Almost exactly 10 years ago, I published an essay in these very pages about the environmental impacts of computationally-intensive technologies such as Bitcoin (Ensmenger 2016). Titled “The dirty parts of the computing world,” my article argued that “the dirty little secret” of the digital economy was that “computing power” was inextricably linked to electrical power, and by extension to the consumption of coal, oil, water, uranium, and all the other natural resources required to generate that electrical power.

A decade later, that dirty little secret is no longer so secret.

Attend a county planning commission meeting almost anywhere in the American data center belt—Loudoun County in northern Virginia, central Ohio, the outskirts of Memphis—and you will find the room packed. Northern Virginia, in particular, is now so full of data centers that it has been nicknamed “Data Center Alley.” By one estimate, as many as 1,000 facilities could occupy nearly 20,000 acres across Northern Virginia and nearby parts of Maryland by the 2030s (Aversa 2026). Ten years ago, almost no one in such a room would have recognized the word “hyperscale,” let alone the acronym PUE (power usage effectiveness). Today, residents cross-examine the developers’ attorneys about cooling-tower evaporation rates, backup-generator emissions permits, and the difference between contracted and actual megawatt draw—the kind of fluency in industrial hydrology and grid economics that it took earlier generations of fenceline communities decades to acquire.

That earlier essay ended with a plea: that the environmental consequences of our information technologies become part of our collective conversation, “and quickly.”

They have.

I was right that these costs needed to be counted. I was less clear about who would do the counting, to what end, and with what authority. This essay is about what that conversation has become—and about what it still cannot see.

The victory

By any reasonable measure, the past decade has delivered everything an environmental critic of computing could have asked for. The material world of the Cloud is now, of all things, an electoral issue. Data-center energy costs also became a major issue in Virginia’s 2025 gubernatorial campaign; the winner had pledged to make the industry pay “its own way and its fair share” (Office of the Governor of Virginia 2026). That question has since gone national: State legislatures have considered ratepayer-protection measures, the White House has secured a voluntary pledge from technology companies to cover the energy and infrastructure costs associated with their facilities, and Congress is considering bipartisan legislation that would require regulators to consider recovering the full incremental cost of grid upgrades from very large-load customers.

The conversation is no longer confined to the places where the buildings actually stand. The United Nations now issues warnings about the water footprint of artificial intelligence (Shah 2026). Consumer Reports (a publication better known for helping Americans choose dishwashers and automobiles) now explains to its readers what a data center will do to their electric bill. Questions that a decade ago were confined to the pages of academic journals are now the stuff of county elections.

All of this is, on its face, exactly what I asked for. Looking back, however, I can see that I treated three different tasks as though they were one: making the environmental costs of computing visible, measuring them, and governing the industrial system that produced them. My 2016 essay called for more attention, more responsibility, and more regulation, but it did not ask clearly enough what institutions would make those costs knowable, decide who should bear them, or determine what obligations the industry owed to the public. A decade later, that omission matters. We have learned to see the data center as a source of environmental impacts. We have not yet learned to understand the kind of institution it is.

The smokestack and the factory

In 2016, the challenge was explaining to readers the hard nuts and bolts that underlie the hazy, amorphous web of computers, servers, storage, databases, connections, and software. The metaphor of the “Cloud” had so thoroughly dematerialized the digital economy that the first task of the critic was simply to insist that the Cloud was a place: a building, in a particular town, drawing on a particular aquifer, with a particular diesel generator out back. That argument has been won—in the hearing rooms of Loudoun County, Virginia, if not yet in the marketing departments of Silicon Valley.

But we have learned to see the data center without learning to understand it.

Which brings me to the claim I have been pressing for a decade now: The data center is a factory (Ensmenger 2021). And a factory is never merely a machine for producing pollution; it is a way of organizing labor, capital, land, energy, and political power. The smokestack was always the most visible and most photographed feature of the industrial landscape. It was also incidental to the factory’s essential function.

When nineteenth- and early twentieth-century reformers confronted industrial capitalism, smoke was often what they saw first. But the institutions that eventually made industry more governable were not principally smoke-abatement leagues. They included factory acts, labor laws, utility commissions, antitrust enforcement, and public-health agencies—interventions aimed at the social relations of production as well as its exhaust. Our critique of the Cloud has too often stopped at the exhaust, or more recently, at the vast amounts of freshwater consumed.

The environmental critique is accurate as far as it goes, but it is partial, and its partiality is convenient: A complaint denominated entirely in megawatts and gallons of water used for cooling is a complaint that the industry can recast as a problem of efficiency metrics, offsets, and procurement contracts. The problem is not that we have learned to count. It is that counting has begun to stand in for governing.

The wrong genre

The new environmental awareness circulates most widely in a familiar form: the per-prompt statistic. A nuanced estimate that the water contained in one 500-milliliter bottle might support roughly 10 to 50 responses on the internet has been compressed in public discourse into “one query, one bottle of water.” The “personal carbon footprint” owes its ubiquity not to the environmental movement but to the advertising agencies of fossil fuel companies like British Petroleum (Klein 2021), which two decades ago spent millions teaching the public to calculate its own culpability—on the sound theory that a public preoccupied with its own consumption habits is a public not asking questions about the mega-emitters of carbon dioxide, such as refineries (Watts 2024).

The per-prompt water statistic inherits both the genre’s appeal and its defects. The most obvious problem is that the numbers are shaky. Published estimates of the water cost of an AI query span orders of magnitude, and everything depends on where one draws the boundary of the system: the water evaporated on site, or also the water consumed in generating the electricity?

The industry has noticed that this is a game it can win. Google now reports that a median Gemini prompt consumes about a quarter of a milliliter of water (Crownhart 2025)—a figure produced by the company’s own accountants, using the company’s own boundaries, verifiable by no one (Ludvigsen 2025). The danger is real, however: If the bottle-of-water claim collapses under scrutiny, the credibility of the materialist critique may collapse with it. It took the better part of a century—factory inspectors, industrial hygienists like Alice Hamilton, ultimately the monitoring apparatus of the EPA—before industrial pollution became reliably knowable. No comparable apparatus yet exists for computation, and in its absence critics and defenders alike are arguing from the industry’s own press releases—a dubious source at best.

A second problem is worse, and no amount of better measurement will fix it: namely, that the genre depoliticizes even when the numbers are right. Per-prompt accounting renders an industrial phenomenon as the sum of individual consumer choices, with abstention as the remedy and guilt as the currency. No one, after all, proposes to address the environmental sins of the steel industry by making consumers feel guilty about paperclips.

The relevant unit of analysis should not be the individual prompt, but rather the larger system that includes the facility, the firm, and the grid. Per-prompt figures may help engineers compare systems, but they tell policy makers very little about the facilities, firms, and grids they are being asked to govern. The environmental effects of artificial intelligence unfold at several scales: the resources consumed by the computation itself; the consequences of particular applications; and the systems of production and consumption those applications reorganize. A low-energy model used to locate new oil reserves, intensify advertising, or accelerate commodity logistics may have a much larger environmental consequence than a high-energy model used for weather prediction or grid management.

The NIMBY paradox

The most visible political expression of the new awareness is the siting fight—and it is here that we should tread most carefully, because the meaning of the data center rebellion is double.

There is a generous reading of this rebellion, and it deserves to be made in earnest: For many residents, the zoning hearing is the first direct venue in which they have been able to contest the material impacts of the internet. For 50 years, computing infrastructure was built with essentially no public process at all; the metaphor of the Cloud kept it off the political map, and the industry used that invisibility to place itself outside the regulatory history of industrialization (Ensmenger 2021). The county hearing room is the first venue in which ordinary citizens have ever been asked whether they consent to the physical expansion of the digital economy. After a half-century of construction without consent, this is a development worth taking seriously.

And yet it matters that the venue is a zoning board. The label “NIMBY” (short for “Not In My Back Yard”) obscures a more basic institutional problem: The zoning board is often the only public body with the authority to say anything at all about the project. A county planning commission is being asked to perform, with a special-use permit, the work of a national industrial policy, an energy regulator, and an international trade regime.

This it cannot do; what it can do is refuse. And refusals are distributed according to wealth: The well-lawyered counties prevail, and the facilities migrate down the gradient of political power—to poorer counties, to states with weaker institutions and greater fiscal desperation, and increasingly to other countries, where the same firms negotiate with national governments on still more favorable terms. Refusing a data center in Northern Virginia does not reduce its harms; it re-routes them. This was precisely the twentieth-century career of the factory itself, and of the electronic waste I described in these pages a decade ago.

But who, exactly, is saying “not here,” and from where? The phrase means one thing in Loudoun County, Virginia, routinely ranked as the wealthiest county in America, and another near Boxtown, the Black neighborhood of South Memphis where xAI installed dozens of methane-burning turbines (Paddison and Marsh 2025)—initially without permits—to power what its owner boasts is the world’s largest supercomputer (Hilt 2025). Opposition in Loudoun can express environmental privilege, even when it also raises legitimate questions about costs and public process. In Boxtown, resistance emerges from a much longer history of cumulative industrial burden.

The pattern—indeed, the very vocabulary—should be familiar: The term “NIMBY” entered common usage in the 1980s, in the fights over the siting of hazardous and nuclear waste (Welsh 1993). Those fights taught a hard lesson: local veto without national settlement produces neither consent nor capacity, only stalemate and maldistribution. The question the data center rebellion has not yet learned to ask is the the one the US nuclear-waste siting system has never satisfactorily answered: What would planning look like that preserved meaningful local consent without permitting every locality to export its burdens?

The broken bargain

What the factory offered its host community, historically, was a bargain. It was often a coercive and unequal bargain—company scrip, captive housing, the foreman’s arbitrary power—but it was recognizable as one: local costs in exchange for local benefits. Jobs, wages, a tax base, and around them, over time, a durable industrial community with institutions of its own. The environmental burden and the economic benefit were fused in a single object. When early twentieth-century boosters looked at a smokestack, they saw a payroll.

The data center dissolves this bargain. The burdens remain stubbornly local: the land, the water, the transmission capacity, the diesel particulates, the permanent claim on the county’s future. But the benefits have been dispersed into the ether. The product is exported instantaneously, at the speed of light; the profits accrue in Seattle and San Francisco; and the labor force—the true labor force of the Cloud—is scattered across the construction trades, the utilities, the semiconductor fabs of Taiwan, the data-labeling shops of Nairobi and Manila, and the cobalt mines of the Congo. What remains on site, once the construction crews depart, is a skeleton crew of technicians. By one accounting (Smolaks 2016), state subsidy packages for data centers now run to more than $1,000,000 per permanent job (Tarczynska 2026)—and that is before the discounted electricity, the water commitments, and the non-disclosure agreements under which county officials negotiate away their own tax base without being permitted to name the company across the table (Garofalo 2025).

Historians of extraction have a name for this arrangement: the enclave. Like the mining camp, or the company banana towns of the United Fruit era, the data center extracts what is cheap and local—power, water, land, tax capacity, grid priority—and exports the value elsewhere, leaving remarkably little behind. And like the enclave, it retains its leverage indefinitely, because although the building is fixed, the workload is not. Computation can be shifted to another facility in another state overnight, and every host community knows it. The industrial factory, whatever its sins, was anchored to its machines and its workforce; it could not credibly threaten to leave over the weekend.

These are the questions, then, that a mature politics of the Cloud would be asking: What may a host community reasonably expect in exchange for a permanent infrastructural commitment? Should tax exemptions be made contingent on heat re-use, on water protection, on community-benefit agreements, on local hiring? What happens to a county that becomes fiscally dependent on an industry that can leave without moving? These are precisely the questions that utility commissions and factory inspectorates were invented to ask. No such institution yet exists for the Cloud. A decade into our new awareness, we have recovered the anger of the Progressive Era—but not, so far, its institutional imagination.

The energy jujitsu

Meanwhile, the environmental critique has suffered the strangest fate of all: It has been absorbed as a justification for the buildout.

The industry has answered “artificial intelligence uses too much energy” not with restraint but with a promise of abundance. The undamaged Unit 1 reactor at Three Mile Island, site of the worst commercial nuclear accident in American history, is being returned to service, rechristened the Crane Clean Energy Center, its entire 835-megawatt output contracted to Microsoft for 20 years (World Nuclear News 2024). Google and Amazon have signed for small modular reactors that do not yet exist, and which have a dismal track record of economic failure on a par with their much larger cousins (Makhijani and Ramana 2021). The waiting list for new gas turbines now stretches years, which is why xAI simply trucked its own into Memphis. And the federal government has declared “energy dominance” the policy of the United States, with the data center as its favored customer. A sentence that would have been unintelligible when I wrote in these pages a decade ago is now a plain description of fact: The data center has become the institution reviving the American nuclear industry, for better or worse.

Readers of this journal will appreciate the symmetry. The Bulletin was founded by physicists who understood that a technology of unprecedented power had escaped the institutions meant to govern it. Eighty years later, artificial intelligence and atomic energy have become curiously entangled: they now share a grid, an apocalyptic vocabulary, and as of this year a Doomsday Clock which stands at 85 seconds to midnight with “disruptive technologies” counted among the reasons (Doomsday Clock Statement 2026).

Nor will efficiency rescue us, and the industry knows it. Every gain in computation-per-watt is immediately reinvested in more computation; economists call this the “Jevons paradox,” after the phenomenon that William Stanley Jevons first described in 1865, when it was coal that was being economized. Efficiency is the industry’s favorite environmental metric precisely because it is the one metric guaranteed never to constrain growth.

Perhaps even more importantly, a megawatt is also a claim on a shared system. Every megawatt committed to a hyperscale campus is a megawatt the grid must generate and deliver rather than use elsewhere. When generation, transmission, transformers, capital, and time are scarce, data-center expansion is therefore a problem of allocation as well as emissions. Who gets the megawatt, on what terms, and for what purpose? When the environmental question is posed as “Where will the power come from?”, the industry always has an answer.

A factory for what, for whom, and for what purposes?

And so we return to the questions the smokestack obscures. Whose labor sustains this factory—the annotators in Nairobi, the moderators in Manila, the invisible human workforce behind everything we are pleased to call “automation”? Who owns the means of computation, and what does it mean that the capacity to produce machine intelligence is being concentrated in a half-dozen firms? By what right does an industry claim the water, the grid, and the public purse of the communities it settles among? And what, in the end, is all of this computation for?

None of these questions is new. In the 1960s, when commercial computing was young, it was widely discussed in the language of the “computer utility”: computation as public infrastructure, regulated like electricity or the telephone, priced and governed in the public interest. Martin Greenberger, writing in The Atlantic in 1964, took it more or less for granted that the information utilities of the future would operate under exactly this kind of public governance (Greenberger 1964).

To be sure, the computer utility always remained more aspiration than achievement; as Alexander Mirowski has shown in the Annals of the History of Computing, the time-sharing industry that adopted the label was happy to wear the utility’s mantle of universal service and fair pricing while pursuing neither (Mirowski 2017). But the vision was never refuted—merely abandoned, and then erased altogether by the metaphor of the Cloud. The computer utility matters now because it reminds us that these questions were once considered askable, by quite serious people, and could be again: What would a “good” data center be? Rate-regulated, disclosure-bound, heat-recycling, contracted to its host community rather than concealed from it—perhaps, in some places, publicly owned? A conversation that knows only how to say “not here” cannot even pose such questions; a conversation about factories could.

Ten years ago in these pages, I asked that we learn to see the Cloud. We have. Now we must learn to see past the smokestack to the factory behind it—its workers, its owners, its bargains, and its purposes. And once again, we need to do so quickly.