Fire Detection, Suppression & Life-Safety

Fire protection is where the data center industry's two cultures collide hardest. To the fire engineer, a data hall is a high-value, high-power, occupied space governed by codes that predate the AI era. To the cluster owner, it is a machine that must not stop, and most legacy fire responses (dump the room, kill the power, soak the racks) are exactly the events the cluster is built to avoid. This chapter resolves that collision at design time, because the alternative is resolving it in an insurer's review after the slab is poured, when every option is expensive and some are impossible.

The consequences are non-linear and partly out of your hands. A mis-chosen cooling fluid can fail an FM Global approval and strand the whole cooling scheme. A battery chemistry picked for energy density can fail a thermal-runaway test and force the energy-storage room to move. The most expensive surprise of 2026 is finding that the property insurer your lenders made a condition of funding adds binding design conditions on top of the adopted fire code.

Every defensible fire strategy has four layers, and the design error is treating any one of them as the whole answer. Prevent is the layer most operators skip and the cheapest to own: arc-flash-safe power design, qualified-worker programs, clean cabling and firestopping, and, at the extreme, oxygen reduction that makes ignition hard. Detect is the layer AI density stresses hardest, because residual air heat, rack-fan behavior, containment and the return path decide where smoke travels. Suppress carries the most contested fork (water, clean agent, mist or none) and is the layer most entangled with insurability. Contain is the architectural layer that decides what a fire can't do whether or not detection and suppression work: compartmentation, smoke control, egress and firefighter access.

The cascade runs the other way from how budgets are usually spent. Money flows to suppression hardware because it is visible and biddable, but the layers that bound your worst day are prevention, which removes ignition sources, and containment, which caps the loss when the other three fail. A hall with immaculate clean-agent coverage and poor compartmentation has bought a fast first response and an uncapped maximum loss. Fund the four layers as a portfolio, weighted to prevention and containment.

Point smoke detectors at the ceiling assume smoke rises and collects there. In a liquid-cooled AI hall, the rack's residual air heat, its fan behavior, the containment and the supply and return paths decide whether smoke reaches the ceiling or is pulled into the return. Obtain those inputs, then use transport analysis and witnessed smoke testing to select, locate and commission the detection. The guide recommends layered, very-early-warning aspirating (VESDA-class) detection for this airflow boundary. The aim is the earliest possible alarm, so a human can investigate and intervene before any suppression event is needed. Chapter 6.1 owns the hall geometry the sampling network must follow.

The decision inside ASD is where to sample and at what sensitivity, and it is hard in a high-airflow hall. Sampling at the ceiling alone misses smoke the airflow has already entrained. Current practice samples where the air actually goes: the hot-aisle / return path and ceiling return plenum, where the room's airflow concentrates combustion products before exhaust. Under-floor sampling matters far less in a slab-and-liquid hall than it did in a raised-floor air hall. Sensitivity is the harder half. Set too sensitive, a dusty commissioning period or a normal hot-spot off-gas event nuisance-trips a live cluster. Set too coarse, dilution defeats the premise. The resolution is cascaded alarm thresholds (Alert / Action / Fire) tied to airflow-aware sensitivity and validated by real smoke testing at commissioning, not a datasheet number.

Suppression is the chapter's sharpest decision, and the wrong choice shows up in the loss column. Four families trade equipment safety, life safety, re-occupancy time, water and collateral damage, and insurability against each other. None wins on every axis; the right answer depends on density, fluid choice, the insurer and whether the room is occupied.

Water (wet-pipe or, the data-center reflex, pre-action / double-interlock). Sprinklers remain a design candidate under the NFPA 75 edition that applies to the project, the adopted code and the insurer's conditions, and the AHJ must accept the actual IT-room basis. The adopted fire code usually brings in the rest: NFPA 72 for detection and alarm, NFPA 13 for sprinklers, and NFPA 2001 or ISO 14520 for clean-agent systems, each in the edition the jurisdiction adopted (Appendix A). Adding a clean-agent system does not by itself remove an applicable sprinkler obligation. A double-interlock pre-action system keeps the pipe dry until both a detection event and a sprinkler-head fusing occur. That makes water tolerable over live electronics, because it sharply cuts the inadvertent-discharge risk that makes operators wary of water in the first place. The cost is paid twice: collateral damage if it does discharge, and delayed water delivery from a more complex system that is impaired more often. FM 5-32 accordingly ranks wet, non-interlocked and single-interlocked systems ahead of double-interlock, and fire growth during the delay belongs in the trade beside inadvertent discharge. Water and energized racks also do not coexist gracefully, which is why detection-driven power interlocks matter.

Clean agent (FK-5-1-12 / Novec 1230, HFC-227ea / FM-200, or inert gases IG-55/IG-541). Clean agents flood the room and extinguish open flame without water damage, and they leave no residue. They suit a sealed IT room where re-occupancy speed and low discharge-related collateral damage are worth a premium. Halocarbon agents still form corrosive decomposition products such as hydrogen fluoride when they meet fire, and discharge noise can damage spinning disks. The catch is scope: an agent dump protects against an open-flame event in a sealed enclosure. It does nothing for a re-igniting energy source, it requires room integrity (a failed door-fan test means the agent leaks out before it works), and it does not arrest a lithium-ion thermal runaway.

Agent choice now carries a supply and regulatory dimension alongside fire performance. 3M's exit from PFAS manufacturing at the end of 2025 ended Novec 1230 production, but FK-5-1-12 is a generic agent that several manufacturers still supply with UL-listed and FM-approved hardware. Specify a listed system and a qualified fill, not a brand that has left the market. Europe is where the timing bites. Regulation (EU) 2024/573 has prohibited placing on the market fire-protection equipment that contains or relies on Annex I F-gases, HFC-227ea among them, since 1 January 2025 (Article 11 with Annex IV point 11(c)), except where safety requirements at the site of operation demand it. That is a first-placement restriction, not a ban on operating or servicing equipment already lawfully placed. The wider EU PFAS restriction is still a proposal: ECHA's scientific assessment (SEAC's final opinion) is expected by the end of 2026, after which the Commission, with Member States, decides the restriction and any fire-suppression derogation. Choose HFC-227ea for a new European installation only on that safety exception, and price refill and retrofit exposure into any agent you select today.

In the US, as of 15 October 2026, the pressure comes from supply and from state rules. The AIM Act caps total HFC production and consumption, not use. The ceiling is 60% of baseline through 2028 and 30% of baseline from 2029 to 2033, and it is no quota on HFC-227ea alone. Under 40 CFR 84.110, servicing and repairing HFC fire-suppression systems must use recycled HFCs from 1 January 2026, and initial installations must from 1 January 2030. Secure a recycled-agent refill and service route when you specify FM-200. FK-5-1-12 falls under Minnesota's broad PFAS definition. Covered manufacturers had to file initial reports by 15 September 2026, or by 14 December if granted an extension. Minnesota's sales prohibition starts on 1 January 2032 and reaches only products with intentionally added PFAS, minus statutory exemptions and any currently-unavoidable-use determination.

Water mist. Water mist works by cooling and by displacing oxygen locally with a fine droplet spray. Water demand comes from the approved design for the specific hazard. To compare it with sprinklers, total the discharge over the same hazard, protected area and duration. Mist sits between water's reliability and clean agent's gentleness, and its cooling action suits high-density halls. It has approved battery applications, but there is no generic mist-over-sprinkler ranking for lithium. FM's data sheet excludes approval of mist for Li-ion UPS cabinets while permitting other mist applications. What decides is the configuration, the tested protection objective (cooling an exposure and limiting propagation is not arresting a cell), and whether the approval covers the installation you are building. The cost is system complexity (high-pressure pumps, fine nozzles, water quality) and a smaller base of approvals to lean on.

Oxygen reduction / hypoxic (continuous prevention, not suppression). The most radical option holds the hall's oxygen permanently below the 20.9% of ambient air by injecting nitrogen-enriched air. EN 16750 sets the level per protected fuel: the lowest applicable ignition threshold for the fuels and arrangement, minus a 0.75 percentage-point safety margin. Operating setpoints also leave room for sensor tolerance and control margins. Performance means tested ignition prevention for the tested fuels at that design level; it does not make the room incapable of fire. When those fuels do not ignite, the design avoids an agent discharge and its collateral damage and re-occupancy delay. A lithium cell's internal runaway can continue without ambient oxygen, so any rack BBU or facility ESS keeps the configuration-specific controls described below. The costs are real and recurring: continuous nitrogen-generation energy (a PUE and opex hit), occupational-health, respiratory-protection and entry requirements for an oxygen-deficient space, sealing requirements, and AHJ unfamiliarity in some jurisdictions. Choose it where the test basis covers the actual fuel package and the avoided discharge justifies the running cost. It cannot replace battery hazard controls.

The rightmost two columns carry the decisions. The lithium-runaway column corrects a common assumption: a clean-agent hall is not a protected battery room, and conflating the two is a recurring and dangerous error. The insurability column overrides everything else, so it gets its own section.

Law sets the minimum: building and fire laws, the codes a jurisdiction adopts, local amendments, and any standards those codes reference. NFPA 75/76 apply on top only where that legal basis incorporates them or the owner, insurer, agency or contract specifies them. The body that actually governs a hyperscale or institutionally financed AI data center is usually the property insurer, because the lenders make a specific insurer's approval a condition of funding and that insurer's standards exceed code. In North America the instrument written into those conditions is typically FM Global's Approval Standards and Data Sheets, which add contract-specific property-protection requirements without replacing adopted law. Elsewhere the roles separate rather than substitute. VdS (Germany) is an insurer-backed approval body. LPCB (UK/Red Book) is a third-party product and system certification scheme whose listings an insurer or AHJ may require. FM Global is also present. Four actors decide and none replaces another: the AHJ enforcing adopted law, the property insurer setting policy conditions, the certification body that lists the product or system, and the owner or lender's contractual criteria. Chapter 2.6 owns the insurability economics; this chapter owns what they do to the engineering.

The consequence is direct and expensive: the required FM product approval and insurer acceptance can gate the cooling-fluid system and battery installation you chose. A dielectric or coolant without the right approval, an immersion fluid that fails a flammability or materials-compatibility criterion, a BESS that has not passed the runaway-propagation tests the insurer recognizes: any of these can be vetoed by the absence of a certificate rather than by an engineer's judgment. That is the channel by which the two-phase immersion fluids (3M's Novec family), already stalled on PFAS health and liability concerns, also became an insurability problem. It is also part of why new NVL72-class deployments usually run single-phase direct-to-chip, the cooling those racks are designed around: the case rests on fire and fluid as well as heat. The fluid and chemistry decisions in Chapter 5.4 and Chapter 5.5 are, in effect, co-signed here.

The AI build brought a fire problem the legacy data center never had at scale: large lithium-ion energy storage, both as rack-level battery backup units (BBUs) and as facility-scale BESS for ride-through and power-transient absorption. Chapter 4.5 owns the electrical role of that chain; this chapter owns the fire consequence. A lithium cell in thermal runaway is an internal exothermic reaction that can continue without ambient oxygen. Smothering the external flame or removing room oxygen does not remove the cell's internal heat, so heat and propagation must be controlled separately. Runaway also propagates from cell to cell and emits a flammable, toxic off-gas (hydrogen, CO, electrolyte vapors) before any visible fire, which creates an explosion hazard distinct from the fire hazard.

In most US jurisdictions the battery rules take legal force through the adopted fire code: Section 1207 of the International Fire Code (IFC) in its 2021 and 2024 editions (Section 1206 where a state still enforces the 2018 edition), or NFPA 1 in states that adopt that code instead. The fire code sets the stored-energy thresholds above which the ESS requirements apply, and it recognizes UL 9540A large-scale fire testing for installations beyond its limits, so confirm which code edition the AHJ enforces before reading NFPA 855 as binding. NFPA 855 itself does not assign one generic battery-room prescription to every installation, so the control stack splits three ways. Rack-level BBUs inside an occupied data hall stay within the applicable stored-energy thresholds and the listed rack or UPS configuration, with the monitoring and isolation that listing and its tested installation require. A BBU shelf inside an NVL72-class rack (Chapter 7.13) cannot be turned into a dedicated deflagration-vented room; if the quantity or listing basis is exceeded, move the storage out of the occupied hall. Facility ESS in dedicated indoor rooms need rated compartmentation and separation, module thermal and off-gas detection, ventilation, explosion control where the gas-hazard analysis requires it, water-based fire control and cooling, and fire-service access. Outdoor listed ESS enclosures use the listed cabinet or container with site and unit separation, enclosure gas management and explosion control as tested, exposure protection, water supply and fire-service access. Do not import an indoor-room detail into an outdoor enclosure.

A fire on 26 September 2025, while UPS batteries were being prepared for relocation at Korea's NIRS Daejeon center, disrupted 647 government systems. At a briefing on 29 September, Korea's Ministry of the Interior and Safety confirmed both the relocation work and the outage. Treat a battery move or replacement as its own fire-risk state, and include the residual stored energy in the work boundary. Chapter 6.9 owns the procedure for isolation, verification and handling.

The chemistry choice remains a fire decision. LFP (lithium iron phosphate) runs cooler and is more abuse-tolerant than NMC, which is one reason it dominates facility BESS. UL published the sixth edition of the 9540A test method in March 2026. The sixth edition changes the installation-level test that already existed. Cell and module results no longer end the protocol: testing continues to a pass/fail large-scale fire test, with vented gases deliberately ignited to create the post-deflagration scenario in NFPA 855-2026 Annex G.11. Unit-level testing remains for residential systems and can also evaluate an active system that prevents thermal-runaway propagation. UL has announced a 1 January 2027 effective date, so verify the report's edition and whether it applies to your installation. The installation-level result gives configuration-specific fire, gas, heat-release and propagation evidence, which the designer and AHJ use to determine or justify spacing, enclosure and venting, and suppression under the adopted code and listed-system conditions. Evidence for one of the three configurations does not establish the basis for the other two.

Containment decides the size of your worst day, and it is set architecturally, in Chapter 6.1's layout and Chapter 6.3's envelope, long before any suppression hardware is specified. Compartmentation subdivides the hall into rated fire compartments so that a fire (or an agent dump) stays in one zone instead of spreading across the facility. The larger and denser the hall, the more this matters, because the value at risk per compartment is enormous. It also cuts against cluster design: a tightly coupled training cluster wants one large contiguous hall for fabric reasons. Resolving that is a real cross-discipline fork between fire containment and network topology.

Smoke control and egress are life-safety obligations that the density ramp complicates. A hall with hot-aisle containment, high airflow and an oxygen-reduced or clean-agent atmosphere is hard to evacuate and hard to ventilate after an event, and smoke-control design must account for the same airflow that challenges detection. Egress paths must stay valid as racks get heavier and aisles get plumbed: a move route or escape path that worked for 5 kW racks can be blocked by the manifolds and CDUs of a liquid retrofit (Chapter 6.7 owns the rigging and floor-loading interactions).

The firefighter-access problem is the under-appreciated one. Responders entering a dense, energized, liquid-cooled hall face hazards a legacy hall never presented: hundreds of kilowatts per rack, lithium energy storage that re-ignites, dielectric and glycol coolants, and DC-shock risk from 800 VDC distribution. Manual firefighting may be neither safe nor effective. That shifts the strategy toward early detection, automatic suppression and orderly power isolation, and toward giving responders what they need to make a safe stay-out decision: zone status, power state and battery-room status. Detection, suppression, power and the BMS must speak one language here.

Carry the passive-fire design as an assembly schedule: required rating and supporting construction, doors and hardware, penetrations, movement joints and any ducts or dampers. Match each installation to its tested and listed system, including permitted substrates, annular space, insulation and movement. Mark concealed inspections before closure, and keep a penetration register for later cable and pipe changes.

Coordinate fire-service approach, appliance setup, hydrants, fire connections, control-room information and safe isolation with the authority and responders. The heavy-haul corridor in 6.3 is not automatically the approved fire-apparatus route. Submit its usable width, loading, turns, gradients, overhead clearance and all-phase obstructions against the adopted access criteria.

The interlock matrix is where fire safety stops being independent systems and becomes a control problem, and it is the layer most likely to either save the cluster or take it down unnecessarily. The protective links are detection → power isolation, detection → the approved suppression sequence, and the cooling and BMS actions that follow. In an AI hall every one of these links is fraught.

Detection → power. The approved sequence coordinates power isolation with water or agent protection, while required sprinkler and preaction functions stay independent of an IT-breaker acknowledgment. A nuisance trip that de-energizes a live training cluster is a multi-hour, multi-million-dollar goodput loss (Chapter 12.2 has the goodput-versus-availability framing). The design answer is cascaded, confirmed alarms (Alert/Action/Fire) and zoned isolation that drops the smallest possible domain, not the hall.

Detection → cooling. Liquid cooling complicates the picture. Shutting pumps in a thermal event can be exactly wrong, because you may want to keep cooling, or specifically cool a battery. A coolant leak is itself a detected event with its own interlock (leak detection → CDU isolation), distinct from fire. The cooling interlocks coordinate with the CDU and secondary-loop logic in Chapter 5.6.

Everything → BMS. The building management system is the integration point that must show the human responder a coherent state: fire zone, power state, cooling state and battery-room state. Required fire and life-safety logic runs on its approved controls, independent of ordinary BMS availability. A fire strategy whose interlocks are not modeled, tested and tuned against the live cluster's failure modes will eventually trip the cluster for the wrong reason or fail to trip it for the right one.

Trace what the report actually covers. First, identify the cabinet, cells, modules, quantities, ventilation, spacing and protection being purchased. Second, compare those fields with the supplied report and listing. A cell-level test does not establish cabinet-to-cabinet propagation or the installation’s gas hazard, and a UL 9540A report is not a UL 9540 system listing. Third, compare the proposed location with the adopted installation requirements and FM contract conditions. The cell report covers neither the cabinet configuration nor the room controls, so hold the in-hall substitution.

Keep the independent battery-room design while procurement obtains the missing test data, and accept no cabinet delivery that silently uses occupied-hall space, egress width or separation. This preserves the hall’s layout at the cost of keeping the dedicated room with its detection, ventilation and applicable fire and explosion controls. The decision flips on test data and listings: the in-hall option becomes eligible only when the same supplied configuration meets every applicable listing, quantity, separation, protection and authority condition and has the insurer's acceptance. A report for a different cell or enclosure does not count.

The fire-panel cause-and-effect test then challenges the actual boundary. Inject the approved alarm condition in an isolated test scope, with the ordinary BMS link unavailable and IT-breaker feedback withheld. Required sprinkler and preaction functions must proceed through their approved fire controls; failed breaker feedback produces a fault and escalation, not an indefinite veto on required water protection. Trace the isolation command to every AC feed and UPS output in the affected zone, keeping required fire and life-safety supplies available. Cooling follows its approved hazard-specific sequence, and closing a leak valve is a different action from the cooling an ESS fire demands. The matrix fails if either protective action depends on a BMS acknowledgment.

UL Solutions distinguishes test evidence, system certification and the sixth-edition transition. FM 5-32 addresses configuration-specific protection and power isolation. This chapter owns the fire decision; Chapter 13.6 witnesses integrated acceptance and Chapter 6.9 authorizes de-energized maintenance.

Choose protection for the room, battery and fluid you are actually installing, then prove that its alarm, water and isolation actions work independently. Approval stops at a missing listing, or at water that waits on breaker feedback; buying the equipment first turns that gap into a costly room redesign.