Data Centre · AI server power supply
Power supply for AI servers and GPU racks
A single GPU server with eight graphics processors draws as much power as an entire office floor. If you want to run AI on your own premises, you need to check early whether the grid connection, distribution, UPS and cooling can carry it. This page shows the power levels to plan for and how we build the supply.
How much power an AI server needs
What counts is the maximum system power from the manufacturer’s data sheet, not the average load. During training, GPUs run close to their maximum for hours; supply and cooling must be designed for that. Some examples from NVIDIA’s specifications:
| System | Power per manufacturer | Context |
|---|---|---|
| DGX Spark | 240 W power supply | Desktop device, standard socket |
| DGX Station (GB300) | 1.6 kW | Workstation, dedicated circuit recommended |
| RTX PRO 6000 Blackwell Server Edition | up to 600 W per GPU | Server with several cards, a few kilowatts depending on configuration |
| DGX H100 / H200 | 10.2 kW max. | 8-GPU server, air-cooled, six 200–240 V power supplies |
| DGX B200 | approx. 14.3 kW max. | 8-GPU server, air-cooled, six 200–240 V power supplies |
| DGX GB200 / GB300 NVL72 | approx. 120 kW per rack | Full rack, liquid cooling of the chips required |
Source: NVIDIA data sheets and user guides, retrieved 24 October 2026. Servers from other manufacturers with the same GPUs are in the same range; the data sheet of the specific system always applies.
From server to rack: kilowatts per cabinet
A conventional server rack in a corporate data centre typically draws 10 to 30 kW. Two to three GPU servers of the DGX B200 class in one cabinet already reach around 29 to 43 kW, a GB200-class rack around 120 kW. In between lies a limit that matters for planning: according to cooling manufacturers, air can remove roughly 20 to 30 kW per rack; above that you need water-cooled rear-door heat exchangers or direct-to-chip liquid cooling.
For mid-sized companies this means: one or two GPU servers often still fit into the existing server room once circuits, UPS and air conditioning are upgraded. From several racks or a round-the-clock load of a few tens of kilowatts, a dedicated room or a container data centre on site becomes the more economical option.
- 10–30 kW
- conventional rack, common in practice, air cooling
- ≈ 29–43 kW
- two to three 8-GPU servers of the DGX B200 class per rack, beyond conventional room air cooling, in practice often with rear-door heat exchangers
- ≈ 120 kW
- GB200 NVL72-class rack, liquid cooling of the chips required
Electrical design: protection, distribution, redundancy
Power supplies and redundancy
A DGX B200 has six power supplies of 3.3 kW each at 200–240 V; five are enough to run, one is a spare. The DGX H100 uses four plus two. It makes sense to split the power supplies across two separate feeds (A/B) so that the failure of one distribution path does not shut the server down.
Circuits and sub-distribution
GPU servers do not run from a household socket. They need rack power distribution units (PDUs) on three-phase outlets, correctly sized cables and protective devices, and a sub-distribution board that carries the continuous load. The detailed design is done by the electrical engineer based on the manufacturer’s specifications.
Grid connection and main distribution
Using your load profile, we check whether the grid connection, transformer and main distribution can carry the additional continuous load alongside ongoing operations. If not, we apply to the grid operator for a capacity increase or a dedicated connection.
UPS and backup power
An online UPS keeps the servers running without interruption during grid disturbances; a standby generator takes over for longer outages. A battery storage system complements both but does not replace a UPS unless it is designed for that purpose.
Cooling as an electrical load
Cooling itself needs power: air conditioners, pumps and dry coolers run alongside. When sizing the connection, what counts is IT load plus cooling and losses, not just the servers.
Metering and monitoring
Separate meters for IT, cooling and the building show where the power goes. This is the basis for sizing PV and storage and, for larger installations, for the indicators required by the German Energy Efficiency Act (EnEfG).
More on uninterruptible supply: UPS and backup power for data centres.
Cooling and waste heat
Almost every kilowatt-hour an AI server draws becomes heat. A 14 kW server heats like several large radiators, without a break. Air-cooled servers release the heat into the room air; air conditioning and airflow must cope with it. Liquid-cooled systems remove the heat via a water circuit, at significantly higher temperatures.
This waste heat can be used, for example via a heat pump for halls, offices or process heat. Whether this pays off depends on the temperature level and the heat demand on site. For data centres from 300 kW of non-redundant rated connection capacity, the German Energy Efficiency Act (EnEfG) prescribes efficiency and waste-heat indicators; smaller installations in mid-sized companies are often below that.
PV and battery storage for the continuous load
An AI server draws almost constant power day and night. For solar power that is an ideal consumer: as long as the PV output stays below the continuous load, every kilowatt-hour from the roof is used on site. A battery storage system shifts the midday surplus into the evening and night and caps load peaks, for example when large compute jobs start.
For scale: in Berlin and Brandenburg, one kilowatt of PV capacity typically generates around 950 to 1,060 kilowatt-hours a year, depending on orientation. A server with a 14 kW continuous load needs about 123 MWh a year. A PV system covers part of this; the grid remains the base supply.
- 14 kW × 8,760 h ≈ 123 MWh
- annual demand of an 8-GPU server at full load, excluding cooling
- 1 kWp ≈ 950–1,060 kWh
- annual PV yield in Berlin/Brandenburg, depending on orientation (PVGIS)
- 60 kW × 8,760 h ≈ 525 MWh
- annual demand of a container data centre with a 60 kW continuous load
Worked example and process for a data centre on your own premises: On-prem data centre with PV and battery storage.
Timeline: power first, then hardware
In practice, GPU servers are often delivered faster than the capacity at the grid connection. A capacity increase or a new medium-voltage connection takes many months with the grid operator. According to manufacturers, medium-voltage transformers currently have delivery times of around six to twelve months, plus switchgear. That is why we settle the connection question before hardware is ordered.
01
Requirements
Which systems, how many, which expansion stages? This gives the continuous load plus cooling.
02
Existing site
Load profile, grid connection, main distribution, server room and installation space.
03
Energy concept
Connection, distribution, UPS, backup power, PV and storage for your site.
04
Application
Capacity increase or grid connection with the grid operator, through to approval.
05
Installation
Electrical integration, testing and commissioning, then the IT moves in.
Frequently asked questions on AI server power supply
How much power does an AI server use?
According to manufacturers, a server with eight current-generation GPUs draws up to around 10 to 15 kW; a DGX B200, for example, about 14.3 kW. At full load around the clock, that is a good 120 MWh a year, excluding cooling. Smaller systems with individual cards draw a few hundred watts to a few kilowatts.
How much power does a GB200 NVL72 rack need?
NVIDIA specifies around 120 kW for a DGX GB200 or GB300 NVL72-class rack. The chips are liquid-cooled; such a rack needs dedicated power and cooling infrastructure and for most mid-sized companies is a container or data centre project rather than a server room installation.
Is a standard socket enough for a GPU server?
For desktop devices drawing a few hundred watts, yes. Not for servers with eight GPUs: they have several 3.3 kW power supplies at 200 to 240 V and need rack PDUs on three-phase outlets, suitable cables and protection, and a UPS.
Can an AI server go into the existing server room?
One or two servers often can, once circuits, UPS and air conditioning are upgraded. Heat removal and the continuous load at the grid connection are decisive. From several racks or a few tens of kilowatts around the clock, a dedicated room or a container on site is usually the better solution.
When is a medium-voltage connection needed?
The grid operator sets this in its technical connection conditions. In practice, a medium-voltage connection with a dedicated transformer station becomes common from a few hundred kilowatts of connection capacity. We check this for your site and submit the application.
Can a PV system supply an AI server?
Partly. A PV system generates during the day and mainly in summer; the server runs around the clock. With a continuous load, the solar power is almost entirely used on site and lowers electricity costs, but the base supply remains with the grid. A battery storage system shifts surpluses into the evening.
Does the German Energy Efficiency Act apply to our AI server?
For most installations in mid-sized companies, no. The obligations for data centres currently apply from 300 kW of non-redundant electrical rated connection capacity. A change to this threshold is in the legislative process; the Bundestag has not yet decided.
Does StromWert also supply the servers?
That depends on the project. You can buy or lease servers and data centre containers from the manufacturer or an IT service provider; we are happy to discuss how we set up procurement. We work vendor-neutrally and in any case take care of the grid connection, energy supply with PV and storage, electrical integration and coordination through to commissioning.
Read more
Data centre on your own premises: On-prem AI data centre. Uninterruptible supply: UPS and backup power. Larger data centres up to high voltage: AI Data Center Power. Capping load peaks: Peak shaving.
Planning GPU servers on site? Let’s spend 15 minutes on power, connection and cooling.