Rack BBU and in-rack energy storage for data center ride-through

What a rack BBU is and why the battery moves into the rack

A rack BBU (battery backup unit) is a lithium battery placed inside the rack, or in a sidecar shelf next to it, that holds up the IT load during a power disturbance and carries it for the few seconds it takes the generator to start and accept load. Putting the battery in the rack instead of in a central battery room is the whole idea, and it changes the backup architecture. The ride-through stops being one large block of stored energy in the electrical room and becomes many small blocks distributed across the floor, each carrying its own rack.

That shift came out of the hyperscale world and got standardized in the Open Compute Project's Open Rack, where a BBU shelf sits alongside the rack power shelf and feeds the rack DC busbar during an upstream AC disturbance. The approach distributes the ride-through, can shorten or replace the central UPS, and increasingly does double duty smoothing the power spikes that dense AI racks throw at the supply. None of that is free. Lithium cells sitting in the white space, in the same room as the servers and the people, raise fire-safety questions that a central battery room kept at arm's length.

This guide covers the rack-level, distributed approach. The central UPS topology it is measured against, N+1 through 2N and the dual bus, is the subject of the UPS topology and redundancy design guide, and the rack DC busbar and power shelf the BBU feeds are covered in the rack DC power distribution guide. Read those alongside this one; they are the two halves the rack BBU sits between.

The BBU as a device: lithium energy in the rack

Strip it to the hardware and a rack BBU is a managed lithium battery pack sized for short, high-rate discharge, packaged to fit the rack or a shelf adjacent to it, with a battery management system and protection built in. It is not a runtime battery. It is an energy buffer that sits on the rack DC bus, charged and ready, and dumps its energy fast the instant the upstream source sags or drops.

The defining trait is power density over energy density. A BBU has to deliver the rack's full load, which on a dense AI rack can be tens to over a hundred kilowatts, for a short window, so the cells are chosen and rated for a high discharge rate rather than for capacity. That is the opposite emphasis from a stationary energy-storage system that exists to run a load for an hour.

In the OCP Open Rack arrangement the BBU is modular. A BBU shelf carries several modules with built-in redundancy and feeds the same busbar the rack power shelf feeds, so the ride-through is integrated into the rack power chain rather than wired in from a separate room. The exact module count, shelf rating, and backup window are set by the design and the manufacturer's specification, and that is where you confirm them, not from a rule of thumb.

What does a rack BBU actually do in an outage?

A rack BBU does one job: it carries the load for seconds during a power disturbance and bridges the gap until the generator starts and takes over. It is a bridge, not a runtime battery. The instant the utility sags, swells, or drops, the BBU supplies the rack from stored energy with no break, and it keeps supplying only long enough for the standby generator to crank, build voltage and frequency, and accept the block load. After that the generator runs the building and the BBU recharges.

Read that as the hard limit it is. The BBU is sized to cover the transfer to the generator, not to ride out a long outage on its own. If the generator fails to start, the BBU buys a short window to alarm, to shed load, or to move applications to another site, and then the rack goes down. Anyone treating a rack BBU as hours of runtime has misread what it is for.

The same short ride-through also covers the smaller events that never reach the generator at all: a sub-second sag, a brief transfer between sources, the kind of disturbance the rack should never feel. Most of what a BBU absorbs over its life is these small blips, not full outages. Either way the device is doing the same thing, holding the rack up across a gap it must not see.

Rack BBU vs central UPS: what changes?

A central UPS is a large uninterruptible power supply in the electrical room, with its battery or flywheel store, feeding the critical load through the building distribution. A rack BBU pushes that stored energy out to the rack, so the ride-through lives next to the load instead of in a room upstream. That is the core tradeoff, and it cuts both ways.

The distributed approach shortens the power path between the energy store and the server, which trims conversion and distribution losses, and it lets the ride-through scale rack by rack as the hall fills instead of buying central UPS capacity ahead of load. Done at scale, it can shrink the central UPS to a smaller conditioning role or, in some hyperscale designs, remove it from the critical path and lean on the rack BBUs plus the generator. The central UPS guide covers what that central plant is and how its redundancy is built.

The cost is that you have moved lithium energy storage into the white space, multiplied it by every rack, and distributed the fire-safety problem across the whole floor instead of containing it in one battery room. You have also multiplied the number of battery management systems, the number of state-of-health records, and the number of cells that age and eventually need replacing. The question is not which is better in the abstract. It is which failure and maintenance model the operator wants to own, and the energy-storage codes apply either way.

Why distribute the ride-through at all

The pull toward distributed BBUs is part efficiency, part scale, and part the hyperscaler operating model. A shorter path from stored energy to silicon means fewer conversion stages and less copper carrying current across the building, which shows up as lower loss and, at hyperscale, real money. The rack DC power guide covers why dense racks are collapsing conversion stages in the first place; the BBU rides on the same logic.

Scaling matters as much as efficiency. A central UPS is bought in large blocks, so a hall that fills over years either pays for capacity it is not using or runs short of redundancy until the next block lands. Distributed BBUs track the load: each rack brings its own ride-through when it is deployed, so the backup grows with the compute rather than in big steps ahead of it.

Then there is the failure model. When the ride-through is distributed, a BBU problem is a rack problem, not a hall problem, which is the same logic that makes hyperscalers comfortable running compute with a different risk profile than a payment system. The catch is that distributing the energy distributes the maintenance and the hazard with it, so the efficiency case has to be weighed against the fire-safety and lifecycle load it puts on operations.

The OCP Open Rack BBU shelf

The Open Compute Project standardized the rack BBU in its Open Rack architecture, where a BBU shelf sits in the rack alongside the power shelf and feeds the same DC busbar at the Open Rack distribution voltage, the 48V class. The two shelves together form the rack power chain: the power shelf rectifies and distributes, the BBU shelf holds the ride-through energy, and both land on the common busbar the IT gear taps. The busbar and the power-shelf side are covered in the rack DC power distribution guide.

The shelf is modular and internally redundant. Public Open Rack V3 documentation describes a BBU shelf built from several modules with redundancy at the shelf, a common arrangement being five active modules plus one spare, so a single module can fail or be pulled without losing the rack's ride-through. A single shelf is rated to back up the rack power up to its specified limit for a defined window. The exact voltage, module count, shelf kilowatts, and backup time are set by the spec revision and the manufacturer, so confirm them against the actual hardware rather than a remembered figure.

Calling it a sidecar is literal in some designs: the BBU energy can sit beside the rack rather than inside the compute envelope, which keeps the cells out of the hottest airflow and makes them easier to service and isolate. Where the energy physically sits also drives the fire-safety design, because containment and detection follow the cells.

How long does a rack BBU last in an outage?

A rack BBU lasts seconds to a couple of minutes at full rack load, not hours, because it is sized to bridge to the generator and nothing more. The runtime is short on purpose. The design target is the worst-case time for the standby generator to start, build, and accept the block load, plus margin for a start that does not go cleanly, and the energy store is sized to that number rather than to any idea of outage duration.

Historically the lead-acid string in a UPS gave several minutes of cushion, which masked a slow generator. A rack BBU sized tight to the transfer gives far less margin, closer to the flywheel end of the range, so the generator start time is no longer a comfortable assumption. It is the design input. Size the BBU shorter than the real generator start and acceptance time and the rack drops in the gap the BBU was supposed to cover, which is the exact failure the whole arrangement exists to prevent.

Treat any specific runtime as a design and manufacturer value, not a constant. The number depends on the cell capacity, the rack load, the discharge rate, the depth of discharge the BMS allows, and how the cells have aged. Confirm the as-installed runtime against the generator start time by witnessed test, and re-confirm it as the cells age, because the runtime you commission is not the runtime you have in year five.

Bridging to the generator

The BBU and the generator are one system, and the only number that matters between them is whether the ride-through outlasts the start. When utility drops, the BBU picks up the rack instantly while the generator gets the start signal, cranks, builds voltage and frequency, the transfer switch moves the building load to the engine, and the engine accepts the block. That whole sequence has to finish inside the BBU's window, with margin.

Generator start and load acceptance is commonly in the seconds range, but it varies with the engine, the controls, the fuel system, the ambient, and whether the set has to come up cold. A set that starts in ten seconds on a good day can take longer on a cold morning or stumble on the first crank, which is why the BBU window has to cover the bad day, not the brochure. Count the ride-through against the real, tested start time.

The coordination runs deeper than timing. The generator is a softer source than utility, and the moment it lands, the BBUs across the floor begin recharging at once, which is a step load the generator has to swallow on top of the IT load. A recharge that slams the engine the instant it accepts load can swing its voltage and frequency, so the recharge behavior, often a soft or walk-in charge, and the generator sizing have to account for it. The central UPS guide covers the generator interaction in more depth.

Lithium chemistry: LFP versus NMC

The two lithium chemistries that show up in rack BBUs are lithium iron phosphate (LFP) and lithium nickel manganese cobalt oxide (NMC), and the choice is a safety-versus-density call. LFP has lower energy density, so the pack is larger for the same energy, but it is more thermally stable, with a higher thermal-runaway onset temperature and less violent failure behavior. NMC packs more energy into less space and weight, which is why it dominated where density ruled, but it runs hotter into runaway and releases more energy when it fails.

For data center backup the trend has moved toward LFP, and the reason is exactly the fire-safety problem of putting cells in the white space. When the energy lives in the room with the servers and the people, the more stable chemistry is worth the size and weight penalty to many operators, and a BBU does not need NMC's density the way a phone or a car does because it discharges fast for a short time rather than carrying range. That is a lean, not a law. Some designs still use NMC where space or weight is the binding constraint, and the cell rating for high-rate discharge can matter more than the chemistry label.

Treat the chemistry as a design decision tied to the fire-safety case, the manufacturer's cell, and the AHJ's view, not a default. The failure behavior, the onset temperature, and the vent-gas makeup differ between chemistries and between manufacturers' cells, and those differences feed straight into the detection, suppression, and containment design downstream.

Lithium versus lead-acid

Lithium displaced valve-regulated lead-acid (VRLA) in data center backup for reasons that are easy to state and hard to argue with on space and life. A lithium pack is smaller and lighter for the same power, lasts longer before replacement, tolerates higher temperatures, and reports its own state through the BMS instead of failing quietly the way a VRLA string does. Those advantages are what make a battery small enough to live in the rack in the first place; you could not distribute VRLA the way you distribute a BBU.

VRLA had one real advantage worth naming: its failure mode was familiar and far less energetic. A bad lead-acid cell dries out, swells, or shorts. It does not propagate a self-sustaining thermal runaway through the string the way lithium can. The shift to lithium bought density and life and took on a more serious fire problem in exchange, which is the honest trade and the reason the codes tightened around lithium energy storage.

The lead-acid string also lived in a battery room, kept away from the IT load. Moving to lithium and moving into the rack are two separate changes that happened together, and the fire-safety thinking has to account for both: a more energetic chemistry, now distributed through the white space.

The battery management system

The battery management system (BMS) is the protection and the brains of a rack BBU, and on lithium it is not optional. Lithium cells are intolerant of overvoltage, overcurrent, over-temperature, and over-discharge in a way lead-acid is not, and an unmanaged lithium pack is a fire waiting for a trigger. The BMS monitors every cell or cell group, balances the cells so none drifts high or low, enforces the voltage and current limits, watches temperature, and trips the pack offline before a cell goes outside its safe window.

Cell balancing is the part that earns its keep over the life of the pack. Cells age at slightly different rates, and without balancing the weakest cell sets the usable capacity and the most stressed cell drives the failure. The BMS equalizes them so the pack ages evenly and the runtime holds longer. In the OCP modules the protection is built down to the module, self-protecting against overcurrent and short circuit with set thresholds and timing, so a fault is contained at the module rather than propagating up the shelf.

The BMS is also the data source for everything downstream: state of charge, state of health, cell temperatures, fault and alarm history. That telemetry is what lets an operator see a degrading module before it fails and is the feed the monitoring and maintenance program runs on. A BBU whose BMS data nobody watches is a BBU whose failure you find at the worst time.

Absorbing the AI power spike

Dense AI racks created a second job for rack-level energy storage: smoothing the power spikes the GPUs throw at the supply. A large training cluster runs tens of thousands of GPUs in step, so they surge and idle together instead of averaging out, and the draw can jump tens of percent above baseline in a fraction of a second and just as fast collapse. Industry figures put those surges in the range of 30 to 50 percent above baseline lasting on the order of 0.2 to 2 seconds. Pushed onto the grid, synchronized swings like that stress the upstream equipment and can interact badly with generation.

A buffer at the rack absorbs the transient so the supply never sees the full swing. The BBU, or a dedicated capacitor or supercapacitor shelf, charges during the low-demand phase and discharges into the spike, shaving the peak and filling the valley so the power shelf, the UPS, and the grid see a smoother draw. This is peak shaving at the rack rather than at the meter.

Worth being precise about the tool. Batteries are well suited to the ride-through job of carrying a load across an outage; the fastest, most frequent GPU transients are better matched by capacitors and supercapacitors, which tolerate constant sub-second charge and discharge cycling without the wear that would degrade a battery doing the same. Some designs use a capacitor shelf for the high-frequency smoothing and the BBU for the outage ride-through, which are related but not identical functions. Whether a given BBU is rated to do spike absorption on top of ride-through is a manufacturer and design question, and the rack power decisions sit alongside the rack DC power guide.

Distributed redundancy and the failure domain

Distributing the ride-through also distributes the redundancy, and it changes what a single failure costs. In a central UPS, redundancy is built at the plant: N+1 modules on a bus, or 2N independent systems, sized so a module or a whole side can be lost with the load still up. The UPS topology guide covers that in full. In the rack model, the redundancy moves into the shelf: the OCP BBU shelf carries spare modules, a five-plus-one arrangement being common, so a module can fail without losing the rack's ride-through.

The important difference is the size of the failure domain. With a central UPS, a plant-level failure is a hall-level event. With distributed BBUs, a BBU failure is a rack-level event, contained to the rack it serves. That smaller blast radius is one of the attractions, and it pairs with the hyperscale habit of building resilience in the software across many racks rather than buying 2N under every one.

The tradeoff is that you now have hundreds or thousands of small redundant systems to keep redundant, each with its own spare module that can quietly be consumed by a failure nobody replaced. Distributed redundancy is real redundancy only while every shelf still has its spare, which makes the state-of-health monitoring across the fleet the thing that keeps the design intent alive. It is not enough to commission it redundant. It has to stay redundant rack by rack.

Are lithium batteries in data centers a fire risk?

Yes. Lithium cells store a lot of energy in a small space and can fail into thermal runaway, a self-heating reaction that vents flammable and toxic gas and can ignite, and putting them in the rack puts that hazard in the same room as the servers and the people. This is the honest center of the rack BBU decision, and it is not hype. The fire services and the standards bodies treat lithium energy storage as a distinct hazard for good reason, and the design has to address it from the start rather than bolt protection on after.

What makes it a design problem rather than a simple no is that the hazard is manageable with the right chemistry, the right detection, the right containment, and the right separation, all of which the energy-storage codes now spell out. The wrong way to handle it is to treat a rack full of lithium like a rack full of servers and assume the building's ordinary fire protection covers it. It does not. Lithium thermal runaway can keep producing its own oxygen-independent heat, so it resists conventional suppression and can reignite after it looks out.

Be blunt about the consequence. If the fire-safety design is an afterthought, the failure is a battery fire in the white space, with toxic off-gas, possible flashover or deflagration, and a hall you cannot safely re-enter. That is why the chemistry choice leans to LFP, why detection and containment are part of the design, and why NFPA 855 and the AHJ sit over the whole thing. Hedge nothing here. The hazard is real and it is designed for, or it is an incident waiting.

Thermal runaway and propagation

Thermal runaway is the failure mode the whole fire-safety design exists to contain. A cell that is overcharged, over-discharged, overheated, physically damaged, or carrying an internal defect can start heating itself faster than it can shed the heat. Past an onset temperature, the reaction accelerates on its own, the cell vents hot flammable gas, and the heat it throws can push the next cell over its own onset. That cell-to-cell, then module-to-module spread is propagation, and it is what turns one bad cell into a pack fire.

The off-gas is its own hazard before there is any flame. A cell going into runaway vents a mix of flammable and toxic gases, and in an enclosed space that vapor can build to a concentration that deflagrates if it finds an ignition source. That is why off-gas detection matters: the vented gas shows up before ignition and is the earliest warning that a cell is failing, the window in which an operator can act.

Containment design works the runaway at three points: keep a single cell from starting it, keep one cell's failure from propagating to the next, and keep a pack fire from spreading to the rack and the hall. The cell-level protection is the BMS. The propagation resistance is the chemistry, the cell spacing, and the module construction. The hall-level protection is the separation, the detection, and the suppression in the sections below. LFP's higher onset and lower energy give more margin at every one of those points, which is the safety half of why it is trending in the white space.

Which codes govern lithium energy storage?

Lithium battery backup in a data center falls under the energy-storage system codes, and the one to know by name is NFPA 855, the Standard for the Installation of Stationary Energy Storage Systems. It covers location, sizing and separation, ventilation, fire suppression, explosion control, gas detection, and commissioning and decommissioning of the battery system, and recent editions have specific guidance for these installations. The product-listing side runs through UL: UL 1973 lists the battery modules and packs, UL 9540 lists the complete energy-storage system, and UL 9540A is the test method that characterizes thermal-runaway fire behavior and propagation, which NFPA 855 references to set protection requirements.

A figure that drives the room layout: NFPA 855 has thresholds for how much stored energy can sit in a space and when the battery has to be separated from the occupied area, with a commonly cited threshold around 600 kWh of stored energy in a room before separation and additional protection kick in. Distributing BBUs across the floor is exactly the kind of arrangement that has to be totaled against those limits, because a hall full of rack BBUs can add up to a large aggregate even though each shelf is small.

Cite all of this to the topic and confirm the specifics, because editions change and the numbers move. The adopted edition of NFPA 855, the applicable UL listings and their current designations, the local fire code, and the manufacturer's listed installation instructions all govern, and above them sits the authority having jurisdiction. The AHJ's interpretation of separation, suppression, and detection for lithium in the white space is the call that controls the project. Confirm the design with the AHJ early; finding out at inspection is the expensive way.

Detection, suppression, and containment

The fire protection for rack lithium works in layers, and detection comes first because the earliest sign of a failing cell is gas, not heat or flame. Off-gas detection, sensing the vapor a cell vents before it ignites, gives the earliest possible warning and a chance to de-energize and isolate before the event becomes a fire. NFPA 855 calls for gas detection in certain indoor installations for exactly this reason, paired with explosion control such as deflagration venting to relieve the pressure if the vented gas does ignite.

Suppression on lithium is harder than on ordinary combustibles, and it is worth being honest about why. Thermal runaway is self-sustaining and can keep generating heat without outside oxygen, so a clean-agent or gaseous system that would knock down a normal fire may not stop runaway, and a pack can reignite after it appears out. Water, applied in volume, is effective at cooling cells and limiting propagation, which is part of why suppression and cooling strategy for lithium differ from the gaseous systems common elsewhere in a data hall. The right suppression for a given installation is a fire-protection-engineering decision tied to the listing and the AHJ.

Containment is the last layer and the one the rack architecture has the most say over. Where the cells physically sit, in a sidecar away from the airflow, in a module that resists propagation, with spacing and separation from other racks, decides how far a failure can spread before the suppression and the building separation take over. Design the containment with the cells, not around them after the fact.

Ventilation and off-gas

Ventilation belongs to the same problem as detection: the gas a failing cell vents has to be handled, not allowed to pool. A cell in runaway releases a mix of flammable and toxic gases, and in an enclosed rack or room that vapor can reach a concentration that will deflagrate if it finds a spark. Ventilation keeps the concentration below that point, and explosion control gives the pressure somewhere to go if it does ignite anyway.

Worth noting the gas is not the hydrogen story from the old lead-acid room. VRLA strings off-gassed hydrogen on overcharge, and the ventilation rule was about keeping hydrogen below its flammable limit. Lithium vent gas is a different and more complex mixture, and the detection and ventilation are designed for that mix, not for hydrogen. Carrying over the old battery-room assumption is a mistake.

The room design, the airflow, the detection placement, and the explosion control are an energy-storage and fire-protection question governed by NFPA 855 and the AHJ for the specific installation. Size and lay them out to the standard and the manufacturer's instructions, and confirm the approach with the authority having jurisdiction before it is built, because retrofitting ventilation and explosion control into an occupied hall is far harder than designing it in.

Monitoring, life, and end of life

A rack BBU is only as good as its state of health on the day it is needed, and lithium degrades whether or not it is ever discharged. The BMS reports state of charge, state of health, cell temperatures, and fault history, and that telemetry is what a monitoring program watches to catch a fading module before it fails. Across a distributed fleet of hundreds or thousands of shelves, this is not a spot check. It is a continuous data feed, and the value of the BBU's self-reporting is lost if nobody is watching the trend.

Battery life is finite and the warranty is written to a specific service condition. A lithium BBU has a usable life set by cycle count, calendar age, temperature history, and depth of discharge, and the runtime fades as the cells age, which is why the as-installed ride-through margin has to be re-confirmed over time rather than assumed to hold. Plan the replacement against the manufacturer's stated life and the measured state of health, not against a fixed calendar guess.

End of life is a real obligation, not a footnote. Lithium cells are a hazardous waste and a fire risk in storage and transport, so spent BBU modules have to be handled, stored, and disposed of or recycled under the applicable regulations and the manufacturer's instructions, and a pile of removed lithium modules waiting for pickup is its own hazard in the building. Build the disposal path into the replacement plan, and treat removed modules with the same care as installed ones.

Testing the ride-through and the transfer

A ride-through that has never been tested is a specification, not a guarantee. The BBU has to be proven under load: discharge it at the rack's real load and confirm it holds the rack for its rated window, then run the integrated test that drops the utility and proves the BBU carries the rack while the generator starts, transfers, and accepts the load, with no break seen by the IT gear. Do it at design load before live compute arrives, because the failure you want to find is the one that only shows up when you force the outage on purpose.

This is where the timing assumptions get caught. The generator that starts in ten seconds on the commissioning agent's stopwatch, against a BBU window sized tight to that ten seconds, leaves no room for the cold morning. The recharge step that swings the generator when every BBU on the floor reloads at once. The module that was supposed to be the spare in a five-plus-one shelf but was never replaced after the last failure. None of these show in a power-up. They show only in a witnessed failure test against the spec.

Acceptance and the integrated systems test are a commissioning discipline of their own, covered with the rest of the critical-power proving in the commissioning guides referenced from the UPS topology guide. The BBU's part is to be testable: instrument it, document the expected ride-through and transfer sequence, and witness the discharge and the transfer to the generator rather than trusting the datasheet.

What to document

The rack BBU program is only defensible if the record states, shelf by shelf, what is installed, what it is rated to do, and how the fire-safety case is met. A fleet of distributed batteries with no inventory and no state-of-health history is a fleet you cannot prove is still redundant or still safe, and that is exactly the record the AHJ and the insurer will ask for.

Record the chemistry and the cell, the rated ride-through against the generator start time, the BMS state-of-health baseline, the redundancy arrangement at the shelf, the detection, suppression, and containment design with its NFPA 855 and listing basis, and the replacement and disposal plan. Tie each entry to the manufacturer's listed instructions and the AHJ approval, so a later change gets checked against the intent and the code, not against habit. A field tool such as FieldOS keeps the per-shelf inventory, the SOH trend, the test records, and the fire-system checks in one place instead of scattered across spreadsheets.

Common mistakes

• Treating lithium fire safety as an afterthought and ignoring the energy-storage codes, so detection, containment, and separation get bolted on instead of designed in.
• Undersizing the ride-through to the generator start, so the rack drops in the gap the BBU was supposed to bridge.
• Installing no off-gas detection or containment, so the earliest warning of a failing cell is missed and a runaway has nothing to stop its spread.
• Choosing the wrong chemistry for the fire risk, putting a denser, less stable cell in the white space to save space it did not need to save.
• Running BBUs with no BMS state-of-health monitoring, so a degraded module is found at the moment it is needed.
• Treating distributed BBUs like one central UPS, missing that the redundancy, the maintenance, and the hazard are now spread across every rack.

Field checklist

Standards and references

The defining standard for the rack BBU itself is the Open Compute Project Open Rack specification, which standardizes the BBU shelf, the module arrangement, the 48V-class busbar it feeds, and the redundancy at the shelf. Confirm the spec revision against the actual hardware, because the Open Rack versions differ and the ratings move between them. The manufacturer's documentation governs the specific BBU: the cell, the chemistry, the rated discharge and ride-through, the BMS behavior, and the listed installation instructions.

The fire-safety and energy-storage requirements run through NFPA 855 for the installation, and through the UL listings for the equipment: UL 1973 for the battery modules and packs, UL 9540 for the complete energy-storage system, and UL 9540A as the thermal-runaway fire test method that NFPA 855 references. Cite each by the topic it governs and confirm the current designation and adopted edition, because these documents change on their own cycles and the thresholds, including the stored-energy separation figures, are revised.

Above the standards sit two authorities. The basis of design and the owner's project requirements set the redundancy, the ride-through target, and the role the BBU plays relative to any central UPS. The authority having jurisdiction governs the fire-safety design: the separation, the detection, the suppression, the ventilation, and the explosion control for lithium in the white space. Where a standard and the AHJ differ, the AHJ controls the project. Confirm the runtimes against the design, the chemistries against the manufacturer, and the codes against the AHJ before anything is built. The central UPS topology this is measured against is in the UPS topology and redundancy design guide, and the rack DC busbar it feeds is in the rack DC power distribution guide.

Units, terms, and synonyms

The rack-backup vocabulary overlaps with the central UPS world and with stationary energy storage, and the same idea travels under several names across an OCP spec, a manufacturer sheet, and a fire code, so a short glossary keeps the design conversation straight.

A rack BBU is also called an in-rack battery backup unit or rack-level energy storage, and the OCP packaging is the BBU shelf or sidecar. Ride-through is also called autonomy or hold-up time. The central alternative is the central or centralized UPS. LFP is lithium iron phosphate, LiFePO4; NMC is lithium nickel manganese cobalt oxide. Thermal runaway is the self-sustaining cell failure the fire codes are written around, and NFPA 855 is the energy-storage installation standard that, with the UL listings, governs lithium in the white space.

Rack BBU

Battery backup unit: a lithium battery in the rack or a sidecar shelf that carries the load through a blip and bridges to the generator

Ride-through / hold-up

The short window the BBU carries the load, sized to the generator start, not to long runtime

Central UPS

The large uninterruptible power supply in the electrical room, the centralized alternative to distributed rack BBUs

OCP BBU shelf

The Open Compute Project Open Rack battery shelf, modular and redundant, feeding the 48V-class rack busbar

LFP

Lithium iron phosphate (LiFePO4): lower energy density, more thermally stable, trending in data center backup

NMC

Lithium nickel manganese cobalt oxide: higher energy density, less thermally stable than LFP

BMS

Battery management system: monitors and balances the cells, enforces limits, and protects the lithium pack

Thermal runaway

A self-sustaining, self-heating cell failure that vents flammable gas, propagates cell to cell, and resists conventional suppression

NFPA 855

The Standard for the Installation of Stationary Energy Storage Systems, governing location, separation, detection, ventilation, and suppression

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What does N+1 redundancy mean?
• Why are data centers moving to DC power for AI racks?
• What is the difference between cold and hot commissioning of a BESS?
• What kind of batteries do data centers use?
• Why do battery rooms need ventilation?