For the past two decades, data centers were the best investment in the eyes of American local governments. They don’t emit wastewater like chemical plants, don’t blast through mountains like mines, take up relatively little land, build quickly, and generate thick tax bases. State governments raced to offer them tax breaks; local governments raced to approve their sites. A single data center investment easily reaches over a billion dollars—an astronomical figure for any small town.
But the wind has shifted. From Maine to Seattle, from Arizona to Georgia, a systemic backlash is unfolding across the United States. In the first six weeks of 2026, legislatures in over thirty states introduced more than three hundred bills restricting data centers. Seattle residents sent fifty-four thousand protest letters to city hall, forcing the city council to propose a one-year ban on new data center construction. A Gallup poll this March found that 71% of Americans oppose AI data centers being built in their communities (GeekWire).
This is an invisible shock to the AI industry. When people discuss AI bottlenecks, they typically focus on chips and algorithms. But the real constraint may be more fundamental: when the growth rate of computing demand exceeds what the power grid can handle, and when local governments start saying no to data centers, the physical expansion of AI runs into an invisible wall.
To understand this backlash, we need to first understand what role data centers played in the eyes of American governments.
In the 1990s and early 2000s, data centers drew almost no attention. They were extensions of corporate IT—like oversized server rooms. Governments saw no need to craft dedicated policies for them.
The real change came in the 2010s. Cloud computing exploded. Google, Amazon, and Microsoft began building hyperscale data centers around the world. These projects carried enormous price tags—a single campus could cost billions or tens of billions of dollars. U.S. state governments quickly realized this was a zero-sum competition: Google wasn’t going to build a data center in every state; whoever won the bid, won.
And so began a decade-long tax incentive race. Take Virginia as an example. In 2010, it expanded a policy exempting data centers from the 5.3% sales tax on purchases of servers, generators, and cooling equipment. The threshold: invest over $150 million and create at least fifty jobs. On the state’s ledger, this looked like a great deal—sacrifice a bit of tax revenue in exchange for massive capital investment and high-paying jobs.
But the math later fell apart. When Virginia first disclosed the cost of this tax exemption in 2017, the figure was $65 million. By 2023, it had soared to $750 million. In 2024, it crossed $1 billion. For fiscal year 2025, it reached $1.6 billion—sixteen times the initial annual projection (Good Jobs First).
The problem wasn’t just runaway costs. People began asking harder questions: Were the job promises being kept? Who benefited from the forgone tax revenue? Who was footing the bill for the grid upgrades that data center expansion demanded?
Before the AI data center controversy erupted, cryptocurrency had already taught Americans a lesson.
In January 2018, the small upstate New York town of Plattsburgh was hit by a cold snap. Residential electricity use surged, pushing the city beyond its cheap hydropower quota and forcing it to buy expensive replacement power. Residents saw their monthly electric bills jump by thirty to forty dollars. The culprit was traced to a cryptocurrency mine in town. It had rented out an abandoned Family Dollar store, packed it with servers, and was drawing as much electricity as four thousand homes.
Plattsburgh’s mayor, Colin Read, was an economics professor at the local university. He tallied the numbers and found that from 2016 to 2018, crypto mining had increased annual electricity bills for upstate New York small businesses by about $165 million, and for residents by about $79 million. He subsequently enacted the first cryptocurrency mining ban in the United States (MIT Technology Review).
The lesson was this: when a new type of massive electricity consumer suddenly appears in a small community, the cost of grid upgrades doesn’t automatically fall on that consumer. It spreads across every ratepayer through the electricity pricing mechanism. Crypto mines made this lesson tangible—residents saw the numbers change on their bills, not just abstract discussions in the newspaper.
In 2022, New York State passed the nation’s first statewide moratorium on proof-of-work cryptocurrency mining. This was the first time the narrative of “large electricity consumers shifting costs onto residents” entered the legislative arena. It provided a complete narrative template and legislative precedent for the later backlash against AI data centers.
By 2023, electricity demand from AI training and inference began exploding. This wasn’t gentle growth—it was an order-of-magnitude jump.
According to the U.S. Department of Energy, nationwide data center electricity consumption grew from 76 TWh in 2018 to 176 TWh in 2023, more than doubling. Projections for 2028 range from 325 to 580 TWh, potentially reaching 7% to 12% of total U.S. electricity consumption (LBNL, 2024). Individual data centers are also swelling in scale: from 2023 to 2024, the average electricity load per data center jumped from 150 MW to 300 MW. For reference, a 300 MW data center uses roughly as much electricity as three hundred thousand homes.
But what really creates the conflict isn’t the total consumption. It’s the temporal mismatch in construction speed.
A data center can go from groundbreaking to operation in eighteen to twenty-four months. The new transmission lines needed to support it, from planning, permitting, land acquisition, to construction, take seven to fifteen years. In 2024, the entire United States built only 322 miles of new high-voltage transmission lines—the third-lowest year in the past fifteen years. By comparison, the Department of Energy estimates roughly 5,000 miles per year are needed (ACEG/Grid Strategies).
Put another way: data centers can be built quickly, but connecting them to the grid takes many years. When dozens of large data centers queue up simultaneously for grid connection, there are two ways to handle it: make them wait, or accelerate transmission construction and spread the costs across all users. Historically, the second approach happened quietly. In 2024, within just the PJM regional grid (covering thirteen northeastern states), 130 transmission projects were built solely to connect data centers, at a cost of $43.6 billion—95% of which was allocated across all electricity users in the region (UCS).
When residents find themselves paying dozens of extra dollars a month on their electric bills while the data center next door enjoys tax breaks and cheap power, the political backlash becomes inevitable.
By 2025 and 2026, responses across the United States began to diverge. They can be grouped into roughly five categories.
The first group consists of states that are aware of the cost but haven’t yet tightened policies systematically. Texas is still attracting data centers, but state auditor data shows that the sales tax exemption alone will cost $3.2 billion over the 2025–2026 biennium (Texas Tribune). Ohio’s data centers received $2.5 billion in state and local tax incentives between 2017 and 2024 (Signal Ohio). These numbers are entering legislative discussions. The question has shifted from “should we attract them?” to “who bears the cost of attracting them?”
The second group is where the backlash is fiercest—places that have already built too many data centers. Virginia is home to the world’s densest concentration of data centers, but its Loudoun County made a landmark decision in March 2025: eliminating automatic approval rights for data centers. Previously, data centers in certain zones were approved by default. Now every new project requires public hearings and a vote. The same Virginia lost $1.6 billion from its data center tax exemption in 2025. Georgia faces a similar situation, with bipartisan legislators jointly proposing a pause on new data centers.
The third group doesn’t ban construction, but reprices it. In 2025, Oregon passed a law that pulls data centers out of the standard industrial electricity rate and creates a separate rate class. Developers must sign power purchase agreements of at least ten years, lock in minimum consumption commitments, and pay even if actual usage falls below projections. The logic is straightforward: over the past five years, Oregon’s residential electricity rates rose nearly 50%, while the large industrial rate class—where data centers sit—barely moved (Oregon Legislature). The added costs can’t fall solely on residents.
The fourth group has hit pause directly. Both chambers of Maine’s legislature passed a bill to suspend construction of large data centers. Although the governor ultimately vetoed it, the act itself signals the political direction. Seattle’s city council is advancing a one-year ban. The trigger was a specific event: five companies proposed building five large data centers in Seattle, with total demand reaching 369 MW—roughly one-third of the city’s average daily electricity consumption. Community organizations launched mass protests. Two developers withdrew their plans. The mayor publicly stated she was “exploring a moratorium” (Seattle City Council).
The fifth group hasn’t yet formed clear policies. Eleven companies have expressed interest in building data centers in Montana, but the state-level regulatory framework hasn’t been established. Alaska’s governor is even actively inviting tech companies—one of the few places still treating data centers as a purely positive investment.
These five categories aren’t random. They resemble stages on an evolutionary path. Kansas just established a data center tax exemption program in 2025—roughly equivalent to where Virginia was in 2010. Virginia itself is transitioning from the second category to the third, already requiring large data centers to pay higher grid connection fees.
This backlash is not being driven by a single group.
In Chandler, Arizona, residents have been complaining since 2014 about the 24-hour low-frequency hum from data center cooling systems. In Wisconsin, Microsoft tried to convert 244 acres of farmland into an industrial zone for a data center, only to withdraw amid community opposition. In a farming township in Michigan, a developer backed by OpenAI and Oracle attempted to rezone 575 acres of farmland for a data center. The township board voted no, 4-to-1. The developer then sued the township (Fortune).
The opposition comes from multiple groups that barely overlap. Neighbors plagued by noise. Taxpayers worried about rising electric bills. Farmers unwilling to see their land permanently converted to industrial use. Government auditors questioning the fairness of tax incentives. Transparency advocates opposing secret developer negotiations. And, yes, environmental groups. At least nineteen townships in Michigan have passed local moratoriums, but the driving force often came from the townships’ own defensive reaction: they realized their zoning codes were designed assuming office-building levels of development intensity—not industrial behemoths capable of consuming a third of a city’s electricity.
Gallup’s poll number—71% of Americans opposing AI data centers in their communities—is something policymakers can’t ignore. This number means the issue is not niche. It has become an electoral variable in mainstream politics.
If you compress all the evidence above into one judgment, the answer is: AI construction won’t stop, but where it gets built, how it gets built, and how much it costs to build will systematically change.
First, site selection flexibility is narrowing. As Virginia, Oregon, and Washington raise the barriers to data center construction to varying degrees, new training clusters will concentrate in states with abundant power and permissive approvals. Texas and Indiana are the most obvious beneficiaries. But this also means a handful of low-friction states will bear disproportionate grid and water resource pressure, potentially triggering their own backlash cycles.
Second, costs are systematically rising. Oregon’s long-term power purchase agreements and separate rate class, and Virginia’s minimum consumption commitments, essentially shift stranded-asset risk from ordinary ratepayers to data center developers. This isn’t about whether you can build—it’s about construction getting more expensive and financing getting more complicated. Projects that once could break ground as soon as they secured a cheap power line now need ten-plus-year power contracts, locked-in minimum consumption commitments, and acceptance of exit penalties.
More importantly, the impact of policy restrictions differs between AI training and inference. Training clusters are more sensitive to electricity costs and less sensitive to network latency; they’ll concentrate in places with cheap power and permissive approvals. Inference nodes are the opposite: they’re latency-sensitive and need to be near users, but individual sites are smaller. And inference needs to be distributed near urban population centers—exactly where resident opposition and zoning resistance are strongest. Loudoun County’s elimination of automatic approvals directly impacts inference deployment, because inference requires placement in dense population areas where approvals are hardest to obtain. Boiled down: government restrictions on data centers affect training by shifting it elsewhere, and affect inference by slowing it down.
The industry chain is already responding. Microsoft signed a twenty-year agreement to restart Pennsylvania’s Three Mile Island nuclear plant. Oracle is deploying fuel-cell solutions deliverable within ninety days to skip the grid connection queue. Google spent $4.75 billion to acquire a renewable energy developer. Meta is working with a nuclear company to develop a 1.2 GW nuclear campus. The common thread across these moves is that they aren’t responses to policy restrictions—they’re bypassing the ultimate bottleneck, which is the speed of power system construction.
Improvements in model efficiency are also partially offsetting demand growth. Techniques like quantization, distillation, and sparse attention can reduce inference energy consumption by 70% to 80%. But historical experience suggests efficiency gains tend to unleash more demand rather than reduce total consumption. Total electricity use is still growing rapidly.
Taken together, the bigger impact of policy resistance is not to obstruct AI development, but to redistribute the costs of AI infrastructure, alter its geographic distribution, and filter for a set of builders capable of handling power, land, regulation, and community relations simultaneously. The number of companies that can train large models will continue to grow. The number that can coordinate local permits, long-term power contracts, and community relations will be much smaller.
Back to Seattle. After fifty-four thousand protest letters and a proposed moratorium, the local paper GeekWire documented a public hearing. Residents expressed fear of AI, called data centers “gifts to the rich,” and shared worries about rising utility bills and environmental damage. City Councilmember Joy Hollingsworth’s response: “We’re not trying to hinder growth in our city. But we do need to slow down and understand the impacts of these facilities.”
That posture of slowing down may be the most accurate description of how America will treat data centers in the coming years. Not a ban. Not a welcome. Just “we need to figure out the bill first.”