The Price of Consent: What Data Center Neighbors Get Paid, and the Risk Nobody Asks Them to Price

The median American who opposes a local data center says $10,000 would change their mind. That is the price of the noise. The price of the risk the building carries is set somewhere else entirely, and it reaches the same households through their power bills and their pensions without anyone asking them to name a number.

In June, researchers at the University of Virginia asked 3,234 American adults a blunt question about data centers: what would it take? Respondents who did not support a data center in their area were asked to name the price of their support. The median answer was $10,000, as a one-time payment. About 22% wanted more than $1 million, and 7% said no amount would do it. The question is not academic. Gallup found in March that 71% of Americans would oppose a data center in their own community, more than the 53% who would oppose a nearby nuclear power plant.

So the AI build-out now has two prices attached to the same building. One is the price of local consent, paid in cash to neighbors, school staff and fire departments. The other is the price of the risk the building represents, set in the credit and power markets. They are paid by different people, quoted in different places, and differ in one important respect: only the first one is ever negotiated with the people who live next door.

The Running Tally of Local Consent

The consent side of the ledger is visible, local and remarkably specific. In Richland Parish, Louisiana, sales tax from construction on Meta's data center funded end-of-year checks of up to $50,935 for eligible certified school staff this June, scaled by seniority, against teacher salaries in the parish that run from roughly $29,500 to $52,300. Officials expect the windfall to shrink once construction ends, because it comes from building the facility, not from running it.

In Saline Township, Michigan, a community of 2,277 people, the board voted 4-1 in September 2025 to deny rezoning for a data center project, citing the master plan and a fire department not equipped for the facility. The developer and landowners sued, alleging exclusionary zoning. The settlement that let the project proceed included $8 million for area fire services, $4 million for a farmland preservation trust and $2 million for a community investment fund. In Memphis, the city passed an ordinance directing 25% of property tax revenue from qualifying AI facilities, including xAI's Colossus campuses, to communities within five miles of the sites. The first year's amount came to about $3.3 million; the $100 million figure often quoted is a long-run target, not a payment.

Where consent was not sought, the nuisance arrives anyway. Across the state line in Southaven, Mississippi, xAI has run 69 temporary gas turbines to power its Memphis operations without an air permit for those units, which the state treats as exempt portable equipment. Under a state order issued at the end of July, the last of them are to be retired by July 2027.

What every item on this list has in common is that it compensates for nuisance: noise, land, traffic, the changed character of a town. Every payment is one-off or tied to a construction phase. None of it compensates for what happens if the buildings stop paying for themselves.

The Bill That Arrives Through the Meter

The electricity question is where the evidence genuinely points two ways. A working paper from the Electric Power Research Institute estimates that in the average state, residential power rates were about 6% lower than they otherwise would have been because of the data centers built between 2019 and 2024. The mechanism is fixed-cost spreading: new, durable demand pays for part of a grid that ratepayers were already financing. The authors are explicit that the effect can reverse if supply bottlenecks bite or if the expected demand fails to materialize.

The forward-looking number is less comforting. PJM's independent market monitor attributes about $6.3 billion of the $16.4 billion in charges from the 2028/2029 capacity auction, roughly 38%, to data center demand, and about $29.4 billion across the last four auctions. In the monitor's view, those costs are not paid by data centers directly but are spread across all customers in the PJM footprint. Both findings can be true at once: the early data centers used spare capacity that already existed, and that spare capacity is now largely gone.

Utilities have responded with large-load tariffs. Indiana Michigan Power, for example, requires an initial contract term of at least 12 years and charges an exit fee if a customer cuts its contracted capacity by more than 20%. That protection is real, but it is only as long as the contract and only as strong as the counterparty. Transmission lines and substations built for a 20-year load do not leave when a tenant does. If demand falls short of the forecast, the fixed cost is recovered from the customers who stayed.

The Price Nobody Asks the Neighbors to Quote

The second price is set in credit markets, and it has been covered on this page before: protection on the most leveraged builders has repriced sharply, as the piece on CoreWeave's CDS spread laid out in October. The part worth adding here is a structural point Bloomberg's Matt Levine has made about why that protection is so expensive. The risk is almost entirely systematic. If one data center operator fails through bad management, someone else buys the building and re-lets it; that is idiosyncratic risk, and it diversifies away. If paying demand for AI computing comes in well below projections, every data center has the same problem at the same time, and no amount of diversification across operators helps. The natural sellers of protection, banks, hedge funds and insurers, already carry a great deal of that exposure, so they only take on more at a high price.

Investment-grade ratings then do the routing. They are what allow this debt to be held by pension funds, annuity writers and life insurers, the channel traced in earlier work on who ends up holding AI credit risk. Put the two prices side by side and the asymmetry is plain. A household near a new campus may be offered a one-off payment for the noise, priced at a median of $10,000 among those who object. The same household can hold the tail of the trade through three channels at once: the power bill that recovers grid costs if the load disappears, the pension or insurance policy that holds the investment-grade paper, and the local economy whose tax base was built around the facility. Nobody has asked it to name a price for that.

A Nineteenth-Century Version of the Same Trade

The scale invites comparison with the railroads, and the comparison is closer than it looks. Columbia economist Stijn Van Nieuwerburgh estimates the US data center and AI infrastructure build-out at roughly $10 trillion through 2032, around 3.6% of GDP a year. The nineteenth-century railroad build-out, by the same comparison, ran at a little over 2% of GDP.

Railroad finance had its own version of the consent payment, running in the opposite direction. Towns and counties bought railroad shares and bonds in the expectation of large local benefits, which often did not arrive for the places that paid. States responded by writing limits on local aid to railroads into their constitutions, Ohio in 1851 and Illinois in 1870 among them, and courts later voided bonds issued after those limits took effect. Midwestern farmers angry about railroad rates organized into the Granger movement, which helped produce the Interstate Commerce Act of 1887 and the Interstate Commerce Commission, the first federal regulator of an entire industry. The railroads got built. Many of the communities that financed them did not get their money back.

The difference today is how the exposure is acquired. A nineteenth-century town voted to buy railroad bonds at a public meeting. A twenty-first-century town acquires its exposure through a pension allocation and a rate case, neither of which is put to a local vote.

What This Means for a Portfolio

The strongest counterargument deserves its place: compute is currently rationed, demand exceeds supply at today's prices, and a bubble in the classic sense requires construction to outrun demand. The open question is what that demand looks like when most users pay the full cost rather than nothing. That is exactly the kind of uncertainty that shows up as a single common factor rather than as many independent ones.

For an allocator, the practical lesson is about where that factor hides. An investment-grade credit sleeve, regulated utility holdings in high-load grid regions, municipal credit in host communities and index-weight exposure to the largest technology companies can look like four separate risks on a statement. Underneath, they share one variable: whether paying demand for AI computing arrives at the scale already financed. Counting line items overstates how diversified that book is, a version of the AI concentration problem showing up in places that do not carry an AI label. The question worth asking of any portfolio is not whether the next data center gets approved. It is who absorbs the fixed cost if the tenant leaves, and how much of that cost the portfolio is already holding without having been asked.