What is a custom PDU for AI hardware?

A custom PDU for AI hardware is a rack Power Distribution Unit specifically configured to power GPU servers, AI accelerators, high-density compute, storage, and networking equipment within an AI or HPC rack.

Instead of adapting the rack design to the limitations of a standard PDU, the PDU is tailored to the infrastructure.

This may relate to, among other things:

• single- or three-phase power supply;
• 10A, 16A, 32A or 63A input;
• specific input plugs and cables;
• customized cable lengths and cable entry;
• type, number and position of outlets;
• phase-balanced outlet layouts;
• branch security;
• energy metering at input, branch, and outlet levels;
• remote outlet switching;
• environmental monitoring;
• communication and network architecture;
• integration with DCIM, BMS and energy management software.

Schleifenbauer PDU 5.0 supports single- and three-phase configurations and current ratings from 10A up to and including 63A, with configurable energy metering at input, branch, and outlet levels and optional remote switching.

The basic principle is simple:

Design the rack PDU around the AI hardware, not the AI hardware around the PDU.

Why does AI hardware require a different approach to rack power?

AI infrastructure increases the power density in the rack.

Traditional enterprise racks often contain various systems that function relatively independently of each other. Modern AI racks, on the other hand, can contain highly concentrated GPU compute that operates as a single cohesive system.

As a result, the architecture of the rack power infrastructure is becoming increasingly important.

NVIDIA documents, for example, designed rack capacities of 120 kW for GB200 NVL72 and 135 kW for GB300 NVL72. At the same time, the sector is already working on new power architectures for future AI factories with even higher power densities.

That means not that every AI rack requires a 120 kW PDU.

It primarily means that “AI rack” no longer equates to a single standard power requirement.

The correct PDU is determined by the hardware and power architecture that are actually deployed.

AI rack power starts with the hardware specification.

The first question when specifying a PDU for AI hardware should therefore not be:

Which PDU do we normally use?

But:

What does this rack actually need?

Important starting points include the maximum and expected energy consumption of the IT equipment, the number of power supplies, the redundancy architecture, the required supply voltage, the number of feeds, available upstream capacity, connector types, and the physical rack layout.

Based on that, the PDU can be configured.

This is becoming increasingly important as AI infrastructure evolves rapidly. A fixed PDU catalog designed around yesterday's rack density does not automatically offer the best solution for tomorrow's GPU infrastructure.

32A or 63A PDU for AI hardware?

There is no universal answer to that.

A 32A PDU may be suitable for one AI rack architecture, while another environment requires multiple 32A feeds, three-phase 63A PDUs, or a completely different distribution architecture.

The right question is therefore not whether 32A or 63A is 'better' for AI.

It concerns the combination of:

voltage × current × number of phases × number of feeds × redundancy

and whether it offers sufficient usable capacity for the specific rack.

Schleifenbauer can configure PDU 5.0 in 10A, 16A, 32A and 63A, including three-phase versions.

A three-phase 63A PDU at 400 V, for example, can be a maximum of approximately 43.5 kVA to deliver.

Multiple feeds can subsequently be deployed to realize the required rack capacity and redundancy architecture.

Why three-phase power is important for high-density AI racks

As rack capacity increases, three-phase power distribution becomes increasingly relevant.

Instead of concentrating the load on a single phase, the power can be distributed over:

L1 → L2 → L3

The outlet configuration of the PDU can then be designed so that connected equipment is distributed across the available phases.

For AI infrastructure, however, phase balancing is not just something that needs to be calculated during the design.

You want it too can measure during operation.

With an intelligent PDU, the operator can see how the actual load develops across the phases, instead of relying solely on the original design calculation.

AI workloads are dynamic, not static.

Nameplate power and average energy consumption do not tell the whole story.

AI computing can cause significant changes in electrical load when systems switch between training, inference, idle states, and other operating conditions.

As a result, headroom and real-time measurement more important.

A rack may appear to have sufficient capacity based on average consumption, while individual phases, branches, or feeds approach their operational limits during peak load.

For a data center operator, the question therefore becomes:

How much capacity do we have available — and where is that capacity actually available?

Measure AI power where it is actually consumed

The closer the measurement takes place to the IT equipment, the more specific the data becomes.

An intelligent PDU can provide insight into various levels:

Rack → PDU → Phase → Branch → Outlet → IT device

Depending on the selected Schleifenbauer PDU functionality, measurements can be taken at the input, branch, and outlet levels. PDU 5.0 supports power monitoring with a specified accuracy of 0.5% at the input and outlet levels.

For an AI rack, this can answer practical questions such as:

How much power does the rack use?

How is the tax distributed across L1, L2, and L3?

Which sector is approaching its capacity limit?

How much energy does an individually connected device use?

How much headroom is still available before additional AI hardware can be installed?

That is significantly more useful than just knowing the total energy consumption of the data center.

Outlet-level monitoring for AI and GPU infrastructure

Outlet-level monitoring adds an extra level of granularity.

Instead of only knowing how much energy enters the PDU, the operator can monitor consumption more closely at individual connected loads.

This is relevant, for example, in racks with a combination of:

GPU computing systems, CPU servers, network switches, storage, management equipment, and other supporting IT hardware.

The PDU thus effectively forms the measurement boundary between rack power distribution and the individual IT loads.

In Managed PDU configurations, outlet-level monitoring can also be combined with remote outlet switching.

Custom outlets for AI hardware

Power capacity alone does not determine whether a PDU is suitable for an AI rack.

The physical connections must also fit.

Different IT systems may require different plug types and outlet configurations. Moreover, in high-density racks, the available physical space for power distribution is increasingly under pressure.

Therefore, at Schleifenbauer, they can The type, number, position, and layout of the outlets are configured project-specifically..

One of the available solutions is the CX Combination Outlet C13/C15/C19/C21.

A single CX outlet supports C14, C16, C20, and C22 plugs, while IEC Lock helps prevent accidental disconnection. Custom outlet combinations and load-balancing layouts can also be specified.

For AI racks with various types of equipment, this can significantly simplify the physical power design.

Redundancy is essential because an AI rack functions as a system.

A high-density AI rack should not necessarily be considered a collection of independent servers.

Within clustered AI environments, the failure of a single component can affect a much larger workload.

The precise redundancy architecture is determined by the AI platform used.

That could mean, for example:

A/B feeds

or

multiple independent power paths

or another architecture prescribed by the hardware supplier.

The design of the rack PDU must support this redundancy architecture, not restrict it.

Do not confuse installed PDU capacity with usable AI capacity

This distinction becomes increasingly important at higher power densities.

A rack can have multiple power feeds, but their combined nominal capacity is not automatically the power that can be safely allocated to IT equipment.

Redundancy, breaker ratings, phase load, upstream infrastructure, and required failover headroom all influence the actual usable capacity.

In AI deployments, capacity management must therefore answer two different questions:

What has been installed?

and

What can we safely use?

Power quality becomes more important as rack density increases.

AI power management is not just about kW and kWh.

With high-density electronic loads, the properties of the electrical load are also becoming increasingly relevant.

Schleifenbauer PDU 5.0 and EnerTree can Power Quality Monitoring offer, including measurements such as Total Harmonic Distortion (THD), Crest Factor, and voltage/current peaks.

This creates context that regular energy monitoring does not provide.

For an AI infrastructure operator, energy monitoring answers the question:

How much energy is consumed?

Power Quality Monitoring helps provide insight into:

What happens electrically while that power is being consumed?

The two complement each other.

From custom AI PDU to Data Centre Energy Management

One intelligent PDU can provide very detailed information about a single rack.

But an AI data center can contain hundreds or thousands of PDUs.

Then the next challenge arises:

How do you turn thousands of electrical measurements into usable operational information?

That comes EnerTree in pictures.

With Schleifenbauer PDU 5.0, the desired PDU functionality is independent of the way the intelligent PDU environment is managed.

The communication architecture can be chosen independently via a Controller Module, Gateway Module or Daisy Chain Module.

Controller Module + EnerTree Lite

The Controller Module offers a conventional intelligent PDU architecture.

EnerTree Lite Runs embedded on the Controller Module and provides a local web interface with direct access to PDU measurements.

In addition, the Controller can be directly integrated with an existing DCIM environment.

EnerTree Lite supports environments up to 100 PDUs per IP address.

For smaller AI installations or data centers that already have an existing management platform, this can provide the desired architecture.

Gateway Module + EnerTree Platform

For larger AI deployments, the Gateway architecture moves the intelligence from the individual PDU to EnerTree Platform.

EnerTree Platform runs as a virtual machine outside the PDU.

The PDU functions as a real-time measurement and control point, while processing, analysis, configuration, and management take place centrally.

One EnerTree Platform environment can scale up to 10,000 PDUs.

For large GPU clusters and AI data centers, this offers a fundamentally different operating model than individually managing thousands of intelligent rack PDUs.

EnerTree offers, among other things, real-time energy monitoring, alerts, historical analysis, reporting, hierarchical infrastructure view, environmental monitoring, and PUE calculation.

Daisy Chain for scalable PDU connectivity

Not every PDU in an AI rack or data center row needs to have its own Ethernet connection.

Multiple PDUs can be connected behind a Controller or Gateway via Daisy Chain.

Within one ring, a maximum of 100 PDUs via a single IP address be managed.

The dedicated Daisy Chain Module functions as a cost-effective data relay without its own Ethernet port. As a result, fewer network connections are required in the rack, while simultaneously limiting the Ethernet attack surface at the rack level.

This can become a significant advantage in large-scale AI deployments.

From a single GPU outlet to full insight into the AI data center

The value of this architecture becomes clear when we look at the entire hierarchy:

AI device
↓
Outlet
↓
Branch
↓
Phase
↓
PDU
↓
Rack
↓
Row
↓
Data center

At the bottom of this hierarchy, the PDU distributes electricity to the AI hardware.

At the same time, the measurement data moves in the opposite direction:

Device → Outlet → Branch → Phase → PDU → Rack → EnerTree

The physical power infrastructure thereby becomes a source of structured energy data.

That is the difference between simply supply power to AI hardware and actually understand what is happening electrically in the AI rack.

Integrating AI rack power with DCIM and BMS

AI power data must not become a new data silo.

EnerTree Platform and EnerTree Lite support integration with external data center systems via interfaces and protocols such as HTTP/HTTPS, REST API, MODBUS/TCP, SNMP v1/v2c/v3, IPv4/IPv6, SMTP, Syslog and Command Line Interface. Additionally, data can be exported to databases such as MS SQL, MySQL, and MariaDB.

This allows electrical data at the rack level to become part of a broader DCIM, BMS, or data analysis environment.

Custom AI PDUs must also be physically configurable.

AI infrastructure is developing rapidly.

A PDU with the correct electrical specifications that does not physically fit properly in the rack is still the wrong PDU.

Therefore, Schleifenbauer produces PDUs build-to-order.

Configurable properties include, among others, cable length, connector type, outlet layout and position, measurement functionality, position of the communication module, identification and labeling, assembly, and other mechanical requirements.

Depending on the project, the PDU 5.0 can also be equipped with various security options, residual current sensing, environmental sensors, and additional functionalities.

This is particularly relevant for OEMs, system integrators, data center designers, and operators implementing AI hardware, where the rack is increasingly designed as a single integrated system.

What should you pay attention to when specifying a custom PDU for AI hardware?

Don't start with the PDU. Start with the rack.

A good AI PDU specification must take at least the following points into account:

Element	What needs to be determined?
AI hardware	GPU/server platform and configuration
Maximum rack capacity	Expected load and design power
Power supply	Voltage and frequency
Phases	Single- or three-phase
Current intensity	16A, 32A, 63A or project specific
Redundancy	A/B, N+1, 2N or hardware-specific architecture
Number of feeds	Required independent power paths
Connectors	Input plug and plug types of IT equipment
Outlets	Type, quantity, position and grouping
Phase distribution	Desired distribution over L1/L2/L3
Branch security	Required breaker/fuse architecture
Metering	Input, branch and/or outlet
Switching	Is remote outlet control necessary?
Monitoring	Energy, capacity, Power Quality, environment
Network	Controller, Gateway or Daisy Chain
Integration	DCIM/BMS/API/SNMP
Physical design	0U/19″/21″, dimensions, mounting and cable entry

The most important principle is:

Do not specify the AI rack PDU separately from the AI hardware, upstream power architecture, and redundancy architecture.

Can an existing data center be adapted for AI hardware?

Often yes, but the available electrical capacity and distribution architecture must be assessed first.

Not every AI deployment requires an entirely new data center.

Existing facilities can potentially support GPU infrastructure by modifying rack layouts, increasing power density in specific zones, adjusting the distribution, or upgrading the rack-level power infrastructure.

For existing rack infrastructure, Schleifenbauer also offers Inline Meters with which input-level energy metering can be added to existing Basic or older intelligent PDUs without replacing the entire installed base.

What is the best PDU for AI hardware?

There is no single PDU that is the best choice for every AI deployment.

The right PDU aligns with the electrical, mechanical, redundancy, monitoring, and management requirements of the specific AI rack.

For one project, that could be a three-phase 32A Monitored PDU.

Another project might require multiple 63A feeds, outlet-level monitoring, custom CX outlets, and a Gateway Module with the EnerTree Platform.

And future AI architectures can eventually move beyond conventional rack-based AC distribution.

That is precisely why configurability is important.

So don't just ask:

“Which standard PDU is suitable for AI?”

The better question is:

“Which rack power architecture does this AI hardware need?”

The PDU can then be designed around that answer.

Read also the Supermicro case study : https://www.schleifenbauer.eu/nl/cases/case-study-supermicro/

Read also the Nvidia case study: https://www.schleifenbauer.eu/nl/cases/case-study-nvidia/

Schleifenbauer custom PDUs for AI hardware

Schleifenbauer develops and manufactures rack PDUs in the Netherlands and does not work exclusively with fixed catalog configurations.

For AI and high-density compute environments, this means that the PDU can be tailored to the infrastructure: 10A to 63A, single or three-phase, custom inputs, cable lengths, outlet configurations, protection, input/branch/outlet metering, remote switching, and various management architectures.

PDU 5.0 adds a modular intelligence layer to that.

Choose Controller + EnerTree Lite for conventional local PDU intelligence, Gateway + EnerTree Platform for centralized Data Centre Energy Management or use Daisy Chain to create scalable PDU networks with fewer Ethernet connections.

Because the communication modules are hot-swappable, the management architecture can be modified later without replacing the underlying PDU.

The result is not just a high-power rack PDU.

It is an architecture that uses AI rack power:

configurable → measurable → visible → manageable → scalable

makes.

Building an AI rack? Start with the power requirements.

Are you designing a new AI, GPU, or HPC rack? Then send Schleifenbauer the hardware specification, the rack design, the tender text or the desired electrical configuration.

Based on this, a custom rack PDU can be configured around the actual infrastructure requirements, instead of forcing the project into a predetermined PDU model.

From the incoming rack feed to the individual AI device: design the power distribution around the hardware.

The post Custom PDUs for AI hardware: designing rack power for high-density AI infrastructure appeared first on Schleifenbauer – PDUs.

Source: https://www.schleifenbauer.eu/nl/custom-pdu-ai-hardware/