行业分析梳理同址数据中心内10类不同气体与化学品隐藏风险

A year ago this month, during maintenance at South Korea’s National Information Resource Services (NIRS) in Daejeon, a UPS battery failure triggered a fire. The fire destroyed 384 lithium-ion batteries and took 647 government systems offline. Recovery was projected to take four weeks.

Nothing about a facility like that would have told you it was dangerous: filtered air, locked doors, no flare stack – nothing that looks remotely like a process plant. Yet the very equipment that makes a data center dependable also makes it hazardous. UPS strings, chillers, standby generators, switchgear, suppression, and the water treatment skid are all essential; together, they introduce something close to 10 gas and chemical hazards hidden behind doors that mostly stay shut.

Batteries, Then Everything Else

Nine times out of 10, when an operator asks me about “gas,” they are asking about the batteries. Lead-acid on float can evolve hydrogen, and flammable concentrations begin at just 4% by volume. Hydrogen is invisible and odorless, and it migrates through gaps you did not know existed. Lithium-ion removes the hydrogen risk, but it introduces another: thermal runaway can vent flammable and toxic gases within minutes, often adjacent to live DC equipment.

Refrigerant is typically the largest chemical volume on site, sitting behind more joints than anyone has counted. Because most refrigerants are heavier than air, leaks settle into plant rooms and pits, and when they do, they can take your cooling with them.

Generator exhaust goes where the wind and the building let it. With short stacks, nearby intakes, and still air, exhaust can re-enter “fresh air” systems. Every test or run of a standby generator puts carbon monoxide and nitrogen dioxide somewhere.

Sulfur hexafluoride (SF6) is an asphyxiant; under arcing, it decomposes into hazardous byproducts that accumulate in low, poorly ventilated areas. Clean-agent and inert-gas suppression systems require concentrations for occupied spaces, which limits tactics when people may be present. In Daejeon, responders used gas-based suppressants such as carbon dioxide rather than water to protect critical state infrastructure. Even so, the fire’s intensity meant it took firefighters 22 hours to extinguish it completely.

Water treatment adds its own risks. Chlorine is a familiar one, and at an Apple data center in Maiden, North Carolina, in 2015, a spilled cooling-system cleaning chemical put five people in the hospital. Underneath many of these scenarios is the same mechanism: oxygen depletion.

Why There’s More Than There Used to Be

Data center capacity is on course to go from 103 GW to 200 GW this decade, according to JLL’s 2026 Data Center Outlook report. The gases have not changed. There is simply more of everything now, and it arrives faster than it used to. Lower-GWP blends, some of which are mildly flammable, are being retrofitted into plants designed for something else. Liquid cooling, unusual five years ago, is now assumed, which means a great many more joints right beside servers.

The building itself can work against you. Data centers are sealed, secure, and pressure-managed for the equipment’s benefit, and this containment slows leaks rather than letting them dissipate. In a room nobody visits, a leak can sit for a week. The person who opens that door may be the first to encounter the hazard.

The Uptime Argument

Uptime Institute lists power and cooling as the leading causes of significant outages, with cooling alone responsible for 19% of outages; more than half of serious incidents cost over $100,000. Daejeon was both a safety incident and an availability incident. No one separates the two afterward because, in practice, they are one and the same.

What I’d Check First

Nobody has a single list of what is in the building. The battery data lives with the UPS vendor. Refrigerant quantities are stamped on the chiller. Suppression information is in a fire file somewhere. Dosing chemicals are handled by a contractor who turns up once a month. Consolidating that information is a dull afternoon’s work, and it can go out of date the moment a contract changes.

Mounting height is the first thing I would look at. Too often, detectors end up at head height because installers can reach that easily. Hydrogen does not stop at head height.

These incidents above occurred during planned work, not because of equipment “breakdowns.” The Daejeon event happened during a job to make the site safer. The Maiden case was scheduled maintenance. It pays to know who is on the schedule next month and what substances and procedures they bring.

And buy detectors you can actually access for routine maintenance. If a sensor requires a permit, an escort, and a scheduled access time, calibration will slip. Eight years later, it can read several percent off, with the paperwork still in order. Notably, the Daejeon batteries reportedly passed an inspection in June of that year.

Finally, treat rooms for what they are, not for the roof they share. A switch room and a chiller hall pose different questions, require different detection strategies, and entail different failure consequences. They just happen to be in the same building.

About the Authors

Crowcon Detection Instruments

Yasin Yehya is a Product Manager at Crowcon Detection Instruments with over 10 years of experience in gas detection, hazardous-area safety, and industrial flame and gas systems. His background spans product management, technical sales, and engineered safety solutions, giving him a practical perspective on both the technology and the challenges customers face in the field. He works across global product strategy, certification, portfolio development, and market applications, with a particular interest in functional safety, hazardous-area certification, and emerging industrial safety technologies. He is passionate about turning complex technical challenges and customer needs into practical safety solutions—because safety-critical technology doesn’t have to be complicated or boring.