Reading · Sonoran Think Tank

There is a conversation happening in communities across the country right now, and it keeps getting stuck in the same place. Residents near operating hyperscale data centers report a persistent hum, ear pressure, and a vibration they feel in their chest at night. They push back at public meetings. Officials and developers respond by citing noise studies indicating the facility meets local decibel limits. Everyone walks away frustrated, and nothing gets resolved.

The reason this keeps happening is straightforward: the measurement being used to satisfy the ordinance is not measuring the sound people are actually complaining about. This is a measurement problem, and once you understand what the numbers actually mean, the policy gap becomes very hard to ignore.

Think about light for a moment. When sunlight passes through a prism, you can see that it contains every color at once: red, orange, yellow, green, blue, and violet. What looks like one beam of white light is actually a wide spectrum of wavelengths occurring simultaneously.

Sound works the same way. Any noise you hear in the real world is not a single tone. It is a mixture of many different frequencies happening simultaneously. Some of those frequencies are high, like a whistle or a bird call. Some are low, like a bass drum or the rumble of a passing truck. The human ear can detect frequencies roughly between 20 and 20,000 cycles per second, which scientists call Hertz (Hz). Below 20 Hz is called infrasound. The range of 20 to 200 Hz is called low-frequency noise.

Data centers are busy across this entire spectrum, but they are especially active in the low end. The large cooling systems, chillers, cooling towers, backup generators, transformers, and uninterruptible power supplies that keep these facilities running are all sources of significant low-frequency energy. These are not quiet machines. They run continuously, every hour of every day, and they push substantial amounts of energy into the ground, through walls, and into the air at frequencies that are not well captured by the most common noise measurement standard in use today.

When a noise consultant visits a site and measures sound levels in decibels, the result is almost always reported as dB(A) or dBA. The A stands for A-weighting. This is a mathematical filter that was originally developed in the 1930s and refined through the mid-twentieth century to approximate how the human ear responds to moderate sound levels in the range most relevant to speech intelligibility and to general annoyance from traffic and machinery at the time.

The filter works by reducing the weight assigned to low and very high frequencies before calculating the final result. This made good sense for the problems engineers were trying to solve when it was designed. If you want to know whether a factory floor is damaging workers’ hearing at normal working volumes, or whether a highway is too loud for a residential neighborhood in terms of traffic noise, A-weighting gives you a reasonable approximation.

Here is the critical detail: A-weighting applies a significant reduction to low frequencies. At 63 Hz, a common frequency for analyzing low-frequency noise, the A-weighting filter reduces the measured level by approximately 26 decibels relative to the actual level. At 31.5 Hz, that reduction is approximately 39 decibels. At 16 Hz, it is approximately 56 decibels. Because the decibel scale is logarithmic, a 10-decibel difference represents roughly a doubling or halving of perceived loudness relative to the actual sound level. A 39-decibel reduction represents about a 16-fold reduction in how the sound would be heard by a human and is an enormous mathematical discount.

The World Health Organization and a substantial body of acoustic research have clearly documented this limitation. A-weighting was designed for a different problem. It does not reliably represent human exposure to low-frequency noise or infrasound. (Berglund, Hassmén, and Job, 1996; World Health Organization Environmental Noise Guidelines for the European Region, 2018)

Large data center campuses are significant sources of industrial noise across multiple frequency ranges, but their acoustic signature is particularly concentrated in the low-frequency bands that A-weighting discounts most heavily.

The primary sources include:

Cooling infrastructure. Chillers and cooling towers are among the largest contributors. Cooling towers use large fans, often operating at low rotational speeds, that generate substantial tonal energy in the low-frequency range. Chillers contain compressors that produce broadband noise with prominent low-frequency components. These systems operate continuously and are among the loudest elements of a typical campus.

Generators and transformers. Backup generators, which at large facilities may number in the dozens, produce significant low-frequency noise when operating. More relevant on a daily basis, transformers in the switching yards hum at twice the frequency of the electrical supply they are connected to, which in the United States means a fundamental tone of 120 Hz, with harmonics extending well into the low-frequency range. Large transformers at utility scale can produce measurable vibration and sound well beyond their immediate surroundings.

Mechanical resonance and structural transmission. Low-frequency sound waves are long. A 40 Hz tone has a wavelength of approximately 28 feet. These long waves pass through walls, floors, and the ground much more easily than high-frequency sound does. Common building insulation that blocks speech-range frequencies does relatively little to attenuate low-frequency energy. This means that even when a data center appears quiet on a meter standing outside the fence line, significant low-frequency energy may be entering nearby homes through the structure itself.

Researchers who have specifically studied low-frequency noise from industrial sources, including Bolin, Nilsson, and Leventhall, have documented that A-weighted measurements routinely and substantially underestimate the total acoustic exposure experienced by nearby residents when the source has significant low-frequency content. (Bolin, Nilsson, and Leventhall, 2004; Leventhall, 2004)

The health effects of low-frequency noise and infrasound are an active area of research, and the science has grown considerably more specific over the past two decades.

The most consistently documented effects in the peer-reviewed literature are sleep disruption, annoyance, and difficulty concentrating. Chronic sleep disruption carries well-documented downstream consequences for cardiovascular health, immune function, and cognitive performance. The annoyance response to low-frequency noise is also disproportionately strong relative to what A-weighted measurements would predict. People often find low-frequency noise more disturbing than higher-frequency noise at the same A-weighted decibel level, a phenomenon researchers have documented across multiple studies. (Persson Waye, 2011; Bolin, Nilsson, and Leventhall, 2004)

Some individuals appear to be more sensitive to low-frequency noise than others. The mechanisms behind this variability are not fully understood, but the phenomenon itself is well supported in the literature, meaning that community-level average exposure assessments can miss significant harm concentrated among a subset of residents.

Infrasound, the range below 20 Hz, presents additional considerations. While research in this area evolves, the assumption that infrasound is harmless because it lies below the threshold of normal hearing has been challenged by multiple lines of evidence. Infrasound can be perceived, particularly as a pressure sensation rather than as sound, at levels well below those historically assumed. The World Health Organization has noted that exposure to infrasound warrants further study and that existing regulatory frameworks are inadequate for its proper evaluation.

Structural effects are also relevant. Low-frequency energy can cause resonance in building components, including windows, walls, and floors. Even when this resonance does not cause physical damage, it can generate secondary rattling, buzzing, or vibration that compounds the annoyance experienced by residents and makes it difficult to identify the original source through standard measurement.

Most noise ordinances across the United States set limits in dB(A) or dBA. A developer submits a noise study showing the facility will comply with those limits. The study uses A-weighted measurements. The A-weighting filter mathematically discounts the low-frequency energy, which is the dominant feature of the facility’s acoustic signature. The facility passes the test.

This is not a hypothetical. It is the standard workflow seen across the United States.

The consequence is that residents who experience real, documentable, low-frequency noise exposure from a compliant facility have no recourse under the existing regulatory structure, because the measurement standard used to define compliance was not designed to detect what they are experiencing.

When concerned residents hire their own acoustic consultants and pay for additional noise testing using the same A-weighted methodology, they reproduce the same gap. A-weighted tests will not find low-frequency noise problems that A-weighted ordinance limits were not designed to address. This is money spent that cannot answer the question being asked.

The appropriate tools exist. Full octave band analysis shows sound levels across the entire frequency spectrum simultaneously. The C-weighting scale, reported as dB(C) or dBC, applies far less attenuation to low frequencies than A-weighting and is better suited for evaluating sources with significant low-frequency content. The G-weighting scale, reported as dB(G) or dBG, is specifically designed to evaluate infrasound. Several countries, including Denmark and Sweden, have developed specific regulatory frameworks for low-frequency noise that require octave-band analysis and set separate limits for low-frequency bands rather than a single A-weighted number.

The United States Environmental Protection Agency produced guidance on noise and its effects that recognized the limitations of A-weighted measurements for certain source types. (EPA Office of Noise Abatement and Control, 1974; EPA, 1978) That framework was largely defunded in 1982, and noise regulation was effectively handed over to state and local governments, most of which adopted or retained A-weighted standards without the technical expertise to assess their adequacy for emerging industrial source types.

If you are a resident living near a data center campus and you are experiencing symptoms consistent with low-frequency noise exposure, an A-weighted noise test will not validate or invalidate your experience. It is measuring a different thing. Requesting, paying for, or relying on A-weighted measurements alone is unlikely to produce useful information about your specific situation.

If you are a local official considering a noise ordinance or evaluating a conditional use permit, A-weighted decibel limits alone are not an adequate standard for regulating hyperscale data center facilities. Requiring full octave-band analysis, specifying separate limits for low-frequency bands, and requiring measurements at the nearest receptor locations rather than at a facility's property line would bring your ordinance into alignment with the actual acoustic character of these facilities.

If you are a developer or operator, the long-term regulatory risk of facilities that technically comply with inadequate standards while generating documented community complaints is real and growing. As communities become more technically informed and as legal frameworks catch up to the science, facilities permitted under A-weighted-only standards will face increasing scrutiny.

The measurement problem is solvable. The science is not new. What is needed is a willingness to apply the right tools to the actual problem rather than the convenient tools to the wrong one.

Measuring the Wrong Sound

Comments