AI data centers are now facing a complaint that ordinary noise rules are not built to handle: sound people say they feel more than hear.
The next infrastructure fight around artificial intelligence may not start with a utility bill or a water permit. It may start in a quiet bedroom at 2 a.m., with a resident trying to explain why the meter says the noise is legal while their body says something is wrong.
That is the problem now gathering around large data centers, especially the AI-scale facilities filled with cooling systems, fans, generators and electrical equipment that run all day and night. Residents near some sites are not only complaining about the audible hum that has become familiar in data center towns. They are also talking about low-frequency vibration, pressure sensations and an uneasy feeling that does not always show up cleanly on conventional decibel readings.
As Heatmap recently noted, local governments are beginning to hear more complaints about infrasound, the very low-frequency sound below the range most people consciously hear. That matters because many noise ordinances were written around dBA measurements, a scale that roughly follows the sensitivity of the human ear. It is useful for plenty of everyday noise disputes. It is less satisfying when the complaint is that a facility is producing a constant low-frequency presence that feels like a rumble, a pulse or a pressure wave.
The most visible recent flashpoint is Vineland, New Jersey, where residents have complained about a persistent hum around a large AI data center project tied to DataOne and Nebius Group. Local reports have described a 300-megawatt facility built for AI computing demand, with Microsoft linked through a major compute supply agreement with Nebius. Neighbors have raised concerns about sleep disruption, daily irritation, property values, water use and electricity demand, while the Cumberland County Department of Health has said it is investigating the source of the sound after multiple complaints.
Virginia has its own version of the problem. Loudoun County, and Ashburn in particular, already carries the label of data center capital because so much of the internet runs through its campuses. Residents near large facilities have described low hums and helicopter-like noise, sometimes from more than a mile away. CloudHQ is one of the operators with major Ashburn facilities, and the broader county has become a live test of what happens when hyperscale infrastructure sits close to homes, schools and ordinary suburban life.
The core dispute is not simply whether data centers are loud. It is whether the rules measure the right thing. A facility can comply with a local dBA limit and still produce a tonal, low-frequency or nighttime sound that people find intrusive. That is not a small technical distinction. A steady hum at night can be more disturbing than a louder sound during the day because background noise drops and the brain has fewer other signals to ignore.
Infrasound complicates this further. Below about 20 hertz, sound is generally described as inaudible to humans, but sufficiently strong low-frequency energy can be felt as vibration or pressure. Regulators and acoustic consultants can measure low-frequency sound with different methods, including frequency-band analysis or dBC weighting, which gives more weight to lower frequencies than dBA does. But many local rules still rely on simpler audible-noise limits, and enforcement is often complaint-driven rather than continuous.
The health claims need careful handling. Residents report sleeplessness, headaches, anxiety, concentration problems and physical unease. Those symptoms are real to the people experiencing them, but proving causation from infrasound is hard because the symptoms are common and can overlap with stress, uncertainty and ordinary noise annoyance. That does not make the issue disappear. It makes it harder for counties to resolve with a single reading from a handheld meter.
For AI companies, the business risk is obvious. The industry has spent the past two years treating compute as the bottleneck. Get the GPUs, secure the power, build the campuses and the market will reward speed. But local tolerance is becoming another constraint. If residents believe AI infrastructure is making their homes less livable, planning boards will respond with tighter conditions, longer hearings, more monitoring requirements and more expensive mitigation demands.
A new compliance market may open
This is where the story becomes more than a neighborhood complaint. Cooling design, fan selection, equipment placement, acoustic barriers, generator testing schedules and real-time monitoring could become competitive infrastructure tools. A developer that can show low-frequency and nighttime noise control before a project is approved will have an easier conversation than one that arrives with a generic compliance report after neighbors are already angry.
There is also a startup opportunity hiding in the friction. AI infrastructure has created demand for power optimization, liquid cooling, grid software and modular construction. Low-frequency noise mitigation could join that list. Communities will want independent monitoring. Operators will want predictive acoustic modeling before they buy land. Insurers and lenders may eventually ask whether a project has measurable exposure to noise disputes, especially in states where data center growth is moving closer to residential areas.
The practical lesson for founders is that AI is not weightless anymore. Every model depends on buildings, substations, cooling systems, pipelines, backup generators and local politics. Compute buyers may never see the hum, but their supply chain now reaches into neighborhoods that will increasingly ask what they are being asked to absorb.
The next phase of AI infrastructure will not be judged only by megawatts and GPU clusters. It will be judged by whether the industry can build at scale without making nearby residents feel trapped inside the machine. Noise rules are likely to get more specific, and the companies that treat low-frequency sound as an engineering problem early will spend less time defending it as a public relations problem later.
Also read: Meta's email mishap shows why AI agents need real controls • Cerebras is testing how hot the AI IPO market can run • Anthropic says Claude learned the wrong stories about AI
