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.
AI inference hardware in comms rooms draws multiples of what standard IT equipment consumes. A single GPU inference server can draw 3 to 5kW, compared to 0.5 to 1kW for a typical 1U server. A high-density rack of AI compute can draw 20 to 40kW, against 8 to 12kW for a comparable traditional IT rack. At the building level, AI-powered devices, WiFi 7 access points, AI CCTV cameras, edge compute nodes, smart building sensors, all draw power from the network via PoE. The aggregate PoE budget across a large floor plate can easily exceed what the switching infrastructure was sized to deliver.
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
Uninterruptible power supplies in comms rooms are typically sized for the equipment load at the time of installation. As equipment is added over time, particularly as AI inference hardware is introduced, the UPS load increases. Many commercial building comms rooms are operating with UPS systems that are already at or above their rated capacity for the legacy equipment installed, before any AI hardware is added.
The first step in any AI infrastructure project involving comms room compute is a UPS audit: current load measurement, rated capacity assessment, and a forward-looking capacity calculation based on the planned AI hardware additions. A UPS running at 80% capacity with legacy hardware will fail to support a graceful shutdown when AI servers are added without increasing the UPS rating. The consequences of a UPS failure during an AI model run or database transaction are significantly more costly than a properly sized replacement UPS.
AI hardware does not just draw more power than standard IT equipment, it draws it less predictably. Inference loads spike when requests arrive in bursts. Power draw can jump from 20% to 90% of rated consumption in milliseconds.
A UPS sized for average load, not peak load, will fail during exactly the conditions that matter most. AI inference loads require UPS systems sized for peak draw, not average consumption.
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: WiFi 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.
In a building deploying WiFi 7 access points (which draw 40 to 55W each), AI CCTV cameras (15 to 25W each), access control systems (5 to 15W per door), and smart building sensors (2 to 8W each), the aggregate PoE demand can substantially exceed what a standard PoE+ switch budget provides. An 8-port PoE+ switch with a 120W budget can power four WiFi 7 access points, and nothing else. A 24-port PoE++ switch with a 740W budget provides significantly more headroom, but at a cost that must be planned for.
PoE budget design is not a switch specification exercise. It is a whole-building power planning exercise that starts with the list of powered devices and works backwards to the switch specification. ACCL includes PoE budget analysis as a standard part of any infrastructure design for a building deploying AI-enabled devices.
A high-density rack of AI compute can draw 20 to 40kW, against 8 to 12kW for a comparable traditional IT rack.
PDU design and power distribution
Power distribution units (PDUs) in comms rooms distribute mains power to the equipment in the rack. Standard PDUs are adequate for legacy IT loads. AI inference hardware requires monitored, managed PDUs that can report per-outlet power consumption in real time, allowing the facilities team to track total rack power draw, identify equipment approaching rated limits, and manage load distribution across circuits to prevent overloading individual breakers.
The cabling implications of PDU design are straightforward but frequently overlooked: the structured cabling infrastructure must bring data connectivity to the PDU management interface as well as to the compute hardware. A managed PDU that cannot be reached on the network cannot be monitored or controlled.
Power consumption and heat generation are directly related. A comms room running 20kW of AI compute generates 20kW of heat that must be removed. Standard commercial building air conditioning systems are not designed for this density of heat generation. Before AI inference hardware is installed, the comms room cooling capacity must be assessed against the anticipated thermal load.
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 802.3 Ethernet and PoE standardsIEEE 802.3bt Power over Ethernet standard
- IET Wiring Regulations (BS 7671)UK electrical installation standards
Frequently asked questions
How much power does AI inference hardware draw compared to standard IT equipment?
A single GPU inference server typically draws 3 to 5kW, compared to 0.5 to 1kW for a standard 1U server, roughly 5 to 10 times more power per unit. A high-density rack of AI compute can draw 20 to 40kW, against 8 to 12kW for a traditional IT rack. This increase in power density has significant implications for UPS sizing, PDU design and comms room cooling capacity.
What is PoE budget and why does it matter for AI-enabled buildings?
PoE budget is the total wattage a switch can deliver across all its ports simultaneously. In a building deploying WiFi 7 access points (40 to 55W each), AI CCTV cameras (15 to 25W), access control systems and smart building sensors, the aggregate PoE demand can exceed what standard PoE+ switches provide. PoE budget planning must account for every powered device across the building and work backwards to the switch specification, to ensure adequate power delivery without overloading switch hardware.
What is PoE++ and when is it required?
PoE++ (IEEE 802.3bt) delivers up to 90W per port, compared to 30W for PoE+ (IEEE 802.3at). PoE++ is required for WiFi 7 access points, high-performance AI CCTV cameras, and some edge compute devices that exceed the 30W limit of standard PoE+. Any building deploying WiFi 6E or WiFi 7 infrastructure should specify 802.3bt PoE++ switches to ensure devices receive sufficient power for full performance operation.
How should a UPS be sized for a comms room with AI hardware?
A UPS for an AI comms room should be sized for peak load, not average consumption, because AI inference loads can spike rapidly from low to high utilisation. The standard recommendation is to size UPS capacity at no more than 60 to 70% of rated load for normal operation, providing headroom for peak demand spikes. Before adding AI hardware to an existing comms room, the current UPS load should be measured and compared against rated capacity to determine whether an upgrade is required.