The Backlash in My Parents' Backyard

Communities blocked $130 billion in data centers in three months. The real constraint on AI isn't chips or models anymore — it's whether the neighborhood will let you plug in.

EXECUTIVE SUMMARY

For two years the AI scaling story has been about silicon and capital: who has the most GPUs, who can raise the most money, who ships the best model. That story is hitting a wall nobody priced in — the local zoning meeting. Across the country, communities are blocking the data centers that AI runs on, and they're winning. The binding constraint on AI is shifting from compute to the physical and social license to build the power that compute requires. Every leader betting their strategy on infinite cheap inference needs to understand why, because it's about to show up in their cloud bill and their siting timelines.

// In the first four months of 2026, communities stalled or blocked at least 75 data center projects worth roughly $130 billion, and at least 69 local governments had enacted outright bans by May.

// This is a policy wave, not a few angry town halls: lawmakers in more than 30 states introduced over 300 data-center-related bills in 2026, and the opposition is bipartisan — a 40,000-acre project in deep-red Box Elder County, Utah is facing the same revolt as projects in blue suburbs.

// The grievance is concrete and personal: data centers could jump from 4.7% of U.S. electricity in 2024 to as much as 15.3% by 2030, with ordinary ratepayers footing much of the bill — even as 51 major utilities raise their build-out plans 27% in a year, to roughly $1.4 trillion, making the collision between AI demand and household costs structural.

// And this is a uniquely American constraint, which is what makes it a race: China already generates more than twice as much electricity as the United States and has been expanding total generation nearly 6% a year — over half of it from clean sources — so its AI buildout faces no equivalent brake.

The leaders who win this won't be the ones who assume compute is a solved, ever-cheapening input. They'll be the ones who treat power, water, and community consent as first-class variables in their AI strategy — and who plan for a world where inference has a rising floor price and a social cost attached.

// The Deep Dive

I visited my parents near Harrisburg, Pennsylvania earlier this year, and watched the entire AI energy story play out over a single weekend. Their neighborhood is fighting a data center — not a zoning squabble over a strip mall, but a hyperscale campus proposed down the road, and the whole community is organized against it: yard signs, a packed county meeting, a Facebook group with more members than the town has voters. My mom, who could not tell you the difference between a GPU and a garden hose, had questions, as most people do.

That same week, I drove over the Susquehanna River past Three Mile Island. My parents live a short drive from the reactor that gave America its worst nuclear scare in 1979 — and that Microsoft has now effectively reserved, signing a 20-year deal for Constellation to restart the shuttered plant and buy its entire output to feed AI data centers. The same area fighting one data center is watching a mothballed nuclear plant come back to life to power other data centers. That was the whole story in one weekend: enormous demand, a community pushing back, and a scramble for electricity so intense it's reviving reactors we had left for dead.

That stopped me. Because I've spent two years telling business audiences that AI's constraints were melting away — models getting cheaper, compute getting cheaper, the whole thing bending toward abundance. And here was the actual frontier of the AI buildout, and it wasn't a lab in San Francisco. It was a county commission meeting where regular people had decided the answer was no.

I've watched the internet's physical layer provoke this before, but never like this. In the dot-com era we laid fiber and nobody noticed — it went in the ground and the boom was abstract, financial, invisible. The open source movement won precisely because it had no physical footprint to fight; you couldn't picket a GitHub repo. AI is different. AI is the first software revolution with a body — enormous, power-hungry, thirsty, and parked in somebody's backyard. And that body is now the thing standing between the AI industry and its growth projections.

This is the part of the AI story that the model-benchmark coverage misses entirely. You can have the best model in the world and a balance sheet full of GPU orders, and none of it matters if the local grid can't carry the load or the county won't grant the permit. The constraint moved. It used to be talent, then it was chips. Now, increasingly, it's megawatts and consent. If you're making AI bets for your business, that shift should reshape how you think about cost, reliability, and risk.

For a long time, data centers were the most boring buildings in the economy — windowless boxes that created a handful of jobs and a lot of property tax. Communities competed to land them. What changed is scale and visibility. AI inference and training need vastly more power than the cloud workloads that came before, and the demand curve went vertical.

The numbers explain the panic. A January 2026 analysis projected U.S. data center power demand would nearly double between 2025 and 2028, from 80 to 150 gigawatts. Lawrence Berkeley National Laboratory put data centers at 4.7% of all U.S. electricity in 2024, heading toward 9.5% to 15.3% by 2030. To serve that, 51 utilities covering 250 million customers raised their collective capital plans to about $1.4 trillion, a 27% jump in a single year.

Here's the political detonator: who pays for that grid expansion. When new transmission and generation get built to serve a data center, the cost often lands in everyone's rates. In the PJM market — the grid operator that runs the electricity system for 13 mid-Atlantic and Midwest states plus Washington, D.C., and the largest such market in the country, serving roughly 65 million people — the cost of providing capacity for new data centers was estimated to add roughly $9.3 billion in energy-market costs, about $18 a month on some households' bills. In Virginia, Dominion proposed its first base-rate increase since 1992. When a retiree's power bill goes up so a trillion-dollar company can train a model, the politics write themselves.

That's why this backlash is bipartisan in a way almost nothing else is right now. The opposition spans a 40,000-acre fight in Box Elder County, Utah — where Trump won nearly 80% — and progressive suburbs alike. Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez introduced an AI Data Center Moratorium Act in late March 2026. When Bernie and a red rural county want the same thing, you're not looking at a fringe — you're looking at a coalition.

Let me argue the optimistic side honestly, because the AI industry has real answers and a lot of money behind them.

First, the companies see the problem and are moving to defuse it. In a March 2026 pledge, Microsoft, Meta, OpenAI, and Amazon agreed to secure their own power and pay for the transmission infrastructure their data centers require, rather than socializing those costs onto ratepayers. If that holds — a real if — it pulls the teeth from the single sharpest grievance.

Second, there's a lot of slack in the existing system. Analysts point out that grid-enhancing technologies and demand response can absorb a meaningful share of near-term data-center load without a wave of new plants — data centers can be made flexible, throttling during peak demand. Add on-site generation, new nuclear and geothermal deals, and efficiency gains in the chips themselves, and the demand curve may bend.

Third, the economic pull is enormous and bipartisan too. The same politics that produce angry town halls also produce governors who desperately want the investment, the construction jobs, and the tax base. Capital this large tends to find the jurisdictions that say yes. The buildout may slow and relocate, but "slow and relocate" is friction, not a wall.

Now the other side, which I find harder to dismiss than I'd like.

The pledge to self-fund power is a promise, not a mechanism, and it doesn't address the deeper issue: there is only so much power, water, and transmission, and it takes years to build more. You can't demand-response your way out of a 70-gigawatt gap, and the community concerns aren't only about who pays — they're about water scarcity, noise, land use, and a basic sense that the trade is lopsided. Even a perfectly funded project still has to win the permit, and 69-plus local bans say the permits are getting harder, not easier.

And the friction has a price even when projects eventually get built: delay. A data center deferred two years by litigation and rezoning is compute that doesn't exist when the model needs it. Multiply that across $130 billion in stalled projects in one quarter, and the aggregate effect on supply — and therefore on the price of inference — is real. The honest synthesis: the bull case is probably right that AI infrastructure keeps growing, and the bear case is probably right that it grows slower, costs more, and gets built in different places than the spreadsheets assumed. For a business planning around AI, "slower and more expensive than the spreadsheet" is the number that matters.

There's a dimension to this fight my parents' county meeting never surfaced, and it's the one that keeps AI executives awake. The power crunch throttling data centers is largely an American and European problem — China is running the same race without the same brake. OpenAI's leadership even coined a name for the danger, the "electron gap." And the gap is concrete: Chinese data centers pay less than half the electricity rates their U.S. counterparts do, and projects there move from planning to operation in months rather than the years American permitting requires.

The scale gap is not subtle. As recently as 2005, the United States generated roughly twice as much electricity as China; two decades later the positions have completely reversed, with Chinese output up nearly fivefold while the mature U.S. grid grew less than 1% a year. China now produces more than twice the electricity the U.S. does — the exact inverse of the picture a generation ago.

That lead is a function of how fast China builds. In 2024 alone it added 277 gigawatts of utility-scale solar and more than 50 gigawatts of thermal generation, against roughly 40 gigawatts of solar and 2.5 gigawatts of thermal in the United States — a reminder that the ability to pour concrete and energize turbines is itself a competitive advantage. Its combined wind and solar capacity more than doubled in the three years to 2024, from 635 to 1,408 gigawatts, and it installed more battery storage in 2024 than the U.S. and EU combined.

The starkest symbol of the gap isn't a statistic — it's a specific reactor. The molten salt reactor, a design originally developed by the U.S. Department of Energy at Oak Ridge National Laboratory in the 1960s, was abandoned in America and left to gather dust. China picked it up: its TMSR-LF1 thorium molten salt reactor in the Gobi Desert reached criticality in 2023 and in late 2025 became the first in the world to convert thorium into usable reactor fuel, a feat no other country has managed. So while the U.S. races to resurrect the 1970s reactor near my parents' home to power data centers, China is piloting the reactor America invented and then walked away from.

None of this means China is winning outright. The U.S. still leads in the chips, the models, and the capital, and its data-center capacity — on track to roughly 95 gigawatts by the end of 2027 — still outweighs China's projected 60-plus gigawatts by 2030. The American constraint isn't ambition or money; it's the grid and the permit. And that's the uncomfortable synthesis of this whole story: the same local democracy that gives my parents a real voice over their neighborhood is the thing China never has to negotiate with. Whether that reads as a bug or a feature depends on whether you're the retiree getting the power bill or the executive watching the electron gap widen.

Most companies will never build or site a data center — but you still inherit the fallout from this fight. When the supply of compute is capped by power and permits while demand keeps climbing, the long, smooth decline in inference prices that everyone has baked into their budgets gets bumpier. Cheap, abundant AI is a planning assumption, not a law of nature, and the grid is where that assumption collides with reality.

There's also a reputational and operational dimension. As communities sour on the AI buildout, "we run everything on frontier AI in the cloud" stops being a pure efficiency story and picks up an externality — energy, water, local opposition — that customers, employees, and regulators are starting to notice. This connects directly to last week's theme: the efficiency move of routing routine work to smaller, cheaper, more efficient models isn't just about your bill, it's about your exposure to a resource constraint that's tightening. The companies that treat compute as scarce and precious will be better positioned than the ones who treat it as free and infinite.

Misstep 1: Treating compute cost as a one-way downward bet. Plans that assume inference gets monotonically cheaper are betting against a power and permitting crunch that's already raising costs. Build sensitivity analysis for flat or rising compute prices into any multi-year AI plan.

Misstep 2: Ignoring the energy externality in your AI narrative. Leaders happily tout AI efficiency while ignoring that the public increasingly associates the AI buildout with higher electric bills and water stress. If your story has no answer on energy, you're exposed the moment a customer or reporter asks.

Misstep 3: Assuming your cloud region is immune. Capacity, reliability, and pricing in a given region depend on local grid and permitting fights you don't see. Concentrating workloads in a contested region is a hidden single point of failure; diversify and ask your provider where the power is actually coming from.

Misstep 4: Confusing the pledge with the fix. The hyperscaler promise to self-fund power is encouraging but unproven at scale. Don't plan as if the ratepayer fight is over; plan as if siting and energy remain contested for years, because the structural gap says they will.

// Key Takeaways

Price compute as a variable that can rise, not just fall. Stress-test your AI roadmap against flat or higher inference costs driven by power, permitting, and a global race in which the cheapest power — and eventually the cheapest inference — may get built offshore where it can be energized fastest. The plan that only works if AI keeps getting cheaper, and cheapest at home, is a fragile plan.

Make efficiency a strategy, not a footnote. Routing routine work to smaller and open source models, caching, and batching aren't just cost tactics — they reduce your exposure to a tightening resource constraint. Treat compute as scarce and you'll be ahead of the firms that assume it's free.

Get an honest answer on where your AI's power comes from. Ask your cloud or model provider about the grid, water, and community status of the regions serving your workloads. Concentration in a contested region is operational risk hiding as a line item.

Build an energy answer into your AI story now. Before a customer, employee, or regulator asks, know what your AI consumption costs the grid and what you're doing about it. The companies with a credible answer will keep their social license; the ones caught flat-footed won't.

The playbook for your next planning cycle is concrete. Add one slide to your AI strategy review titled "What if compute gets more expensive?" and actually model it. Ask your providers, in writing, where the power serving your regions comes from and how exposed those regions are to permitting fights. Audit your own workloads for the routine 80% that could run on cheaper, more efficient models. And decide, before you're forced to, what you'll say when someone connects your AI usage to their electric bill. The fight in my parents' neighborhood isn't a local-news curiosity. It's the AI industry's growth model meeting the grid, and the grid is pushing back. Plan like the constraint is real, because in 75 communities and counting, it already is.