Data Center Fire Protection
Detection, suppression and lithium-ion batteries

Fire protection in a data center has four jobs, in this order: detect a fire while it is still smoke, confirm it before acting, put it out without destroying the equipment it protects, and get people out. Lithium-ion batteries add a fifth: keep flammable vent gas below the level at which it can explode. Every number on this page links to its source: a row of the site’s fire datasets [S1]–[S3], the fire engine [S4], a named standard, or an incident dossier.

1. What burns in a data center, and where

Almost every data-center fire starts in electrical equipment. What differs between zones is the fuel next to it and what the fire costs. In the eight fires in section 7, six started in UPS, battery or power rooms, one in a rack, and one in a power compartment; in most of them the outage came from cutting the power to fight the fire, not from the flames.

Fires in energised equipment are electrical fires (NFPA Class C) until the power is off; after that they burn as ordinary combustibles (Class A). Clean-agent design concentrations are quoted for Class A [S1].

2. Detection

The goal of detection is to find smoke early enough that staff can act before any automatic release. Three detector types do this job; release logic then decides when one alarm becomes a discharge.

2.1 Aspirating smoke detection (ASD, often called VESDA)

A fan draws air through a pipe network with sampling holes over the racks, in the return-air path and in the floor and ceiling voids. A central laser chamber measures the smoke in that air. EN 54-20 rates aspirating detectors in three sensitivity classes: Class A (very high sensitivity, for very early warning), Class B (enhanced) and Class C (normal). The site’s fire engine uses a pre-alarm sensitivity of 0.005 % obscuration per metre [S4]; the design band in the commissioning checklist is 0.005–0.05 %/m [S5].

The thresholds are set per site from the measured background, not from a fixed number. High airflow in a data hall dilutes smoke, which is why air sampling is preferred over ceiling detectors alone.

2.2 Spot detectors

Point-type smoke detectors on the ceiling and in voids. NFPA 72 sets a nominal spacing of 9.1 m (30 ft), about 84 m² per detector [S3] [S4], and requires the spacing to be reduced where air changes are high, which is normal in a data hall. Each level that contains combustibles (room, floor void, ceiling void) is counted separately. Example: a 600 m² hall needs at least 600 / 84 = 7.1, so 8 detectors per level before the airflow reduction (derived).

2.3 Cross-zoned release

A gas release is expensive and, for some agents, unsafe for people still in the room, so one detector is never allowed to release it. Cross-zoned detection requires two independent detectors or zones to alarm before the panel starts the release sequence [S5]. The sequence then sounds a pre-discharge alarm, stops the air handlers, closes the dampers, and runs a time delay (typically 30 s [S5]) so occupants can leave. An abort switch holds the delay while it is pressed. A manual release station at the exit releases the agent directly, and an abort switch does not cancel a manual release.

2.4 Lithium-ion off-gas detection

A lithium-ion cell heading into thermal runaway vents hydrogen, carbon monoxide and electrolyte vapour before it burns [S2]. Smoke detectors see this late. NFPA 855 requires gas detection that alarms at 25 % of the lower flammable limit (LFL) and starts exhaust [S3]. The LFL of hydrogen is 4.0 % by volume and of carbon monoxide 12.5 %; the mixed vent gas is about 6 % [S2]. So the alarm point is 1.0 % hydrogen (10,000 ppm), or about 1.5 % of mixed vent gas (derived). Off-gassing is covered in detail in section 4.

3. Suppression

Two families protect IT rooms: gaseous agents, which fill the room and leave no residue, and water systems, which cool and are the building-level backstop. Many sites use both: a clean agent for the room and pre-action sprinklers because the building code or insurer requires sprinklers.

3.1 Clean agents (NFPA 2001, ISO 14520)

Halocarbon agents (Novec 1230, FM-200) absorb heat and interrupt the flame chemistry; inert gases (IG-541, IG-55) lower the oxygen level. The rules that matter:

• Design concentration: the agent percentage by volume the system must reach. It sits above the minimum extinguishing concentration by a safety factor [S1].
• NOAEL and LOAEL: the highest concentration with no observed adverse effect, and the lowest with one. In a normally occupied room the design concentration must be at or below the NOAEL [S1] [S5].
• Discharge time: 10 s or less for halocarbons, 60 s for the inert gases in the table [S1].
• Hold time: the concentration must be retained for at least 10 min [S3] [S4].
• Room integrity: the hold time depends on how tight the room is. A door-fan test (NFPA 2001 Annex C) measures the leakage area and predicts the hold time; the room is sealed and re-tested until it passes [S5].

Agent mass for halocarbons, from NFPA 2001 [S3]:

W = (V / s) · C / (100 − C)      s = k1 + k2 · T

W  agent mass, kg          V  protected volume, m³
C  design concentration, % T  minimum room temperature, °C
s  specific vapour volume, m³/kg   (k1, k2 per agent, table below)

Worked example, a 500 m³ room at 20 °C (derived from [S1]):

• Novec 1230 at 4.7 %: s = 0.0664 + 0.0002741 × 20 = 0.07188 m³/kg; W = (500 / 0.07188) × 4.7 / 95.3 = 343 kg.
• FM-200 at 7.0 %: s = 0.1269 + 0.0005132 × 20 = 0.13716 m³/kg; W = (500 / 0.13716) × 7.0 / 93.0 = 274 kg.

Inert gases are sized by volume: X = ln(100 / (100 − C)) m³ of agent per m³ of room [S4]. IG-541 at 37.5 % needs ln(1.6) = 0.47 m³/m³, or 235 m³ for the same room, before the temperature correction (derived). The oxygen left in the room is about 20.9 × (1 − C/100): 13.1 % at 37.5 % IG-541 and 12.5 % at 40 % IG-55 (derived). The fire calculator runs these equations, the cylinder count and the NOAEL check.

Reading the table: FM-200’s design concentration (7.0 %) is close to its NOAEL (9.0 %); Novec 1230 has more margin (4.7 % against 10.0 %). FM-200 has a GWP of 3,220 and is being phased down under HFC rules; Novec 1230 has a GWP of 1 [S1]. The comparison is worked through on FM-200 vs Novec 1230. Two further points: a halocarbon discharge first drops and then raises the room pressure, and an inert-gas discharge raises it strongly, so NFPA 2001 requires the enclosure pressure to be checked and most inert systems need a pressure relief vent.

3.2 Water-based systems (NFPA 13, NFPA 750)

NFPA 75 permits sprinklers in IT rooms, and many building codes require them regardless of a gas system [S3]. The type of sprinkler system decides how likely an accidental discharge is.

The trade-offs between wet and pre-action systems are set out on wet vs pre-action. Sprinklers respond to heat, so they act late; the Digital Realty LAX12 fire in 2023 was stopped by sprinklers, and the water damaged servers (dossier).

3.3 Emergency power off

NFPA 70 Article 645 covers IT rooms and the emergency power off (EPO): a guarded, two-stage station that de-energises the room so firefighters can work [S3]. The EPO sequence is tied to the suppression release and tested with it [S5]. Accidental EPO operation takes rooms down, which is why the guard and the second stage exist. In the X Oregon fire of 2025 the burning UPS cabinet was still energised and firefighters had to wait for it to be isolated (dossier).

4. Lithium-ion batteries

Lithium-ion UPS and battery backup units have replaced most lead-acid (VRLA) strings. They need a different plan because a cell in thermal runaway heats itself and spreads to the next cell, and no gaseous agent stops that reaction. In the SK C&C, Digital Realty Singapore and NIRS fires the batteries burned or re-ignited for hours while water cooled them (section 7). The order of protection is: monitor the cells, detect gas and ventilate, control an explosion, separate the units, and then suppress the surrounding fire.

How much gas: one 1,333 kWh NMC pack at 4.0 L/Wh can vent 1,333,000 × 4.0 = 5,332,000 L, or 5,332 m³ (derived from [S2]). At a 6 % LFL that is enough to make 5,332 / 0.06 = 88,900 m³ of flammable mixture. An LFP pack of the same size vents half as much (2,666 m³, 44,400 m³ of mixture). No battery room is that large, so the room volume does not dilute a full runaway; exhaust started at 25 % LFL, and the BMS stopping the event earlier, have to.

The fire calculator computes the off-gas volume and runaway energy for a given pack, and the fire checklist lists the battery-room requirements item by item.

5. Standards and the number each one fixes

Twelve documents set the numbers on this page. The AHJ (the authority having jurisdiction: fire service or building authority) and the insurer decide which edition applies to a site.

Also cited on this page: EN 54-20 (aspirating detector classes), NFPA 12 (CO₂ systems), NFPA 25 (inspection of water systems), NFPA 750 (water mist), NFPA 68 and 69 (explosion venting and prevention). The NFPA fire and risk lab works through NFPA 75 and 76 room by room.

6. Inspection, testing and commissioning

A fire system that is not tested drifts: detectors clog, cylinders lose pressure, and new cable penetrations open the room. The table lists the minimum intervals from the standards and the shorter intervals the site checklist adds.

Commissioning gates before the system is accepted [S5]:

1. Room integrity: door-fan test shows the concentration is held for at least 10 min; seal and re-test until it passes.
2. Agent quantity: charged weight or pressure matches the design calculation; cylinder serials and fill weights recorded.
3. Discharge (where required): full discharge in 10 s or less (halocarbon) or 60 s or less (inert), and the measured concentration reaches design.
4. Cross-zoning: no single detector releases the agent; pre-discharge delay, abort and manual release all work.
5. Interlocks: on a confirmed alarm the air handlers stop, dampers close and the EPO sequence runs.
6. Li-ion gas detection: sensors calibrated; alarm at or below 25 % LFL starts the exhaust; the BMS shutdown sequence is confirmed.
7. Integrated systems test: detection to alarm to HVAC to EPO to release to BMS, end to end.

The full item list, the symptom-cause-action table and a printable service record are on the fire checklist.

7. Fires in data centers, 2021–2026

Each row links to a dossier with the sequence of events, the sources and an FMEA. Common pattern: the burned area was small, and the outage came from the power cut needed to fight the fire, from water reaching IT equipment, or from batteries that kept re-igniting.

All incident dossiers are listed on the incident library.

8. Tools, data and terms

Glossary: cross-zoned detection · design concentration · NOAEL / LOAEL · room integrity · off-gassing · thermal runaway · NFPA 855 · UL 9540A · VESDA · pre-action · Inergen · EPO

Frequently asked questions

Sources

• [S1] Clean-agent properties dataset — data/fire/clean-agent-properties.csv, 5 rows, compiled from NFPA 2001 (2022) tables, the 3M Novec 1230 datasheet, IPCC and NFPA 12. Taken: design and extinguishing concentrations, NOAEL/LOAEL, k1/k2, discharge times, GWP.
• [S2] Lithium-ion chemistry dataset — data/fire/li-ion-chemistry.csv, 5 rows, compiled from Nature Scientific Reports (2024), UL 9540A, IEEE 1188 and NFPA 855. Taken: runaway onset, off-gas volume and species, LFL values.
• [S3] Standards register — data/fire/standards-references.csv, 12 rows with links to each publisher. Taken: the number or rule each standard fixes.
• [S4] Fire calculation engine — js/fire-model.js and data/fire/README.md, validated by tools/test-fire-calc.mjs. Taken: discharge ≤ 10 s, hold ≥ 10 min, 84 m² per spot detector, ASD pre-alarm 0.005 %/m, inert-gas volume equation.
• [S5] Fire design, commissioning and Li-ion checklist — fire-checklist.html, sections 02, 04 and 07. Taken: ASD band 0.005–0.05 %/m, cross-zoning, 30 s pre-discharge delay (typical), commissioning gates, site test intervals.
• [S6] Standards cited by clause or chapter in the text: NFPA 2001 (2022) ch. 7 and Annex C; NFPA 72 (2022) ch. 14; NFPA 25 ch. 5 and 13; NFPA 13 (2025); NFPA 855 (2023); NFPA 750; NFPA 68; NFPA 69; NFPA 12; EN 54-20 (aspirating smoke detectors). Links to the publishers are in section 5.
• [S7] Incident dossiers on this site, one per row of section 7, each with its own source list.