At a glance
The cooling units proposed for a Perth data centre are rated at 89 decibels each, two below the refrigeration on a suburban Woolworths assessed under the same rules.
On a quiet NSW street the whole site must stay under about 35 decibels at the nearest house, and the quieter the street the tougher that limit gets.
The recording driving Australian objections was made outside a facility in Michigan, in a town that permits 55 decibels at night.
A 72-hour double-blind trial found the inaudible low-frequency sound blamed for illness had no effect on sleep, while ordinary traffic noise did.
A A$600 million Perth data centre was withdrawn in May over a generator scenario the regulations never defined.
The hum going viral in Australia was recorded in Michigan
No one has published a recording of an operating Australian data centre. The audio that circulates with Australian objections was captured outside a building in Dowagiac, a small town in south-west Michigan, most of the way around the world from Sydney. A Detroit television crew recorded it, and it has since passed 700,000 views on YouTube alone, before counting the times it has been clipped and reposted over footage of Australian paddocks.
Dowagiac adopted a 65 decibel daytime and 55 decibel night limit in March 2026 and has fined the operator under it, though the operator disputes the city's readings and its method.
The loudest evidence available to an Australian objector was recorded fifteen thousand kilometres away, under an ordinance that permits 55 decibels at night where a quiet NSW street would allow 35. That gap between the microphone and the objection has now been documented by a national masthead. The Guardian reported on 26 July 2026 that the petition against Syncline Energy's proposed 2.4GW campus at Plumpton, north-west of Melbourne, had passed 3,600 signatures. Its author lives more than 250 kilometres from the paddock in question, first heard of the project on TikTok, and told the paper she had done some research into American data centres and concluded Australian ones would be much the same.
Syncline's managing director Phil Galloway told the Guardian his company would spend at least a year talking to neighbours before lodging anything, and that locals would need to be shown there would be no noise, heat island, traffic or ecology problems. He also made the case for the scale: a single precinct on a strong point of the 500kV grid, with buffers and green space, rather than facilities scattered through the suburbs.
Other objections come from people who live beside the sites. In Lane Cove West, residents formed a planning group and hired their own acoustic consultancy, Acoustic Dynamics, to go through the noise report lodged with a proposal in their business park. It disputed where the developer's noise loggers had been placed and how a neighbouring building had been classified, both of which change the limit that applies. Neither point has been settled. The proposal drew 374 submissions, nine of them supportive, and is still awaiting assessment. In Melbourne's west, a resident living about 300 metres from an operating facility told Star Weekly in April that the constant hum was a problem.
Machine for machine, a supermarket’s fridges are louder
Marshall Day’s Hazelmere assessment, lodged on 28 November 2025, rates each of the 64 cooling units at 89 decibels of sound power, a measure of the machine itself and not of what a neighbour hears. The acoustic report for a Woolworths at Eglinton, on the other side of Perth and assessed under the same regulations fourteen months earlier, rates its refrigeration condensers at 91. The heat pumps specified for Temora Hospital in regional NSW also rate 91. An ordinary commercial chiller of the sort bolted to office towers and shopping centres rates 101 on Trane’s own published data, and it is the smaller machine.
Data centres buy the quiet version and then insulate it. Marshall Day specified another 4 to 10 decibels of sound reduction across the units at Hazelmere and designed the site to clear its limits by at least five decibels, in case a regulator later decided the hum counted as tonal.
Even the generators are unremarkable company. They rate 111 decibels. A warehouse assessment at Huntingwood in western Sydney puts the air brakes on a truck at 118, and a heavy truck pulling away at exactly 111. Those are the everyday numbers of a suburban logistics estate.
A supermarket needs two condensers to keep its fridges cold. A data centre needs sixty-four because its servers make heat every second they are running, and that heat has to be pushed outside at four in the morning as much as at midday. Sound adds up, so sixty-four units at 89 decibels come to roughly 107 decibels of sound power across the site, against about 94 for the supermarket’s two.
That site total is not the number the rules use. The NSW Noise Policy for Industry and its state equivalents set the limit at the listener’s house whatever sits on the other side of the fence, so distance, barriers and attenuation do most of the work between the plant and the neighbour. At Hazelmere that came out at 31 to 37 decibels for the building’s day-to-day services. The same policy applies the same tonal penalty to a data centre, a cold store, a battery site or a warehouse, and under it the Temora hospital’s heat pumps were predicted to breach their night limit by one to three decibels, and a four-container battery installation at Glen Innes by three to four. The 64-unit data centre was predicted to comply.
Australia's noise rules are stricter than America's
The limit an Australian data centre answers to is not a fixed number. NSW sets it at whatever the neighbourhood already sounds like, plus five decibels, measured at the nearest house. On a quiet street that lands at 35 decibels at night.
Jurisdiction | Night limit | How it is set |
|---|---|---|
NSW | About 35 dB(A) on a quiet street | Whatever the neighbourhood already sounds like, plus 5 |
Western Australia | About 35 dB(A) | Fixed level, adjusted for surrounding land use |
Victoria | Up to 55 dB(A) | Derived from zoning, measured against background |
Dowagiac, Michigan | 55 dB(A) | Flat municipal limit |
Prince William County, Virginia | 68 dB(A) | Flat limit in industrial areas |
Source: NSW EPA Noise Policy for Industry 2017, WA Environmental Protection (Noise) Regulations 1997, EPA Victoria Publication 1826.5, Prince William County and City of Dowagiac ordinances.
The rows are not measured identically, and the Virginia figure applies at an industrial boundary rather than at a house, so this compares how the regimes are built. Even allowing for that, an Australian limit tightens as a street gets quieter and an American ordinance does not. Prince William County's supervisors settled on 73 decibels by day and 68 at night on 30 October 2025, overruling their own consultants, who had asked for 67 and 62.
The WHO limit everyone quotes was written for traffic
Objections reliably reach for the World Health Organization's 45 decibel night figure. It is the road traffic value.
The 2018 guidelines cover roads, railways, aircraft, wind turbines and leisure noise. The WHO's own guideline development group recorded that it agreed not to write recommendations for occupational and industrial noise, and ventilation systems appear on the excluded list by name. There is no WHO figure for infrasound or low-frequency noise anywhere in the document.
The infrasound health claim failed a double-blind trial
The other staple of the campaign is that a sound too low to hear is making people ill. It has been tested directly, and it failed.
A 2023 trial in Environmental Health Perspectives put 37 noise-sensitive adults through three separate 72-hour stays in a laboratory, blinded, giving them simulated inaudible infrasound, recorded traffic noise, and silence dressed up as the real thing. Traffic noise cost them 6.07 minutes of sleep a night. The infrasound did nothing measurable, to their sleep or to any of the psychological, neurological and physiological measures taken. Australia's NHMRC concluded in 2015 that there is no consistent evidence that wind farms cause adverse health effects in humans. A 2013 study in Health Psychology found that people shown alarming footage beforehand reported the same symptoms during genuine exposure and during a sham that delivered nothing at all. The Australian energy writer Ketan Joshi applied that literature to data centres in May.
The complaint that does hold up is the ordinary one. A 2016 review in Science of the Total Environment found 10.5 per cent of people living near audible low-frequency sources reported high annoyance, along with disturbed sleep and headaches. Losing sleep to a hum is a real harm whether or not a meter agrees, and it is the thing an assessment exists to prevent.
One noise test ended a A$600 million Perth proposal
Stirling Crescent runs through an industrial estate at Hazelmere, fifteen kilometres east of the Perth CBD, past freight yards and the back of the airport. On two lots there, GreenSquareDC proposed a A$600 million, 120MW data centre. The City of Swan received 1,897 submissions on it. All but sixty-one objected.
On 14 May 2026, five days before Western Australia's outer metropolitan development assessment panel was due to decide, GreenSquareDC withdrew the application. City of Swan officers had recommended refusal, and the ground was noise.
The dispute was not about how loud the building would be on an ordinary Tuesday. It was about a scenario. Western Australia requires a proponent to model its backup generators running at 100 per cent load under full emergency conditions, and the Environmental Protection (Noise) Regulations 1997 say nothing at all about how often that is assumed to happen. A data centre's generators exist so that a hospital's records or a bank's ledger survive a grid failure. They run for minutes a year.
Chief executive Charles Penny set out the reasoning in the withdrawal letter. GreenSquareDC had been required to assess generator operations under a scenario involving 100 per cent load during full emergency conditions, with the regulations silent on the frequency of noise emission events. In his words, that creates an unrealistic and impractical benchmark for the operation of modern data centres, and it was unfortunate that the assessment had focused on a highly conservative emergency case instead of the technical merits of the proposal.
Marshall Day had the building comfortably inside its limits, and the emergency case landed under its criterion too.
What to watch
The assessment scenario. Hazelmere died on a question no Australian instrument answers: how often a data centre should be assumed to be running every generator at full load. Until a regulator settles it, every large application in the country carries the same unpriced risk.
The NSW inquiry. The Legislative Council held hearings on 1, 8, 22 and 29 May and 2 July 2026 and has not reported. Submissions already argue that noise is assessed one project at a time rather than across a precinct, which is the live question in clusters such as Lane Cove West and along the western Melbourne corridor.
Whether anyone reaches for a meter. Until an Australian operator publishes third-party measurements from a running site, the Michigan clip stays the loudest evidence in the room, and the industry has nothing a phone can play back. The state-by-state policy comparison shows how little of this is settled, against a pipeline the AEMO connection queue puts at 5.4GW and an industrial land market that keeps putting these buildings near houses.