How it feels for people living near it:
Digital profits. Real-world costs.
Our communities are not server farms.7
The industry's vocabulary for AI build-out — sustainable, essential, quiet — describes the brochure. It does not describe the bedroom at 2 a.m. with the windows closed and the pressure in the ears. The poster below has been circulating; we've reproduced it without its photographic backdrop and annotated the rhetorical moves.
Cloud. Sustainable. Innovation. Quiet. Efficient. Essential. Those six words are the entire pitch deck. They aren't false, exactly — they are the kind of thing that becomes true only after you've decided what doesn't count.
"The cloud" is a metaphor that works by hiding a building. Every photo of "the cloud" is a stock image of fluffy white cirrus over a wheat field. Every actual cloud is a 350,000-square-foot warehouse with a fence around it, a substation behind it, and a row of cooling towers on top of it. The metaphor is doing siting work. It is telling you not to picture the building.
"Sustainable" usually means the operator has purchased renewable energy credits matched annually to consumption, while the actual electrons feeding the racks come from whatever the grid happens to be burning that hour. In Saskatchewan and Alberta, that's natural gas. In Memphis, that has been 27 methane gas turbines running without an air permit half a mile from homes and a mile from an elementary school.
"Quiet" is the word that most reliably fails on contact with reality. It is true only in the narrow sense that a 96 dB hum at 8 Hz doesn't register on a meter calibrated for a barking dog. The meter is wrong about that bedroom. The body is right.
"Quiet, efficient, essential digital infrastructure."
A hum below the threshold of meters but inside the threshold of bodies.
"Sustainable, climate-conscious operations."
Gas turbines, REC offsets, and water draws in regions already in drought.
"Powering the future."
Powering a chatbot, on a grid your residential rate is subsidizing.
"Local economic development."
Fewer than 100 permanent jobs. Tax abatements on most of them.
A municipal noise ordinance is usually written in dBA — A-weighted decibels — which is a measurement scale that deliberately discards the low end of the spectrum because human ears don't consciously hear it. That weighting was designed for occupational hearing damage in factories, where the harm is in the audible range and the question is "will this make you go deaf." It was never designed to capture the kind of harm a data center produces.
A hyperscale facility is, mechanically, three loud things stacked on top of each other: cooling systems rejecting heat from tens of thousands of GPUs; backup generators (or, increasingly, always-on natural gas turbines); and electrical equipment — transformers, switchgear, UPS — humming at the grid's harmonic frequencies. Most of the noise lands below 200 Hz. A lot of it lands below 20 Hz, which is the threshold of conscious hearing — what's called infrasound.
Schematic. The dBA curve under-weights frequencies below ~100 Hz by 20 dB or more — meaning sound that a meter logs as "30 dBA" can be physically present at 80–96 dB SPL in the bedroom. The body responds to physical pressure, not to the weighting curve.
So you get a class of complaint that looks, to a regulator, like hypochondria, and looks, to the resident, like the building is doing something to them that the inspector refuses to see. Both descriptions are accurate. The inspector is using the wrong instrument. The resident is using their nervous system, which is the right instrument but isn't admissible at a planning hearing.
Pre-2023 data centers used air-cooled CPU racks at moderate power densities, roughly 5–10 kW per rack. AI training clusters push 40–120 kW per rack. That heat has to go somewhere, and the somewhere is bigger cooling towers, more fans, more chillers, and — when the grid can't deliver power fast enough — on-site gas turbines. The audible part of the problem and the infrasound part of the problem are both increasing, together, on the same curve as model size.
The complaints reported around AI data centers in 2026 are not new. They are the same complaints reported around wind turbines in the 2010s, around gas compressor stations in the 2000s, around highway tunnel ventilation shafts in the 1990s. Different sources, same acoustic signature, same body response.
The cluster, every time:
The point is not that infrasound is a uniquely mysterious health hazard. The point is that chronic, unavoidable, low-frequency mechanical vibration inside a residential building is a well-understood category of harm, and the only thing standing between the resident and a finding of harm is the instrument the regulator chose to bring to the property line.
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. — acoustic consultant, cited in Startup Fortune, 2026
A non-exhaustive index of communities currently in active dispute with the AI build-out. Each was, at one point, told the project would be "quiet, efficient, essential."
Residents of the Brittany Heights subdivision adjacent to a CyrusOne data center reported being woken nightly by a low-frequency hum. Households tried noise-cancelling headphones and earplugs with little effect. The complaint eventually contributed to a second proposed data center in the area being blocked — likely the first time a North American hyperscale build was stopped on noise alone.
xAI installed 35 methane gas turbines to power Colossus 1 in South Memphis without filing for Clean Air Act permits, by characterizing them as "temporary-mobile" units exempt from review. The turbines have now run for more than a year. For Colossus 2, a further 27 turbines were installed across the state line in Southaven, MS — half a mile from homes, a mile from an elementary school, in a community already failing federal smog standards.
In April 2026, the NAACP, represented by Earthjustice and the Southern Environmental Law Center, filed suit. The plant has the potential to emit 1,700 tons of NOx, 180 tons of fine particulate, and 19 tons of formaldehyde annually.
Ten residents filed suit to halt the expansion of a 350,000-square-foot data center, alleging regulators and developers bypassed the statutory public-review process. The structural complaint is now a template across multiple jurisdictions: hyperscale operators increasingly use "expansion" or "by-right" classifications to skip the planning hearings where noise objections would otherwise be raised.
Approximately 50 people live in Indus, east of Calgary. Beacon AI Centres proposed an off-grid 1,494 MW natural-gas-fired data centre campus. The Impact Assessment Agency of Canada determined in March 2026 that no federal review was required. The application sits before the Alberta Utilities Commission. Local opposition has been organized largely by a returning resident with two decades of data centre engineering experience.
Water draw is the second front: a single hyperscale facility can consume more water per year than the entire neighbouring town of Langdon. The region was under drought restrictions in 2025.
Synapse Data Centers proposed a 1,400 MW gas-fired plant adjacent to Olds, marketed as "Canada's largest" data centre. In March 2026 the Alberta Utilities Commission dismissed the application, citing "errors, insufficient or incomplete information, and internal inconsistency." Local opposition centered on noise, water draw, and the proximity of the proposed site to the town itself.
Kevin O'Leary's proposed $70 billion Wonder Valley project would run on 7.5 GW of BYOP (bring-your-own-power) natural-gas generation. The Sturgeon Lake Cree Nation has launched legal challenges citing inadequate consultation and the environmental footprint of building a generation facility of that scale in a culturally and ecologically sensitive corridor. The District of Greenview and the provincial government have publicly supported the project.
A proposed data centre near Saint John would consume the equivalent of nearly half the output of a nearby natural gas plant — for one facility. Opposition has formed around the implications for grid reliability, emissions, and the question of whether residential ratepayers will subsidize the necessary transmission upgrades.
The pattern across these files is not coincidental. It is structural. Hyperscale data centers are sited where land is cheap, where regulators are under-resourced, and where the political cost of saying no to a "tech investment" announcement is high. The communities that get the noise are the communities the planning process was already not built to protect.
When Canadian politicians and business-council releases talk about data centres, the picture in their head is Quebec or BC — hydroelectric, cold, abundant. 60% of Canada's installed generation capacity is hydro. The pitch writes itself in those provinces.
But that is not where most of the new AI capacity is being proposed. Alberta and Saskatchewan run on natural gas. They are the provinces with the cheapest grid power for new industrial loads, the most deregulated siting regimes, and the most enthusiastic provincial governments. Three quarters of the data centre sites planned in Alberta are in regions under high or extreme water stress. Saskatchewan's grid is roughly 80% fossil-fuelled.
Meanwhile, the major AI companies announced in March 2026 that they would cover the residential power-cost impact of their US data centres in response to public pressure. That commitment was not extended to their Canadian operations. In Alberta, Bill 8 — the Utilities Statutes Amendment Act — would require data centres to pay for grid upgrades and would levy facilities using 75 MW or more. Off-grid (BYOP) facilities, which is the model most of the new projects are choosing, would be exempt from the levy.
Several of the largest proposed Canadian sites — Wonder Valley most prominently — sit on or adjacent to treaty territory, and have proceeded through provincial approval channels with consultation that affected nations describe as inadequate or absent. The Sturgeon Lake Cree Nation's legal challenge to Wonder Valley is the most advanced case, but it is not the only one. The federal Impact Assessment Agency's decision not to require review of the Beacon Indus project signals a posture: Ottawa will defer to provincial energy regulators even at the gigawatt scale.
The Canadian version of this story is, in other words, not "we have clean grids and quiet neighbourhoods, so we are immune." It is "the build is being directed at exactly the provinces and exactly the lands where the harms — emissions, water, noise, inadequate consultation — are worst."
The data centers exist. They will keep being built. The honest question is who absorbs the externalities — the noise, the water, the emissions, the strain on the grid, the strain on the people downwind of all of it — and who gets to keep calling those externalities "sustainable" while it happens.
The corporate vocabulary — cloud, sustainable, innovation, quiet, efficient, essential — is not neutral. It is a siting strategy. It pre-emptively classifies the resident's complaint as a nuisance and the operator's emissions as overhead. It says, in advance: the meter is the truth, your body is the noise.
A community that can describe the gap between the meter and the body — in its own words, on its own record — is much harder to site against. That's the work. Document the brochure. Document the bedroom. Note when they don't match. Insist that the mismatch is the story.
An editorial. No author byline. Cite, fork, redistribute. If your community is fighting a siting decision and you want the spectrum chart, the case file index, or the lexicon table in a format you can use at a planning hearing, take them — they are not under copyright.