Data centres don’t have gas incidents at steady state. They have them during maintenance
4 min read
Article
A running data centre is a comparatively benign place. Cold air, sealed rooms, very few people, plant sitting exactly where the design intended. The gas hazards are present – hydrogen, refrigerant, carbon monoxide, SF₆, CO₂ – but they are contained, and containment is what the building is good at.
Then someone opens something.
Across high-hazard industries, more than 30% of incidents occur during maintenance, and data centres are no exception. Read the published record and the pattern is hard to miss: the hazard was known, the physics behaved exactly as the chemistry said it would, and the intervention was what turned an inventory into an exposure.
The activities that raise the risk
Battery replacement and relocation is the single highest-risk routine task in most facilities. Cells at end of life, disconnection under load, energetic modules handled by hand in a live building. It is also the activity behind most of the published incident record. The 2025 failure at a government data centre in Daejeon, South Korea occurred during relocation work on a lithium-ion UPS system: 384 batteries destroyed, 647 systems offline, recovery measured in weeks rather than hours.
Charging faults and configuration errors are quieter but no less serious. A charger left at the wrong float voltage, or a string reconnected with a cell removed, drives gassing far beyond anything the ventilation was sized for. One documented US case involved a hydrogen explosion in a UPS battery room that opened a 400 sq ft hole in the roof, after a prolonged charging period with no effective ventilation.
Cooling and refrigerant work puts hot work alongside an open system holding a gas that collects at floor level. Charging, brazing and circuit alteration are all release opportunities, and with liquid cooling now past half of all new high-performance deployments, every make and break of a coupling is a release point in volumes that are no longer trivial.
Water treatment and chemical handling is the activity least likely to be considered part of the gas safety case, and it produced one of the most serious data centre gas exposures on record. At a large facility in Maiden, North Carolina in 2015, a spill of a chemical used to clean the cooling system released chlorine; five workers were taken to hospital with dizziness and respiratory symptoms.
Generator testing and refuelling runs combustion plant deliberately, at length, and usually when the fewest people are on site to notice. Weather can push exhaust back towards intakes and openings. At a US telecoms facility, exhaust from eight ground-level diesel units during a monthly generator test was carried up to a roof air intake five storeys above and distributed across a floor by the HVAC system, hospitalising 17 employees.
Switchgear and suppression maintenance means breaking into a pressurised, heavier-than-air inventory in a room that is often below ground.
What makes all of it worse
One factor appears in incident findings more consistently than any other: ventilation isolated for the works while gas generation carries on regardless. Extract systems switched off, unpowered, or never interlocked to detection in the first place, while charging or cooling continues unchanged.
Construction alongside live plant compounds it. Temporary services, altered airflow, propped fire doors, and detection isolated to stop nuisance alarms – each individually defensible, collectively removing the protection at exactly the moment it is most needed. On an expanding site, that condition is close to permanent.
The failure modes repeat
Strip the published cases back and the same handful of causes recur.
Loss of containment through corrosion, vibration, mechanical damage or a failed joint – a cracked switchgear insulation flange at Blackhillock in Scotland released 33.92 kg of SF₆ over a matter of hours, assessed by Ofgem as equivalent to 798.53 tonnes of CO₂e.
Air paths nobody modelled, putting the hazard somewhere the design never anticipated.
Detection gaps and sensor drift: detectors mounted in the wrong plane, isolated during works, poisoned pellistors, expired cells, or calibration long overdue in a room nobody enters.
Alarms that reached nobody – a local sounder in an empty plant room, with no route to a manned position and no automatic action.
And change without reassessment. New refrigerant blends, liquid cooling, lithium-ion, added suppression. Where detection, ventilation and hazardous area assessment are not revisited alongside those changes, protection quietly falls behind the risk.
The Uptime Institute’s own finding is that failure to follow established procedures remains the leading driver of human-error outages. The same discipline gap shows up in the safety record.
Designing for the intervention, not the steady state
If the risk concentrates around activity, the protection has to be present during activity, which has practical consequences.
Fixed detection needs to be interlocked to something that acts, and the whole chain (sensor, alarm, ventilation, notification) tested at commissioning and again after any change to either system. Isolating detection during works should be a managed decision with compensating measures, not a default response to nuisance alarms.
Portable monitoring covers what fixed systems cannot: pre-entry checks on plant rooms, refrigerant and battery work, confined spaces. And temporary area monitoring covers the long window where fixed systems are being installed or are isolated, which on a live expanding site is most of the time.
Normal running is not where the harm happens. The maintenance window is.
If you would like to review how your gas detection arrangements hold up during maintenance and construction activity, contact a Crowcon specialist to start the conversation.
INDUSTRY INSIGHTS
Subscribe to get the
latest insights
in your inbox
Read about Crowcon‘s Privacy and Cookie policy here. If you change your mind, you can unsubscribe at any time