- Use the actual hardware schedule and vendor power data; “AI hardware” does not have one standard load.
- Calculate PoE by device class and quantity, then check both per-port capability and the switch’s total budget.
- UPS, PDU, electrical distribution and cooling should be designed together, with M&E input where required.
The power problem nobody is talking about
When IT Directors and Facilities Managers plan an AI deployment in a commercial building, the conversation typically focuses on software licences, hardware specifications and network bandwidth. Power rarely features until something fails. That is the wrong sequence.
Power demand varies widely by server, accelerator, configuration and utilisation. The design should use the manufacturer’s maximum or design power values for the actual equipment, plus network, storage and auxiliary loads. Across the building, access points, cameras, controls and gateways may also increase the aggregate PoE demand beyond the capacity of legacy switches.
Power is not an afterthought in AI infrastructure planning. It is a constraint that must be understood before a design decision is made.
UPS sizing for AI comms rooms
Existing UPS systems are often carrying more equipment than the original design anticipated. Their present load, battery condition, runtime, bypass arrangement and maintainability should be verified before additional compute or PoE switching is introduced.
For projects involving local compute, start with measured load, the proposed hardware schedule, required runtime, redundancy target and the UPS manufacturer’s sizing method. A unit already operating at a high percentage of its rating may not provide the required runtime or growth capacity after new equipment is added.
Compute loads can vary with utilisation and may change quickly. The UPS and upstream electrical design should therefore be checked against credible peak demand, startup behaviour and the required autonomy, not only a casual average reading.
A UPS should be selected using the actual load profile, required runtime, redundancy and manufacturer guidance. Average load alone is not enough for a dependable design.
Power is not an afterthought in AI infrastructure planning. It is a constraint that must be understood before a design decision is made.
PoE budget management across an AI-dense building
Power over Ethernet is the delivery mechanism for power to the majority of AI-enabled devices in a commercial building: Wi-Fi access points, IP CCTV cameras, access control readers, occupancy sensors, room booking displays, and edge compute nodes. Each device draws power from the switch it is connected to. The switch has a total PoE power budget, the maximum aggregate wattage it can deliver across all ports simultaneously.
Device demand must be taken from the approved equipment schedule. As a simple check, four devices allocated 40W each require 160W before any design allowance, so a switch with a 120W total PoE budget cannot support that load. Per-port class, total switch budget, power-supply configuration and redundancy all need to be checked together.
PoE budget design starts with the powered-device list and works back to the switch, power supplies and cabling. ACCL can produce the passive-infrastructure and PoE schedule, with electrical and active-network design coordinated with the appropriate specialists.
The rack power figure must come from the actual server, storage and network schedule. Use measured or manufacturer design values rather than a generic “AI rack” allowance.
PDU design and power distribution
Power distribution units (PDUs) distribute mains power within the rack. The appropriate PDU depends on load, socket type, resilience and monitoring requirements. Metered or managed PDUs can provide useful visibility where the operating model requires per-rack or per-outlet data.
Where a managed PDU uses an Ethernet management interface, provide an appropriate network connection and security zone. Monitoring may also be delivered through another interface or management architecture, so confirm the selected PDU rather than assuming every rack outlet needs a data point.
Most electrical power consumed by IT equipment becomes heat within the room. A sustained 20kW IT load therefore creates a cooling requirement of roughly the same order before allowances and system losses. The existing cooling system should be checked by a competent M&E specialist against the actual load, airflow and resilience target.
ACCL coordinates with M&E consultants and cooling specialists on any project involving significant AI compute hardware, to ensure the comms room thermal design is aligned with the power consumption planning.
Standards and sources
- IEEE 802.3btFour-pair Power over Ethernet and power-class management
- Fluke Networks PoE installation guidePoE negotiation, load testing and cable considerations
- IET BS 7671Requirements for electrical installations
Frequently asked questions
How much power does AI inference hardware draw compared to standard IT equipment?
There is no reliable single multiplier. Power varies by processor, accelerator, server density, storage, utilisation and cooling design. Use the approved equipment schedule and manufacturer design values, then verify the upstream circuit, UPS, PDU and cooling capacity. Generic rack figures are useful only for early screening, not final design.
What is PoE budget and why does it matter for AI-enabled buildings?
PoE budget is the total power a switch or power-supply configuration can allocate across its powered ports. Check the required class for every device, the number of devices expected to operate simultaneously, cable and environmental considerations, and the total budget. For example, four devices allocated 40W each require 160W before design allowance, so a 120W switch budget is insufficient.
What is PoE++ and when is it required?
PoE++ commonly refers to IEEE 802.3bt, which supports higher power classes using four pairs. It is required only when the selected powered device and operating mode need an 802.3bt class. Some Wi-Fi 6E, Wi-Fi 7, camera and edge devices use it, while others operate within lower classes. Confirm the data sheet and total switch budget.
How should a UPS be sized for a comms room with AI hardware?
Size the UPS from the measured existing load and approved new equipment schedule, then account for peak demand, required runtime, battery condition, efficiency, redundancy, bypass and planned growth. Use the UPS manufacturer’s sizing method and competent electrical input. A blanket 60% or 70% loading rule is not a substitute for design.