Quick Answer: Data centres range from a few hundred kilowatts to hundreds of megawatts in total capacity, but that headline number is not what determines whether a facility can support your workload. The figure that matters is per-rack power density: standard enterprise racks draw 5-10kW, AI inference racks require 20-50kW, and training clusters exceed 100kW. Pair that with the facility's Power Usage Effectiveness (PUE) score, which determines how much of your electricity bill actually reaches your servers, and you have the two numbers that should drive every data centre procurement conversation.
The number circulating most often in data centre coverage is total megawatts. Providers advertise it, analysts track it, and headlines use it to frame the scale of the industry. But total megawatts is a facility-level figure, and you are not buying a facility. You are buying rack space, and the power question that matters is how many kilowatts your provider can reliably deliver to the rack your servers live in.
That difference between the headline number and the operational reality trips up infrastructure teams more than almost anything else in procurement. A 50MW campus that caps rack density at 8kW cannot support an AI inference deployment that needs 40kW per cabinet.
Layer on top of that the facility's PUE score, which determines how much of your bill goes to cooling and overhead rather than your workload, and data centre power consumption becomes a much more layered topic than any single number suggests.
Data centre power gets discussed in three units: watts, kilowatts, and megawatts. The unit used depends on what is being measured, and using the wrong reference frame from the start leads to poor infrastructure decisions downstream.
A watt is the base unit of electrical power. A kilowatt (kW) is 1,000 watts, and a megawatt (MW) is 1,000 kilowatts. In a data centre context, kilowatts are the unit of choice when discussing individual racks or smaller deployments. Megawatts describe total facility capacity or large wholesale deployments purchased by a single tenant.
When a colocation provider quotes a rack with a 10kW power allocation, that means the cabinet is wired to draw up to 10,000 watts continuously. When a facility announces a 20MW expansion, it means the building can support 20,000 kilowatts of total IT load distributed across all customers.
How much power does a data centre use per day? For a 10MW facility running at 85% utilisation, that figure approaches 204,000 kWh daily. Pricing in colocation is almost always expressed in kilowatts per rack, not total megawatts, so matching your workload requirements to the per-rack allocation is where the real planning begins.
Total facility capacity tells you the ceiling. It says nothing about what a single rack can receive. A 40MW colocation facility could support thousands of low-density racks drawing 5-7kW each. That same building, if its power distribution and cooling infrastructure were not designed for high-density loads, could be completely unsuitable for a 30kW AI rack even with megawatts to spare at the utility connection.
This brings our attention to the biggest problem companies face when power is a specific requirement. The binding constraint for AI workloads is rarely total facility capacity; it is per-rack density.
A provider claiming to support AI data centres in Canada needs to back that up with both electrical infrastructure and the cooling systems to match. A facility's rated capacity is also not the same as its committed available capacity. Power infrastructure carries overhead, redundancy headroom, and reserved expansion capacity. The figure a tenant actually has access to is the committed available power, and that number belongs in any contract you sign.
Rack power density is measured in kilowatts per rack. For most of the last two decades, a standard enterprise server rack drew somewhere between 5 and 10kW. That number was stable enough that facilities, cooling systems, and power distribution units were all engineered around it. AI has shifted the baseline by a lot, and providers that have not kept pace with that shift are selling capability they cannot consistently deliver.
A conventional server rack loaded with CPU-based compute draws 5-8kW under sustained load. A GPU-accelerated rack configured for AI inference commonly draws 20-40kW. A dense AI training rack carrying eight high-performance GPUs alongside supporting network and storage infrastructure can approach or exceed 100kW per rack. That is a 10x to 20x increase in power demand from the same footprint of raised floor.
The reason behind this change is the GPU itself. Traditional CPU-based servers run at 150-200 watts per chip. Modern AI GPUs now draw 700 watts or more per chip, with next-generation processors expected to exceed 1,400 watts.
Populate a rack with eight of those processors plus dual CPUs, networking cards, and NVMe storage, and the power demand becomes something a standard colocation cage was never designed to handle. The implications cascade through every layer of the facility, from raised floor load ratings to power distribution architecture.
Getting 50kW to a single rack requires a fundamentally different infrastructure approach than delivering 8kW. Power distribution units capable of handling those loads, three-phase circuits rather than single-phase, and dedicated busways that bypass standard in-row distribution are all part of what a genuinely high-density environment requires. The electrical design has to be engineered for it from the start, not retrofitted after the fact.
Beyond the electrical side, there is the question of cooling. Air cooling works reliably up to roughly 20kW per rack, depending on containment design and airflow management. Above that threshold, which many enterprise AI deployments now exceed, liquid cooling becomes a serious consideration, whether that is rear-door heat exchangers, direct-to-chip liquid cooling, or full immersion systems.
Data centre cooling and power density are inseparable at these load levels. A facility advertising high-density power allocation but running legacy air cooling is presenting an incomplete picture. The thermal limits in high-density AI environments are just as real as the electrical ones.
Not all AI workloads draw the same power. AI training and inference have meaningfully different power profiles, and conflating the two leads to infrastructure decisions that do not fit the actual use case.
Training runs are intensive, sustained, and power-hungry. A large model training cluster can operate at full GPU utilisation for days at a time, with power draw staying near the maximum throughout.
A 50,000 GPU training cluster approaches 35MW of sustained draw, a requirement that calls for hyperscale purpose-built infrastructure and dedicated utility connections. Inference is different. Enterprise AI inference workloads are bursty, variable, and more manageable in their peak-to-baseline ratios. A rack running inference models might average 30-40kW but spike during peak query periods.
For most organisations migrating GPU workloads out of public cloud or standing up their first on-premise AI environment, inference is the primary use case and it sits within a range that a well-designed colocation facility can genuinely support.
|
Workload Type |
Typical Rack Power Draw |
Cooling Requirement |
Notes for Buyers |
|
Standard enterprise compute |
5–10 kW/rack |
Air cooling |
Compatible with most standard colocation |
|
Virtualisation and cloud compute |
8–15 kW/rack |
Air cooling with containment |
Requires hot/cold aisle management |
|
AI inference |
20–50 kW/rack |
Advanced air or liquid cooling |
Needs upgraded PDUs and three-phase circuits |
|
AI training clusters |
60–120+ kW/rack |
Liquid or immersion cooling |
Requires purpose-built facility design |
Power Usage Effectiveness (PUE) is the ratio that describes how much of a data centre's total electricity consumption actually reaches the IT equipment inside it. Everything else, cooling systems, power distribution losses, lighting, security infrastructure, represents overhead.
The formula is PUE equals total facility power divided by IT equipment power. A PUE of 1.0 is theoretical perfection, meaning every watt entering the building powers a server. Real-world values are always higher because supporting systems always consume some share of total draw.
According to the Uptime Institute's 2026 Global Data Centre Survey, the industry-wide annual average PUE sits at 1.52. Newer large-scale facilities with modern power and cooling designs routinely report values around 1.3 or lower.
To put that in practical terms: a facility with a PUE of 1.5 consumes 1.5 kWh for every 1 kWh delivered to IT equipment. The remaining 0.5 kWh goes to cooling, lighting, power conditioning, and distribution losses.
At scale, that overhead adds up quickly. A data centre drawing 10MW of IT load at a PUE of 1.5 draws 15MW at the utility meter. Data centre power usage effectiveness is relevant to your billing, not just the sustainability section of a provider's website. When evaluating colocation costs, PUE should be in the conversation alongside per-rack pricing and power commitment terms.
PUE is an energy efficiency ratio, not a sustainability metric and not a complete cost predictor. Two facilities with identical PUE scores can have very different electricity costs depending on the rate their utility charges. Two facilities with different PUE scores can have similar carbon footprints depending on the energy source feeding the grid they draw from.
As analysis of PUE measurement standards shows, a facility with a PUE of 1.1 on a high-carbon grid can emit significantly more CO2 per unit of compute than a facility with a PUE of 1.6 on a clean grid.
The energy source dominates the carbon calculation, and PUE does not capture that. There is also a measurement nuance that matters when comparing providers: PUE fluctuates with season and load. A facility quoting only a best-case reading taken during mild weather is presenting a figure that looks better than it performs across a full year. Always ask for the annualised figure, not the headline number.
|
PUE Score |
What It Means in Practice |
Typical Facility Type |
|
1.0 |
Theoretical ideal, not achievable |
N/A |
|
1.1–1.2 |
Very high efficiency |
New hyperscale with advanced cooling |
|
1.3–1.4 |
Strong performance |
Modern colocation with free-air cooling |
|
1.5–1.6 |
Industry average |
Mixed-age colocation fleet |
|
1.8–2.0 |
Older or less efficient design |
Legacy enterprise or edge facilities |
Most organisations entering a colocation procurement conversation know to ask about redundancy, uptime certification, and physical security. Fewer know the right power-specific questions, and the ones left unasked tend to create problems once a contract is in place. A structured RFP for AI-ready colocation can help formalise this process, but the core concepts here apply to any provider evaluation regardless of formality.
Rated capacity is the total power a facility is designed to distribute across all its customers. Committed power is what a provider is contractually obligating itself to deliver to your rack. These are not the same number, and the difference matters when you need guaranteed performance rather than best-effort availability.
Before signing any data centre agreement, get written answers to these questions:
The committed vs. rated distinction is most consequential for AI workloads, where sustained high utilisation means the power needs to be there reliably, not just when the facility happens to have headroom.
"AI-ready" has become one of the most overused terms in data centre marketing. Providers apply it to everything from a minor power density upgrade to a purpose-built high-density environment. The label alone tells you very little. The specs behind it tell you everything.
A colocation environment that can genuinely support enterprise AI inference should confirm the following in writing before any agreement is signed:
Data centre power requirements have shifted faster than most enterprises anticipated. The gap between a standard colocation rack and an AI-capable one is not a matter of modest increments. It is a fundamental infrastructure distinction that affects every layer of a facility's design, from the electrical panel to the cooling plant to the floor structure underneath the cabinet.
Choosing a provider based on total megawatt capacity or available floor space, without verifying per-rack density, cooling architecture, and annualised PUE, leaves organisations at real risk of deploying into infrastructure that cannot keep pace with workload growth.
Qu Data Centres offers a national footprint of nine facilities built for the demands of modern Canadian enterprise IT. Our colocation solutions support medium and high-density compute environments, with power and cooling infrastructure sized to match the workload. High-availability connectivity runs through carrier-neutral interconnection across all locations, and our managed services team provides around-the-clock technical support for organisations that need more than a cage and a power circuit.
Whether you are colocating existing infrastructure, standing up an AI inference environment, or building a disaster recovery architecture, we have the specifications and the staff to support it across facilities in Toronto, Ottawa, Calgary, Edmonton, and London, Ontario.
Qu is built from the ground up to support medium-to-high density compute environments the right way. Book a tour to learn more today.
Power draw varies with facility size and load. A mid-size enterprise data centre in the 1-5MW range draws between 24,000 and 120,000 kWh per day. A hyperscale facility at 100MW can consume over 2.4 million kWh daily. The more useful question for colocation buyers is how much power their specific rack allocation draws and what PUE multiplier applies to their billing, not the facility's total consumption figure.
The industry-wide annual average PUE is 1.52, according to the Uptime Institute's 2026 Global Data Centre Survey. A score below 1.4 is considered strong for an enterprise colocation facility. Hyperscale operators can achieve 1.08 to 1.2 through advanced cooling and scale efficiencies. Always ask for annualised PUE rather than a point-in-time reading, as seasonal variation can make a facility appear more efficient than it performs year-round.
It depends on the workload type. A large-scale model training cluster can require 35MW or more of sustained draw. Enterprise AI inference is far more manageable, typically running within a standard colocation footprint at elevated rack densities of 20-50kW per rack. Most organisations do not need a dedicated AI campus. They need a colocation provider whose power and cooling infrastructure genuinely supports high-density racks at the per-rack level.
Total facility capacity describes the megawatts a building can receive from the grid and distribute across all tenants. Rack power density describes how many kilowatts a single cabinet can reliably receive. A facility with large total capacity can still cap individual racks at 5-8kW if its power distribution and cooling were not designed for high-density loads. For AI workloads, per-rack density is the more critical number to verify before signing any agreement.