AI is no longer being judged only as software. Once the buildout arrives as a massive industrial facility, new transmission lines, continuous power demand, and possible pressure on utility rates, the argument changes. Communities are being asked to absorb costs they can see immediately for benefits that remain distant, uncertain, and spread across people and companies somewhere else.
I think the industry keeps treating this as a communications problem when it is really a legitimacy problem. Secrecy, aggressive dealmaking, and repeated warnings about job displacement have made people question not only whether a particular facility is worthwhile, but whether the broader AI project is being pursued on terms they would ever choose. The fact that disruption is arriving before the promised cures and abundance makes that problem much harder.
There is no simple policy answer. Blocking projects locally may move compute to less accountable places and make the best systems even more exclusive. Building without reform will deepen resistance. The more durable path is to make the deals transparent, protect communities from the downside, and demonstrate value that ordinary people can recognize and expect to share.
Related content
- Jasmine Sun in conversation with Ezra Klein
- The “Data Center Rebellion” is here
- The Bear Case for AI Data Centers
Subscribe to our weekly newsletter
Appendix: The New Politics of AI Infrastructure
Table of contents
- Why Data Centers Have Become the Face of AI
- How the Industry Turned Development Into a Trust Crisis
- The Environmental Argument Is Uneven
- The Economic Case Is Stronger Than Many Opponents Admit
- AI Populism Is About Power and Consent
- The Industry’s Own Contradictions Are Fueling the Backlash
- The Benefits Are Arriving Too Slowly
- Labor Risk Is Primarily About Bargaining Power
- The Race Narrative Has No Clear Finish Line
- The Builders Rely on Three Moral Rationales
- Moratoria Create Their Own Tradeoffs
- China Suggests a Different Model of AI Adaptation
- What Silicon Valley May Be Learning Too Late
Why Data Centers Have Become the Face of AI
- AI’s abstract promise now has a highly visible physical footprint. Hyperscale data centers are enormous, windowless industrial buildings surrounded by transmission lines, towers, and electrical infrastructure. In Port Washington, one complex takes one minute and 42 seconds to pass while driving at 70 miles per hour. Nearby residents also hear continuous humming, buzzing, whirring, and occasional rattling. That sensory experience makes the AI buildout much harder to treat as an abstract debate about software.
- The costs are concentrated while the benefits are diffuse. The people living near a data center experience the noise, construction, altered landscape, and potential grid pressure. The benefits are spread across distant users, AI companies, and large corporate customers. That is a politically difficult arrangement because the people bearing the most visible costs may not believe they will receive a meaningful share of the upside.
- This is broader than conventional NIMBYism. Opposition is not limited to people who live beside a proposed facility. Opposition appears to be closer to 70 percent, including among people with no nearby project. My read is that the data center has become a proxy for a larger question: Do people want AI development to proceed this quickly and on these terms?
- AI lacks a natural constituency willing to defend the infrastructure. Solar farms have environmental advocates. Auto plants have workers, unions, suppliers, and families who can point to a tangible product. Data centers are mainly defended by utilities, developers, and AI companies, all of which begin with low public trust. There is no equally visible group of ordinary people saying that their lives depend on building more of them.
How the Industry Turned Development Into a Trust Crisis
- Nondisclosure agreements created the appearance of secrecy and collusion. Local officials were sometimes unable to disclose the developer, customer, project size, or anticipated electricity demand. The information still leaked through contractors and workers, so rumors spread while government officials remained silent. By the time the facts emerged, many residents had concluded that the company and city council had been working behind closed doors.Counterpoint. The use of nondisclosure agreements may have been more reflexive than conspiratorial. Companies routinely use them and may not have anticipated the political consequences. That does not change the result, which was a severe and largely avoidable loss of trust.
- Low trust makes even legitimate benefits difficult to sell. Companies can promise to pay for grid upgrades, prevent household electricity rates from rising, treat wastewater responsibly, and create hundreds of jobs. In many communities, the recurring response has typically been, “I don’t believe them.” Once credibility collapses, better presentations and more impressive projections do very little.
- Every new developer inherits the failures of earlier corporations. Wisconsin residents remember Foxconn, which promised 13,000 jobs after receiving substantial public support but produced closer to 1,000. Janesville residents remember the GM plant that closed and left behind an estimated $30 million in contamination. A data center developer may have had nothing to do with those failures, but local communities evaluate new promises through that history.
- Legal victories can become political defeats. A small township voted 4 to 1 against rezoning land for a data center. The developer sued, and the small township eventually settled rather than continue fighting a much richer opponent. Even if the settlement provided money for schools and public services, the sequence reinforced the idea that a sufficiently wealthy company can override a democratic no.
- More money can make people more suspicious. Silicon Valley often assumes that opposition means the offer is not generous enough. In the current political environment, a larger payment can look like evidence that the company is purchasing consent. Compensation may be necessary, but it cannot substitute for transparency, enforceable protections, and a genuine ability to influence the decision.
The Environmental Argument Is Uneven
- Water has become the most memorable symbol, but not necessarily the largest problem. Many newer data centers use closed-loop cooling systems that recycle water. Their consumption may be far below what opponents imagine and, in some locations, below the usage of nearby golf courses. Yet water remains politically powerful because it represents the broader fear that an outside company is consuming local resources for a technology residents do not value.Counterpoint. Closed-loop cooling does not mean zero water use, and every facility is different. Communities are justified in demanding project-specific information rather than accepting an industry-wide assurance.
- Electricity demand is the more substantial infrastructure issue. AI clusters require enormous amounts of continuous power for processors, networking, and cooling. That can require new transmission lines, generators, and power plants, with natural gas likely to supply at least part of the near-term increase. The facility itself may produce little pollution while still driving emissions elsewhere on the grid.
- Who pays for the grid may matter more than how much power the facility consumes. Residents want to know whether the developer will cover the full cost of substations, transmission, and new generation. They are skeptical of promises that household rates will remain unchanged, especially in places where utilities have raised rates repeatedly.
- Correcting environmental exaggerations will not eliminate the backlash. Some claims about water may be outdated or overstated, but opposition is rooted in a much broader judgment about the industry. People frequently use an environmental concern to express a deeper objection involving trust, power, fairness, and consent.
The Economic Case Is Stronger Than Many Opponents Admit
- Property-tax revenue can be transformative. Microsoft was expected to pay approximately $19.6 million in 2026 property taxes in Mount Pleasant, a village of around 28,000 people. Revenue on that scale can materially improve schools, public safety, infrastructure, and local government finances.
- Construction jobs are real economic benefits. A project may provide roughly 500 well-paid jobs for two to six years, including apprenticeship-based positions that do not require a four-year degree. That is enough time for workers to support families, purchase homes, and establish careers. Temporary employment should not automatically be dismissed as meaningless.Counterpoint. Once construction ends, a data center generally supports far fewer permanent employees than an auto plant, semiconductor facility, or similarly large industrial project.
- Some data centers are being proposed for sites that badly need investment. The former Foxconn property already had cleared land and expensive infrastructure. The abandoned GM property in Janesville contained contamination that had deterred other buyers, while the data center developer was willing to pay for cleanup. A facility on a damaged industrial site presents a much stronger public case than one built on productive farmland.
- Communities reasonably fear being left with a stranded asset. AI demand could disappoint, a developer could lose the competitive race, or newer hardware could make an existing facility less attractive. Residents worry that the company can leave while the community remains responsible for roads, power infrastructure, an unfinished project, or an obsolete building.
- The bargaining power of communities is increasing. Two years ago, companies were looking for sales-tax exemptions, subsidies, and other incentives. As opposition has grown, politically welcoming sites have become scarcer. Some local officials now believe they could reject subsidies and demand larger community benefit packages. Resistance may not stop construction, but it can materially change the economics of the deal.
- The central policy problem may be deal design rather than data centers themselves. A small town should not have to independently negotiate electricity protections, cleanup guarantees, tax arrangements, noise limits, and decommissioning obligations with a trillion-dollar industry. State or federal frameworks could establish minimum disclosure, benefit, and risk-sharing standards while still allowing communities to reject projects that do not fit.
AI Populism Is About Power and Consent
- People increasingly see AI as an elite political project. The recurring complaint is not that AI has no use. It is that a small number of executives and investors appear able to impose it on workplaces, schools, communities, and public institutions without meaningful consent. “Why is this being forced on us?” may become a more important political question than “How capable is the model?”
- The public does not yet see AI as essential. Many people have used ChatGPT or Gemini to draft an email or create a meme. That does not mean they view it like housing, transportation, electricity, or healthcare. The enormous valuations and infrastructure demands appear disconnected from the modest value many people currently experience in daily life.
- The backlash appears organic rather than manufactured. Foreign governments may have attempted to amplify opposition, but those efforts appear to have attracted negligible engagement. Many organizers understood technical distinctions such as closed-loop versus open-loop cooling. They were not necessarily confused. They had simply decided that the project was not worth the trade.
- Data center opposition is producing unusual political alliances. Environmental activists, labor-oriented populists, social conservatives, and AI safety advocates are converging on a shared desire to slow development. Their reasoning differs, but the policy conclusions increasingly overlap.
- Local fights are giving people a sense of political agency. Residents may have little influence over national AI policy, but they can attend zoning hearings, organize neighbors, and pressure local officials. Some liberal activists said data centers were the first issue in years that allowed them to work productively with Trump-voting neighbors.
The Industry’s Own Contradictions Are Fueling the Backlash
- AI companies are warning about harms while building the systems that could cause them. Industry leaders discuss mass white-collar displacement, severe inequality, loss of control, and catastrophic risk. At the same time, their companies are selling coding, banking, design, and administrative agents intended to reduce reliance on human labor. The business model depends on the disruption they describe as dangerous.
- “This technology may be dangerous, so we must build it faster” is not a persuasive public argument. Executives may privately wish development were slower, but they continue accelerating because they believe competitors or China will do the same. That logic may make sense inside the industry, but it gives communities little reason to help.
- The companies originally communicated with recruits and investors, not the public. Claims about artificial general intelligence, the end of work, existential danger, and the transfer of power from labor to capital signaled seriousness and ambition within Silicon Valley. The companies appear not to have anticipated that workers, voters, politicians, and community activists were listening too.
- The attempt to soften the rhetoric may deepen suspicion. Leaders now have reasons to seek political goodwill, infrastructure approvals, and public-market investors. When their warnings become less dramatic, the public will reasonably ask whether the underlying beliefs changed or whether only the communications strategy changed.
- The industry may have a product problem rather than a marketing problem. Better branding cannot fix a future that people do not want. If companies truly expect large-scale displacement, widening inequality, and meaningful catastrophic risk, public opposition is not simply the result of poor communication.
The Benefits Are Arriving Too Slowly
- The public is seeing disruption before abundance. The industry has promised medical breakthroughs, cheaper goods, scientific discovery, and radically higher productivity. Many people instead see threatened jobs, new industrial facilities, higher power demand, and extraordinary wealth accruing to a small group. If the losses arrive first, the political coalition supporting rapid deployment may not survive long enough to see the gains.
- Technical achievements do not necessarily correspond to public priorities. AI researchers may be thrilled by advanced theorem proving or scientific reasoning. Most people want lower living costs, improved health, less unpleasant work, and more time with family and friends. Technical impressiveness will not automatically produce political legitimacy.
- Distribution is part of the promised benefit. A medical breakthrough does not create broad enthusiasm if people believe only billionaires, large corporations, or governments will be able to afford it. People worry that their employer will gain access to systems capable of automating their work while they receive no comparable access or bargaining power.
- The AI industry’s imagined utopia may not match what most people want. Some builders describe a world of universal basic income, no need to work, radically extended life, and machines that discover all of mathematics and physics. Many people derive meaning, status, relationships, and independence from work. They may not regard permanent dependence on transfers as an attractive outcome.
Labor Risk Is Primarily About Bargaining Power
- The permanent-underclass scenario is a loss-of-leverage scenario. If machines can perform most economically valuable work, capital owners can purchase machine labor rather than hire people. Workers may still receive welfare payments and inexpensive goods, but they would lose much of the bargaining power, mobility, and independence that employment provides.
- Even industry insiders generally expect inequality to increase. Almost nobody building frontier AI appears to believe it will reduce inequality by default. Some expect living standards to rise even as the gap between owners and workers widens, but that is a different promise from broadly shared power or prosperity.
- Complete job replacement may be less likely than the most dramatic predictions suggest. Human jobs combine many tasks, relationships, judgments, physical activities, and forms of unstated knowledge. AI capabilities remain uneven. A model that excels at coding or mathematics may still struggle with the full complexity of an ordinary job.Counterpoint. AI does not need to replace every task to reshape employment. Automating enough valuable work can reduce hiring, compress wages, or eliminate entry-level roles even when humans remain necessary.
- The pace of change may be more damaging than the eventual endpoint. Workers can adapt to technological change when they have time to retrain and move into new roles. The difficult question is whether a person can learn a new specialty faster than the next generation of models improves at that same specialty.
- Entry-level jobs may become the first structural casualty. Senior engineers can become more productive with AI tools while companies hire fewer junior engineers. That may improve short-term output but weaken the pathway through which inexperienced workers become experts. The industry could consume a talent pipeline it still depends on.
The Race Narrative Has No Clear Finish Line
- Artificial general intelligence has no accepted definition. It may mean a system that can perform all human jobs, generate a certain amount of economic value, or independently develop the next generation of AI. Different companies use different milestones, which means the race has multiple and constantly moving finish lines.
- Uneven capabilities complicate the idea of a single winner. Models can be excellent at mathematics or cybersecurity while remaining weak at physical work, social judgment, or unfamiliar games. That makes it difficult to identify one moment when a company has decisively crossed into general intelligence.
- A race that cannot be judged may never end. Every laboratory can always find a benchmark, capability, or competitor that makes it appear behind. The metaphor therefore creates a permanent justification for acceleration, even in the absence of evidence that any one company is pulling decisively ahead.
- The strongest race argument is recursive improvement. The major laboratories are attempting to build systems that can contribute to AI research and improve the next generation of models. If one company achieved a compounding advantage, being first could matter enormously.Counterpoint. Leading laboratories are all pursuing similar methods, and there is no clear evidence that one has already established a decisive, self-reinforcing lead.
- The industry faces a genuine collective-action problem. Many executives want slower progress but will not slow down unilaterally. They fear that competitors or Chinese laboratories would continue. Without enforceable coordination, every participant can describe acceleration as a reluctant defensive action.
- Technological inevitability provides moral cover. Many builders believe advanced AI will be created regardless of what any one person or company does. From that assumption, the only question becomes who builds it and whether they can make the process marginally safer. The belief in inevitability does much of the work of justifying continued acceleration.
The Builders Rely on Three Moral Rationales
- The inevitability rationale. Some believe superintelligence cannot be prevented, so they want to influence who develops and controls it. Working on the technology becomes an attempt to ensure that a more responsible organization reaches it first.
- The expected-value rationale. Some believe even severe risks are justified by the possibility of curing disease, extending life, eliminating scarcity, and creating extraordinary abundance. The potential upside is treated as large enough to offset even a meaningful probability of catastrophe.
- The technical-fascination rationale. Some builders are deeply motivated by the challenge of discovering whether advanced intelligence can be created. The technical question is so compelling that social and political consequences remain secondary.
- None of these arguments asks the public for meaningful consent. Founders and researchers may voluntarily accept extreme risks in their careers and companies. That logic becomes much harder to defend when the risks involve other people’s communities, livelihoods, political power, and long-term security.
Moratoria Create Their Own Tradeoffs
- Local restrictions may relocate compute rather than reduce it. If one town or state blocks construction, developers can move to Texas, Tennessee, the Dakotas, another country, or eventually a less conventional location. A patchwork of local bans may alter where AI is built without substantially changing the overall pace of development.
- Moving infrastructure can reduce democratic oversight. If projects become difficult to build in communities with strong public review, companies may favor jurisdictions with weaker environmental standards or less political accountability. Efforts to increase democratic control locally could unintentionally push development into more authoritarian settings.
- Compute is becoming geopolitical leverage. Countries may offer land, power, and favorable regulation in exchange for access to frontier models, cybersecurity capabilities, or influence over AI policy. Data center construction is therefore becoming part of foreign policy as well as energy and land-use policy.
- Scarce compute may deepen inequality. Demand currently appears to exceed supply. If capacity remains constrained, large banks, technology companies, and governments will be able to outbid startups, small businesses, and individual users for the most capable systems.Counterpoint. Additional capacity does not guarantee broad access. Companies may still reserve their best models for the largest customers because of operating costs, safety concerns, or commercial strategy.
- The frontier may be closing even as AI becomes more capable. The most advanced models are expensive to run and increasingly likely to be offered selectively. That raises the possibility that employers will gain access to systems capable of replacing work while workers and small companies receive only weaker versions.
China Suggests a Different Model of AI Adaptation
- Lower visible resistance may reflect fatalism rather than optimism. The prevailing attitude in China is not necessarily that AI will create abundance. It was closer to a belief that technological development is unstoppable once the state makes it a national priority. The individual response is therefore to adapt rather than resist.
- High adoption does not necessarily mean enthusiastic adoption. In a fiercely competitive labor market, workers may use AI because they fear falling behind colleagues or job applicants. Usage numbers alone cannot tell us whether people welcome the technology.
- The Chinese state has paired acceleration with more visible social controls. China has imposed restrictions on some companion chatbots, labeling requirements for AI-generated content, and legal limits on dismissing a worker simply because AI can perform the role. Some citizens may therefore expect the state to manage social harms more actively than the US government has.Counterpoint. Public concern is difficult to measure in an environment where dissent and polling are constrained. The absence of visible backlash should not be treated as proof of public acceptance.
What Silicon Valley May Be Learning Too Late
- “AI will solve it” is not an implementation plan. AI can contribute to climate research, medicine, robotics, and productivity. But the industry often skips the steps connecting a smarter model to a solved real-world problem. Each of those steps usually requires institutions, capital, infrastructure, and political agreement.
- AI researchers may overgeneralize from unusually automatable work. Software, quantitative research, and some scientific fields contain large amounts of explicit, machine-readable context. Many jobs depend far more heavily on tacit knowledge, interpersonal trust, physical environments, and organizational negotiation.
- Politics cannot be optimized like a model. The industry has succeeded at technical problems that once looked impossible. It is now confronting problems that remain difficult because people disagree about values, risks, and who should benefit. There may be no technical solution that causes everyone to accept the same trade.
- Relational competence is becoming a strategic capability. Several of the most important failures involved trust, communication, negotiation, coalition building, and judgment rather than technical intelligence. Companies that treat those skills as secondary will continue to misread both communities and governments.
- The AI industry is losing control of its own story. Two years ago, companies could largely define AI through innovation, competitiveness, and future abundance. Workers, communities, populist politicians, safety advocates, and regulators are now developing their own accounts of what the technology represents. The only durable response will be visible, widely shared benefit rather than better messaging alone.
