What does a data center actually do to a neighborhood?

Why this chapter exists

Local opposition to data centers is now the rate-limiting step on the AI buildout in several US states. The objections are real, the numbers behind them are mostly knowable, and the public conversation is mostly not based on those numbers. This page is the honest accounting — the strongest version of each objection, then what the data says.

The conclusion is not that the concerns are wrong. Several of them are well-founded and the industry has earned its reputation in specific places. The conclusion is that an informed resident, planner, or policymaker can sort the durable concerns from the inflated ones and ask for the specific concessions that move the needle.

1. Water

The objection: “A data center will drink the aquifer dry. They use millions of gallons a day.”

What is actually true: A large evaporatively-cooled data center can consume 1–5 million gallons of water per day in summer. That number is real. Whether it matters depends entirely on where the water comes from and what is next to the facility.

The comparison that matters: a 100 MW data center uses roughly the water of a 30,000-person town, or one mid-sized semi fab, or ten desert golf courses. That is meaningful, but it is not exotic. Agriculture in the same region almost always uses one to two orders of magnitude more. In the Phoenix metro area, agriculture accounts for roughly 70 percent of water use; municipal and industrial demand combined is about 25 percent; data centers are a small fraction of that 25 percent.

The legitimate version of the water concern is local: an aquifer under stress, a small municipal system, or a region in multi-year drought. In those places a new evap-cooled data center is genuinely net-bad and the operator should be forced onto closed-loop or air-cooling. In water-abundant regions (Virginia, Ohio, the Pacific Northwest most years) the same facility is closer to a rounding error.

The other thing to know: water cooling is a choice. Closed-loop liquid cooling, direct-to-chip systems, and air cooling all exist. They cost more, sometimes use more electricity, but consume near-zero water on an ongoing basis. When an operator picks evap, they are picking a $/MW number over a gallons/day number. That is a negotiable trade.

2. Electricity bills

The objection: “Hyperscalers will get sweetheart power deals and residential ratepayers will subsidize the new transmission.”

What is actually true: Some version of this is happening in real time in Virginia, Ohio, Georgia, and Oregon. Utilities are filing rate cases to recover billions in transmission and generation upgrades driven primarily by large-load AI customers, and the cost-allocation methodology is genuinely contested.

The honest read: large industrial customers pay lower per-kWh rates than residential customers everywhere, for legitimate reasons (they take power at higher voltage, with flatter load shapes, and need less distribution). But when a new transmission line is built specifically to serve one hyperscaler campus, allocating any of that cost to residential ratepayers is hard to defend. Several state public utility commissions (PUCs) are now requiring large new loads to be on dedicated tariffs that recover the marginal infrastructure cost from the customer that caused it. That is the right outcome and it is happening unevenly.

The resident-facing answer: ask the local PUC whether new data center load is being served on a Large Load Tariff with cost-causation principles, or on a general industrial tariff that socializes the buildout. The first protects you. The second does not.

3. Noise

The objection: “The data center sounds like a jet engine that never stops. Property values are dropping.”

What is actually true: Outdoor chiller plants, dry coolers, and emergency diesel test runs generate continuous low-frequency noise in the 60–75 dB range at the property line, sometimes higher in older designs. That is roughly the volume of a vacuum cleaner running constantly. For homes within 500–1000 feet, the impact is real and well-documented.

The Chandler, Arizona case and several Loudoun County, Virginia properties have measurable nighttime noise impacts. Residents are not imagining this. The mitigations exist — sound walls, lower-RPM fans, indoor chiller plants, larger setbacks — and add 5–15 percent to construction cost. Whether a jurisdiction requires them is a zoning decision that usually was not made before approval.

4. Land use and tax abatements

The objection: “The county gives away the tax base, the data center pays almost nothing, and the only thing the community gets is a windowless box.”

What is actually true: Tax abatements vary wildly. Some are net-positive for the host community; some are obvious giveaways negotiated by counties that did not understand what they were trading away.

The honest reference point is Loudoun County, Virginia. Loudoun is the largest data center cluster in the world and the numbers are public. Data centers generate roughly 30 percent of total county tax revenue through the personal property tax on computer equipment, on a relatively small land footprint. That funds schools, roads, and lowered residential property taxes. The deal is good for Loudoun residents in any honest accounting.

The Loudoun model is replicable, and it is also not replicated everywhere. A rural county that grants a 30-year sales tax exemption plus a property tax abatement on the equipment is giving up the entire economic case. The resident-facing question is: what fraction of the property tax base does this facility actually become, after abatements, in year 5?

5. Property values

The objection: “Nobody wants to live next to a data center. Home values within a mile crash.”

What is actually true: The peer-reviewed evidence is thin but converging. Studies of data-center adjacency in Loudoun County, Quincy WA, and the Phoenix metro show statistically significant negative effects on home prices within roughly 0.25–0.5 miles, in the range of 4–10 percent, mostly driven by noise and visual impact. Beyond 1 mile the effect is usually statistically indistinguishable from zero.

The bigger picture: in growing metros, total residential property values often rise because the tax base expands and schools improve, even as immediately-adjacent homes take a localized hit. That redistribution is real and it is unfair to the closest neighbors. Setback requirements, sound mitigation, and direct compensation to the closest properties are how other jurisdictions have addressed this. Most US data center zoning has not adopted those tools.

6. Secrecy

The objection: “They show up under a shell company name, refuse to say who the customer is, and the county approves before residents know.”

What is actually true: This is almost entirely true and it is the most legitimate complaint on the list. Hyperscalers routinely route site selection through shell LLCs (Project Sapphire, Project Mountain, etc.) and require NDAs from local officials before sharing the end customer. The stated reason is competitive — knowing where Amazon is building tells Microsoft and Google where to build next. The effect is that residents discover the project at the rezoning hearing, after months of negotiation, with no leverage.

There is no defensible argument for this practice once the project is approved. Disclosure of end customer at the public hearing stage is a low-cost ask that several jurisdictions are now requiring. It does not stop the project. It restores the basic civic premise that residents know who is moving in.

7. Jobs

The objection: “They promise hundreds of jobs and deliver fifty technicians who don't live in the county.”

What is actually true: This is largely accurate. A 100 MW data center typically supports roughly 50–100 permanent operations jobs once built — security, technicians, facilities, network operations. Construction jobs are real but temporary (2–4 years per phase) and predominantly filled by traveling specialty trades.

A data center is not a manufacturing plant. It is closer in employment density to a power station: high capital intensity, low headcount, jobs that pay well but do not multiply through the local service economy the way a factory does. Anyone who pitches it as a major jobs program is overselling. The honest pitch is tax revenue plus a small number of good technical jobs, not employment writ large.

8. Pollution and diesel backup

The objection: “The diesel generators run constantly and the air quality near the facility is bad.”

What is actually true: Backup diesel generators are tested monthly or quarterly. During normal operation they do not run. Annual diesel runtime is typically 25–50 hours per generator. Emissions over a year are real but small compared to highway traffic in the same airshed.

There are two legitimate exceptions. First, during grid stress events (heat waves, capacity emergencies), some data centers run generators voluntarily to relieve the grid; this can produce concentrated emissions over a few days per year. Second, a small but growing number of sites use natural-gas turbines as primary or supplemental power because grid interconnect is delayed. Those turbines run continuously and the air-quality impact is genuinely larger. The Memphis xAI Colossus site became the case study; the EPA filed enforcement notices in 2024-2025 over the turbines installed without permitting.

The resident-facing question: is the facility grid-connected with diesel backup, or is it running on-site combustion as primary power? The first has marginal air-quality impact. The second is a real ongoing emissions source and requires the same scrutiny as any new gas plant.

9. Grid reliability

The objection: “If a hyperscaler tanks the grid, my house loses power first.”

What is actually true: Large data centers do put new stress on regional grids and the recent July 2024 PJM event demonstrated how quickly synchronous load loss can cascade. In that event, a fault near a Loudoun County data center cluster caused multiple sites to trip offline within seconds, removing roughly 1.5 GW of load from the grid almost instantly. The grid handled it, but the analysis afterward made clear that data-center load behavior is now a system-stability factor that the grid operator did not have models for.

FERC and the regional reliability councils are now writing rules requiring data centers to ride through faults rather than trip, similar to what generators have always been required to do. This is a real new constraint on facility design and it is the correct response. The risk is real; it is being addressed; the addressing is several years from being complete.

10. The bigger objection: “AI is bad for society”

The objection: “Even if every local impact were managed, why should we build infrastructure for a technology that mostly automates white-collar jobs, concentrates wealth in five companies, and gives my kid a worse internet?”

What is actually true: This is the hardest concern to answer with data, because it is not really a data-center question. It is a society question that is showing up at the zoning hearing because the data center is the visible piece. A rural Virginia resident has no leverage over Anthropic's product strategy, but they can show up and oppose the substation.

The honest response is not to dismiss this. AI's near-term impact on labor markets, concentration of wealth, and quality of public information are all legitimate concerns that the industry is doing poorly at engaging with. But the data center itself is approximately neutral on these questions — it is the building that hosts the computation. Killing the data center does not change the AI roadmap; it moves the data center to a friendlier jurisdiction. The leverage on AI policy is elsewhere: at the federal level, in labor law, in antitrust, in liability rules. The zoning hearing is the wrong forum for the right question.

That said: when a community is being asked to absorb concentrated local costs (noise, water, transmission rate impact) for diffuse global benefits (cheaper AI for everyone, profit for distant shareholders), the calculus deserves real compensation. Host-community benefits agreements, dedicated tariffs, and binding noise/water commitments are the negotiable answer. “Trust us, it's for AGI” is not.

What honest data-center development looks like

The pattern that holds up under scrutiny is roughly:

• Dedicated large-load tariff with cost-causation principles — residents do not subsidize transmission built for the facility.
• Closed-loop or air cooling in water-stressed regions; evap acceptable in water-abundant ones with disclosure.
• Sound mitigation engineered into the design, not added after complaints — indoor chillers, low-RPM fans, generous setbacks (1000+ feet from residential).
• End-customer disclosure at the public hearing — no NDAs with elected officials before approval.
• Negotiated property-tax floor rather than open-ended abatement; a public schedule of payments over the facility life.
• Grid-connected primary power with diesel backup for emergencies only; any on-site combustion permitted as the new gas plant it actually is.
• Direct compensation to the closest neighbors for measurable property-value impact; this is cheap and almost never offered.

None of this stops the buildout. All of it is happening somewhere already. The places that get it right are the places where the host community had organized representation at the table early. The places that go badly are the ones where the project arrived faster than the local governance did.

The bottom line

Of the ten objections above, three are well-founded and largely unaddressed (secrecy, ratepayer cost allocation, noise near homes), three are real but mitigatable with off-the-shelf engineering (water in stressed regions, property-value impact at the property line, on-site combustion as primary power), and four are mostly wrong at the magnitude they are being argued (jobs, blanket water alarm, diesel backup emissions, grid collapse).

A resident reading this should walk into the next public hearing with the specific asks — large-load tariff, end-customer disclosure, sound engineering specs, closed-loop cooling commitment — rather than a binary up-or-down vote. The data centers will get built somewhere. The question is whether they get built well.

Sources

• EPA WaterSense Program data on residential water use.
• USGS National Water Use Compendium, 2020-2024 reports.
• Loudoun County, Virginia FY2024 Comprehensive Annual Financial Report; Data Center Coalition revenue disclosures.
• Lawrence Berkeley National Lab (LBNL) 2024 United States Data Center Energy Usage Report.
• International Energy Agency (IEA) Electricity 2024 data-center analysis.
• Virginia State Corporation Commission, Dominion Energy 2024 Integrated Resource Plan filings on large-load tariff proposals.
• PJM Interconnection event analysis, July 2024 Loudoun County voltage disturbance.
• EPA Region 4 enforcement notices, Memphis xAI Colossus turbine permitting, 2024-2025.
• Peer-reviewed property value studies: Journal of Real Estate Research and Energy Policy, 2021-2024 issues covering data-center proximity effects.