Data Centre · UPS and backup power
UPS and backup power for data centres and AI servers
A voltage dip of a few milliseconds is enough for servers to reboot and running compute jobs to be lost. The power supply of a data centre therefore rests on three building blocks: grid connection, uninterruptible power supply (UPS) and standby generator. A battery storage system can complement them. This page explains what each one does.
Three building blocks, three jobs
UPS: without interruption
The UPS sits between the grid and the IT. An online UPS (class VFI under IEC 62040-3) supplies the load continuously via rectifier and inverter; if the grid fails, its battery takes over with no transfer time. It bridges minutes, not hours.
Standby generator: for longer outages
A generating set, usually with a diesel engine, starts when the grid fails and takes over the supply after a few seconds. The UPS bridges this time. With a fuel reserve, the generator lasts hours to days.
Battery storage: a complement
A commercial battery shifts solar power, caps load peaks and, with a backup function, can form an island grid during an outage. That usually takes a few seconds. In front of IT loads it therefore does not replace a UPS, but it can complement the generator.
UPS classes: VFD, VI and VFI
The standard IEC 62040-3 (DIN EN IEC 62040-3 in Germany) classifies UPS systems by their dependence on the grid. For servers and GPU systems, in practice only the online UPS is suitable, because it supplies the load without transfer and with clean voltage.
| Class | Transfer | Typical use |
|---|---|---|
| VFD (offline / standby) | up to about 20 ms | Desktop computers, non-critical loads |
| VI (line-interactive) | about 4–10 ms | Smaller server cabinets, network equipment |
| VFI (online / double conversion) | none | Servers, data centres, GPU systems |
Context: the transfer times are typical values from technical literature, not values from the standard. The UPS data sheet and the requirements of the server power supplies apply.
What matters in the design
- Capacity – the UPS must carry the full IT load, for AI servers the maximum system power per data sheet, with headroom for expansion.
- Bridging time – in practice a few minutes until the standby generator takes over stably; without a generator, as long as an orderly shutdown takes.
- Redundancy – two separate supply paths (A/B) to the rack, so that maintenance or failure of one UPS does not shut the servers down. GPU servers have several power supplies for this.
- Cooling – cooling must keep running during a grid outage, otherwise the servers overheat even though they have power. Pumps and fans therefore belong on the secured supply.
- Load steps – GPU clusters change their power draw quickly, for example when training runs start. UPS and generator must cope with these steps; this is part of the design.
- Testing – regular tests under load, battery replacement as specified by the manufacturer and documented maintenance.
Availability classes under EN 50600
The EN 50600 series describes four availability classes for data centres, from a single supply path without redundancy (class 1) to fault-tolerant, multiply separated supply paths (class 4). For large AI installations, NVIDIA recommends data centres to Tier 3 or EN 50600 availability class 3, i.e. with paths that can be maintained during operation.
For a data centre in a mid-sized company, we set the class according to need: what does an hour of downtime cost, which applications must keep running, what may shut down in an orderly way? UPS size, redundancy and whether a standby generator is worthwhile depend on this.
Can a battery storage system replace the UPS?
In front of IT loads, generally not. A commercial battery with a backup function must first detect the outage, disconnect from the grid and build its own island grid. According to manufacturers, this takes a few seconds, typically around 5 to 20. Servers without a UPS would have failed long before.
The combination makes sense: the UPS supplies the IT without interruption, the battery then takes over as backup for a limited time or complements a generator, and in everyday operation shifts solar power and caps load peaks. Whether and for how long a battery can act as backup depends on its capacity and the load.
Context: this is derived from the standard and manufacturer specifications, not a general rule. Some storage systems are designed for UPS operation; their data sheet then applies.
Standby generator: registration and operation
A standby generator is installed in accordance with DIN VDE 0100-551 and coordinated with the grid operator. An important question is whether the generator may run in parallel with the grid, for example for test runs under load or uninterrupted switch-back. If it runs grid-parallel for more than a short time, grid operators treat it as a generating plant under their technical connection conditions, with corresponding requirements for protection and registration.
We plan backup power, UPS, storage and PV as one concept: with a metering concept that cleanly separates all sources and a single point of contact for the grid operator application.
Redundant power supply in the data centre: N, N+1, 2N
Redundancy means there is more equipment than the load needs. If one part fails or is being maintained, the rest takes over. The usual shorthand describes how much reserve is built in. N stands for the number of units that must carry the load on their own, such as UPS modules, generators or power supplies.
- N
- Exactly as many units as the load needs. If one fails, capacity is missing.
- N+1
- One unit more than needed. One may fail or be under maintenance without the load being affected.
- 2N
- The entire supply duplicated in two separate paths. If a whole path fails, the load keeps running on the other.
- N+N
- Term used in manufacturer documentation, for example by NVIDIA for power supplies: two equal groups, each able to carry the load alone.
Rule of thumb: two devices on the same cable, switch or distribution board are not yet a redundant path. What matters is whether a single fault shuts the load down.
Separate supply paths: what the standard and manufacturers say
EN 50600-2-2 distinguishes between primary supply, secondary supply and additional supply. By definition, a secondary supply is “independent of the primary supply” and “available at all times”. An additional supply, such as a standby generator, delivers power when the primary and/or secondary supply fails. According to the summary of the standard, the four availability classes range from a single-path design (class 1) to a fault-tolerant multi-path design (class 4).
Two feeds from the same grid are therefore not automatically two independent paths. Technical authors recommend two standby generators if the independence of two feeds cannot be proven beyond doubt. On this reading, a connection point for a mobile generator is not sufficient; the generators must be permanently installed and ready for use at all times. In practice, availability class 3 is widespread.
The servers themselves are prepared for two paths. According to NVIDIA documentation, a DGX B200 has six power supplies in 5+1 redundancy, a DGX H100 six power supplies in 4+2 redundancy, and a DGX B300 in the power supply version twelve power supplies in N+N redundancy, of which at least six must be working. For the DGX H100, NVIDIA states: if three power supplies lose power, for example due to a fault in the power distribution, the system continues to run at reduced performance. More on power draw and protection: AI server power supply.
Context: for planning, this means splitting a server’s power supplies across two separate power distributions (A and B feed). This is our conclusion from the data sheets. How much performance remains if one side fails is stated in the manual of the respective system.
Transfer times under DIN VDE 0100-560
How long may a load be without power? For safety services, DIN VDE 0100-560:2022-10 (the German version of HD 60364-5-56) divides automatic supplies into classes A to F according to their maximum changeover time. The standard explicitly does not apply to standby power systems, i.e. not to the generator that keeps a data centre running for business reasons. The classes are, however, a useful common language for comparing building blocks.
An online UPS (VFI) supplies power with no transfer time and therefore corresponds to class A. According to the manufacturer, a commercial battery with a backup function takes around 5 to 20 seconds and, depending on the unit, falls into class D to F. For IT loads, only class A is therefore suitable.
| Class | Transfer time |
|---|---|
| A | no break |
| B | ≤ 0.15 s |
| C | ≤ 0.5 s |
| D | ≤ 5 s |
| E | ≤ 15 s |
| F | > 15 s |
Context: classes A to F are documented in the official summary by DIN Media. The figures come from the Austrian implementation of the same harmonisation document (OVE E 8101); we have not checked the wording of the German standard text.
UPS autonomy: how long the battery lasts
Autonomy or bridging time is the time during which the UPS battery carries the full load on its own. We have not found a standard that prescribes a fixed time for data centres. The time follows from the concept: with a standby generator, the battery must bridge the start and stable takeover by the generator, with a reserve. Without a generator, it must last long enough for an orderly shutdown.
Values from practice: a planning guide by Schneider Electric gives 8 minutes to 1 hour or longer. A case study by the battery manufacturer Hoppecke describes a colocation operator with 13 data centres bridging several megawatts for 5 to 10 minutes.
In practice, GPU clusters change their power draw quickly and sharply, for example when a training run starts or ends. The UPS must carry these steps on battery just as on grid power, and the generator must regulate them after takeover. We therefore size the UPS and generator for the maximum system power per data sheet, not for the average.
The battery energy required is load multiplied by time. A worked example without allowances:
- 57.2 kW
- four NVIDIA DGX B200, each up to 14.3 kW per data sheet
- 10 min
- required autonomy until the standby generator takes over (assumption)
- ≈ 9.5 kWh
- usable battery energy, calculated: 57.2 kW × 10/60 h
On top come conversion losses, battery ageing and, if it is connected to the UPS, the cooling. The allowances are given in the data sheets of the UPS and the battery.
Data centre standby generator: standards, grid operator, electricity tax
- Installation – DIN VDE 0100-551:2017-02 covers low-voltage generating sets for three cases: supply as an alternative to the grid, in parallel with the grid and without a grid connection. For feeding into installations entitled to standby power, the pre-standard DIN VDE V 0100-551-2:2023-10 also applies.
- Standby rather than safety power – A generator in a data centre keeps loads running for business reasons. Technically, this is standby power. Safety power supplies, for example in hospitals and places of assembly, are subject to their own standards such as DIN 6280-13.
- Grid-parallel operation – At low voltage, a generator that runs in parallel with the grid for longer than the 100 ms permitted for synchronisation counts as a generating plant with parallel operation; this is stated in N-ERGIE Netz’s notes on its technical connection conditions (TAB), with reference to VDE-AR-N 4105. At medium voltage, VDE-AR-N 4110 has its own chapter on standby generators, including the duration of grid-parallel operation.
- Test runs – Grid operators can limit test runs in their TAB. One example: at medium voltage, Stadtwerke Bayreuth allow at most one start per month with a test run of at most 60 minutes at 50 % of rated load. We clarify your grid operator’s rules before planning.
- Electricity tax – Electricity from installations that serve as a temporary supply when the usual supply fails or is disrupted is exempt from electricity tax under section 9 (1) no. 4 of the Electricity Tax Act (StromStG). The act calls them emergency power installations (Notstromanlagen).
Battery storage as backup power: the sequence
A commercial battery with a backup function can take over part of the supply, but only after a changeover pause. A design guide by the storage manufacturer INTILION describes the sequence as follows:
01
Grid-parallel operation
In everyday operation the battery runs on the grid. It shifts solar power and caps load peaks.
02
Changeover, approx. 5–20 s
The battery detects the outage, disconnects the installation from the grid, interlocks the disconnection and earths the star point. Only then does it build the island grid.
03
Grid-forming operation
The battery supplies the connected loads until the grid returns or the reserved capacity is used up.
04
Switch-back
When the grid returns, the system switches back with another interruption of approx. 5–20 s or resynchronises without interruption; this can take up to 180 s.
Registration: at low voltage, data sheet E.3 of VDE-AR-N 4105 asks whether the battery is island-grid-forming under VDE-AR-E 2510-2 and requires all-pole disconnection from the public grid in backup mode. At medium voltage, VDE-AR-N 4110 contains its own chapters on storage and island operation.
Sizing battery storage as backup power
It starts with two questions: which loads are on the backup supply, and how long should they run? Two figures follow from this. The inverter power must carry the load. The capacity that must remain free for emergencies is load multiplied by time.
In everyday operation, this reserve is not available for solar power and peak shaving. How much to reserve is, in practice, a trade-off between resilience and yield. A worked example without allowances:
- 80 kW
- load on backup power, for example cooling and building services of a server room (assumption)
- 2 h
- required bridging time (assumption)
- 160 kWh
- usable capacity that must remain free in the battery for backup
The UPS stays in front of the IT loads. Manufacturers of data centre batteries such as Vertiv also present their systems paired with a UPS, not as a replacement. How a battery works in operation: Commercial battery storage.
Testing and maintenance
Backup power that fails to take over in an emergency is useless. In practice, regular tests with records are therefore part of every concept. What we take on during operation: Maintenance and operations.
01
UPS
Usual practice is tests under load and battery replacement as specified by the manufacturer, before capacity falls below the planned autonomy.
02
Standby generator
Test runs according to the manufacturer and the grid operator’s TAB, usually with checks of fuel reserve and starting system.
03
Battery storage
Test the backup function itself, not just everyday grid operation: does the system switch over, carry the load and return cleanly?
04
Interaction
Test the whole chain at fixed intervals: grid off, UPS takes over, generator or battery takes over, switch-back. Every test is documented.
Frequently asked questions on UPS and backup power
What is the difference between a UPS and backup power?
A UPS supplies the connected equipment without interruption from its battery, typically for a few minutes. Backup power in the narrower sense comes from a standby generator, usually a diesel set, which starts within a few seconds and then runs for hours to days. In a data centre the two work together.
Which UPS does a data centre need?
For servers and GPU systems, generally an online UPS of class VFI under IEC 62040-3. It supplies the load continuously via rectifier and inverter and switches to battery without interruption when the grid fails.
How long does a UPS need to bridge?
With a standby generator, a few minutes until the generator has started and takes over the load stably, plus a reserve. Without a generator, as long as an orderly shutdown of the systems takes. We set the time according to your applications and the availability class.
Can a battery storage system serve as a UPS?
A typical commercial battery cannot, because it needs a few seconds to switch to island operation. It can, however, step in as backup after the UPS, complement a generator and in everyday operation shift solar power and cap load peaks. Systems designed specifically for UPS operation are an exception.
Which availability class does an AI data centre need?
That depends on what an outage costs. For large AI installations, NVIDIA recommends data centres to Tier 3 or EN 50600 availability class 3. For a data centre in a mid-sized company, we set the class according to your applications.
Does the cooling also need to be on the UPS?
It must keep running during a grid outage, otherwise the servers overheat despite having power. Whether pumps and fans go on the UPS or on the generator depends on the heat load and how long the servers can manage without cooling.
Does a standby generator have to be registered with the grid operator?
Yes, it is coordinated with the grid operator. This is particularly important if the generator is to run in parallel with the grid, for example for test runs under load; the technical connection requirements for a generating plant then apply.
What do N+1 and 2N mean for UPS and backup power?
N is the number of units that must carry the load on their own. N+1 means one more unit, so that one can fail or be maintained. 2N means the entire supply duplicated in two separate paths, so that a complete path may fail.
Is electricity from a standby generator subject to electricity tax?
No, provided the installation serves as a temporary supply when the usual supply fails or is disrupted. The electricity is then exempt under section 9 (1) no. 4 of the Electricity Tax Act (StromStG); the act calls such installations emergency power installations.
How quickly does a battery storage system take over during a power cut?
According to the manufacturer, a commercial battery with a backup function needs around 5 to 20 seconds to build the island grid. It switches back to the grid with another interruption of similar length, or without interruption by resynchronising, which can take up to 180 seconds. A UPS is therefore still needed in front of IT loads.
Read more
Power draw of AI servers and racks: AI server power supply. Data centre on your own premises: On-prem AI data centre. Storage in operation: Commercial battery storage. Connection with the grid operator: Grid connection & registration.
Want to protect your servers against grid outages? Let’s spend 15 minutes on UPS, backup power and storage.