Fire Detection, Suppression & Life-Safety

Fire protection is the discipline where the data center industry's two cultures collide hardest. To the fire engineer, a data hall is a high-value, high-power, life-safety-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 precisely the events the cluster is engineered to avoid. The job of this chapter is to resolve that collision at design time, because the alternative is resolving it during an insurer's review after the slab is poured, when every option is expensive and some are impossible.

The consequences here are non-linear and partly out of your hands. A mis-chosen cooling fluid does not just cool poorly; it can fail an FM Global approval and strand the entire cooling scheme. A battery chemistry picked for energy density can fail a thermal-runaway test and force a re-siting of the energy-storage room. And the most expensive surprise in the 2026 era is discovering that the body which actually controls your design is not the national fire code at all, but the property insurer whose approval your lenders made a condition of funding.

Every defensible fire strategy is 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 thermodynamically hard. Detect is the layer AI density stresses hardest, because the same airflow that cools a 130 kW rack also dilutes and disperses the smoke you are trying to sense. Suppress is the layer with the most contested fork — water vs clean agent vs mist vs none — and the one most entangled with insurability. Contain is the architectural layer that decides what a fire can't do regardless of whether detection or suppression works: compartmentation, smoke control, egress, and firefighter access.

The cascade runs the other way from how budgets are usually allocated. Money flows to suppression hardware because it is visible and biddable; but the layers that actually 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 is a hall that has bought a fast first response and an uncapped maximum loss. The four layers are a portfolio, not a checklist.

Point smoke detectors at the ceiling assume smoke rises and accumulates. A liquid-cooled AI hall violates both assumptions: air changes per hour are extreme, supply and return are engineered to move heat as fast as possible, and a smouldering connector or a charring cable can have its smoke plume diluted below a spot detector's threshold and swept into the return before it ever reaches a ceiling device. The standard answer is aspirating smoke detection (ASD) — air-sampling systems, of which VESDA is the genericized brand name — which actively draws air through a pipe network and analyses it with a laser nephelometer sensitive enough to catch the incipient (pre-visible) stage. NFPA 75's 2024 edition leans on this layered, very-early-warning philosophy: get the earliest possible alarm so a human can investigate and intervene before any suppression event is needed. Chapter 6.1 owns the hall geometry that the sampling network must follow.

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

Suppression is where the chapter has its sharpest decision, and where the wrong choice is most visible in the loss column. There are four families, and they trade equipment safety, life safety, re-occupancy time, water/collateral damage, and insurability against each other. No option wins on every axis; the right answer is a function of density, fluid choice, the insurer, and whether the room is occupied.

Water (wet-pipe, or — the data-center default — pre-action / double-interlock). Sprinklers are the code baseline almost everywhere; NFPA 75 treats automatic sprinkler coverage as the expected default and most AHJs and insurers require it regardless of what else you add. A double-interlock pre-action system keeps the pipe dry until both a detection event and a sprinkler-head fusing occur, which is what makes water tolerable over live electronics: it dramatically reduces the inadvertent-discharge risk that makes operators fear water in the first place. The cost is collateral damage if it ever does discharge, and the physics that water and energized racks do not coexist gracefully — which is precisely 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, then leave no residue — ideal for a sealed IT room where re-occupancy speed and zero collateral damage are worth a premium. The catch is that 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 the discharge itself has caused hard-drive damage from acoustic over-pressure. Clean agent is a finish, not a cure — and notably it does not arrest a lithium-ion thermal runaway.

Water mist. Fine-droplet mist suppresses by cooling and local oxygen displacement with one to two orders of magnitude less water than sprinklers, bridging the gap between water's reliability and clean agent's gentleness. It is attractive for high-density and for battery rooms specifically, because its cooling action is what a thermal event actually needs. The cost is system complexity (high-pressure pumps, fine nozzles, water quality) and a smaller installed base of approvals to lean on.

Oxygen-reduction / hypoxic (continuous prevention, not suppression). The most architecturally radical option lowers the hall's oxygen concentration permanently to ~15–17% (vs 20.9% ambient) by injecting nitrogen-enriched air, so that most materials simply cannot sustain combustion — fire is prevented rather than fought. Governed in Europe by EN 16750, it eliminates the discharge event, the collateral damage, and the re-occupancy delay entirely. The costs are real and recurring: continuous nitrogen-generation energy (a PUE/opex hit), occupational-health limits on personnel exposure and time-at-altitude, sealing requirements, and AHJ unfamiliarity in some jurisdictions. It is the prevent-layer answer that some hyperscale and high-value halls now choose precisely because it removes the suppression dilemma rather than resolving it.

The table's rightmost two columns are the ones that catch teams off guard. The lithium-runaway column kills the lazy assumption that a clean-agent hall is also a protected battery room — it is not, and conflating the two is a recurring and dangerous error. The insurability column is the one that quietly overrides everything else, and it deserves its own treatment.

Here is the fork that surprises strategists. The national fire code (NFPA in North America, EN-aligned national codes in Europe) sets the floor — the minimum to obtain a certificate of occupancy. But the body that actually governs the design of a hyperscale or institutionally-financed AI data center is usually the property insurer, because the lenders behind the project make a specific insurer's approval a condition of funding, and that insurer's standards exceed code. In North America that body is overwhelmingly FM Global, whose Approval Standards and Data Sheets function as a private, stricter, prescriptive code. In Europe and much of the rest of the world the equivalent gatekeepers are VdS (Germany) and LPCB (UK/Red Book), with FM Global also present. Chapter 2.6 owns the insurability economics; this chapter owns what it does to the engineering.

The consequence is direct and expensive: an FM Global approval can gate your cooling fluid and your battery chemistry. A dielectric or coolant that lacks 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 not by an engineer's judgment but by the absence of a certificate. This is the channel by which the two-phase immersion fluids (3M's Novec family) that stalled on PFAS health and liability concerns also became an insurability problem, and why single-phase direct-to-chip became the 2026 default for reasons that are partly fire-and-fluid, not just thermal. The fluid and chemistry decisions made in Chapter 5.4 and Chapter 5.5 are, in effect, co-signed here.

The AI build introduced 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. The electrical role of that chain is owned by Chapter 4.5; the fire consequence is owned here. A lithium cell in thermal runaway is a self-sustaining exothermic reaction that produces its own oxidizer — it does not need ambient oxygen to keep going, which is exactly why the IT hall's clean-agent flood is the wrong tool. You cannot smother what makes its own oxygen. Worse, runaway propagates cell-to-cell and emits a flammable, toxic off-gas (hydrogen, CO, electrolyte vapors) before visible fire, creating an explosion hazard distinct from the fire hazard.

The contemporary containment strategy is therefore a different stack from the IT hall, and it is layered: off-gas detection (catch the vented electrolyte gases before flame), thermal monitoring at module level, dedicated, compartmented battery rooms separated from the IT load (NFPA 855 governs the separation and siting in North America), explosion venting / deflagration management to relieve the gas pressure safely, and water-based cooling — because cooling, not flame suppression, is what an energy-producing runaway actually requires. The chemistry choice is itself a fire decision: LFP (lithium iron phosphate) runs cooler and is more abuse-tolerant than NMC, which is one reason it dominates facility BESS, and it interacts with the UL 9540A propagation testing that the insurer will demand. UL released the 9540A Test Method's most significant revision in 2025, tightening how runaway propagation is characterized — and that test result is frequently the artifact that decides whether a BESS clears the insurer's review.

Containment is the layer that 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) is bounded to one zone rather than the whole facility; the larger and denser the hall, the more this matters, because the value-at-risk per compartment is enormous. The tension with cluster design is real: a tightly-coupled training cluster wants one large contiguous hall for fabric reasons, and compartmentation cuts against that. Resolving it is a genuine 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 a difficult environment to evacuate and a difficult one to ventilate post-event; smoke-control design must account for the same airflow that challenges detection. Egress paths must remain valid as racks get heavier and aisles get plumbed — a move route or escape path that worked for 5 kW racks can be obstructed by the manifolds and CDUs of a liquid retrofit (the rigging and floor-loading interactions are owned by Chapter 6.7).

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, which shifts the entire strategy toward early detection, automatic suppression, and orderly power isolation — and toward giving responders the information (zone status, power state, battery-room status) to make a stay-out decision safely. This is where detection, suppression, power, and the BMS must speak one language.

The interlock matrix is where fire-safety stops being a set of 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 classic interlock is detection → power isolation → suppression: confirm a fire, drop the affected power, then (if needed) discharge. In an AI hall every one of those links is fraught.

Detection → power. Dropping power to energized racks before a water or agent event is the safe sequence, but a nuisance trip that de-energizes a live training cluster is a multi-hour, multi-million-dollar goodput loss (the goodput-vs-availability framing lives in Chapter 12.2). 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 (you may want to keep cooling, or specifically cool a battery), and a coolant leak is itself a detected event with its own interlock (leak detection → CDU isolation), distinct from fire. The cooling interlocks coordinate with Chapter 5.6's CDU/secondary-loop logic. Everything → BMS. The building management system is the integration point that must present a coherent state — fire zone, power state, cooling state, battery-room state — to both the automatic logic and the human responder. A fire strategy whose interlocks are not modeled, tested, and tuned against the live cluster's failure modes is a strategy that will eventually trip the cluster for the wrong reason or fail to trip it for the right one.