AI Data Center Boom and Location
The AI Data Center Boom Is Becoming a Question of Place
A new study suggests the next phase of AI infrastructure may depend less on how many data centers America can build than on where electricity, water, land and communities can support them.
For much of the artificial-intelligence boom, the data-center story has been about scale.
Technology companies need more computing power. More computing power requires more servers. More servers require larger data centers. And larger data centers require extraordinary amounts of electricity, cooling equipment and supporting infrastructure.
But a new academic paper suggests that this way of looking at the problem misses something important.
The environmental and economic consequences of an AI data center cannot be determined simply by looking at the building itself. They depend heavily on **where the building is located and what systems surround it**.
The paper, *Environmental and Economic Implications of Artificial Intelligence Data Centers in the United States*, was posted to arXiv on August 10 by Johanna Bolaños-Zuñiga and Alberto J. Lamadrid. It examines electricity demand, cooling requirements, backup power, water use, noise, land use, utility pricing and regulation. Its central argument is that data centers need to be understood not as isolated buildings but as parts of larger electricity, water and land-use systems.
That distinction could have major implications for the enormous data-center buildout now underway.
The paper notes that new AI data-center interconnection requests commonly exceed 100 megawatts and can approach 1 gigawatt for a single facility. It also summarizes outside projections suggesting that U.S. data centers could consume between 6.7% and 12% of the nation's electricity by 2028 and between 9% and 17% by 2030. Those figures are projections compiled from other research rather than forecasts produced independently by the paper's authors, but they illustrate the scale of the challenge.
The consequences point toward seven changes in how future data centers may be planned and approved.
- Where a Data Center Is Built May Matter as Much as How Efficient It Is
For years, much of the technology industry's environmental discussion has focused on efficiency: faster chips, better servers and improved cooling systems that perform more computing with less energy.
Those improvements matter.
But the new paper argues that efficiency inside the facility tells only part of the story.
Most of a data center's operational greenhouse-gas footprint comes indirectly from the electricity used to run it. What matters therefore is not simply how much electricity the facility consumes, but **which power plants respond when that additional electricity is demanded**.
That is the difference between average and marginal electricity emissions.
A regional grid may have a relatively clean average generation mix because it receives substantial electricity from nuclear, wind, solar or hydroelectric power. But if the additional electricity required by a new data center is supplied primarily by natural-gas or coal plants operating at the margin, the emissions associated with adding the facility can be substantially different from the grid's average.
Time matters too.
A computing workload performed at 3 a.m., when electricity demand is low and renewable power is abundant, can have a different system effect from the same workload performed during an extreme summer afternoon when the grid is already strained.
Location matters because transmission constraints also determine which generators can respond.
The practical implication is significant:
Two equally efficient data centers can have very different environmental consequences simply because they are built in different places.
That shifts the policy discussion away from asking only, "Is this an efficient data center?" toward a more difficult question:
"Is this an efficient place to put this data center?"
The authors ultimately recommend coordinating data-center deployment with electricity-system characteristics, water availability and land-use planning rather than evaluating each of those factors separately.
- The Cheapest Location for the Developer May Not Be the Cheapest Location for Society
Companies choosing data-center sites naturally look for favorable economics.
They want affordable land, reliable electricity, fiber-optic connections, tax incentives, permitting certainty and proximity to other infrastructure.
But the new paper identifies a problem familiar to economists: some of the costs created by a project may fall on people who are not parties to the transaction.
Data-center developers may need new substations, transmission lines, distribution equipment and additional generating capacity. Large water systems may require expansion. Electricity production can create emissions. Cooling equipment and generators can create noise. New campuses can consume significant amounts of land.
The authors describe these as questions of **externalities and cost allocation**.
Their electricity example is especially important.
Transmission and distribution systems are shared networks. If billions of dollars of infrastructure must be constructed to serve enormous new electricity users, regulators must decide how those costs will be recovered.
If a utility spreads them broadly across its customer base, households and existing businesses can effectively subsidize part of the infrastructure needed by the new development.
The paper warns that electricity tariffs do not always perfectly reflect which customers caused particular infrastructure costs. Depending on how rates and interconnection agreements are designed, the arrival of very large loads can therefore shift some costs from data-center operators to other classes of customers.
That issue is already moving from academic discussion into public policy.
On August 18, Pennsylvania Governor Josh Shapiro issued an executive order requiring data-center developers seeking state permits and benefits to meet new standards. Among them is a requirement that developers pay the full cost of new electricity generation, transmission, distribution and other infrastructure needed to power their projects rather than shifting those costs to households and businesses.
The principle is straightforward:
Growth should pay for growth.
If that idea spreads, it could fundamentally change data-center economics. A site with seemingly cheap electricity may look less attractive once the developer must also account for the infrastructure needed to deliver that electricity reliably.
The relevant calculation may therefore become not merely the cost of buying electricity but the **total system cost of adding the facility**.
- Water Could Become a Major Factor in Deciding Where Data Centers Go
AI computing generates large amounts of heat.
That heat has to go somewhere.
Some data centers use evaporative cooling systems that can consume significant amounts of water. Others rely more heavily on air cooling or other technologies, which can reduce direct water consumption but sometimes increase electricity demand.
The new study emphasizes this tradeoff.
Water-cooled systems can lower electricity requirements while increasing water consumption. Dry cooling can dramatically reduce on-site water consumption but require more electricity. Advanced or hybrid systems may improve the balance, but there is no single cooling method that is environmentally optimal everywhere.
That means water availability matters in much the same way as electricity availability.
A cooling design that makes sense in a region with plentiful water may be inappropriate in an arid or drought-prone area where farmers, cities, ecosystems and industry are already competing for limited supplies.
There is another complication that is easy to overlook.
A data center's water footprint includes not only water used directly at the facility but potentially water consumed in producing its electricity. Thermal power plants themselves may require substantial water for cooling.
So switching from water cooling at the data center to a more electricity-intensive cooling system does not necessarily make the water problem disappear. Some water use can effectively move from the data center to the power plants supplying it.
The authors therefore treat water as a regional resource problem rather than merely an engineering problem inside the building.
Their recommendation is that policies account for scarcity conditions and competing uses when considering data-center development.
The implication for future construction is clear.
Developers may increasingly have to demonstrate not just that water can physically be supplied, but that the proposed level of consumption makes sense **for that watershed and that community**.
Water-rich regions could gain an advantage.
Water-stressed regions may begin requiring recycled water, closed-loop systems, strict consumption limits or other technologies as a condition of approval.
- Electricity Availability May Become the Biggest Constraint on AI Expansion
The AI boom is arriving much faster than major electric infrastructure can usually be built.
The paper highlights an important mismatch.
A data center can sometimes be developed and connected within one or two years. Large transmission projects generally take much longer. Meanwhile, generating plants, substations and other infrastructure face their own permitting, financing and supply-chain delays.
That creates a new problem for electric-grid planners.
Historically, utilities could forecast relatively gradual increases in electricity use from population growth, businesses and ordinary development.
AI data centers can arrive as enormous concentrated loads.
According to figures summarized in the paper, PJM—the regional grid operator serving all or parts of 13 states and the District of Columbia—projects summer peak demand rising by roughly 56 gigawatts by 2035, with data centers accounting for much of the increase.
PJM itself has spent much of 2026 developing special rules for integrating large new electricity users. Its board has said that rapidly arriving data-center demand presents challenges involving both grid reliability and consumer affordability. PJM has been developing mechanisms allowing some new large loads to connect under conditions in which they could be curtailed earlier during shortages if sufficient new generation is not available.
This represents a striking change in the traditional relationship between economic development and the electric grid.
For much of modern American history, a large company deciding to build a facility generally assumed that the electric system would find a way to serve it.
For some of the largest AI projects, that assumption may no longer hold.
A developer could have land, financing, permits and customers and still face a fundamental obstacle:
There may not be enough electricity available at that location at the time the company wants it.
If so, the AI infrastructure race becomes partly an electricity-infrastructure race.
Communities with sufficient generation and transmission capacity could attract projects more easily.
Those without it may face a choice between delaying data centers, building substantial new infrastructure or asking the developers to supply more of their own power.
- "Bring Your Own Power" Helps—but Creates New Questions
One solution seems obvious.
If an AI data center needs the electricity equivalent of a small city, why not make the developer bring its own power plant?
That idea is already moving into actual grid policy.
PJM has encouraged a "Bring Your Own New Generation" approach and has been developing faster pathways for large customers that secure new generating resources. The goal is to allow growth without requiring existing customers to bear all the reliability consequences of rapidly increasing demand.
But the new study demonstrates why bringing private generation does not eliminate the broader environmental problem.
It simply changes it.
On-site generation might reduce demands on the grid. But depending on the technology, it can also create local air pollution, greenhouse-gas emissions, noise and additional land requirements.
Natural-gas turbines have a different environmental profile from solar panels.
Fuel cells differ from diesel generators.
Battery storage can help manage short-term fluctuations but is not itself an unlimited source of electricity.
Renewable microgrids require land and storage or backup resources if round-the-clock power is expected.
The authors specifically caution that co-located generation and hybrid energy systems can improve reliability while simultaneously creating new local environmental impacts or shifting infrastructure costs among consumers.
So "bring your own power" should not be confused with "solve your own problem."
The deeper question becomes:
- What kind of power are you bringing, where will it be built, what resources will it consume, and what happens when it is unavailable?**
That may make future data-center proposals resemble power-development projects as much as technology projects.
- Communities May Demand Much More Information Before Approving Data Centers
There is another issue running through the paper that receives less public attention: **information**.
Grid planners need reliable forecasts of when large data centers will actually be built and how much electricity they will use.
But proposed projects do not always become real projects.
Developers may submit multiple requests while considering different sites. Project sizes can change. Construction dates can move. Some proposals disappear entirely.
The paper warns that speculative applications, incomplete information and congested interconnection queues can cause utilities and grid operators to overbuild, underbuild or delay needed infrastructure.
The same transparency problem exists at the community level.
Residents considering a new data center may reasonably want to know:
How much electricity will it consume?
How much water?
What kind of cooling will it use?
Will new transmission lines be required?
Will backup generators operate?
How much noise will neighboring properties experience?
What public subsidies or tax exemptions will the project receive?
How many permanent jobs will it create once construction is finished?
And who will pay if roads, water systems or electric infrastructure must be expanded?
Pennsylvania's new rules provide an early example of what greater disclosure could look like.
The state is requiring operators to report information including total energy use, peak hourly electricity requirements, natural-gas consumption, annual water use and maximum daily water demand. The order also bars nondisclosure agreements for data-center projects under the covered state process and requires more community engagement.
Pennsylvania also removed AI data centers from its expedited permitting program and requires local approval before certain state permits can be issued.
That policy does not prove every state will follow the same model.
But it illustrates a broader change already underway.
Data centers are increasingly being treated not merely as private real-estate developments but as projects that can significantly affect **public infrastructure and shared resources**.
And when private projects depend heavily on shared resources, public demands for transparency tend to increase.
- The Logic Points Toward Something Like a "Data Center Impact Statement"
The authors do **not** propose creating something formally called a "Data Center Impact Statement."
That is a policy inference from their analysis.
But their findings make the idea easy to understand.
At present, different pieces of a large data-center project are often examined by different institutions.
A planning department looks at zoning and land.
A water utility examines water demand.
An electric utility studies service requirements.
A regional grid operator considers transmission and reliability.
Environmental regulators evaluate particular permits.
Economic-development agencies calculate investment and employment.
Tax agencies determine whether incentives apply.
Each may be doing its job correctly while still seeing only part of the picture.
The study repeatedly argues that those pieces interact.
Cooling decisions affect both water and electricity.
Electricity choices affect emissions.
Location affects transmission requirements.
Power systems can create noise and land impacts.
Utility pricing determines who bears infrastructure costs.
And the value of each resource varies geographically.
That suggests a case for evaluating major data centers as integrated infrastructure projects.
For a sufficiently large proposal, a comprehensive assessment might ask:
How much electricity will the facility require at full operation?
What generation will supply the additional demand?
What transmission and distribution infrastructure must be built?
Who will pay for it?
What happens during electricity shortages?
How much water will the facility consume?
Where will that water come from?
How scarce is it locally?
What cooling technology will be used?
What backup or on-site generation will operate?
What are its emissions and noise effects?
How much land will the entire development—including supporting energy infrastructure—require?
What public tax benefits will the project receive?
How many construction jobs will it create?
How many permanent jobs will remain?
And what happens if the project is canceled after infrastructure has already been built for it?
The purpose would not necessarily be to stop data centers.
It would be to answer a question that is becoming increasingly important:
- Of all the places where this facility could be built, is this actually a sensible place to build it?**
- A Different Way to Think About the AI Buildout
The emerging data-center debate is often portrayed as a conflict between people who support artificial intelligence and people who oppose it.
The new study points toward a more useful distinction.
The issue is not simply whether America should have more computing infrastructure.
It is whether that infrastructure is being located and operated in ways that reflect the real costs of electricity, water, land and supporting infrastructure.
Some locations may be able to accommodate very large data centers with relatively modest additional costs.
Others may require new power plants, transmission lines, water infrastructure and extensive public investment.
Some cooling technologies may be appropriate in one climate and wasteful in another.
Some AI workloads may be flexible enough to move to times when the grid is less strained.
Some developments may be able to supply new low-carbon electricity.
Others may increase dependence on existing fossil-fuel generation.
These differences mean there may be no meaningful answer to the question, "Are data centers environmentally sustainable?"
The better question may be:
- Which data centers, operating in which places, under which conditions?**
That may eventually transform the competition among states and communities.
During the early AI boom, governments often competed to attract data centers by offering inexpensive land, tax exemptions and faster approvals.
In the next phase, the bargaining power could begin to shift.
Communities with abundant electricity, adequate transmission capacity, sustainable water supplies and appropriate industrial land may become increasingly valuable to developers.
Instead of communities competing simply to attract data centers, **data centers may increasingly compete for communities capable of supporting them.**
That could also change what qualifies as a "good" site.
The best location may not be the one offering the largest tax incentive.
It may be the one where additional electricity can be produced cleanly, transmission already exists or can be expanded efficiently, water is available without displacing higher-value uses, neighboring communities can be protected from noise and other impacts, and developers—not existing residents—bear the infrastructure costs they create.
In that sense, the biggest implication of the new paper is not a new estimate of AI's environmental footprint.
It is a change in perspective.
A data center should not be judged solely by what happens inside its walls.
Its true footprint extends outward into power plants, transmission lines, water systems, land, utility bills and surrounding communities.
And as AI computing grows from a technology industry into one of America's largest new infrastructure demands, that broader footprint may increasingly determine **where the next generation of data centers can actually be built**.
The Bolaños-Zuñiga and Lamadrid paper should be interpreted with appropriate caution. It is an arXiv preprint submitted on August 10, 2026 and has not gone through the peer-review process associated with publication in a scientific journal. Much of its analysis synthesizes existing research, projections and regulatory-economic concepts rather than reporting a nationwide field study of individual data centers.
But events are already moving in the direction the paper describes.
Pennsylvania is demanding that developers internalize infrastructure costs and disclose more information. PJM is creating special mechanisms for enormous new loads and projects bringing their own generation. Grid planners are confronting electricity growth occurring faster than traditional infrastructure can be constructed.
The AI data-center boom is not ending.
It may instead be entering a more difficult phase—one in which computing power is plentiful, but **suitable places to put it are not.**
