The Apocalypse Is Not on the Agenda

In the space of a few weeks, the argument about artificial intelligence has changed.

Researchers and executives at some of the companies building the most advanced systems are warning publicly that development is moving faster than our ability to control it. OpenAI disclosed that during an internal cybersecurity evaluation this summer, models operating with reduced safeguards circumvented isolation controls, exploited vulnerabilities and gained access to outside systems. In September, several leading figures in the industry called for slowing the development of frontier systems. President Donald Trump responded by dismissing fears of AI taking over the world as a “hoax” and claiming a “SICK conspiracy” against AI and data centers.

That is a serious debate. It is also not the one facing local governments across Pennsylvania.

In Hanover Township in Washington County, supervisors have spent the summer hearing testimony on a proposed campus of seven data-center buildings alongside a large natural-gas power plant. In Butler County, developers have now proposed a million-square-foot data-center campus in Forward Township, a municipality that does not have a zoning ordinance. These projects will not determine whether humanity can control superintelligence. They will determine what happens to specific pieces of land, how infrastructure costs are allocated, what communities collect in return, and what protections residents have when an enormous new industrial use arrives next door.

Where I sit

I should be clear about my own vantage point. Earlier this year I wrote that I did not believe AI was likely to end the world. I still think catastrophic outcomes are less likely than the alternatives, although the events of this summer have made me less comfortable treating the risks as speculative. My professional work also puts me inside the AI industry, helping regulated financial institutions figure out how to use these systems responsibly.

What that work has taught me is less dramatic than either side of the national debate. Most useful AI is not science fiction. It reads documents, reconciles information, finds patterns and prepares a first draft. The difficult part is not simply making the system more powerful. It is making the system answerable. Who authorized it? What information did it use? What did it do? Who remains responsible when it is wrong?

Banks demand those controls because their regulators, boards and customers expect them. The controls do not make the technology useless. They are what make the technology deployable.

That same idea applies to data centers.

Two questions, not one

The first AI question is whether increasingly capable systems can remain under meaningful human control. That belongs principally to the companies developing them, Congress, federal regulators and, ultimately, international institutions.

The second question is much more tangible: whether a particular data center, on a particular piece of ground, is a fair bargain for the people who live around it. That belongs to state regulators, county planners, borough councils and township supervisors.

Treating opposition to a specific data center as opposition to AI itself makes little sense. A zoning decision in Pennsylvania will not stop a global technological race. The capacity will move somewhere else, and the borough will still pay its solicitor. But dismissing local concerns as obstructionism or conspiracy makes just as little sense. Residents are asking legitimate questions about electricity prices, noise, water, tax revenue, infrastructure and land use. Those are exactly the questions local government is supposed to ask.

The test is whether it pays its way

I support data centers in Pennsylvania on one condition: that they pay their way. The test is simple. A data center should not depend on everyone else carrying costs the project creates.

Start with electricity. PJM’s independent market monitor estimates that data-center demand accounted for $29.4 billion of the $63.6 billion in capacity charges produced by the region’s last four capacity auctions, roughly 46 percent. Those costs do not remain inside the fence around a data center. They move through the regional electricity market and ultimately reach customers far beyond the communities hosting the facilities. Mt. Lebanon households are helping to pay for data centers that will never be built in Mt. Lebanon. That matters in a state where 57 percent of registered voters told Franklin & Marshall pollsters in August that paying their utility bills was somewhat or very difficult.

When a project creates a cost, the economics of the project should reflect that cost.

Pennsylvania has begun moving in that direction. Governor Josh Shapiro’s August executive order requires developers seeking certain state permits to make legally binding commitments under the state’s GRID standards. Those commitments include paying the full cost of new generation, transmission, distribution and other electric infrastructure their projects require, rather than shifting those costs to households and businesses. The order also requires greater public transparency and prohibits nondisclosure agreements in covered data-center projects. But an executive order lasts as long as the governor who signed it. This standard belongs in statute.

Ohio shows what happens when speculative demand meets a real price. AEP Ohio’s data-center tariff requires large customers to pay for at least 85 percent of the power they reserve even if they ultimately use less. It also requires evidence of financial viability and imposes exit costs if projects disappear. After the tariff was approved, the utility’s pipeline of data-center requests fell from more than 30 gigawatts to about 13. The projects that disappeared were never going to be built.

That is not anti-growth. Like the controls banks insist on, it is how a functioning market separates serious projects from placeholders.

The revenue side needs the same discipline. Advocates often point to Loudoun County, Virginia, where data centers produce close to half of local tax revenue. The comparison does not travel. Virginia localities tax the computer equipment inside the buildings, which is where most of a data center’s value sits. Pennsylvania’s local taxes reach land and buildings, not servers. A host community here should model its revenue on that basis, not on Loudoun’s.

The same discipline should apply to public subsidies. Pennsylvania currently exempts qualifying data-center equipment from the state’s sales and use tax. State budget estimates put the exemption’s cost at $41.1 million in fiscal year 2024-25 and project the annual cost to rise to $517.2 million by 2030-31. In June, the House voted 197 to 5 to repeal the exemption. The Senate passed its own repeal language inside a larger bill, then let it stall. Both chambers have voted to end the break, and it is still law. An industry that can pay its way does not need Pennsylvania to pay part of the way for it.

Data centers can create real economic value, but the case should be made on that value rather than on incentives or inflated promises of employment. Construction activity can be substantial. Permanent employment is much smaller. Brookings researchers examining roughly 1,500 facilities found that a typical county receiving a data center gained approximately 100 to 200 jobs over a decade, while wages were essentially unchanged.

That does not make a data center a bad project. It makes it primarily an infrastructure, land-use and fiscal project rather than a traditional jobs program. Communities should negotiate accordingly.

Noise, water use, setbacks, screening and eventual decommissioning belong in the same framework. The question is not whether these impacts make a project inherently unacceptable. It is whether the rules identify the impacts in advance, assign responsibility for them and give local government the tools to enforce the bargain.

That last part is where Pennsylvania’s municipal structure creates a much larger problem.

The rules have to exist before the application arrives

Springdale provides the clearest lesson.

Last December, the borough council voted 5 to 2 to approve a conditional-use permit for a proposed 180-megawatt data center on the site of the former Cheswick Generating Station. The approval included 19 conditions covering issues such as noise, landscaping and screening. Several council members made clear that they were uncomfortable with the project. One said publicly, “Do I want to see this data center come in? No I don’t,” before explaining his fear that a denial would expose the borough to litigation.

The borough manager explained the legal constraint more precisely. Once an applicant demonstrates compliance with the criteria contained in the zoning ordinance, approval of a conditional use is generally not discretionary. A rejection could have produced litigation and, potentially, a court-approved project without the conditions the borough had negotiated.

That is not principally a story about whether Springdale’s council members were brave enough. It is a story about timing and institutional capacity.

By the time a sophisticated developer arrives with engineers, lawyers, consultants and a completed application, much of a municipality’s leverage has already been determined by the ordinance on the books. The most important work may therefore happen months or years before anyone files an application. That work means defining where the use is permitted, establishing performance standards, addressing noise and water, creating setback requirements, specifying financial guarantees and deciding what technical evidence an applicant must provide.

Doing that well is expensive. Township supervisors who have been through it put the cost of professional help at $30,000 to $150,000, a real sum for a borough of a few thousand people. It makes little sense to expect every one of Pennsylvania’s small, lightly staffed municipalities to become expert in electrical infrastructure, acoustics, water systems, data-center engineering, municipal finance and land-use law every time a project appears.

This is precisely the kind of problem regional institutions should exist to solve. Allegheny County alone has 130 municipalities. I have argued before that it makes little sense for each of them to reinvent the same trash contract. It makes even less sense for each municipality to independently write a data-center ordinance, retain specialized engineers and negotiate against companies whose technical and legal resources can dwarf the entire staff of a small borough.

The response does not require eliminating local control. It requires giving local officials more capacity to exercise it. Counties and councils of governments can develop shared ordinance language, maintain pools of technical experts, establish common standards and provide negotiating assistance before an application lands. A small borough should be able to call for sophisticated help without first assembling an entirely new professional team of its own.

Local control without local capacity is often control in name only.

The rehearsal

So here is where I land. Data centers can be good neighbors and sound fiscal partners in Pennsylvania on four terms. They pay for the power and wires they cause, with collateral behind the promise. The state stops subsidizing their equipment. Local agreements are made in public and are enforceable. And the revenue communities are promised is calculated on Pennsylvania’s tax base, not Loudoun’s. Where a project cannot meet those terms, the answer should be no, and the ordinance should already say so.

That brings the argument back to the larger debate about artificial intelligence. There is no contradiction between believing AI can create enormous value and insisting that institutions set terms around its deployment. Those terms are not obstacles to technological progress. They are what governing technological progress looks like.

The larger AI debate asks whether democratic institutions can establish meaningful boundaries around technologies whose capabilities may grow faster than our ability to understand them. That problem is extraordinarily difficult. But the physical infrastructure supporting those systems gives us a much more immediate test.

A data center sits on land we zone. It connects to power systems we regulate. It pays taxes under rules we write. If we cannot establish a fair and enforceable bargain around that, we have no business being confident we can govern the parts of artificial intelligence that are harder to see.

The apocalypse is not on the agenda at the next township meeting. A data center might be.

Getting that decision right is not a distraction from the larger question. It is the rehearsal for it.

---

Sources for the figures above:

• OpenAI, The Hugging Face incident and the road ahead, August 2026
• Associated Press, Trump calls AI risks a “hoax”, September 14, 2026
• WESA, Washington County data center draws hundreds to Hanover Township, August 7, 2026; Butler Eagle, Data center proposed in Forward Township, September 18, 2026
• Utility Dive, Data centers drove $6.3B in PJM capacity auction costs: market monitor, July 2026
• Franklin & Marshall College Poll, August 2026
• Office of the Governor, Executive order on data center development, August 18, 2026
• Data Center Dynamics, AEP Ohio slashes data center pipeline by more than half, October 2025
• Loudoun County, Data Center Capital of the World, July 2026
• Pennsylvania Budget and Policy Center, Budget overview: data centers, citing the Governor’s Executive Budget tax expenditures; Governing, Pennsylvania lawmakers regret $1.8B data center tax break, September 4, 2026
• Brookings, New evidence on data center employment effects, August 2026
• PublicSource, Lessons from a Springdale AI data center development, January 2026
• Pennsylvania Capital-Star, Township supervisors urge colleagues to adopt zoning for data centers, April 21, 2026