I spent years working as a municipal grant analyst, which means I spent years translating the distance between what a community needs and what it can actually afford into language a federal agency might fund.
You learn something doing that work.
Infrastructure is never free.
Someone builds it. Someone maintains it. Someone finances it. Someone assumes the risk if the assumptions behind it prove wrong.
And whenever a project is described primarily in terms of investment, tax base and economic development, it is worth looking farther down the ledger to find out where those other costs went.
That is increasingly where the American data-center boom becomes interesting.
The problem is not that data centers exist. Artificial intelligence, cloud computing, financial systems, communications and modern government all require enormous computational infrastructure. Pretending otherwise would be unserious.
The question is what that infrastructure actually costs, what it causes us to build around it, and who pays the bill.
That question is becoming harder to avoid.
Lawrence Berkeley National Laboratory now estimates that data centers could consume about 11.8 percent of all U.S. electricity by 2030, with considerable uncertainty around the final number. U.S. electricity demand is already setting records as AI infrastructure joins broader electrification in pushing consumption higher. [1]
In PJM - the regional grid stretching across 13 states and the District of Columbia - the 2027/28 capacity auction failed to secure its full reliability requirement by 6,517 megawatts as forecast demand continued to grow faster than new generation. [2]
This does not mean every rising electricity bill can be blamed on a server farm. Aging infrastructure, fuel costs, transmission constraints, weather, regulatory decisions and years of underinvestment all matter.
It does mean that adding power demand measured in hundreds or thousands of megawatts is no longer an ordinary economic-development decision.
It is an infrastructure decision.
And infrastructure decisions require an honest ledger.
The cost question is not optional
There is an old principle in public infrastructure finance called cost causation.
The idea is almost embarrassingly simple.
If a development creates the need for new infrastructure, the entity creating that need should bear an appropriate share of its cost.
Build a subdivision that requires new streets: account for the streets.
Build a factory requiring additional water-treatment capacity: account for the additional capacity.
Create a new electrical load comparable to a power plant: account for the generation, transmission, substations and equipment necessary to serve it.
The alternative is to privatize the economic return while spreading the infrastructure cost across people who had no role in creating the demand.
That was the concern I raised with the Dubuque City Council in June while the city was considering its approach to data-center regulation: Who builds what? Who maintains it? Who carries the liability? Property-tax projections alone cannot answer those questions. [3]
Congress is now wrestling with essentially the same principle.
The bipartisan Ratepayer Protection Act would establish a federal standard centered on recovery of the full incremental costs of infrastructure upgrades required to serve large-load customers. The House Energy and Commerce Committee advanced it unanimously this summer, and congressional leaders are now preparing it for floor consideration. [4]
That does not settle every regulatory question.
But it is an important recognition that a megawatt demanded by a hyperscale customer and a dollar paid by a residential ratepayer are connected policy questions.
And the physical bottlenecks are not theoretical.
More than 80 percent of the large power transformers used in the United States are imported. Demand has surged, prices have risen sharply, and lead times for some high-capacity transformers can approach four years. Data centers are not solely responsible for that shortage, but their rapid growth is competing for equipment already needed to replace aging grid infrastructure and connect new generation. [5]
The grid does not care whether a new megawatt is fashionable.
Someone still has to build the equipment that delivers it.
The Pentagon is downstream too
There is another customer on this grid that complicates the discussion.
The United States military.
Most electricity consumed by military installations in the continental United States comes from the commercial grid - a system RAND has described as largely outside Defense Department control and increasingly vulnerable to both natural hazards and deliberate attack. [6]
Congress has directed the Defense Department to provide at least 99.9 percent energy availability for the loads supporting critical missions at every installation by the end of fiscal year 2030. [7]
Those facts matter because the military's dependency does not end at the base gate.
Weapons manufacturers, logistics systems, transportation networks, communications infrastructure and commercial cloud systems supporting government operations draw from the civilian economy surrounding military installations.
A microgrid can help keep a base operating when the commercial grid goes down.
It cannot keep every off-base supplier, warehouse, fiber route, factory and logistics node feeding that base operating.
So the public, hyperscale technology companies and the military are increasingly downstream of the same infrastructure.
They may have very different reasons for needing it.
But they all need it to work.
That makes grid investment a legitimate public-security concern.
It does not mean that every residential customer should subsidize whatever infrastructure the largest commercial customers request.
Those are different propositions, and confusing them is precisely how costs disappear into the rate base.
Then the UAE changed the resilience question
This week, the United Arab Emirates provided a useful glimpse of what happens when another cost is added to the ledger.
The UAE had planned a vast 5-gigawatt AI campus in Abu Dhabi. After Iranian attacks damaged Gulf technology infrastructure earlier this year, Reuters reports that officials are reconsidering the concentrated design and looking instead at a geographically distributed network incorporating measures such as underground infrastructure, blast-resistant construction and air defenses. [8]
There is an obvious reason enormous campuses exist.
Concentration can be efficient.
Put compute, cooling, electricity, connectivity, engineers and security in the same place and the cost of each additional unit can decline.
That is an economy of scale.
But extraordinary concentration can create something else:
an economy of disruption.
The same architecture that lets an operator efficiently concentrate capability can let an outage, physical attack, cooling failure, transmission failure or other event disrupt an extraordinary amount of capability at once.
Suddenly redundancy is not waste.
Separate substations are not inefficiency.
Duplicate fiber routes are not needless expense.
Geographic dispersion is not poor utilization.
They become the cost of preventing a single failure from becoming a systemic one.
Call that resilience causation.
Cost causation asks who created the need for infrastructure.
Resilience causation asks what additional cost exists because the resulting concentration must survive failure.
That distinction is becoming important even absent missiles.
On July 22, nearly 4,000 megawatts of data-center load in Northern Virginia unexpectedly disconnected from the PJM grid and switched to backup generation. The sudden loss of load forced grid operators to manage resulting changes in system voltage and frequency. PJM says it was the third measurable event of its kind in two years and has proposed new reliability standards in response. [9]
That episode turns the conventional discussion on its head.
A data center can present a grid challenge while consuming enormous amounts of electricity.
It can also present a grid challenge when it abruptly stops consuming it.
At sufficient scale, the facility ceases to behave like an ordinary customer.
It becomes part of the operating conditions of the system itself.
And then there is the place where we put it
Once infrastructure reaches that scale, resilience becomes inseparable from land-use planning.
Not every data center is a plausible target for a foreign adversary.
Scale matters. Workload matters. Network importance matters. Government dependency matters. Geography matters.
Pretending every server building is a military target would be alarmism.
Pretending none of those distinctions matter to local planning would be equally foolish.
A sufficiently important data center may contain or support financial systems, communications infrastructure, government workloads, artificial-intelligence systems or other services whose interruption carries consequences well outside its property line.
The UAE is now considering underground structures and air defenses for parts of its AI infrastructure.
That should at least cause American planners to reconsider whether the land-use category 'data center' tells them everything they need to know about the facility being permitted.
This is not an exotic concept.
Land-use regulation exists because what one property owner does can affect people outside the property line.
We regulate setbacks.
We separate incompatible uses.
We examine flood exposure.
We regulate hazardous materials.
We consider traffic, noise, emergency access and environmental effects.
The principle is that private land use can create public consequences.
Strategically consequential digital infrastructure is not exempt merely because the externality is unfamiliar.
The planning question therefore should not be, 'Is a data center permitted in this industrial district?'
It should include a harder question:
What is the project's failure domain?
If this facility loses power, cooling, water, communications or physical access, what else fails with it?
If a substation serving it fails, who else loses service?
If several enormous facilities share a transmission corridor, fiber route or watershed, what happens when planners approve the fourth because the first three were each individually permissible?
And if the scale or function of the infrastructure makes the location more consequential to a hostile actor, what risk - however small in probability - is being imposed on people living beyond the fence?
Those are not reasons automatically to reject the project.
They are reasons to plan it.
Iowa is already running the experiment
You do not have to travel to Virginia or the Persian Gulf to see the governance problem.
It is happening in Linn County, Iowa.
In February, the county adopted special zoning requirements for large data centers covering water studies, traffic, emergency planning, noise, utilities and other long-term impacts.
By July, county supervisors had already concluded that more study was necessary and imposed an 18-month moratorium on new large-scale data-center rezoning applications in unincorporated areas. The county specifically identified emerging questions involving public infrastructure, water resources, electrical infrastructure, emergency response, transportation, land-use planning and cumulative effects. [10]
Then came another question: Should the county change the zoning district containing nuclear generating facilities so large data centers could locate there in connection with nuclear power?
Residents objected that the county had not yet resolved the same water, grid, environmental, decommissioning and land-use questions that prompted the moratorium.
In August, supervisors declined to proceed with the change. [11]
Meanwhile, those county actions do not control incorporated communities within Linn County.
Palo has been conducting its own process after Google announced a roughly 545-acre project there. Residents packed a public meeting and raised concerns about water, traffic, noise, pollution and infrastructure. Public records reviewed by Iowa Public Radio indicated that the proposed facility could draw as much as 14 million gallons of water per day from the Cedar River. [12]
And data-center construction already underway inside Cedar Rapids continues unaffected by Linn County's moratorium. [10]
There, in miniature, is the planning problem.
The jurisdiction is local. The systems are not.
Water crosses boundaries.
The electrical grid crosses boundaries.
Road networks cross boundaries.
Emergency services assist one another across boundaries.
Environmental effects cross boundaries.
Economic benefits and infrastructure costs can fall in different jurisdictions.
And whatever security or resilience consequences arise from concentrating critical infrastructure do not consult the municipal code before spreading outward.
A project can therefore comply perfectly with each individual zoning decision while the regional development pattern becomes irrational.
Zoning asks whether a particular use is allowed on a particular parcel.
Planning is supposed to ask whether we should be creating the resulting system at all.
That distinction matters enormously here.
An honest ledger has two sides
None of this requires pretending that data centers produce no public benefit.
That would make the analysis as shallow as pretending they produce only benefits.
Consider another Iowa development.
Google's electricity demand is helping support the planned restart of the Duane Arnold nuclear plant, which would return roughly 615 megawatts of generation to the grid. The federal government has also committed up to $1.9 billion in loan financing to the restart. [13]
That is not the same economic arrangement as dropping a huge new load onto a constrained grid and asking existing customers to finance the upgrades.
The demand may be helping cause new generation capacity to exist.
Good.
Put it on the positive side of the ledger.
Then put the federal financing, infrastructure requirements, grid effects and risks on the other side.
The point of cost causation is not to reach a predetermined verdict.
It is to prevent costs and benefits from being counted selectively.
A data center that finances new generation, pays the incremental transmission cost, improves infrastructure, provides meaningful public revenue and bears its own development risks should receive credit for those contributions.
A project that consumes scarce capacity, requires expensive upgrades, receives public incentives and leaves ratepayers or taxpayers holding the long-term risk should not be credited as though those things did not happen.
The analytical question is not:
Are data centers good or bad?
It is:
Does the deal add up?
The politics are beginning to reflect that
For a while, opposition to data centers could be dismissed as another local land-use fight.
That is becoming difficult.
Reuters documented coordinated protests across 42 states this summer. In Texas, political leaders who previously embraced data-center development are now reconsidering tax incentives and calling for stronger restrictions amid concerns about utility costs, land use and local benefits. And Congress is preparing to vote on bipartisan legislation explicitly intended to keep large-load infrastructure costs from falling onto ordinary ratepayers. [14]
This should not be surprising.
People generally understand a basic bargain.
If a company earns the return, the company should bear the costs that made the return possible.
If public money is invested because a project creates a genuine public benefit, the public should be able to identify what it purchased.
If a development creates extraordinary infrastructure or resilience requirements, those requirements should be disclosed before approval rather than discovered afterward.
And if concentrating strategically important infrastructure changes the risk borne by surrounding communities, people living there deserve more than an assurance that the parcel was properly zoned.
That is not anti-technology.
It is competent government.
Make the deal add up
The AI infrastructure boom does not need a single national verdict.
Different projects will produce different answers.
Some locations possess abundant generation, adequate water, appropriate land and room for transmission expansion.
Some data-center investments may finance generation or grid improvements that would not otherwise be built.
Some communities may decide that the economic return is worth the tradeoffs.
Others may reasonably conclude that a proposed facility asks too much of resources already under strain or places too much infrastructure inside one failure domain.
The policy framework should be capable of telling the difference.
Start with cost causation: the large load bears the full incremental infrastructure costs it causes unless regulators can identify a genuine shared benefit that justifies broader allocation.
Then conduct the same analysis for water, transportation and emergency services.
Require cumulative-impact planning rather than pretending each new campus is the first one.
Require developers and utilities to disclose the reasonable range of load, infrastructure and resource requirements rather than just the project's tax valuation.
Examine the failure domain: what systems become mutually dependent, and what happens when one stops working?
And when public financing or incentives enter the transaction, make the public benefit explicit enough that somebody can check later whether it actually arrived.
None of those requirements prevent development.
They make development legible.
That is what planning is supposed to do.
For decades, economic-development decisions have been dominated by the numbers visible at the ribbon cutting:
Capital investment.
Construction jobs.
Tax valuation.
Those numbers matter.
But the AI infrastructure boom is exposing the rest of the balance sheet.
Generation.
Transmission.
Transformers.
Water.
Roads.
Emergency services.
Redundancy.
Security.
Public financing.
Decommissioning.
And the consequences of concentrating extraordinary amounts of capability in places where people already live.
Development accounting asks what something costs to build.
Cost causation asks who created the bill.
Resilience causation asks what it costs to keep the resulting system functioning when normal conditions disappear.
And responsible planning asks one more question:
Is the cheapest place to build also an appropriate place to concentrate the consequences of failure?
The UAE is asking that question now because events forced it to. [8]
Linn County is asking versions of it before the next round of development arrives. [11]
The rest of us still have that opportunity.
Historically, societies become very interested in resilience immediately after discovering that something they treated as ordinary infrastructure was critical infrastructure all along.
The servers are already being built.
The grid is already being asked to accommodate them.
The zoning hearings are already happening.
We are still in the before.