Why ‘Invisible Factories’ Suddenly Got Loud

When explaining the generative AI boom, people usually think of ‘models,’ ‘chips,’ and ‘investment amounts.’ But what actually shakes politics and daily life are things that weren’t visible. Gray box-like buildings, windowless warehouses, servers and coolers filling them, and the electricity feeding them all. Data centers were large before too, but now they’re no longer at “nice to have” levels in electrical systems. In the U.S., data centers consumed about 4.4% of power in 2023, with estimates publicly disclosed in official reports and government announcements that by 2028 this could reach 6.7-12% depending on scenarios. Figures estimated by Lawrence Berkeley National Laboratory (2014: 58TWh → 2023: 176TWh, 2028: 325-580TWh) make people tangibly feel “where electricity demand springs from” in daily life units. The U.S. Department of Energy also introduced the same report, nailing down as policy issues that “AI diffusion and data center expansion pull up power demand.”

Here the key is ‘geography’ more than ’totals.’ Data center electricity doesn’t spread evenly nationwide. It concentrates in a few regional clusters. The result is, some states or counties face situations like “our neighborhood suddenly hosting several aluminum smelter-grade factories.” Electrical facilities originally move on 10-year plans, resident complaints and elections run on 2-year rhythms, but data centers deliberately misalign the beats of the two. This off-beat is today’s story’s ‘curse.’

The Scale of AI ‘Electricity Feast’ Shown by the International Energy Agency

Let’s pin down numbers first. The International Energy Agency sees data centers consuming about 1.5% (415TWh) of global electricity in 2024. By country, the U.S. share is largest at 45%, with China at 25% and Europe at 15% following. And this power consumption could grow to about 945TWh by 2030, “more than doubling.” This 945TWh is explained as similar scale to Japan’s current total national electricity consumption. In other words, it’s correct to see data centers as no longer periphery of ’electricity users’ but having entered as an axis changing the texture of global power demand.

The reason AI makes this situation more uncomfortable is also clear. AI-specialized data centers can suck electricity rivaling power-intensive factories (e.g., smelters) at single points, and most importantly they enter “in clusters.” It’s the worst form for electricity. When demand grows evenly, generation/transmission/distribution follow sequentially, but AI data centers compress demand and throw it at specific substations and transmission lines. The industry where “where to build” determines power system bottlenecks.

But this market has thick ‘forecast fog.’ The Electric Power Research Institute (EPRI) presented in 2024 analysis that U.S. data center power by 2030 could be in the “4.6-9.1% of total U.S. generation” range. The range is wide. It’s because uncertainties are large around how fast AI will spread, how much efficiency will improve, and how far demand will grow. EPRI itself says AI is ’estimated’ to use 10-20% of data center power. In a word, the direction “AI devouring electricity era” is mostly agreed on, but speed and ceilings are still fighting.

This uncertainty doesn’t just end as scholars’ debates. The electricity industry has strong instincts to move ‘safely,’ so even if demand is exaggerated, they try to lay facilities. Conversely, underestimating demand causes accidents like blackouts and rate explosions. So the moment forecasts go wrong, the remaining choices are one of two: (1) overinvesting in expensive power grids to make someone bear costs later, or (2) facing regional bottlenecks while unable to follow plans. The World Resources Institute warns that such forecast fluctuation ranges can directly lead to “consumer rates, greenhouse gas emissions, and power reliability.”

Why Virginia’s Data Centers Concentrated in Northern Virginia

Looking at U.S. data center maps becomes a ‘power lesson’ not a ‘geography lesson.’ Data centers can’t be built just anywhere. It simultaneously requires (1) communication infrastructure with densely laid fiber optics, (2) relatively cheap and stable electricity, (3) wide and flat land, (4) distance from customers and urban areas (especially services needing low latency), and (5) policy like tax incentives. Virginia JLARC (Joint Legislative Audit and Review Commission) summarized “factors making northern Virginia the world’s largest data center market” also ultimately to this combination. Strong fiber optic networks, cheap and reliable energy, available land, major customer access, and state-level data center tax incentives operating together.

JLARC also throws more direct numbers. Northern Virginia comprises 13% of reported global data center operating capacity and 25% of Americas region capacity. Meaning “one region is a quarter of the continent.” And growth isn’t confined only to northern Virginia. It warns of potential spread not only to existing markets but other regions of the state, especially central Virginia along the I-95 corridor.

This giant chunk is particularly thick around Loudoun County on maps. Loudoun County describes itself as the “world’s largest data center concentration area,” saying the data center industry diversified county economy and helped maintain low property tax rates thanks to tax revenue. Reports also exist of ’nearly 300’ data centers scattered across northern Virginia as of 2023, with explanations repeating that the central axis is Ashburn in D.C. suburbs.

Here’s a point many people miss. Data centers’ ‘regional economic effects’ have different textures from general factories. JLARC notes a typical 250,000 square foot data center may permanently employ around 50 people (half contractors). Instead, construction stages differ. Construction periods are about 12-18 months, with testimony emerging that construction peaks can attach around 1,500 people on-site. Simply put, the structure is “uses lots of people while building, but after built it’s an electricity-eating building operated by a few.”

The reason this structure ignites politics is that tax and complaint timetables misalign. Construction booms create ‘immediately’ felt traffic, noise, dust, and landscape problems for regions. But stable jobs or regional service demand don’t follow as much as factories. Conversely, tax revenue is large. So from local government perspectives, it becomes a “hard-to-miss cash window even if disliked.” This is where the ‘curse’ starts.

Noise and Transmission Lines, and Prince William County’s Rebellion

The moment data center conflicts turn into ‘political events’ is usually similar. Neighborhood people initially aren’t interested. What’s the big deal about one big windowless building being built? But from some day on, low-frequency ‘wooom—’ noise is heard even at night, and they realize there’s no proper law to block that noise. Or they hear “transmission towers must pass there,” suddenly realizing landscapes will change.

JLARC fairly frontally addresses these ’life impacts.’ The gist goes like this. Data center noise is low-frequency, not loud enough to damage hearing, and often not loud enough to violate noise regulations. The problem is ‘persistence.’ It’s not one-time party noise but spreads 24 hours like drones. But existing noise ordinances struggle to catch such noise, with points emerging that civil fine caps (e.g., $500) are too weak to move giant corporations. So some regions see that handling noise not from noise ordinances but from “zoning/development permits (zoning)” sides may be more effective.

If this sounds abstract, look at Prince William County cases. In this region, ‘Prince William Digital Gateway’ - a massive data center development plan to build up to 37 data centers on 2,000 acres of farmland - amplified controversy. Resident opposition concentrated on noise, aesthetics, power consumption, and high-voltage transmission line problems attached to bring that power. The location itself near Manassas National Battlefield Park also amplified conflicts.

Politics moved next. In the 2023 Democratic primary, a newcomer critical of data center expansion created an ‘upset’ defeating the then-county’s top elected official, which regional opposition groups evaluated like “shockwave to industry.” In other words, data centers are no longer development issues but have become voting issues.

Evidence that this flow isn’t one-time happenings continued afterward too. The Washington Post reported that Prince William County western region’s special election was heavily influenced by data center expansion concerns, with winners fronting data center sprawl suppression, zoning strengthening, and cost burden ‘fairness.’

Ultimately, the route by which data centers enter election battlegrounds is simple. The question “who will be responsible for this neighborhood’s landscape (scenery), sleep (noise), and electricity rates (costs)” becomes votes fastest. And this question doesn’t neatly cut along ideological spectrums like Democrat-Republican. So it’s more dangerous and stronger.

Power Grids Take 10 Years, Elections Take 2 Years

The reason electrical systems struggle with data centers isn’t just “big.” “New” is more the problem. The North American Electric Reliability Corporation (NERC) points out that ’emerging large loads’ like data centers attach to transmission grids faster, with scales also varying widely from several MW to several GW. Especially noting data centers can change consumed power very quickly with Power Electronics Loads-based control, challenging existing planning and operation assumptions. Expressions also stand out that some data center loads have ‘pulsed, nonlinear’ characteristics that can trigger power quality and stability problems.

Here the key sentence is this: “Laying one new transmission line can take up to 10 years.” Combining permits, planning, equipment lead times, and construction. But data center projects usually move faster than that. So from electricity perspectives, large customers line up ‘before knowing,’ and investments must be decided in states uncertain whether that line will all actually enter. NERC also points out that large loads aren’t yet registered entities directly subject to NERC reliability standards, making securing information needed for incident analysis difficult, which itself can become operational risk.

The same dilemma becomes clearer in Virginia. JLARC introduces independent forecasts that while Virginia electricity demand was nearly stagnant from 2006-2020, data center booms pulled up demand, and if “unconstrained data center demand” continues, power demand could double within 10 years. It explicitly concludes that laying generation/transmission infrastructure “sufficiently” to meet such demand is very difficult, and even meeting half is difficult.

Rate problems jump out right here. While JLARC sees current rate structures as “so far” appropriately distributing costs to data center customers, going forward it sees high possibilities of additional generation/transmission investments increasing due to data center demand and energy prices rising to raise other customer costs too. It especially warns of risks that infrastructure costs could be ‘stranded’ and remain to existing customers “if demand doesn’t realize as much as forecast or data centers close.” And it even presents estimates that Dominion Energy customers’ (residential basis) generation/transmission-related costs could increase monthly $14-37 (real terms) through 2040. At this level, it goes straight to living cost politics not just ’environment’ debates.

Numbers inside power companies also show concentration dangers. From materials Dominion Energy disclosed, data center customers are concentrated in a few (7 of 54 customers account for 72% of annual cumulative demand), forecasting 2025 data center ‘billing demand’ at 4.2GW, noting industry load factor around 90% level. High load factors mean closer to “always using lots” not just “using lots only at times.” It’s the most burdensome form from power grid perspectives.

Political Ripples Spreading to Midterms

The mechanism by which data centers become election issues is a simpler cause-effect chain than thought.

First, local governments like data centers. Because tax revenue is large, and narratives establish that they help lower rates without greatly increasing expenditures like schools and welfare. Loudoun County explains in official FAQs that property tax rates were lowered annually over the past 10 years, with the data center industry being one of factors significantly affecting budget planning.

Second, residents start disliking data centers. Especially as data centers ’too close to homes’ increase. JLARC organizes that a third of Virginia data centers are located near residential areas, with inappropriate zoning and exception approvals amplifying residential impacts. Noise is hard to catch with existing ordinances due to low-frequency and persistence characteristics, and local government authority is also unclear, making conflicts easy to prolong.

Third, electricity rates and transmission lines translate into political language. The moment economic development frames called ‘industry attraction’ change to ‘my rent/mortgage/utilities’ frames, issues strengthen everywhere nationwide. JLARC summarizing institutional risks like “data center demand may raise other customers’ costs” and “stranded cost risks exist” is literally content easy to convert into campaign promise sentences.

Fourth, actual votes move. In Prince William County, opposition to data center expansion became background for primary upsets, and super-projects like Digital Gateway exploded public hearings, transmission lines, and landscape damage issues all at once. “The moment data centers entered campaign flyers” was created. Like Washington Post-reported special election cases, this process has high possibilities of continued reproduction.

Fifth, it ultimately rises to ‘federal politics.’ While data centers themselves are mostly state/local jurisdiction as buildings, infrastructure attached to bring electricity (transmission lines, power plants, gas pipelines, sometimes federal land use) becomes entangled with federal permit/reform debates. A Congressional Research Service (CRS) report organizes that data center power demand growth requires additional generation, data centers and related energy infrastructure go through various permits, and power supply can become “core physical constraints” on data center expansion. Simultaneously it explicitly notes data center expansion and related infrastructure permits become policy focus at congressional level, with electricity rate impacts as ’emerging concerns.’ In other words, data centers no longer end as local zoning fights.

So talk of ‘shaking even midterms’ is hard to exaggerate. Data centers start like technology/industry issues, but endings always go to the same place: (1) who takes benefits, (2) who pays costs, and (3) when those costs become life they become votes. The data center curse isn’t supernatural phenomena, but very human accounting problems.

References

• Data center global power consumption (2024: 415TWh, 2030: 945TWh forecast, regional shares)
• U.S. data center power consumption (2014-2028 estimates, 2023: 176TWh/4.4%, 2028: 6.7-12%)
• U.S. data center demand forecast uncertainty and 2030 4.6-9.1% range discussion
• Northern Virginia concentration (global/Americas capacity shares, growth factors, residential impacts/noise/rates/stranded risks, 2040 residential rate impact estimates)
• Prince William Digital Gateway (scale, conflict factors, election ripples)
• Data center/power grid reliability perspectives (large load characteristics, fast ramping/metering/planning risks, 10-year transmission line construction)
• Loudoun County official materials (tax revenue/rates and data center relationships)
• Data center power infrastructure’s federal permits and policy issues (definitions, power/cooling shares, permits/rate concerns)
• Water use (northern Virginia 2023 approx 2 billion gallons, +63% vs 2019, Loudoun approx 900 million gallons)