Where Should America Build AI?

Data Centers, Water Scarcity, and the Economics of Resource Allocation

Abstract: Artificial intelligence is driving an unprecedented US infrastructure boom, with over 5,400 operational data centers and 1,500 more planned. Data centers impose substantial costs on many communities, especially areas with scarce water or electricity. Driven by these physical limits, the issue has escalated into acute political paralysis, ranging from emergency municipal bans to performative federal moratoria. We argue that eliminating tax subsidies and mandating that developers pay the full incremental social cost—inclusive of both localized pollution and the broader costs associated with higher electricity and water prices—will naturally steer the development of AI data centers toward locations where resource constraints and social costs are lowest.

The United States has rapidly become the world’s largest market for digital infrastructure, hosting more than 5,400 operational data centers and roughly 1,500 additional projects in various stages of planning or construction. Driven by explosive growth in artificial intelligence and cloud computing, these facilities have become one of the largest sources of new electricity demand in the country. Their rapid expansion is reshaping regional electricity markets, water systems, and state economic-development strategies.

Unlike most industries, data centers are remarkably mobile. They require abundant electricity, reliable fiber connections, land, and cooling capacity, but they do not need to be located near customers. As a result, geography matters. The social cost of building a hyperscale AI campus varies dramatically depending on local electricity capacity, water availability, environmental conditions, and public infrastructure.

This geographical sorting, however, is heavily distorted by a ubiquitous patchwork of state-level corporate incentives. Data center development is no longer isolated to traditional tech hubs; instead, state governments have engaged in an aggressive, nationwide bidding war. According to data tracked by the economic development watchdog Good Jobs First, at least 32 states have enacted statutory tax exemptions specifically engineered to attract the data center industry. These programs primarily waive sales and use taxes on high-value server equipment, cooling infrastructure, and the massive quantities of electricity these facilities consume.

Because these tax abatements are frequently uncapped and can span decades, the fiscal strain on state budgets has become massive. Disclosed annual revenue losses have reached $1 billion in Texas, $1.02 billion in Virginia, and an estimated $2.5 billion in Georgia. Compounding the issue, a recent 2026 Good Jobs First study revealed that 14 states completely fail to disclose their revenue losses under the guise of taxpayer confidentiality. This baseline policy approach treats data centers as traditional economic development engines, ignoring the structural reality that they generate very few permanent local jobs relative to the immense, multi-decade resource burdens they shift onto local communities.

The current pattern of development reflects these combined tradeoffs of physical limits and aggressive fiscal courtship. Northern Virginia remains the historical center of global internet traffic but is increasingly constrained by transmission congestion and local opposition. Ohio has aggressively pursued data center investment through lucrative tax incentives and infrastructure support. Tennessee benefits from the extensive generating capacity of the Tennessee Valley Authority. Texas offers abundant land, a completely decoupled utility grid, and one of the nation’s largest electric systems.

The most controversial expansion, however, is occurring in the arid West. Arizona and Utah have emerged as major destinations for hyperscale AI infrastructure because they offer inexpensive land, business-friendly regulatory environments, and access to growing western technology markets. Yet both states also face chronic water scarcity, increasing electricity demand, and long-term environmental challenges. Their experience raises a broader policy question: should governments encourage resource-intensive industries in regions where the underlying resources are already scarce?

The economic benefits of data center growth are uneven. Construction creates substantial short-term employment for electricians, engineers, construction workers, and specialized contractors. Once completed, the facilities become highly automated, capital-intensive operations requiring relatively few permanent employees. The strongest economic case exists when projects generate lasting improvements in infrastructure, tax revenue, or complementary business activity rather than simply large construction expenditures.

The principal costs of new data centers involve the use of electricity and water and additional pollution or traditional external costs associated with any industrial project.

Data centers place enormous new demands on electric grids. If utilities must build additional generation, transmission lines, or substations, the central policy question becomes who pays. Residential customers should not subsidize infrastructure constructed primarily to serve private hyperscale facilities.

Water presents an equally important challenge. In humid regions electricity may be the binding constraint. In the arid West, however, water scarcity may be even more significant. Facilities using evaporative cooling consume substantial quantities of water throughout their operating lives because server heat generation is continuous. Switching to dry cooling reduces direct water consumption but substantially increases electricity demand, shifting rather than eliminating environmental costs.

Local political and water authorities often eager to attract investment fail to protect local consumers. The lopsided nature of this development is laid bare: for example, a Blackstone-owned data center in Fayetteville, Georgia, consumed nearly 30 million gallons of unmetered water through unauthorized hookups during a severe drought without having to pay fines.

Data centers also generate localized externalities through diesel backup generators, cooling equipment noise, wastewater management, and construction impacts. These costs are real even when they are not reflected in market prices.

The appropriate response is neither a blanket ban nor an unconditional subsidy.

Instead, states should require data centers to internalize the full cost of the resources they consume.

Utilities should establish separate large-load rate classes so ordinary households are not forced to finance grid expansions serving hyperscale facilities.

Water-stressed regions should require water budgets, recycled or non-potable water where feasible, drought contingency plans, and pricing structures that reflect the true marginal cost of scarce supplies.

Pollution should be addressed through Pigouvian taxes, emissions standards, generator restrictions not subsidization of a private industry. In reality, state and local governments compete and provide tax abatements, infrastructure commitments, utility concessions, and favorable zoning decisions to attract jobs.

The political economy of data center development increasingly resembles the economics of publicly subsidized sports stadiums. The sports stadium literature provides a useful warning. Decades of research conclude that promised economic gains often fail to materialize while taxpayers absorb substantial long-term costs. AI infrastructure is unquestionably more productive than a football stadium, but the underlying lesson remains valid: visible capital investment does not guarantee positive social returns.

Utah and Arizona illustrate the central challenge of AI infrastructure policy.

Both states have become attractive locations for hyperscale developments because of available land and favorable business climates. Both also face chronic water scarcity, increasing electricity demand, and long-term environmental pressures.

Utah is particularly instructive. This scarcity is acute in the west where the seven Southwest basin states have blown past multiple federal deadlines to negotiate usage cuts as a catastrophic 2026 snow drought has pushed Lake Powell and Lake Mead to record-low inflows, prompting unprecedented federal management intervention to stabilize the collapsing river system.

The sheer absurdity of building hyper-scale computing in this fragile arid zone is underscored by Kevin O’Leary’s recent high-profile retreat, where intense public backlash and state pressure forced him to slash his proposed 40,000-acre Utah data center plan in half—vividly proving that the region simply cannot hydrologically or politically sustain the unmitigated footprint of AI.

Unlike many industrial facilities, hyperscale AI campuses generate continuous cooling requirements for decades. Water consumption is therefore not a temporary construction issue, but an ongoing operating requirement directly tied to electricity consumption and server heat generation.

The broader lesson is national rather than regional. AI infrastructure should be located where resource constraints are smallest rather than where subsidies are largest. Regions with abundant water supplies, excess generating capacity, existing transmission infrastructure, or access to hydroelectric or nuclear power can provide the same computing services at substantially lower social cost.

The unmitigated expansion of digital infrastructure has become one of the defining issues of the 2026 political cycle at the local, state and national level. The WSJ has identified 150 digital infrastructure projects that were halted in the last year alone.

In New Albany and Hebron, Ohio, a wave of emergency municipal bans on new data center footprints forced the state legislature to abruptly suspend long-standing corporate sales tax exemptions for the industry while launching an urgent grid-impact study.

City councils in Baltimore, Maryland, and Oklahoma City, Oklahoma, unanimously passed emergency moratoria halting all new applications, rezoning, and building permits to shield over-allocated municipal water supplies.

In Port Washington, Wisconsin, citizens took the resistance a step further, enacting a first-in-the-nation municipal referendum mandating that any future large-scale data center project seeking public tax incentives must first secure a majority vote from local residents on the ballot.

Over a dozen states—including New York (S.B. 9144) and Pennsylvania (H.B. 2533)—have introduced bills to enforce multi-year, statewide freezes on hyperscale development pending comprehensive environmental and grid-resiliency reviews.

In Maine, shortly after vetoing a bill imposing a moratorium on the construction of data centers Governor Janet Millsdropped out of the race for the Democrat nomination for the U.S. Senate.

A moratorium is a blunt, non-market instrument. Rather than pricing a scarce resource (like water or power), it drops the supply curve to zero by fiat. While it temporarily shields an over-allocated grid or aquifer, it creates a massive deadweight loss (lost economic efficiency) because it treats all digital infrastructure equally—whether a facility uses highly efficient, closed-loop water recycling or an outdated, resource-heavy cooling plant.

Economists generally prefer pricing externalities through a cap-and-trade system or targeted resource tariffs (e.g., charging exponential premiums for peak-hour megawatts). This forces the industry to innovate its way out of the bottleneck rather than halting development entirely. However, for local town councils facing immediate resource depletion, a moratorium acts as an emergency circuit breaker when they cannot afford to wait for a multi-year tax structure to phase in.

Opposition to data centers at the federal level is centered on a national moratorium on all facilities over 20 megawatts in a proposal offered by Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez.

Senator Sanders is also proposing a sovereign wealth fund which would own 50 percent of all AI firms over $100 million dollars. This would create an incentive for the federal government to drastically expand AI.

This escalating political debate over AI data centers increasingly mirrors the gridlock of the American health care debate. On one side, a laissez-faire faction argues for giving the private sector unbridled power, claiming that any regulatory friction will stifle innovation and cede technological dominance to global rivals. On the other side, populist critics offer performative resistance tailored for their political base rather than serious policy solutions, proposing sweeping bans and moratoria that would effectively cripple a vital, nascent industry. This polarization leaves a massive vacuum where pragmatic governance should be—trapped between a corporate blank check and an outright technological freeze.

Artificial intelligence infrastructure may well become as vital to the twenty-first-century economy as railroads, interstate highways, and telecommunications networks were to earlier generations. However, economic importance is not a license to transfer massive structural costs onto households, localized utility ratepayers, or future generations. The era of unfettered state-level competition to attract hyperscale data centers through blank-check fiscal sacrifices is rapidly ending, fractured by both hard physical resource limits and an aggressive legislative backlash. The geographic sorting of digital infrastructure is no longer dictated purely by a state’s willingness to forfeit its tax base; it is running directly into the physical boundaries of local energy grids, depleting municipal water supplies, and triggering severe political liabilities at every level of government.

Ultimately, the core objective of modern public policy should not be to maximize the raw number of data centers crammed within a state’s borders, but rather to minimize the total social cost of providing the computing infrastructure the nation requires. Achieving this outcome requires adhering to a straightforward economic rule: data centers must pay the full, unsubsidized marginal cost of the electricity, water, pollution, and public infrastructure they consume. This result will lead developers to search for the location where costs are the smallest.

Authors Note: Did you find this post informative? You will probably also enjoy A Tale of Three Energy Sectors and Trump and Biden on Wind and LNG.