Data centre design reference

Generic reference designs for power, cooling, controls and operations

Power systems

How a data-centre power system is structured: utility feed, backup generation, UPS, distribution and protection.

On this page
  1. The power chain from grid to rack
  2. Site power
  3. The 2N power block
  4. Distribution to the racks
  5. What design models often leave out
  6. Reference parameters
  7. Sources

The power chain from grid to rack

A data-centre power system is a chain: grid connection, transformation to medium and low voltage, backup generation, UPS and batteries, distribution, and the rack's own power supplies. Availability is decided by how many independent paths the chain has, and by which links can be maintained without dropping the load.

Two frameworks name the levels. EN 50600-2-2 applies the four availability classes of EN 50600-1 to power supply and distribution,1 and the Uptime Institute's Tier IV adds fault tolerance to the concurrently maintainable Tier III topology.2 The reference designs below are built so that one item out for maintenance, or one fault while the grid is available, never drops a rack; the site power section says what fault tolerance on generators adds.

Design position. State the availability target of each power block as one EN 50600 Availability Class or one Uptime Tier, and design that block to the class it states. The two frameworks are analogous, not equivalent, so name one.3

Site power

Reference site power: two grid supplies, split medium-voltage switchboard, N+1 generator plant and N power blocks Two independent grid supplies feed two halves of a medium-voltage switchboard through their own transformers. A medium-voltage generator plant with one spare set backs up both halves from a two-section bus, one section per half. Each power block takes one feeder from each half, which gives the 2N block its two independent power trains. An optional battery energy storage system on the medium-voltage board can smooth fast load swings and support the grid. Generator plant, N+1 MV A MV B Generator bus A Generator bus B Grid supply A high voltage HV breaker Main transformer A HV / MV, N+1 sized MV incomer A Grid supply B high voltage HV breaker Main transformer B HV / MV, N+1 sized MV incomer B Bus coupler normally open G~ Gen 1 G~ Gen 2 G~ Gen n G~ Gen n+1 spare Gen tie A Gen tie B Spare tie A Spare tie B Bus-section breaker normally closed BESS optional Feeder Feeder Feeder Feeder Power block 1 2N, see block drawing Power block N 2N, see block drawing 1. Each power block takes one feeder from switchboard A and one from switchboard B, so its two power trains stay independent back to the grid. 2. The generator plant carries the whole site load with one set spare. Its sets are split between two bus sections, one tied to each half of the switchboard and joined by a normally closed bus-section breaker. A generator and storage per block (at low voltage) is the common alternative. 3. A fault on one generator section opens the bus-section breaker, so the other half stays energised. With the grid out, the surviving section cannot carry the whole site load unless each section is sized for it (the fault- tolerant variant). The spare set connects to either section through two interlocked breakers. 4. The incomers, generator ties and bus coupler are interlocked: a half's incomer and generator tie are never closed together, and the bus coupler and both incomers are never closed together, so the two grid supplies are never paralleled, through the bus coupler or through the generator bus. 5. Not drawn: earthing, protection relays and their communications, metering, auxiliary supplies, the grid connection's own switchgear and the network company's equipment.
Reference site power: two grid supplies, split medium-voltage switchboard, N+1 generator plant and N power blocks

Two grid supplies, each through its own main transformer sized to carry the site alone, feed the two halves of a medium-voltage switchboard. The bus coupler between them stays open in normal operation, so a fault on one half cannot reach the other. Each power block takes one feeder from each half.

The generator plant sits on its own bus, built in two sections: one section ties to each half of the medium-voltage switchboard, the sets are split between them, and a normally closed bus-section breaker joins the sections. One spare set can connect to either section. The incomers, generator ties and bus coupler are interlocked so that the two grid supplies are never paralleled, through the bus coupler or through the generator bus, and a fault on one generator section leaves the other half energised. With the grid out, though, the surviving section holds only part of an N+1 plant while the 2N blocks move their whole load onto it, so the hall must shed load within the battery autonomy; for this topology to ride through that fault on generators, each section must be sized for the whole site load. In a generator maker's summary of ISO 8528-5, G3 is the most stringent of the standard transient classes for generating sets: on a load step the voltage may fall by up to 15 % and rise by up to 20 %, and a diesel set's frequency may move by up to 7 % down and 10 % up;4 AI load steps make that class, and a test with a realistic load profile, worth specifying. A major engine maker has approved its standby generator sets for paraffinic fuels to EN 15940, including HVO.5

The common alternative to a medium-voltage generator plant is a generator, and possibly a battery storage system, per power block at low voltage. That keeps each block self-contained at the cost of more, smaller sets.

The 2N power block

Reference power block, 2N: two independent power trains to dual-corded racks One of N identical power blocks. Each power train has its own medium-voltage feeder (from half A or half B of the site switchboard, which the generator plant backs up), transformer, LV switchboard, UPS with battery and UPS output board, and feeds its own overhead busway. Every rack takes one tap-off from each busway. Critical cooling is fed from both UPS output boards. Power train A Power train B Rack row (one of several) LV switchboard A UPS output A LV switchboard B UPS output B Busway A Busway B MV feed A from MV switchboard A MV breaker Transformer A MV / 0.4 kV, P_b + critical cooling LV incomer UPS input AC/DCDC/AC UPS A modular, P_b + critical cooling Battery A 5-10 min UPS output Critical cooling A CDU and pump drives, feed A Busway A feeder MV feed B from MV switchboard B MV breaker Transformer B MV / 0.4 kV, P_b + critical cooling LV incomer UPS input AC/DCDC/AC UPS B modular, P_b + critical cooling Battery B 5-10 min UPS output Critical cooling B CDU and pump drives, feed B Busway B feeder Tap-off Tap-off Tap-off Rack 1 dual-corded Rack 2 dual-corded Rack n dual-corded Tap-off Tap-off Tap-off 1. 2N: each power train carries the whole block load, IT and critical cooling, so any one component can fail or be taken out for maintenance without dropping a rack. 2. Every rack is dual-corded: one tap-off from busway A, one from busway B. P_b is the block's IT design load; set it, and the rack and tap-off ratings, with the sizing calculator. 3. Not drawn: earthing and bonding, surge protection, metering at every incomer and outgoing way, generator synchronising and load-shedding controls, and maintenance bypasses.
Reference power block, 2N: two independent power trains to dual-corded racks

The hall is built from identical power blocks. Each block has two independent power trains, A and B, and each train can carry the whole block alone: medium-voltage feeder (backed by the site generator plant), transformer, LV switchboard, modular double-conversion UPS with its battery, UPS output board and an overhead busway. Every rack takes one tap-off from each busway.

  • Block size is the main parameter. It sets the transformer, UPS, busway and generator ratings, and how much of the hall one fault can touch.
  • Critical cooling, the CDU pumps and facility-water pumps, is fed from both UPS output boards, because liquid-cooled racks have almost no thermal buffer.
  • Battery autonomy only has to bridge the generator start and load transfer, plus margin; it is a parameter in the sizing calculator.

For AI training halls there is a real alternative to 2N UPS: a shared reserve system, or no UPS on the IT load at all. Both are compared on Power for AI halls.

Distribution to the racks

Overhead busway with plug-in tap-off boxes suits AI halls: racks can be added or changed without new cable runs, and each rack's feed is a box on a bar. Rack power now sets the tap-off size. One current rack-scale GPU system is specified at 135 kW, with up to 155 kW at peak;6 at 400 V three-phase, 155 kW draws about 224 A, within a 400 A tap-off at its usual 80 % loading limit (about 220 kW). A 600 kW rack, announced for a later generation, would draw about 870 A at 400 V, which needs larger tap-offs, several per rack, or DC distribution. Treat 600 kW as an upper planning case: a 2026 write-up says the design shown that year appeared to have lost half its GPU density.7

Design position. Install the busway for the final hall and fit tap-off boxes as racks arrive. In the large design models we studied, most tap-off positions were left open for later racks. The busway rating, the open positions and the block size should come from the same growth plan.

What design models often leave out

Detailed 3D design models of data centres usually show where the equipment is and how it repeats. They often do not show the electrical design itself: transformers and medium-voltage switchgear may be missing, transfer switches and metering absent, and every panel's load and breaker rating a placeholder. That is not a fault in the model; the electrical design is a separate deliverable. It is the part that decides whether the topology actually delivers its availability:

  • load lists and demand per board, with growth;
  • short-circuit levels, protection settings and selectivity on every path, in normal operation and on generators;
  • arc-flash studies and labels;
  • earthing and bonding, and the interaction of UPS, generators and storage with it;
  • metering for energy reporting and for the operators.

Reference parameters

ParameterReference valueBasis
Rack power, current GPU systemsabout 135 kW, 155 kW peakPlatform specification6
Rack power, announced next generationsup to 600 kW, as an upper planning casePlatform roadmap8; a later design appeared to have about half the GPU density7
Block loada parameter, typically a few MWDesign assumption: limits what one fault can touch
UPS redundancy2N per blockDesign position; alternatives on the AI power page
UPS battery autonomy5 to 10 minutesDesign assumption: generator start, transfer and margin
Generator plantN+1, ISO 8528-5 class G3Design choice; class G3 as summarised by a generator maker's note on ISO 8528-54
DistributionOverhead busway, dual-corded racksDesign position

Sources

  1. Review of standardisation activities: Energy Management and Environmental Viability of Data Centres (executive summary of the 8th-edition report). CEN/CENELEC/ETSI Coordination Group on Green Data Centres, 2021-10-06. Read 2026-10-18.https://www.cencenelec.eu/media/CEN-CENELEC/AreasOfWork/CEN%20sectors/Digital%20Society/Green%20Data%20Centres/brochuredatacentrestandardizationedition8_2021.pdf
  2. Tiers (Uptime Institute tier classification summary). Uptime Institute. Read 2026-10-18.https://uptimeinstitute.com/tiers
  3. ISO/IEC 22237 & EN 50600: Data Centre Design Standard. GreenCalculus, 2026-06-19. Read 2026-10-18.https://www.greencalculus.com/standards/iso-iec-22237-en-50600-data-centre-design/
  4. ISO 8528-5 and Generator Transient Performance (Kohler Energy, Power Systems Topics 143). Kohler Energy, 2024-06. Read 2026-10-18.https://techcomm.kohler.com/techcomm/pdf/ISO%208528-5%20and%20Generator%20Transient%20Performance_WP.pdf
  5. Cummins high-horsepower diesel generator sets approved for use with HVO. biobased-diesel.com, 2021-12-10. Read 2026-10-18.https://www.biobased-diesel.com/post/cummins-high-horsepower-diesel-generator-sets-approved-for-use-with-hvo
  6. Lenovo NVIDIA GB300 NVL72 Rack Scale AI Product Guide. Lenovo Press, 2026-08-30. Read 2026-10-18.https://lenovopress.lenovo.com/lp2357-lenovo-nvidia-gb300-nvl72-rack-scale-ai
  7. NVIDIA Kyber. Glenn K. Lockwood (personal site), 2026-04-02. Read 2026-10-18.https://www.glennklockwood.com/garden/kyber
  8. Nvidia's Vera Rubin CPU, GPUs chart course for 600kW racks. The Register, 2025-03-19. Read 2026-10-18.https://www.theregister.com/2025/03/19/nvidia_charts_course_for_600kw/