7 The machine hum
The distributed transition has been audited; the concentrated one is next. That one reached a million separate doorsteps, an alerting system bolted to a silent car or a heat pump breathing on a party wall at a time. The digital transition arrives instead as a single large object that never switches off. A data centre does not roar. It hums: a steady tone compounded from cooling fans, chillers, transformers and the periodic exercising of standby generators, as present at three in the morning as at three in the afternoon, on every day of the year. People who live beside these buildings tend to describe a sound they feel as much as hear: a pressure in the room, a rumble that seems to rise through the floor instead of arriving through the window. They report that it is worst at night, once the traffic that masked it by day has thinned away. The operator’s instrument, set to the standard weighting and averaged across the day, frequently returns a figure comfortably inside the local limit. That gap, between an intrusion residents insist is severe and a number the regulation calls acceptable, recurs in every case that follows. It is not the signature of a single badly sited plant; it follows from what the standard metric was built to record and, more consequentially, from the part of the spectrum it was built to leave out. That is the same low band in which the 58 Hz tone of Chapter 2 disappeared from the reading without leaving the room.
7.1 The 24/7 drone
The acoustic problem begins with a difference in kind, not only degree. Road traffic, the dominant burden of Chapter 5, is intermittent and broadband: it rises and falls with the passing vehicle and spreads its energy across the audible range. The plant that cools a large data hall does neither. It runs continuously, and a substantial share of its output is concentrated into a few narrow frequency bands — blade-passing tones from banks of fans, harmonics from transformers, the drone of chillers. The ear resolves them as a definite pitch rather than a wash. Steady, tonal and low: taken together, the three properties describe a sound that human listeners find disproportionately hard to ignore. They also describe one that the built environment does little to hold back, since low frequencies pass through the walls and windows that would stop a higher-pitched noise (Leventhall, 2004). Constancy compounds the effect. A sound that never pauses offers the ear no respite and no contrast, and it removes the intervals of quiet against which a household organises rest. The absence of variation, which the averaged metric treats as benign, is what residents report as intolerable.
At the property line, the numbers look modest. A 2024 review of Virginia’s data-centre corridor, where roughly four-fifths of the state’s capacity is concentrated in three northern counties, placed the sound reaching nearby homes in a band of about 40 to 59 dB(A) (Joint Legislative Audit and Review Commission, 2024). On paper that range reads like a quiet suburban street. The same review likened what reaches those homes to the low, continuous hum of air-conditioning plant (Joint Legislative Audit and Review Commission, 2024). The comparison registers both the character of the sound and its constancy, and marks it off from the intermittent noise the surrounding zoning had been framed to address. The lived account is harsher. In Prince William County a decades-old ordinance caps residential noise at 60 dB(A) by day and 55 by night, but exempts heating and cooling equipment. Residents beside a cluster of data centres recorded their own measurements well above those limits, averaging around 62 dB(A) with night-time peaks near 69, while the facilities remained, on the operator’s reading, within the rules (Prince William Times, 2024). The discrepancy is not mysterious once the character of the sound is admitted. A steady 62 dB(A) tone at three in the morning is a different creature from a 62 dB(A) figure produced by daytime traffic, and an ordinance written to catch parties and modified exhausts cannot tell the two apart (Joint Legislative Audit and Review Commission, 2024).
The extreme cases make the mechanism legible. In Chandler, Arizona, close to a decade of complaints against the air-conditioning tone of a large campus produced first noise-reducing retrofits, then, in 2022, a zoning amendment requiring sound studies before any new facility, and finally, in 2025, a council vote to refuse a proposed data centre outright (Environmental and Energy Study Institute, 2025). The reversal was driven substantially by noise. Cryptocurrency mining, which packs tens of thousands of processors and their cooling fans into open-sided halls, supplies the loudest instances. Beside a 300-megawatt bitcoin operation in Granbury, Texas, residents recorded readings that ran into three figures, one household measuring 103 dB(A). The state threshold treats 85 dB(A) as the point of unreasonableness and caps the penalty at a five-hundred-dollar fine. The sound has been likened to a jet engine that never lifts off, and it carries for miles, waking sleepers who live well outside any plausible buffer zone (Chow, 2024). In the Bono community near Greenbrier, Arkansas, the fans of some seventeen thousand machines a few hundred feet from the nearest house pushed readings to 82 dB(A), around the clock. The state had just passed a law forbidding local governments from regulating such operations at all, a preemption its legislature partially reversed a year later, once the complaints had become impossible to set aside (Lenora, 2024). The bitcoin cases sit at the extreme, but they are the same object magnified: banks of fans cooling banks of processors, and no off switch.
These are American cases because the American build-out came first and largest, but the concentrated transition has now reached Andalusia. In June 2026 the first stone was laid, in the Escúzar technology park south of Granada, of SP01: a campus built expressly for artificial-intelligence training and high-performance computing. Around 100 MW of gross grid capacity has been granted, roughly 70 MW of it destined for the servers themselves, ramping from a first phase of some 10 MW through an intermediate stage of around 25 MW toward full build-out by the end of the decade (DataCenterDynamics, 2026). Its acoustic signature is, for now, a matter of design and not yet of complaint. Two of its stated choices bear directly on it: the campus is to draw on renewable power, and to cool its densest racks with direct liquid rather than with air. Liquid cooling matters here for a specific and non-rhetorical reason. Moving heat in a closed loop of fluid displaces a large part of the fan bank whose blade-passing tones are the usual source of the drone. None of this promises a silent campus, and none of it makes the campus unwelcome. What it establishes is that the tonal, low-frequency character of the hum is set at the drawing board, in decisions about cooling and siting taken before the first server is powered. That character is therefore still open, in a way the metric applied afterwards is not.
7.2 The infrasound controversy
Below the audible drone lies a more contested claim. Infrasound is pressure fluctuation below about 20 Hz, at or beneath the threshold of conscious hearing; the claim is that these installations emit it, and that the inaudible component is itself a cause of ill health. The claim recurs wherever large rotating or pumping machinery sits near homes, and the evidence around it has a characteristic shape, one already met in Chapter 6’s account of wind farms. It divides, roughly, into three postures.
The first minimises. A body of work, some of it produced with support from the industries whose installations are under scrutiny, concludes that measured infrasound levels near such sites fall below the threshold of perception and can therefore be dismissed as a source of symptoms (Berger et al., 2015). The reasoning is not empty. Perception thresholds at very low frequencies are real and steep. But the provenance warrants the caution due to any assessment funded by an interested party, and the conclusion tends to treat below the average threshold of perception as equivalent to without effect, which is a stronger claim than the data support.
The second finds an effect and hedges it. Controlled exposure of volunteers to low-frequency and infrasonic tones has been reported to shift markers of autonomic function in directions consistent with a mild stress response (Chiu et al., 2021). The markers include heart-rate variability and the balance of sympathetic and parasympathetic activity. Such studies are suggestive, but they are typically small, brief and conducted under laboratory conditions that do not reproduce the years-long, sleep-interrupting exposure residents describe. A measurable change in a physiological marker over an hour is not the same as a demonstrated pathway to disease over a decade.
The third is neutral, technical and the most useful of the three. A detailed measurement study of large wind turbines established that as rotating machines grow, the balance of their emitted spectrum shifts downward. The relative share of low-frequency energy rises, and even after A-weighting a substantial part of what remains sits at the bottom of the range: for several machines the loudest third-octave band fell at or below 250 Hz (Møller & Pedersen, 2011). The finding concerned turbines, not data halls, but the physics is general and the parallel is exact: the larger and more powerful the source, the more its acoustic weight migrates into the band that ordinary assessment handles worst.
The defensible position, on the present evidence, is that the complaints are real and the specific attribution to infrasound is not settled. People living beside these facilities are not imagining the sound. The readings confirm a persistent, tonal, low-frequency presence, and the reported symptoms of broken sleep, headache and difficulty concentrating are well-attested consequences of chronic low-frequency noise, whether or not any inaudible sub-20 Hz component contributes (Leventhall, 2004). What remains genuinely open is whether infrasound as such, rather than the audible low-frequency noise that accompanies it, does independent physiological harm. Holding those two questions apart is what keeps the argument honest. The evidence can carry the claim about audible low-frequency noise; it cannot yet carry the claim about infrasound, and the case against the metric depends on the first of these, not the second.
7.3 Why the metre cannot hear it
The reason the operator’s number and the resident’s experience diverge is built into the number itself, and it can be stated without recourse to any disputed physiology. Two operations are applied to almost every regulatory measurement of environmental noise, and each, independently, discards the part of the signal in which the data-centre hum lives.
The first is A-weighting. The standard sound-level reading, dB(A), is not a neutral record of acoustic energy; it is energy passed through a filter that mimics the reduced sensitivity of the human ear at low frequencies, and that filter rolls off steeply below a few hundred hertz (International Electrotechnical Commission, 2013). Transformer hum and the lower blade-passing tones sit at 50 or 60 Hz. A pure tone there is cut by the weighting curve by tens of decibels before it is ever compared with a limit. The effect was made visible in Chapter 2, where a 58 Hz tone dropped out of the A-weighted figure while remaining plainly present in the spectrum. The second operation is temporal averaging. The indicators that anchor most noise law, \(L_\text{den}\) and its night-time counterpart, integrate sound energy across long periods. A perfectly steady source, which is the defining property of a data centre, is thereby smoothed into a single background value instead of being flagged as the relentless, unvarying presence that makes it unbearable. Weighted to discount the frequencies and averaged to discount the constancy, the metre is, in a precise sense, deaf to the two features that define the complaint.
This is the point at which care matters most. The case here is not that infrasound has been proven to sicken people; that claim remains unsettled. The case is narrower and far more secure: the instrument used to certify these facilities as compliant is systematically blind to the band of the spectrum that generates the complaints. Whether infrasound sickens people is a contested question in physiology. That the filter and the averaging window discard the relevant band is a demonstrable property of both, and it holds regardless of how the physiology is eventually resolved. A steady 58 Hz tone can be, at once, the dominant feature of a resident’s night and nearly absent from the dB(A) reading that governs whether anything can be done about it. Even after A-weighting, spectra from large low-frequency sources keep their peak energy at the bottom of the range (Møller & Pedersen, 2011). The weighting does not merely soften the low end; it can misplace where the sound’s energy actually lies.
The corrective tools exist, which is itself telling. Assessment procedures written specifically for low-frequency complaints measure in narrow low-frequency bands and test them against dedicated criteria, not against a single A-weighted figure (Moorhouse et al., 2005). Procedures of that kind were developed for national environmental agencies because the ordinary metric was failing. And where a clearly audible tone is present, some schemes add a penalty of up to six decibels to the assessed level, on the reasoning that a pure tone annoys more than its bare energy would predict. The same logical correction appeared in Chapter 6, applied to the tonal and amplitude-modulated character of wind-turbine noise. The Spanish noise regulation sets exactly that six-decibel ceiling for an emergent tone, and identifies the tone on the unweighted third-octave spectrum rather than on the A-weighted figure that governs compliance (BOE, 2007).
The existence of these corrections is an admission, encoded in the standards themselves, that the default metric under-reads tonal low-frequency noise. Yet it is the default metric, not the corrective procedure, that is written into most compliance regimes. The operator is judged by the instrument that cannot hear the sound, and the burden of demonstrating the sound falls on the residents, who must produce their own measurements to capture what the official method was built to pass over. In Prince William, the officials revising the county’s code said as much in plain terms: the low-frequency noise the data centres emit is not fully captured in A-weighted decibels (Prince William Times, 2024). When a measurement regime reliably yields compliant numbers for a sound that reliably yields complaints, the question of whom that arrangement serves answers itself.
7.4 Distance and the tonal hum
Why the hum reaches homes that look, on a map, comfortably far from the plant is a question of how sound of different pitches survives the journey. Geometric divergence is indifferent to frequency — the wavefront spreads, and the same loss applies to a low tone and a high one alike. Atmospheric absorption is not indifferent. The air itself converts sound to heat, and it does so far more efficiently at high frequencies than at low. A crisp, high-pitched noise loses its edge within a few hundred metres; a deep tone loses almost nothing to the air and is checked only by spreading. Distance therefore acts as a filter, stripping away the treble and leaving the bass, so that the further a listener stands from a broadband industrial source, the more purely tonal and low its remaining sound becomes. This is why a data-centre or mining hum can be reported from a mile or more away as a rumble rather than a roar. It is often the households that moved out for quiet who are worst affected.
Whether a hum survives a given separation depends on its frequency as much as on its level at source. That is what the model below computes, for any combination of the two. A low tone and a high one decay side by side, and the A-weighted figure parts company with the true sound-pressure level as the frequency drops — the same divergence between what is present and what is counted, now expressed in metres of separation rather than in a single reading at the fence.
Source · geometric divergence with frequency-dependent atmospheric absorption computed after International Organization for Standardization (1993) and International Organization for Standardization (1996); A-weighting after International Electrotechnical Commission (2013); representative property-line levels after Joint Legislative Audit and Review Commission (2024). Accessed 21 July 2026.
7.5 The energy of the hum
The hum is, in the end, the sound of electricity being turned into computation and heat. Fans spin because processors run hot; processors run hot because the calculation is vast and unremitting; and cooling, the single largest non-computing draw in most facilities, is the immediate source of the noise. The acoustic signature is thus a fairly direct readout of the energy appetite of artificial-intelligence infrastructure, traced elsewhere (Moreno-Muñoz, 2026). A campus sized in tens or hundreds of megawatts is, by that very sizing, committed to moving a proportionate quantity of waste heat, and the ordinary way to move it is to push air across metal at scale. The design choice on display at Escúzar, liquid cooling in place of air, is significant because it breaks that chain at the fan. It does not lighten the underlying load, since the heat still leaves the building and the energy still enters it. As AI-scale campuses grow denser, the quantity of heat to be shifted grows with them, and with it the pressure on whatever mechanism moves it.
The concentrated transition is a double object, an energy story and an acoustic one told in the same breath. The sound does not stop at the property line. It passes through the walls that would stop a higher-pitched noise and settles into bedrooms and living rooms, where the same low-frequency physics meets sleep, attention and the private spaces least equipped to exclude it. These cases also share a pattern: a real intrusion, a compliant number, an ordinance built for a different kind of noise, and, in Arkansas, a law that briefly forbade the affected communities from regulating the source at all before a partial retreat. The regulatory architecture assembled to contain environmental noise leans on the very metric this source is built, whether by accident or by convenience, to slip past. The machine hum is the clearest case in these pages of a sound loud enough to drive people from their homes and quiet enough, by the measure that counts, to be officially inaudible.