The hum doesn’t stop. That’s the whole complaint.
Microsoft’s own community update acknowledged the “tonal humming sound” reaching homes near its flagship AI campus. A New Jersey county put a nighttime noise violation on the record. Residents describe vacuum cleaners that never shut off and picture frames that rattle. This is the narrowest question in the series, and in 2026 the most litigated.
12 min read · Answers the ninth of the ten questions
What the hum physically is.
A data center's sound is unlike traffic or construction in the three ways that matter: continuous, tonal, and weighted toward frequencies that walls barely stop.
data center class actions filed in 2026 are about noise, by one legal survey's count
Reporting on 2026 filings, retrieved August 2026
measured by a television crew at a Loudoun County facility, against the county's 55 dB residential-line limit
NBC4 / WJLA measurements (2024–26)
Amazon reportedly spent replacing rooftop fans near one Virginia community; the replacement lowered the pitch, and residents then reported rattling frames
Inside Climate News (2023); Prince William Times (2025)
New Jersey's nighttime limit, which a county health inspection found the Vineland facility exceeding, a violation notice on the record
CBS Philadelphia; Cumberland County (2026)
The sources of the sound are cooling and power: chiller and fan banks running around the clock, transformers, and, at sites generating their own power, gas turbines. Typical readings near facilities run 60 to 80 dBA at property lines, with diesel generator tests reaching 80 to 100+. Three properties make the experience worse than the numbers suggest. The sound never stops, and chronic exposure is what the health literature studies. The sound is tonal — a hum at a specific pitch reads as louder and more intrusive than broadband noise at the same meter reading. And much of the energy sits below 100 Hz, where sound travels farthest, penetrates walls best, and registers weakest on the A-weighted meters most ordinances specify.
The health baseline comes from the WHO’s environmental noise guidelines: chronic exposure links to sleep disturbance, annoyance, and cardiovascular and metabolic effects, with roughly 16,600 premature deaths a year in Europe attributed to transport noise. Read that citation carefully: the WHO studied traffic and aircraft, not data centers. What the guidelines treat as most intrusive at night, though, is precisely the continuous tonal pattern data centers produce, and a 2026 review of low-frequency noise around data centers is the literature’s first attempt to close that gap directly.
Measured and litigated.
Not anecdotes: violation notices, acknowledged complaints, and federal class actions, each with a named facility and a date.
Mount Pleasant, Wisconsin.The $7.3B facility Microsoft calls its most powerful AI data center drew a federal class action in July 2026: cooling and generator sound allegedly audible a mile and a half out, on behalf of more than a thousand households. Microsoft’s own April community update had already acknowledged a “tonal humming sound” reaching residents to the north: an operator conceding the phenomenon while the litigation argues its extent.
Vineland, New Jersey. Residents describe loud industrial noise around the clock near the DataOne facility, the same site whose unpermitted generators anchor the enforcement guide. The county health department measured, found the state’s 50 dB nighttime limit exceeded, and issued a violation notice in March 2026; a class action followed in May.
Dowagiac, Michigan. A site that ran as a crypto mine from 2021 converted to AI computing in 2024, and the continuous cooling made the noise worse. Residents compare the sound to a vacuum cleaner running in the living room. The city passed an industrial noise ordinance and issued fines; the operator is contesting the measurements; the CEO publicly offered to buy the neighboring homes if mitigation fails; a class action arrived in May 2026. One site, the full arc, still unresolved.
Loudoun County, Virginia.In the densest data center cluster on earth, a news crew’s meter read 90 dB at one facility, and residents near another, which runs eight gas turbines continuously, kept complaining while official measurements came in below the county’s violation limit. The county’s response was to form a working group to review the ordinance itself, which is what a code admitting its criteria miss the problem looks like in procedural form.
One labeling rule this page enforces on itself: Granbury, Texas is a crypto-mining facility, and also the most heavily litigated noise site in the country, with four legal actions, health documentation, and a failed attempt to incorporate a town solely to gain noise-ordinance power. The acoustics and the law generalize to any 24/7 computing load; the workload attribution does not, and borrowing the case for AI without the label would be exactly the drift this series exists to catch.
The ordinance problem.
Whether a hum is a violation or a fact of life is decided by the plumbing of a local code: the limit, the weighting, the exemptions, and the grandfather clause.
The instructive case is Great Oak, in Prince William County. A 291-home community reported around-the-clock cooling noise from Amazon’s adjacent data centers, with sleep loss and migraines in the record. The county’s ordinance, written in 1989, exempted commercial HVAC noise, so for years there was no violation to find, because the code had defined the noise out of existence. The county removed the exemption in March 2023 and later adopted continuous-noise limits. Then came the two twists that make the case a syllabus: the roughly $40M fan replacement lowered the pitch of the sound, after which residents reported rattling picture frames. The mitigation moved the energy down the spectrum rather than removing it. And the new rules grandfather existing equipment, so the community that forced the reform may never benefit from the reform.
| Jurisdiction | The limit | The catch |
|---|---|---|
| Loudoun County, VA | 55 dBA at residential property lines | Complaints continued while measurements came in under the limit; the county formed a working group to review the ordinance |
| Prince William County, VA | 52 dBA day / 47 dBA night for continuous noise (10+ min) | Adopted 2023; existing facilities grandfathered |
| Dowagiac, MI | 65 dBA day / 55 dBA night, industrial | Adopted mid-fight; operator contests the city's measurement method |
| Vineland, NJ | 50 dBA at night, under state code | The limit the county cited the DataOne facility for exceeding |
| Cherokee County, NC (pre-2023) | No decibel threshold at all in the 1999 ordinance | No lever existed; the county reached for a moratorium instead |
The compliance gap, in one sentence: a facility can meet an A-weighted decibel limit while low-frequency sound still carries into bedrooms, because A-weighting discounts the frequencies that travel farthest. That is why residents and meters keep disagreeing, why the strictest county in the country is reviewing its own ordinance while measurements come in under the limit, and why newer ordinances have started requiring C-weighted measurements alongside.
What has actually worked.
The record has fixes in it, and the pattern across the fixes is consistent: measurement first, hardware second, code third.
Chandler, Arizona is the complete arc. Residents logged over a hundred reports of a high-pitched chiller hum beginning in 2017; three hundred signed a petition. Without litigation, CyrusOne retrofitted sound attenuation onto every chiller by October 2018 and committed to factory-attenuated units going forward; residents reported roughly 60% improvement. The city then wrote sound mitigation into a 2023 ordinance, and in December 2025 its council rejected a proposed AI data center outright. Complaint, retrofit, code, and finally a siting consequence.
New Albany, Ohio is the operator-side counterexample this page carries deliberately: after resident feedback about exhaust fans near Google’s campus, the company had a sound engineer engaged within two weeks and implemented monitoring and dampening. The city’s own FAQ reports four operational-noise complaints in fifteen-plus years of data center development. A city with an interest in saying so, saying so; the tier label on the citation does the disclosure. The point survives the source: outcomes vary enormously by operator and by code, which means neither “always a nightmare” nor “never a problem” survives contact with the record.
And the highest-leverage community move on the record is a measurement: the Prince William coalition commissioned a professional acoustics study, one of the only independent, methodical measurements of data center noise at residential receivers in the public record. Every argument after that study had a number the record was required to carry.
Nobody is counting.
The finding this series keeps producing, in its narrowest-scale form yet.
As of August 2026, CPAI could find no organization anywhere tracking data center noise complaints. The opposition trackers count blocked projects and explicitly exclude complaints; the facility maps count megawatts; the national noise-health nonprofits have flagged data centers as an emerging concern without a ledger. The EPA’s noise office was defunded in 1982, and the federal statutes that created it remain on the books, unenforced. Every number on this page exists because a specific newsroom, county inspector, or homeowners’ association produced it on their own.
So the state of knowledge is a patchwork of cases, and this page has treated the patchwork honestly: named sites, dated measurements, and no fleet-wide claim, because no fleet-wide data exists for anyone to claim from.
“Is the noise actually harming people, or just annoying them?”
The version that goes too far
Imports the WHO's mortality figures directly: data centers are killing thousands. The guidelines studied traffic and aircraft; no epidemiology of data center noise at that standard exists yet, and pretending otherwise sets the claim up to be debunked with its own citation.
The version that waves it away
Calls the complaints subjective: the meters show compliance, and annoyance isn't harm. This ignores that A-weighted compliance systematically discounts the low-frequency component, that sleep disruption is a health pathway in the chronic-noise literature, and that one county's health department found an actual violation when it finally measured at night.
What the evidence supports
Continuous tonal noise from named facilities is measured at and above local limits, acknowledged by at least one operator in writing, cited by a county health department, and described consistently by residents across independent sites: sleep loss, migraines, rattling homes. Chronic-noise health mechanisms are established for the exposure pattern data centers produce, while direct epidemiology on data centers specifically is only now being attempted. Real, local, fixable where the record forced fixes, and uncounted at any scale beyond the fence line.
Sources for this split: fairwaterSuit · vinelandNoise · whoNoise · lfnReview — full citations below.
Action for every level of influence.
For yourself
- Read your local noise ordinance tonight; every code is public. Three things to find: the decibel limit at a residential line, whether continuous industrial noise is exempted, and whether anything addresses low-frequency sound. Prince William's HVAC exemption sat unquestioned for 34 years.
- If a hum is reaching your home, log dates, times, and conditions. Contemporaneous logs are what turned scattered complaints into the Chandler petition and the Vineland citation.
For a community
- Request a measurement from code enforcement in writing, and ask that C-weighting be recorded alongside A-weighting. The low-frequency component that penetrates walls is the part A-weighted meters discount.
- Pool for a professional acoustics study before the decisive hearing. The Prince William coalition's commissioned study gave every subsequent argument a number the record had to carry; smartphone readings gave the operators a rebuttal.
For an organization
- If you operate or host facilities: Chandler's retrofit and Google's New Albany response show the working playbook. A sound engineer within weeks, attenuation hardware, and factory-attenuated units in procurement. The alternative playbook is now a docket of class actions.
- If you advocate: the WHO noise guidelines and the low-frequency review give health arguments a peer-reviewed spine. Anchor testimony there, not in adjectives.
For policy
- Continuous-noise limits distinct from peak limits, since a 24/7 hum at 60 dBA and a passing truck at 60 dBA are different harms. Prince William now sets 52 daytime and 47 nighttime for sustained noise.
- C-weighted or low-frequency criteria in new ordinances, because A-weighted compliance and sleepless bedrooms demonstrably coexist.
- Setback and attenuation requirements at approval, when the leverage exists. Grandfathering is why the community that forced Prince William's reform may never benefit from the reform.
Related
What Are We Arguing About?
The ten separate questions hiding inside “AI and the environment” — where each impact lands, and where each is actually decided. Start here.
Electricity and Emissions
US data centers used 4.4% of national electricity in 2023, headed for 6.7–12% by 2028. What that does to emissions depends on something the number doesn’t contain.
The Water Question
The most-quoted statistic in this debate has been retired by its own authors. What replaced it is more useful — and it points at your watershed, not the planet.
Where this leads
CPAI teaches this in workshops and cohort programs.
We deliver this material to schools, libraries, employers, and community organizations — in person and online.
Research & further reading.
Dealing with a hum, or a hearing about one?
CPAI's community workshops cover reading your noise ordinance, logging what a code officer can use, and what to ask for at approval: the questions that are cheap before construction and expensive after.