Generative AI is being developed by a small handful of firms. They are taking on massive amounts of debt to build computing power on which costly new tools can be offered at scale. Given the size of these investments, the firms are commercially viable only if they sell their wares at scale. This means either tapping hidden wellsprings of productivity growth (unlikely) or being used as a low-marginal-cost substitute for white-collar labor (more likely).

The ensuing debt-fueled displacement of white-collar labor, if successful, will induce a massive shift of income from labor to capital, as even the AI firms concede. The ruination of ecosystems and acceleration of the climate crisis are mere side effects. There is strong historical evidence of technology’s corrosive effect on the labor share: recent data from Europe suggests that AI, no less than other recent tech leaps, causes losses to labor and corresponding gains to capital. Worse, because most democracies tax capital at lower rates than labor, that shift will constrict the state, diminishing its fiscal capacity to cushion job losses and declining wages. In effect, hyperscale AI creates a distributional crisis and robs the state of tools to address it.

Of course, this strategy might not work. AI might not be transformative as its boosters claim. Markets may lose their nerve before profits can be reaped. But if the AI strategy isn’t successful, the leading firms now dominant in public equity markets (and hence our pensions and 401[k]s), will almost certainly demand a “rescue” akin to the 2008–9 federal financial-crisis interventions.

As of August, hyperscale AI firms had accrued more than $500 million in debt from private equity, banks, and institutional investors such as Blackrock and Pimco. If that debt sours, the repercussions of the crisis, as in 2008–9, would likely force the government’s hand. If that happens, a massive transfer of risk to public balance sheets will follow. This is capitalism as a “heads I win, tails you lose” proposition.

Whether or not it founders, hyperscale AI capitalism will cleave open immense lost opportunities for society at large. On one estimate, US AI investments will add up to some $10.3 trillion between 2025 and 2032, or an average of 3.63 percent of US gross domestic product per year. This is $10 trillion that will not and cannot be spend on maternal care, pre-K education, cancer treatment, environmental mitigation, or any other social programs and planetary reparative work we desperately need. It is, by any measure, a staggering divergence of investment away from areas of urgent moral priority.

No wonder, then, that the American public has such profound anxieties. The data center build-outs are simply the most apparent of the many economic, political, fiscal, and environmental crises that hyperscale AI is inducing.

What then might be an effectual response? Except in Harwood, North Dakota, voters have pushed back at a local level by denying construction permission, or in Perry, Ohio, demanding that firms post a bond as insurance against the risk of fiscal implosion. Localities’ responses are circumscribed, though, by their limited powers and their small geographic reach. Cities and counties can do little about the labor market effects or fiscal risk. Many environmental costs associated with new demands on the electric grid are out of their reach. Some cities such as Austin and San Antonio own and operate their own electrical utilities, giving them a lever that doesn’t exist elsewhere in the country. But these are exceptions. Solely local responses are usually tightly constrained.

Since the effects of hyperscale AI capitalism emerge at the national level, it would ideally be the federal government that drives a comprehensive response. Pessimism about the prospect of federal action is wise for now. After Tuesday’s election, Trump reiterated his belief that data centers are “gold . . . only foolish leadership will say NO.” His December 2025 executive order, barring states from regulating AI, remains in force, although the legal theories upon which it rests are flawed. Even if a Democratic-controlled House and Senate can overcome his veto, any resulting AI legislation will likely be weak thanks to the need to assemble supermajorities to ensure bills’ passage.

The difficulty of federal action, however, does not mean that progressives should give up addressing hyperscale AI capitalism. They still have ways to use both their control at the state level and at the regional level to highlight its costsand offer the beginnings of a commensurate response — indeed, one that might be fully nationalized in two years.

In January, Democrats will have “trifectas” in twenty-two states and hold the governor’s office in seven more (Kansas and Alaska are still not called as I write). This is a first platform for action. To date, states’ responses have been reactive and purely negative: New York’s and Oregon’s governors paused data center construction, while laws imposing moratoria have been introduced but not passed in sixteen states. Advocates of successful local ballot initiatives have indicated that they plan to turn to bans in their states.

Yet a pause is not a policy. It does nothing to communicate the complex and interconnected harms of hyperscale AI capitalism. So it leaves the field open for AI’s advocates to advance their tall tales of illimitable gain. Nor does it prevent the building of data centers elsewhere in the nation and so does nothing to rein in its fiscal and labor costs.

An alternate approach would be for states to force tech firms to not just account for the local effects of data centers but also recognize and take responsibility for all the social costs of their hyperscale strategy. It would seek to channel economic development into less costly, more decent forms. Bans, however, are a badly formed tool to this subtle end.

As a start, states could recognize their own complicity in hyperscale AI. In forty states, data centers benefit from exemptions from the sales and use taxes that would ordinarily accrue on building materials, machinery, and equipment. The nominal justification for such exemptions is the jobs new builds purportedly bring. But a data center at best tends to create only two or three hundred jobs, with no long-term effect on wages. Where a landlord like Equinix, Data Realty, or CyrusOne rents computing power to remote clients, even that meager gain goes unrealized.

States’ and localities’ ensuing revenue loss is hard to measure, because only some of the states with these tax exemptions report their fiscal effect. Texas, however, lost some $1 billion in FY2025; in 2024, Illinois lost $370 million; and Arizona lost $19 million. A simple first step is terminating all these tax exemptions on the ground that the local benefits from data centers have been illusory, following Maine’s example. That might also prompt states to reconsider more generally their blind pursuit of corporate investments through tax relief.

Having withdrawn their reckless subsidies, states should follow Perry, Ohio’s lead by forcing tech firms to bear the social costs of their speculative investments. New construction, for example, might be conditioned on bonds tied to the environmental and state-level labor market effects of their business models. Firms would first be required to report not only on expected environmental harms (including from increased energy demand) but also address whether a data center’s expected uses will displace human labor or otherwise have wage-suppressive effects (as a leading study has already found). It would then have to post bonds sufficient to pay for state investments to buffer those effects.

The law might impel an AI-funded guarantee fund to sustain state unemployment payouts and retraining costs generated by AI. States could also require that data center investments be backed by sufficient nondebt financing to mitigate any risk of catastrophic fiscal events. And even if data centers are not fully banned, states might simply target specific uses as verboten — say, the large-scale automation of American jobs. Firms that flout this would see their centers shut down.

At the same time, states need not act alone. An underappreciated part of the American constitutional order is groups of states’ latent authority to enter into binding “compacts.” These are treaty-like agreements, often created without Congress’s approval, to coordinate regulation between a pool of states. Existing compacts address drivers’ licenses, insurance coverage, and medical practice. They allow states to set uniform standards and realize regulation at scale, even if the national government is AWOL.

Imagine Texas, Virginia, California, Pennsylvania, and New York forging an interstate compact on AI and data centers. No firm would be permitted to operate data centers in any state unless they agreed to meet certain environmental standards and avoid use cases that threatened serious job loss. The compacting states could require a common set of disclosures — not only about the data centers themselves but also about the fiscal and labor market risks associated with AI more generally, and perhaps about other AI-related harms.

Unlike simple bans, this arrangement is more likely to be politically self-sustaining. Firms would, in effect, have to document on an ongoing basis the knot of harms that flow from hyperscale AI — hence supplying the grounds for their own regulation. And they could not simply blame states for quashing economic growth. A race to the bottom as firms tried to pit states against each other would be more difficult.

It is true that hyperscale AI capitalists are likely to flee from compacting states. But notice that the creation of a robust compact would limit their flight options, at least within the United States. To see this, consider the political costs to a red state governor who welcomes fleeing data centers that decline to comply with environmental and labor market protections. Recall that some 89 percent of Americans oppose data center siting near them. Isn’t it likely that the governor will meet the same fate as other erstwhile champions of AI, such as outgoing Texas Governor Greg Abbott? Compacting to clarify AI’s harms makes it politically more difficult for noncompacting states to embrace the industry.

No doubt, these solutions are incomplete. But they offer a pathway for dealing with not just the local but also the national harms of hyperscale AI in a politically durable way. They not only signal clearly that voters’ concerns have been heard but also help build virtuous circles of disclosure, action, and revenue so the perils of hyperscale AI can be enduringly addressed — a groundwork that offers a foundation for national action when the time is right.