer, shows use rising from about 76 terawatt-hours to 176, or from 1.9 percent to 4.4 percent of U.S. electricity. Paths after that are ranges: 325 to 580 terawatt-hours by 2028 in that report, and about 200 to more than 1, 000 by 2030 across institutional projections the brief cites from the World Resources Institute.
Where is the strain actually showing up?
In interconnection speed, utility balance sheets, stranded-asset risk, rate design, and forecasts too opaque to plan against. There is no verified break in the capital markets, and no construction stop. In Virginia the evidence is a generation plan measured in tens of gigawatts, a first base-rate case since 1992, and reform bills that stalled or were vetoed while the cranes kept moving.
What should a company ask before it bets on always-on AI?
Four things a buyer can put in a term sheet. Where the project sits in the interconnection queue, and the date on that position. Curtailment and exit rights if the power slips. Whether upgrade costs sit in a special tariff or in the rate base everyone pays. What happens to the power purchase agreement if the load misses. Whether the use case is AI for sales or a back-office agent, chips, models, and AI agents can be procured on a corporate schedule. Substations and generation cannot.
One more number worth sitting with
The ranges in this piece are already wide. The International Energy Agency has been tracking the same load from the other side of the meter.