What is a custom PDU for AI hardware?

A custom PDU for AI hardware is a rack Power Distribution Unit configured specifically around the power requirements of 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 configured around the infrastructure.

This can include:

• single- or three-phase power;
• 10A, 16A, 32A or 63A input;
• specific input plug and cable requirements;
• custom cable lengths and cable-entry positions;
• outlet type, quantity and positioning;
• phase-balanced outlet layouts;
• branch protection;
• input-, branch- and outlet-level energy metering;
• 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 through 63A, with configurable metering at input, branch and outlet level and optional remote switching.

The objective 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 is increasing rack power density.

Traditional enterprise racks may contain relatively diverse and independently operating equipment. Modern AI racks can instead contain highly concentrated GPU compute operating as a coordinated system.

That changes the importance of the rack power architecture.

NVIDIA, for example, documents designed rack power of 120 kW for GB200 NVL72 and 135 kW for GB300 NVL72 in its current power-management documentation.

At the same time, NVIDIA is developing its infrastructure towards 800 VDC architectures for future AI factories, specifically because increasing compute performance and rack density are putting greater demands on traditional power distribution.

This does not mean every AI rack requires a 120 kW rack PDU, nor that every AI deployment has the same electrical architecture.

It means that the term “AI rack” no longer describes one standard power requirement.

The correct PDU depends on the actual hardware and power architecture being 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 model do we normally use?

It should be:

What does this rack actually need?

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

The PDU specification can then be built around those requirements.

This becomes increasingly important as AI infrastructure evolves quickly. A fixed PDU catalogue designed around yesterday’s typical rack density may not provide the best fit for tomorrow’s GPU infrastructure.

32A or 63A PDU for AI hardware?

There is no universal answer.

A 32A PDU may be appropriate for one AI rack architecture, while another may require multiple 32A feeds, 63A three-phase PDUs or a different distribution architecture entirely.

For example, NVIDIA’s earlier DGX H100 SuperPOD design guidance specified 415 VAC, 32A, three-phase as the preferred power configuration for high-density deployment patterns, while allowing designs to be modified for other supply schemes.

The correct question is therefore not whether 32A or 63A is “better” for AI.

It is whether the complete combination of:

voltage × current × phases × number of feeds × redundancy

provides the required usable capacity for the specific rack.

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

For higher-density racks, a three-phase 63A PDU can provide up to approximately 43.5 kVA at a 400 V supply.

Multiple feeds can then be designed around the required rack capacity and redundancy architecture.

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

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

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

L1 → L2 → L3

The PDU outlet configuration can then be designed to distribute connected loads across the available phases.

For AI infrastructure, phase balance is not merely a design consideration at installation.

It should also be measurable during operation.

This is where an intelligent PDU becomes valuable: the operator can see how the actual load develops instead of relying only on the original design calculation.

AI workloads are dynamic, not static

Nameplate power and average consumption do not tell the entire story.

AI compute can produce significant changes in electrical demand as workloads move between training, inference, idle states and other operating conditions.

NVIDIA specifically notes that GPU clusters can create sudden power-demand bursts and that infrastructure based purely on static provisioning may struggle with these runtime fluctuations.

This makes headroom and real-time measurement increasingly important.

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

For data centre operators, the question therefore becomes:

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

Measure AI power where it is actually consumed

The closer measurement gets to the IT equipment, the more useful the data becomes.

An intelligent PDU can provide several layers of visibility:

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

Depending on the selected Schleifenbauer PDU functionality, measurements can be provided at input, branch and outlet level. The PDU 5.0 platform supports power monitoring with stated 0.5% accuracy at input and outlet level.

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

How much power is the rack consuming?

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

Which branch is approaching its capacity?

How much power is an individual connected device consuming?

How much headroom remains before additional AI hardware can be deployed?

That is considerably more actionable than knowing only the total facility load.

Outlet-level monitoring for AI and GPU infrastructure

Outlet-level monitoring adds another level of granularity.

Instead of knowing only how much energy enters the PDU, operators can see consumption closer to individual connected loads.

This becomes useful in racks containing combinations of:

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

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

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

Custom outlets for AI hardware

Power capacity alone does not determine whether a PDU fits an AI rack.

The physical connections must fit as well.

Different IT systems can require different plug types and outlet arrangements. High-density racks can also create pressure on the amount of usable rack space available for power distribution.

Schleifenbauer therefore allows the type, quantity, position and layout of outlets to be configured around the project.

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

A single CX outlet can accept 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 containing different device types, this can simplify the physical power design considerably.

Redundancy matters because an AI rack is a system

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

In clustered AI environments, the failure of one component can affect a much larger workload.

NVIDIA makes this particularly clear in its DGX SuperPOD design guidance: a failure of a single system within a multi-node AI workload can cause the complete job to stop. Its H100 design therefore incorporates specific power-source redundancy requirements.

The precise redundancy architecture depends on the AI platform.

That can mean:

A/B feeds

or

multiple independent power paths

or another architecture prescribed by the hardware vendor.

The rack PDU design needs to follow that redundancy strategy rather than undermine it.

Do not confuse available PDU capacity with usable AI capacity

This distinction becomes increasingly important at high power densities.

A rack may have several power feeds, but their combined nameplate capacity is not automatically the amount of power that can safely be allocated to IT equipment.

Redundancy, breaker ratings, phase loading, upstream infrastructure and required failover headroom all affect usable capacity.

NVIDIA’s current power-management documentation similarly distinguishes between physical power paths and the usable power envelope that remains after redundancy requirements are considered.

For AI deployments, capacity management therefore needs to answer two different questions:

What is installed?

and

What can safely be used?

Power Quality becomes more relevant as rack density increases

AI power management is not only about kW and kWh.

High-density electronic loads also make the characteristics of the electrical load increasingly interesting.

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

This adds context that ordinary energy-consumption monitoring cannot provide.

For an AI infrastructure operator, energy monitoring answers:

How much electricity is being consumed?

Power Quality Monitoring helps answer:

What is happening electrically while that power is being consumed?

The two should be considered complementary.

From custom AI PDU to Data Centre Energy Management

A single intelligent PDU provides detailed information about one rack.

An AI data centre may contain hundreds or thousands of PDUs.

That creates the next challenge:

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

This is where EnerTree becomes part of the architecture.

Schleifenbauer PDU 5.0 separates the required PDU functionality from the way the intelligent PDU environment is managed.

The communication architecture can be selected independently using a Controller Module, Gateway Module or Daisy Chain Module.

Controller Module + EnerTree Lite

The Controller Module provides a conventional intelligent PDU architecture.

EnerTree Lite runs embedded on the Controller Module, providing a local web interface and direct access to PDU measurements.

It can also integrate directly with an existing DCIM environment.

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

For smaller AI installations or environments that already have an established management platform, this may provide the required architecture.

Gateway Module + EnerTree Platform

For larger AI deployments, the Gateway architecture moves intelligence away from each individual PDU and centralises it within EnerTree Platform.

EnerTree Platform runs as a virtual machine.

The PDU becomes a real-time measurement and control endpoint, while processing, analysis, configuration and management are centralised.

A single EnerTree Platform environment can scale to 10,000 PDUs.

For large GPU clusters and AI data centres, this provides a very different operating model from independently managing thousands of intelligent rack PDUs.

EnerTree functionality includes real-time energy monitoring, alerts, historical analysis, reporting, hierarchical infrastructure views, environmental monitoring and PUE calculation.

Daisy Chain for scalable PDU connectivity

Not every PDU in an AI rack or row needs its own Ethernet connection.

A Controller or Gateway can connect downstream PDUs through a Daisy Chain architecture.

Up to 100 PDUs can operate within one ring on a single IP address.

A dedicated Daisy Chain Module functions as a cost-efficient data relay without its own Ethernet port, reducing network hardware and the Ethernet attack surface inside the rack.

This can become particularly relevant when deploying intelligent PDUs at scale.

From one GPU outlet to complete AI data centre visibility

The value of this architecture becomes clearer when looking at the complete hierarchy:

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

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

At the same time, measurement data travels in the opposite direction.

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

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

That is the difference between simply delivering power to AI hardware and actually understanding AI rack power.

Integrating AI rack power with DCIM and BMS

AI power data should not become another isolated dataset.

EnerTree Platform and EnerTree Lite support integration with external data centre systems through interfaces and protocols including HTTP/HTTPS, REST API, MODBUS/TCP, SNMP v1/v2c/v3, IPv4/IPv6, SMTP, Syslog and command-line interfaces. Data can also be exported to databases including MS SQL, MySQL and MariaDB.

This allows rack-level electrical information to become part of a broader DCIM, BMS or data analytics environment.

Custom AI PDUs should also be physically customisable

AI infrastructure changes quickly.

A PDU that has the correct electrical rating but does not fit the rack layout is still the wrong PDU.

Schleifenbauer therefore builds its PDUs to order.

Configuration options include cable length, connector type, outlet layout and position, metering functionality, controller position, identification and labelling, mounting adaptations and other mechanical requirements.

PDU 5.0 can also be configured with different protection options, residual current sensing, environmental sensors and additional features depending on the project.

This is particularly relevant for OEMs, system integrators, data centre designers and operators deploying AI hardware where the rack itself is increasingly engineered as an integrated system.

What should you specify when ordering a custom PDU for AI hardware?

Before selecting the PDU, define the rack.

A useful AI PDU specification should at minimum establish:

Requirement	What to determine
AI hardware	GPU/server platform and configuration
Maximum rack load	Expected and design power
Supply	Voltage and frequency
Phases	Single-phase or three-phase
Current	16A, 32A, 63A or project-specific requirement
Redundancy	A/B, N+1, 2N or vendor-prescribed architecture
Number of feeds	Required independent rack power paths
Connectors	Input plug and IT equipment plug types
Outlets	Type, quantity, position and grouping
Phase layout	Required load distribution across L1/L2/L3
Branch protection	Required breaker/fuse architecture
Metering	Input, branch and/or outlet
Switching	Whether remote outlet control is required
Monitoring	Energy, capacity, Power Quality, environment
Networking	Controller, Gateway or Daisy Chain
Integration	DCIM/BMS/API/SNMP requirements
Physical design	0U/19″/21″, dimensions, mounting, cable entry

The most important principle is:

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

Can an existing data centre be adapted for AI hardware?

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

Not every AI deployment requires a completely new data centre.

Existing facilities can potentially accommodate GPU infrastructure by changing rack layouts, increasing power density in selected areas, modifying distribution or upgrading rack-level power infrastructure.

NVIDIA itself is developing hybrid approaches intended to allow next-generation AI compute to operate with existing AC facility infrastructure during the industry’s transition towards higher-voltage DC architectures.

For existing rack infrastructure, Schleifenbauer also offers Inline Meters that can add input-level energy measurement to existing Basic or legacy intelligent PDUs without replacing the complete installed base.

What is the best PDU for AI hardware?

There is no single best PDU for every AI deployment.

The correct PDU is the one that matches the electrical, mechanical, redundancy, monitoring and management requirements of the specific AI rack.

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

Another may require multiple 63A feeds, outlet-level monitoring, custom CX outlets and a Gateway Module connected to EnerTree Platform.

Another AI architecture may move beyond conventional AC rack distribution altogether.

That is precisely why a configurable approach matters.

Rather than asking:

“Which standard PDU model is designed for AI?”

a better question is:

“What rack power architecture does this AI hardware require?”

The PDU can then be engineered around the answer.

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

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

Schleifenbauer custom PDUs for AI hardware

Schleifenbauer designs and manufactures rack PDUs in the Netherlands and does not rely on fixed catalogue configurations.

For AI and high-density computing environments, this allows the PDU to be configured around the infrastructure: 10A to 63A, single- or three-phase, custom inputs, cable lengths, outlet configurations, protection, input/branch/outlet metering, remote switching and different management architectures.

PDU 5.0 then adds a modular intelligence layer.

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

And because the communication modules are hot-swappable, the management architecture can be changed later without replacing the underlying PDU.

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

It is an architecture designed to make AI rack power:

configurable → measurable → visible → manageable → scalable.

Building an AI rack? Start with the power requirements

If you are designing a new AI, GPU or HPC rack, send Schleifenbauer the hardware specification, rack design, tender specification or required electrical configuration.

A custom rack PDU can then be configured around the actual infrastructure requirements rather than forcing the project into a predefined PDU model.

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