How Loud Is a Data Center?
Noise is the question people here ask most, and it’s the one this project has answered least. The commitment on the record is that operational noise will be no louder than a household air conditioner
at the property lines.1 At the September 9 town hall, University officials said noise levels haven’t yet been determined.
Both of those can be true. They can’t both be reassuring.
Here’s what’s actually known about data center noise, what the equipment manufacturers publish about their own products, and what you’d need to know to judge a claim like that one. Everything is sourced, and where the evidence is thin I say so.
First: there are two different noise problems
Almost every confused argument about data center noise comes from mixing these up.
Cooling equipment runs continuously. Chillers, cooling towers, air handlers and fans operate every hour of every day, at a moderate level. This is the quieter source and the more persistent problem.
Backup generators run occasionally and are very loud. They are the single loudest thing at any data center, but they operate only during outages and scheduled testing — under EPA rules, up to 100 hours per year for maintenance checks and readiness testing.2
They need separate answers. A facility can be completely truthful that its continuous noise is modest while also producing, once a month, a sustained event the neighbors hear clearly. Mixing the two is the fastest way for either side to talk past the other.
How sound is measured, and why the numbers mislead
Three things are worth understanding before any decibel figure means anything.
The scale is logarithmic. A 10 dB increase is roughly twice as loud to the human ear. The difference between 50 and 60 dB(A) is not 20% — it is double.
Sound falls with distance in a predictable way. From a point source in open air, levels drop about 6 dB each time the distance doubles. That single rule lets you check any claim, provided you know the level at a stated distance.
The A-weighting hides the part people complain about. dB(A) filters sound to approximate human hearing, which de-emphasises low frequencies. But low-frequency content — the hum below 100 hertz produced by large fans and transformers — travels much further than the A-weighted number implies and penetrates building walls more easily. This is why residents near data centers elsewhere report a persistent hum at facilities that pass their decibel tests. Some jurisdictions have begun writing property-line limits in octave bands specifically to capture low-frequency noise.3
What the generator manufacturers publish
This is the firmest ground available, because it’s factual data published by the companies that build the equipment. I should be clear that my work is with the gear that lives inside these buildings, not with designing or building them — but these figures are published, and anyone can check them against what gets installed.
Generator noise is conventionally stated as dB(A) at 7 meters (23 feet) at full rated load, averaged across eight microphone positions arranged around the set.4
Before any treatment, the individual sources are extremely loud. Cummins publishes the following at one meter:3
| Source | dB(A) at 1 m |
|---|---|
| Unsilenced engine exhaust | 120–130 |
| Engine block | 100–121 |
| Radiator cooling fan | 100–110 |
| Alternator | 80–90 |
Enclosures and silencers bring that down substantially. Published enclosure performance, measured at 7 meters at full load:56
| Enclosure | dB(A) at 7 m |
|---|---|
| Caterpillar, weather-protective industrial | 84–98 |
| Caterpillar, Level I | 78–95 |
| Caterpillar, Level III | 78–91 |
| Cummins, Level I | 70–89 |
| Cummins, Level II | 63–78 |
| Cummins, Level III | 68–70 |
A concrete example from a published specification sheet: a 100 kW Cummins set reads 86.3 dB(A) at 7 meters unhoused with infinite exhaust, and 71.9 dB(A) with a second-stage mounted muffler.7
The lesson is that enclosure specification, not engine size, drives the result. Two facilities with identical generating capacity can differ by more than 20 dB at the fence depending on what somebody bought. That’s a design decision, and it gets made early.
One technical footnote worth knowing, because it affects every published figure: Caterpillar notes that the eight measurements must be averaged logarithmically rather than arithmetically, and that using an arithmetic mean makes a set appear quieter on paper than it is in practice — about 2.5 dB in their worked example.4
Working the arithmetic for a facility this size
What follows is an illustration, not a prediction. I don’t have this project’s generator specifications, because they haven’t been published. This is how you’d evaluate them if you did.
The project has committed to backup capacity at roughly 10% of load. At a facility in the 100 to 110 megawatt range, that is on the order of 11 MW of on-site generation — roughly four to eight generator sets, depending on whether the units are in the 1.5 MW or 3 MW class.
Multiple identical sources add logarithmically: five running together is about 7 dB above one, and ten is about 10 dB above one. Applying the 6 dB-per-doubling rule from a 7 meter reference, and assuming six sets at 85 dB(A) each running simultaneously during a load test:
| Distance from the generator yard | Approximate combined level |
|---|---|
| 7 m (23 ft) | ~93 dB(A) |
| 56 m (185 ft) | ~75 dB(A) |
| 224 m (735 ft) | ~63 dB(A) |
| 450 m (¼ mile) | ~57 dB(A) |
| 900 m (½ mile) | ~51 dB(A) |
Treat those figures as a rough upper bound. They assume free-field propagation and ignore terrain, tree lines, berms, barriers, atmospheric conditions and the specific layout of the site, all of which matter. A quieter enclosure specification would move every row down; a louder one would move every row up.
For context, Cummins notes that maximum permitted overall noise levels across North American jurisdictions range from 45 to 72 dB(A), depending on location and zoning.3
The point of this exercise isn’t the numbers. It’s that anyone holding the specification sheet can do this arithmetic in ten minutes. The specification sheet is the thing nobody has produced.
What continuous cooling noise actually measures
Here the evidence is genuinely weaker, and I’d rather say so than pretend otherwise. There’s no substantial peer-reviewed literature on data center noise at residential property lines. What exists is trade publications with wide ranges, a handful of independent studies, and one state legislative audit.
Published property-line figures vary from roughly 45 dB(A) at the low end to 80 dB(A) at the high end, depending on the source. That spread is itself the finding: property-line noise is a design outcome, not a property of data centers. Siting, setback, cooling technology, equipment selection and acoustic treatment determine it.
So “no louder than a household air conditioner” isn’t implausible. It sits at the optimistic end of a wide range, and it’s achievable with enough setback and the right equipment. It’s a commitment about what gets bought and where it gets placed — which is exactly why it should be written down and measurable instead of offered as reassurance.
What has gone wrong in other communities
Two findings from Virginia, which has more data centers than anywhere else in the country, are worth knowing here.
The 2024 Joint Legislative Audit and Review Commission report found that almost a third of Virginia’s data centers sit within 200 feet of residentially zoned property — possible because zoning ordinances classify data centers in the same non-industrial category as office space.8
And in Prince William County, where residents have complained of levels routinely exceeding 60 decibels, HVAC systems are exempt from the county noise ordinance. An independent study by a resident with a technical background projected that noise from proposed and existing facilities would exceed the county’s 55 to 60 decibel limits at nearby schools and public safety facilities — but the cooling equipment producing it was not covered by the ordinance in the first place.9
There’s a structural mismatch here too. Noise ordinances were written for intermittent sources. A limit a noisy business might brush up against occasionally is something a data center produces continuously, every hour, forever. Engineer a facility to sit just under the limit and it’s permanently compliant and permanently audible.
The problem specific to this project
In Virginia the argument is about whether a facility meets the local ordinance. Here, we don’t even get that argument.
Ypsilanti Township’s noise ordinance doesn’t reach this facility, because the University is exempt from local regulation. So the question isn’t whether the projected level meets the township’s limit. It’s whether any enforceable limit exists at all.
Pending legislation would help, partially. House Bill 6139 would require an acoustic engineering report from a licensed professional engineer — baseline ambient measurements at the property boundary and surrounding residences, operational noise modelled to the ISO 9613-2 standard, construction noise modelling, a day-versus-night comparison, and a comparison against the local noise ordinance. But that report attaches to a local building permit application, and this project does not require one. As drafted, the bill would miss it.
The fix is one clause: require the acoustic report to be filed with the Public Service Commission as a condition of interconnection, independent of any building permit. That would reach this project.
Two other bills show how it should be done. House Bill 6251 and House Bill 6252 use a different architecture entirely: no construction may begin until the Public Service Commission issues a certificate of completeness, and the duty falls directly on the developer rather than on a local permitting office. They also define a covered facility by load — 25 megawatts of IT load or 50 megawatts of total electrical demand — and name high-performance computing explicitly.
Neither addresses noise directly. But HB 6251 would require disclosure of the cooling systems, heat rejection methods and backup power systems, and HB 6252 requires a thermal assessment covering the heat rejection equipment. Those are the same machines that make the noise. Between them, the equipment that determines this facility’s acoustic footprint would have to be described on the public record before construction — which is more than the noise bill would achieve here.
The questions that would settle this
Every one of these has a specific answer sitting in a design document or on a manufacturer’s spec sheet. None of them is classified.
- What is the baseline ambient noise level at the property boundary and at the nearest residences, measured before construction, at night as well as during the day? Without a baseline, no future measurement can be interpreted.
- What is the modelled continuous operational level at the nearest residence, at night, with cooling at full load on the hottest day of the year — not the annual average?
- How many generator sets, at what rating, in what enclosure specification, and what is the published sound pressure level at 7 meters at full load?
- How many generators run simultaneously during a load test, for how long, on what schedule, and at what time of day?
- What are the projected levels in octave bands, not just as a single A-weighted figure, so that low-frequency content can be assessed?
- What noise limit is the University willing to commit to, at what measurement location, enforceable by whom, with what consequence if it is exceeded?
The last one is the only question that finally matters. A projection is a prediction. A limit is a promise. Saline Township has started an independent noise review of the data center going up there instead of relying on the developer’s modeling, and that’s the standard we should expect here.
Until those numbers exist, “no louder than a household air conditioner” isn’t a specification. It’s a hope, and we’re being asked to take it on trust.
Sources
- University of Michigan Office of the Vice President for Research, LANL FAQ. research.umich.edu
- U.S. Environmental Protection Agency, “Understanding the Stationary Engines Rules.” epa.gov
- Cummins, Power Topic #7015, “Generator set noise solutions.” Source levels at 1 m, permitted range across North American jurisdictions, and octave-band limits. PDF
- Caterpillar, “Generator Sound Pressure Level Calculations.” Measurement convention and the logarithmic mean. cat.com
- Cummins, “Sound-attenuated and weather-protective enclosures,” F-1493. Sound levels at full load at 7 meters, steel enclosures. PDF
- Published Caterpillar and Cummins enclosure ranges as compiled by generator distributors. Trade source; manufacturer specification sheets for the specific units selected should replace this.
- Cummins 100 kW DSGAA specification sheet, sound levels section. PDF (hosted by a distributor).
- Environmental and Energy Study Institute, “Communities Are Raising Noise Pollution Concerns About Data Centers,” citing the 2024 Virginia Joint Legislative Audit and Review Commission report. eesi.org
- Bristow Beat, on Dr. John Lyver’s study of projected data center noise in Prince William County and the county’s HVAC exemption. bristowbeat.com
Last updated September 2026. The distance calculations above are illustrative and assume free-field propagation; they aren’t a prediction of levels at this site. If you work in acoustics and I’ve got something wrong, tell me — I’ll fix it and say that I did.