What Does a Data Center Actually Sound Like? The Hum, the Physics, and How to Measure It

Stand at the fence line of an operating campus and the first thing you notice is that nothing changes. There is no engine revving, no gear shift, no pause. There is a broad, low, continuous mechanical sound, similar in character to a large air conditioner heard from the next yard over, layered with a faint higher pitched whir from fan blades if you are close enough. JLARC's description, drawn from residents and county staff in the largest data center market on earth, is a constant "drone" or "hum," similar to house air conditioning systems but magnified to an industrial scale. Residents told JLARC the sound could be heard both inside and outside their homes.

That constancy is the whole story acoustically. Traffic noise is cyclical: it peaks at rush hour, thins after dinner, and mostly disappears by 2 a.m. A data center's cooling load does not fall at night, because the servers do not stop computing at night, so the sound level at 3 a.m. is roughly the sound level at 3 p.m. What changes is everything else. Neighborhood background noise drops after dark, and a hum that was masked by daytime activity becomes the only thing in the soundscape. That is why the query people actually type is about data center noise at night, and why the phrase residents keep using is not "loud" but "inescapable."

Living near one, then, sounds like a low background pressure that you notice in the gaps: with the television off, in the backyard, lying in bed. JLARC staff interviewed residents near facilities that had drawn complaints and heard reports of migraines, disrupted sleep, difficulty concentrating, and people avoiding their own decks and yards because the sound is louder outdoors. Those are the same complaints that anchor the Fairwater class action in Mount Pleasant, Wisconsin, where Microsoft acknowledged that residents north of the campus had noticed a tonal humming sound and attributed it to cooling fans.

Nearly all of it comes from moving heat and moving electricity. A modern site, in JLARC's inventory, holds trailer sized backup generators, a median of 35 per site, lined up beside the building or housed in sheds; industrial scale cooling equipment such as chillers or water towers on the roof or alongside the building; and one or more electrical substations on or next to the property. Here is what each of those contributes, and when.

• Chillers and their compressors. Continuous. A chiller is a refrigeration machine, and like any refrigerator it runs a compressor and pumps. Compressors are the deepest, most persistent part of the signature.
• Dry coolers and condenser fan arrays. Continuous, and usually the single biggest exterior contributor. These are banks of large fans, sometimes dozens in a row along a roof line or a yard, throwing heat into outdoor air. The Fairwater complaint names condenser fans, chillers, cooling towers, and air handlers as the sources running around the clock.
• Cooling towers. Continuous where used. Fans plus falling water, which adds a broadband rushing component to the tonal fan sound.
• Air handling units. Continuous. These move air through the data halls and are more often heard as a hum through walls and roof vents than as a distinct outdoor source.
• Switchgear, transformers, and the on-site substation. Continuous, and energized whether or not the servers are busy. This is the purest tone on the site.
• Backup generator testing. Intermittent and much louder while it runs. Practices differ by market: Loudoun County reports that the majority of its data centers run backup diesel generators for one hour, once per week, for mandatory testing, while operators told JLARC that maintenance testing typically involves a short monthly test of 10 to 30 minutes plus one long annual test of one to four hours, staggered across a site. Our generator guide covers the permitting side.
• Gas turbines, at some AI campuses. Continuous where they exist, and a different animal entirely. Where an operator cannot get grid power fast enough it may build its own generation on site. The Southaven class action describes a turbine fleet that grew from 3 to 57 units in under a year next to residential streets, with residents reporting readings above 70 dBA at their property lines along with vibration they can feel.

Two practical implications. First, because the dominant sources run continuously, there is no quiet window to wait for and no schedule to negotiate, which is what separates this from a quarry or a rail yard. Second, whether any of it clears the property line depends on the design and the type of cooling system, and on local geography and surrounding buildings, as JLARC found. Two campuses of identical size can produce very different experiences three hundred feet away.

A hum is a tone, and a tone means something is repeating at a fixed rate. On a data center site, two mechanisms repeat with mechanical precision.

Fans. Every blade that sweeps past a fixed point in the housing pushes a small pressure pulse into the air. Those pulses arrive at a steady rate set by how many blades the fan has and how fast it spins, which acousticians call the blade passing frequency. Cooling fan noise concentrates at that frequency and its harmonics, with turbulence and unsteady flow spreading the rest of the energy into broadband hiss. Because a data center runs its fans at nearly constant speed, the tone sits at nearly the same pitch hour after hour. And because a campus runs many identical fans, the same tone arrives from dozens of sources at once.

Transformers. The core of a transformer physically flexes as the magnetic field cycles, an effect called magnetostriction, and the flux density peaks twice in every electrical cycle. That makes the fundamental hum frequency twice the electrical frequency, which on the United States grid at 60 hertz means 120 hertz, with additional harmonics above it from the nonlinear behavior of magnetic materials. That is the deep electrical buzz people notice near substations, and a large data center has substation scale electrical infrastructure on the property.

Why does a tone bother people more than louder, messier sound? Because the ear locks onto it. Loudoun County, which has more data centers than any jurisdiction on earth, puts the point plainly on its own noise page: single frequency tones can be more aggravating to the human ear than broadband background noise at similar volumes, and low frequency sound may pose adverse impacts ranging from general discomfort to potential health complications. That is a government describing why a facility can pass a decibel test and still make a neighborhood unlivable.

Low frequency sound behaves differently from the rest of the spectrum in two ways that matter to anyone living near a campus. It travels farther: the National Park Service notes that low frequency sounds travel farther than high frequency ones, which is why infrasound is useful for communicating over long distances. And at the very bottom of the range, below the roughly 20 hertz limit of audibility, the Department of Energy describes infrasound as typically felt rather than heard, as vibration or pressure, capable of causing structural vibration such as window rattling. That is the vocabulary residents reach for when they say they can feel the hum, and it is what the Southaven complaint alleges when it describes low frequency vibration.

Now the measurement problem. Almost every noise ordinance in the country is written in A-weighted decibels, and the A filter deliberately throws away low frequency energy to mimic the ear's reduced sensitivity at moderate volumes. The Minnesota Pollution Control Agency's measurement guide quantifies it: at 100 hertz, the A-weighting network filters out approximately 20 decibels from the incoming signal before it is combined into the reported level. C-weighting, by contrast, represents the sound pressure level the meter actually received and does not noticeably vary in its compensation across the audio spectrum.

Put those together and you get the central paradox of data center noise. JLARC stated it as a finding: the lower frequency noise data centers emit is not fully captured in A-weighted decibels, so data center noise rarely exceeds the allowable limits set in ordinances, despite the constancy of the sound being problematic for residents. Its Recommendation 8 asked the General Assembly to let local governments set maximum sound levels for data centers using alternative low frequency metrics, such as C-weighted decibels, and to put noise rules in zoning ordinances rather than general noise codes. Loudoun's own page concedes the same gap. A handful of jurisdictions have started closing it, including Prince William County with a C-weighted standard and the Village of Mount Pleasant with per octave band limits; our ordinance table has the exact sections and figures.

Decibels are logarithmic, which is why intuition fails. The Minnesota guide gives the working rules of thumb: a change of about 3 dBA is the threshold of perception, 5 dBA is a clearly noticeable change, and an increase of 10 dBA sounds twice as loud. The National Park Service puts the same relationship the other way around, noting that a 10 dB increase causes a doubling of perceived loudness and represents a ten-fold increase in sound level. Adding sources follows a different rule: doubling the sound energy, for example running a second identical generator, adds about 3 dB, so two units at 70 decibels produce roughly 73, and four produce roughly 76.

Distance is where expectations break down. Minnesota's guide illustrates two cases. Sound from a point source, like a factory, spreads out into a sphere and drops about 6 decibels each time the distance doubles: 70 decibels at 50 feet becomes 64 at 100 feet and 58 at 200 feet. Sound from a line source, like a busy highway, spreads into a cylinder and drops only about 3 decibels per doubling: 70 at 50 feet, 67 at 100, 64 at 200. A quarter mile long building face lined with condenser fans is not a point in space. Geometrically it behaves closer to the line source case, decaying more slowly, which is one reason a facility stays audible much farther out than a single piece of equipment would.

For calibration against ordinary life, these are published reference points, not our estimates. The Minnesota guide's chart puts a bedroom at night around 30 dBA, a quiet urban or suburban night around 40, a dishwasher in the next room or a quiet urban daytime around 50, normal speech at one meter around 60, and a vacuum cleaner at three meters around 70. The National Institute on Deafness and Other Communication Disorders puts normal conversation at 60 to 70 dBA and says sounds at or below 70 dBA are unlikely to cause hearing loss even after long exposure, while long or repeated exposure at or above 85 dBA can cause it. NIOSH sets its recommended workplace exposure limit at 85 dBA averaged over an eight hour day.

Health guidance for sleep sits far below all of that. The World Health Organization's Night Noise Guidelines for Europe set an outdoor night level of 40 decibels as the target to protect the public including children, the chronically ill, and the elderly, with 55 decibels offered only as an interim target. The United States Environmental Protection Agency, in its 1974 identification of safe levels, named 55 decibels outdoors and 45 decibels indoors as the levels that prevent activity interference and annoyance. Set those against JLARC's finding that complaint-generating data center noise in Virginia measured an estimated 40 to 59 A-weighted decibels and the picture resolves: the readings that fail the health guidance and the readings that pass the ordinance are frequently the same readings.

You do not need to be an acoustician to produce a record a professional can build on. You do need to be repeatable. The protocol below follows the general measurement procedure in the Minnesota Pollution Control Agency's guide, which is representative of how agencies do this work.

• Use a real meter, and calibrate it. Minnesota requires a sound level meter and microphone meeting ANSI S1.4 Type 0, 1, 2, or S specifications, along with a calibrator of known frequency and level, with calibration performed before and after each monitoring period. A phone app has an uncalibrated microphone with unknown frequency response, especially at the low end where this sound lives, so treat app readings as a way to spot patterns and decide whether to buy a meter, not as the measurement itself.
• Record A and C together. Log both at every reading. The gap between them is the low frequency signature, and it is the number that explains why your experience and the ordinance disagree. Where a jurisdiction has adopted a C-weighted or octave band standard, the C reading may be independently enforceable.
• Measure at the property line and inside the home. Most codes regulate at a boundary, so a property line reading maps onto the legal test. An indoor reading with the windows closed captures what actually wakes you up. Label every location precisely and use the same spots every time.
• Control the conditions. Minnesota's procedure calls for measurements at least three feet off the ground, taken at least as far from large reflecting objects as from the source and in no case closer than 30 feet to structures, with wind speeds below 11 miles per hour and rainy conditions avoided. Note wind, temperature, and direction on every entry.
• Know your descriptors. L_Aeq is the A-weighted energy average over a stated period and is what most modern limits use. L_max is the loudest instant, useful for a generator test and nearly meaningless for a steady hum. Statistical levels are written as L followed by the percentage of time the level is exceeded: Minnesota's rules use L₁₀, the level exceeded 10 percent of an hour, and L₅₀, exceeded half the hour. By the same convention L₉₀ is the level exceeded 90 percent of the time, which is why it is treated as the residual background.
• Log over multiple nights, not one. A single reading proves nothing about a continuous condition. Weeks of entries at fixed times, including the small hours, prove persistence, and persistence is the entire point.
• Capture a baseline if you ever can. Minnesota's method separates a source from its surroundings by measuring total sound, then measuring again with the source off, and correcting for the difference; it also warns that confidence drops when the background is within 10 dBA of the source. If the facility is ever down, or if you can measure before a new phase energizes, take that reading. Some jurisdictions require the operator to file baseline studies, and those are public records worth requesting.

Why this matters legally, briefly and without any advice about your own situation: a facility can sit under its A-weighted limit and still be a private nuisance, because nuisance law asks whether the interference with your home is substantial and unreasonable rather than whether a number was crossed. JLARC's own finding is that data center noise rarely exceeds ordinance limits while remaining problematic for residents. A recording plus a log is what converts that mismatch from a feeling into an exhibit. Our evidence guide covers the full file, the ordinance table tells you which number governs your address, and the noise lawsuit guide explains how these claims are built. If the hum is affecting your sleep or your health, a free case review connects you with an independent attorney in our network who handles data center cases in your state, at no cost and with no obligation.