How loud is a data center?
Cooling fans, chillers, transformers and backup generators make data centers hum around the clock. Here is how loud they are, why the legal limit does not always match what neighbours hear, and what actually quiets one down.
Where the noise comes from
A data center is loud for the same reason it is hot: thousands of servers in one building produce heat that has to go somewhere, and moving that heat takes machinery that makes sound.
The noise comes from four places. Cooling is the biggest: fans, chillers, dry coolers and cooling towers run continuously, and cooling accounts for about 40% of a data center's electricity use. Air-cooled sites tend to be noisier than water-cooled ones. Inside the building, the combined sound of thousands of servers can reach 96 decibels — that is the level inside the server hall, not at the fence, and it is the reason the building itself is sealed and insulated. Transformers hum constantly. Backup diesel generators are the loudest single source: an industrial-sized unit can reach 105 decibels, as loud as a jet flying overhead — but generators normally run only during monthly tests and real outages. The EPA allows emergency generators up to 50 hours a year for non-emergency testing; during a genuine emergency there is no federal time limit. A newer source is onsite power generation: gas turbines installed to power a data center directly run continuously, not just in emergencies, and their cooling towers can emit up to 70 dBA within 400 feet.
The important point is that the noise does not stop. A data center runs 24 hours a day, 365 days a year, and so does its cooling. That is what separates it from most industrial neighbours: a factory has shifts and quiet nights, and traffic has a rush hour. A data center has neither.
Why it sounds worse than the meter says
If the noise were only as loud as the meter says, there would be fewer complaints. Three properties of data-center sound explain why the meter and the neighbour disagree.
It never stops. Complaints about data-center noise tend to focus on its round-the-clock presence rather than its volume. A sound you can ignore for an hour becomes a different problem when it is still there at 2 a.m., and still there the next night, and the next.
It is tonal. Much of the sound is a hum at a specific pitch — the drone of fans and compressors running at a constant speed. A tonal sound reads as louder and more intrusive than broadband noise at the same meter reading, and it is the quality residents describe most often: a vacuum cleaner that never shuts off, a jet engine winding on the tarmac, a clogged vacuum running in the living room.
It is low-frequency. Much of the energy sits below 100 hertz. Low frequencies travel furthest, pass through walls and windows best, and register weakest on the A-weighted meters most ordinances specify. The World Health Organization's community-noise guidelines say plainly that A-weighting underestimates low-frequency sound, and that where prominent low-frequency components are present, A-weighted measures are inappropriate — C-weighting should be used instead. The WHO also notes that low-frequency noise from ventilation systems can disturb rest and sleep even at low sound levels.
That is why the same reading can mean two different things. A site measured at 55 dBA at the property line may be complying with its local limit — and still be audible inside a bedroom with the windows closed, because the part of the sound that gets through the wall is the part the meter discounts.
Distance helps, but slowly. Sound from a point source drops by about 6 decibels each time the distance doubles, in open ground with nothing in the way. A 70 dBA source at 400 feet needs roughly 2,300 feet of open ground to fall to 55 dBA. That is why setbacks matter more than any single piece of equipment: every doubling of distance is worth as much as a substantial piece of acoustic treatment, and it works on all frequencies at once.
What the law allows
There is no federal noise standard for data centers. The EPA published health-based noise levels in 1974 — 55 decibels outdoors and 45 decibels indoors, as levels that would prevent activity interference and annoyance — but those were recommendations, not regulations, and the EPA's noise office was defunded in 1981. Since then, noise has been a state and local matter, decided by zoning ordinances that were mostly written for block parties and factories, not for a building that hums 24 hours a day.
What those ordinances actually say varies widely. Some examples, all from real ordinances adopted or proposed in 2026:
| Place | Limit at the property line | Notes |
|---|---|---|
| Roswell, Georgia (adopted Aug 2026) | 55 dBA day (7 a.m.–10 p.m.), 45 dBA night; 60 dBC at any time | Noise survey by a licensed acoustics engineer, on both dBA and dBC scales, plus an annual report |
| Dowagiac, Michigan (adopted Mar 2026) | 65 dB day, 55 dB night (residential) | Notices of violation already issued; the operator disputes them |
| New Jersey (state code) | 65 dBA day, 50 dBA night (10 p.m.–7 a.m.) | The limit the Vineland facility was cited for exceeding |
| San José, California | 55 dBA at a shared property line with homes | Same standard in the city's General Plan for new nonresidential development |
| West Hanover Township, Pennsylvania (adopted 2026) | 50 dB(A) day, 40 dB(A) night; no more than 10 dB(A) above pre-development ambient | Absorptive barrier screening required around loud equipment |
| Bensalem Township, Pennsylvania (proposed) | 40 dB(A) and 50 dB(C) | 1,000-foot buffer from residential districts |
| Sellersburg, Indiana (proposed) | 50 dB day, 45 dB night | 1,000 feet from homes; limits on low-frequency rumble and steady tone |
Two things stand out in that table. First, the numbers are not the same from place to place — a site that complies in one town would violate in the next. Second, the newer ordinances measure low frequencies as well as the A-weighted total: Roswell caps dBC as well as dBA, and Bensalem and Sellersburg write limits for tone and rumble directly. That is the law catching up with the physics.
Virginia went further than any other state in 2026. Since July 1, 2026, Virginia law (Va. Code § 15.2-2209.5) requires a site assessment — including a sound profile of the facility's effect on homes and schools within 500 feet — before a locality approves any facility expected to need 100 megawatts or more. The same legislation directs the state's environmental agency to adopt regulations setting the "lowest achievable noise for a data center" by the end of 2029. Virginia is also the state where the problem is most concentrated: the state's legislative auditor found almost one-third of Virginia's data centers sit within 200 feet of residentially zoned land, and Northern Virginia alone holds nearly 300 operating data centers. You can see where facilities sit relative to homes on our data-center map.
Three places where it went to court
Dowagiac, Michigan — the toddler. The case behind our Short about a toddler hearing a data center from his crib is real, and it is in Dowagiac, a town of about 5,300 people in southwest Michigan. The site is a former copper factory where a company called Hyperscale Data began bitcoin mining in 2021 and has since begun moving into AI and cloud computing; residents say the noise got worse with the change. Lindy Valenzuela lives a few houses down the street. Her 1½-year-old son's nursery faces the facility, and the noise can be heard in his bedroom as he sleeps in his crib. "We have no idea how it's affecting them," she told the Detroit Free Press. "He is not talking yet; if it is bothering him in some way, he can't tell us. And that terrifies us." Her neighbour's meter read in the high 50s to low 60s on his porch; the company's own readings were in the 50s on the residents' side and the mid-60s on the other. The town adopted a noise ordinance in March 2026 — 65 decibels by day, 55 at night — and has already issued violation notices, which the company disputes. In late May 2026, Valenzuela and another neighbour filed a class action in federal court, alleging nuisance and loss of property value. The company's chief executive told the city council the firm would work on mitigation and offered to buy the closest neighbours' homes at fair value plus moving costs. Watch the Short: a toddler can hear this data center from his crib.
Mount Pleasant, Wisconsin — Microsoft's Fairwater. In July 2026, three residents filed a proposed class action against Microsoft over its Fairwater AI campus in Mount Pleasant, seeking to represent everyone within a mile and a half — more than 1,000 households. The complaint alleges noise from diesel generators and cooling equipment — chillers, cooling towers, air handlers and condenser fans — running around the clock, with community-facing levels that can exceed 48 to 60 dBA without adequate mitigation. Those are the plaintiffs' allegations, not measurements, and the case is pending. What is on the record is Microsoft's own acknowledgement: in an April 2026 community update the company said a "tonal humming sound" from cooling fans was reaching residents to the north. It added mitigation measures and said on June 18 that the issue was fully resolved — then said on July 20 that it was still working on short-term mitigation. The facility became fully operational in June 2026, with a second building under construction next door.
Vineland, New Jersey — the violation on the record. A 300-megawatt AI data center built by DataOne for Nebius, on South Lincoln Avenue in Vineland, drew complaints of a constant hum from early 2026. On March 10, 2026, the Cumberland County Department of Health inspected the site and issued a notice of violation for noise above 50 decibels between 10 p.m. and 7 a.m. — New Jersey's nighttime limit. Two residents filed a federal class action in May 2026 on behalf of about 1,013 households within a mile; it was voluntarily dismissed in June 2026 on a procedural point, without any ruling on the merits, and the claims can be refiled. The same site was fined $1,072,000 in September 2026 for running 62 natural-gas generators without air permits — a separate issue from the noise, but the same generators are named in the noise complaints. The operator's own acoustical consultant modelled the finished campus at a peak of 49.4 dBA at the nearest home, 230 feet away — just under the 50 dBA limit — behind 14-foot concrete walls on three sides, two earth berms and a planned 30-foot wall. That is how tight the margin is: a few tenths of a decibel, and a wall as tall as a three-storey building.
These are the new cases. The earlier one is Chandler, Arizona, where residents of the Brittany Heights neighbourhood have complained about a constant hum from a data centre since late 2014. In 2022 the city changed its zoning to make data centres harder to site, and in 2025 the council unanimously voted against a new proposed data centre, with noise concerns playing a big role. Chandler now requires a pre-construction baseline sound study, shared with residents, plus annual noise studies for five years after a facility opens.
What the evidence says about health
The honest answer is that the evidence is strong in general and thin in particular. There is a large body of research on environmental noise and health, and almost none of it is about data centres specifically — because the data-centre noise problem is only a few years old, and the studies have not caught up.
What the general evidence says is this. The World Health Organization's community-noise guidelines set 30 dBA indoors as the level below which continuous noise should not disturb sleep, with a still lower guideline where low-frequency sound dominates — and data-centre sound is dominated by low frequencies. The EPA's 1974 levels document put the outdoor threshold for preventing annoyance and activity interference at 55 decibels, and the indoor threshold at 45. The WHO's guidelines also note that the elderly, children, shift workers and people with chronic illnesses are the most likely to be affected. EESI's review notes studies showing that noise above 65 decibels can raise stress and blood pressure.
What that evidence does not say is that a particular data centre caused a particular person's illness. The symptoms residents report — headaches, sleep disturbance, tinnitus, dizziness, high blood pressure — are real reports from real people, and they are consistent with what the noise literature predicts. But they are complaints, not findings tied to a specific facility, and no study has yet measured the health of people living next to a data centre against a comparable group living away from one. Anyone who tells you the health effects are proven for data centres is overstating the evidence; anyone who tells you there are none is ignoring it.
One claim deserves a specific warning. You will read that data centres emit "infrasound" — sound below 20 hertz, too low to hear — that harms the central nervous system and heart. That claim is contested and not established for data centres. Low-frequency sound is real, and it is the core of the noise problem. Infrasound as a distinct health hazard from data centres is not proven, and this page does not present it as one.
What actually quiets a data center down
There is no single fix. The measures that work fall into five groups, roughly in order of how much they help.
Distance. The strongest lever, and the cheapest. Sound from a point source drops about 6 decibels every time the distance doubles, in open ground. A site 2,000 feet from the nearest home is a different problem from a site 200 feet away. That is why newer ordinances require setbacks — 1,000 feet from homes in the Bensalem and Sellersburg drafts — and why Virginia's 2026 law makes the sound profile at homes and schools within 500 feet part of the approval. It is also why the Virginia auditor's finding matters: almost a third of the state's data centres are within 200 feet of residentially zoned land, which is where distance cannot help, because there is none. Facilities end up close to homes for reasons our explainer on why data centers gather in the same places sets out — power, fibre and land in the same place.
Barriers and berms. Walls, earth berms and acoustic screens between the equipment and the neighbours. At Vineland the operator is building 14-foot concrete walls, two earth berms and a planned 30-foot wall. Barriers work best on high-frequency sound; low frequencies bend around and over them, which is why a wall alone rarely solves a data-centre complaint.
Enclosures and attenuators. The equipment itself can be quieted: sound-attenuating enclosures around generators, mufflers on exhaust stacks, sound attenuators on chiller exhaust paths. On one project an acoustic engineering firm needed 10-foot attenuators on every chiller exhaust to meet a 34 dBA target at homes roughly 1,500 feet away. Chillers can be enclosed in plant rooms lined with sound-absorbing material, with acoustically lined intake paths and discharge stacks. These are standard measures — the question is whether the operator installs them before opening, or after the complaints.
Quieter cooling. The most effective change is to make less noise in the first place. Liquid cooling removes heat at the chip instead of blowing air across it, so the large outdoor fan banks shrink or disappear — immersion cooling, where servers sit in a tank of fluid, eliminates the large fans entirely. Direct-to-chip liquid cooling still has pumps and some fans, so "no fans" is true of immersion, not of every liquid-cooled design. Our explainer on liquid cooling sets out how the heat path works; the 120 kW rack page shows the split in practice, with 115 kW removed by liquid and 17 kW still by air.
Generator-test scheduling. Generators have to be tested — at least once a month per unit — and on a large campus with hundreds of generators the testing can consume the entire allowable daytime window. The fix is to schedule it: daytime only, weekdays only, never overnight. Some ordinances now write that into law — Franklin County, Indiana's draft limits routine testing to Monday to Friday, 10 a.m. to 4 p.m., and bans it on weekends and holidays; Watts Township, Pennsylvania's draft bans testing between 8 p.m. and 7 a.m.
And one measure that belongs at the start, not the end: measure before building. A pre-construction sound study — the existing ambient, the predicted levels at every neighbouring property line, on both dBA and dBC scales — is what Chandler, Arizona now requires, and what Virginia's 2026 law requires for facilities of 100 megawatts or more. A study done after the complaints is a dispute; a study done before construction is a design input.
Questions readers ask
Is a data center louder than a factory?
Not necessarily at the fence, but it is louder at 2 a.m. A factory has shifts and quiet nights; a data center's cooling runs 24 hours a day. The complaint is rarely about a single loud moment. It is about a sound that never stops.
My neighbour's data center passed its noise test. Why can I still hear it?
Probably because the test measured what the ordinance measures, and the ordinance measures dBA. Most of a data center's sound energy is low-frequency, which dBA discounts and which passes through walls better than high frequencies. Ask whether the study measured dBC as well, and whether it measured inside a home at night, not just at the property line during the day.
How far away do I need to be?
There is no single distance, because it depends on the equipment, the barriers and the ground between you. As a rule of thumb, sound from a point source drops about 6 decibels each time the distance doubles, in open ground. A 70 dBA source at 400 feet needs roughly 2,300 feet to fall to 55 dBA. That is why setbacks are the strongest protection, and why a site 200 feet from homes — as almost a third of Virginia's data centers are — is hard to fix with equipment alone.
Can liquid cooling make a data center quiet?
It can make it much quieter, but it does not make it silent. Immersion cooling eliminates the large outdoor fan banks entirely. Direct-to-chip liquid cooling shrinks them. Either way the site still has transformers, pumps and generators — and the generators still need monthly testing. Quieter cooling is the most effective single change, but it is not a complete answer on its own.
What can I do if a data center is proposed near me?
Ask for the sound study before approval, not after construction: the predicted levels at your property line, in both dBA and dBC, day and night, including generator testing. Check what your local ordinance actually limits — many were written before data centers existed and say nothing about them. And ask whether the approval requires post-construction monitoring, so the prediction is checked against reality.
Noise is the same question our explainer on who pays for grid upgrades asks about money: when a data center arrives, who bears the cost — and who decides?
Sources and further reading
- Environmental and Energy Study Institute, "Communities Are Raising Noise Pollution Concerns About Data Centers"
- World Health Organization, Guidelines for Community Noise (1999)
- U.S. Environmental Protection Agency, Information on Levels of Environmental Noise Requisite to Protect Public Health and Welfare with an Adequate Margin of Safety, EPA 550/9-74-004, March 1974
- Detroit Free Press, "Constant noise from Dowagiac data center prompts class-action lawsuit," July 16, 2026
- U.S. District Court for the Eastern District of Wisconsin, Ostergaard et al. v. Microsoft Corporation, Case No. 2:26-cv-1169, complaint filed July 1, 2026
- SNJ Today, "Vineland Data Center, Who Owns It and What's Next," November 5, 2026
- Cumberland County Department of Health notice of violation, March 10, 2026
- Virginia Code § 15.2-2209.5 (HB 153 / SB 94, 2026), effective July 1, 2026
- City of Roswell, Georgia, Ordinance No. 2026-08-13 (adopted August 10, 2026)
- West Hanover Township, Pennsylvania, Ordinance 2026-01
- Salas O'Brien, "When data center noise becomes a development risk"
- Marshall Day Acoustics, Brooklyn Data Centre town planning assessment, Rp 002 R03, May 2026
