A colocation power quote tells you how much electricity the site will deliver to your hardware, and how much you are committing to pay for whether you use it or not. The numbers look simple, but each one means something different, and mixing them up leads to either an undersized deployment or a contract you are paying to waste.
IT load versus facility load
Every data center carries two power figures. IT load is the power your servers actually draw. Facility load is IT load plus everything the building needs to run: cooling, lighting, and the losses in the electrical distribution chain.
When a site quotes you a number, ask which one it is. A quote for 1 MW of IT load and a quote for 1 MW of facility load are not the same thing. The first gives you 1 MW for your hardware. The second gives you less, because the building's overhead comes out of that 1 MW first.
The ratio between the two is called PUE (Power Usage Effectiveness). A PUE of 1.4 means that for every 1 kW your servers draw, the facility consumes 1.4 kW in total. You do not need to optimize for a specific PUE, but you do need to know whether the kW figure on your quote is the IT side or the total side.
Contracted versus usable power
Contracted power is the amount the site reserves for you and bills against. Usable power is what you can actually run at any moment without tripping a breaker or triggering an overload alarm.
Sites commonly hold back a margin between the two. Ask the operator to confirm the usable figure before you size your deployment, because the contracted number and the usable limit are often not the same. The gap is a safety buffer, not a billing error. If you fill your cabinets to the contracted figure, the site may flag you for exceeding safe operating limits. Ask the operator to confirm the usable figure before you size your deployment.
Metered versus committed power
Some contracts bill you for the power you actually draw each month. Others commit you to a fixed kW amount from day one, whether your hardware is installed yet or not. A committed contract protects the operator's investment in reserved capacity. A metered contract gives you more flexibility early on.
Either way, the contract locks you in for a term. Colocation agreements for AI infrastructure are multi-year commitments; the terms XIRR brokers run 36 to 60 months. Know which billing model you are signing before you negotiate the term.
From server count to kW on a quote
The fastest way to sanity-check a power quote is to work from your server count. These are planning figures, close enough to size a quote.
- An 8-GPU H100 or H200 server draws about 10 kW.
- An 8-GPU B200 server draws about 14 kW.
- A DGX B300 (8 GPUs, air cooled) draws about 14 kW typical, up to 15 kW maximum.
So 10 B200 servers need roughly 140 kW of IT load. At a PUE of 1.4, those same 10 servers drive about 196 kW of total facility consumption. If a quote gives you a facility-load number and you want to know how many servers it supports, divide by the PUE and then by the per-server draw.
Rack-scale systems like the GB300 NVL72 run at up to 142 kW per rack and require liquid cooling infrastructure the site must be built to support. For those deployments, confirm the site's rack-level power limit before anything else. Liquid vs Air Cooling for AI GPUs: Rack Density Explained covers what the site needs to support that density.
Redundancy and what it costs you in usable power
Redundancy is the second thing that shrinks your usable power. A site built to Tier III standards has no single point of failure and can be maintained without taking your equipment offline. That design requires parallel power paths, which means the site installs more capacity than it can ever run at once.
A 2N design means two fully independent power feeds, each capable of carrying the full load alone. In practice, each feed runs at no more than half its rated capacity so that either can take over instantly. If your contracted allocation is 200 kW in a 2N facility, the site is holding 400 kW of installed capacity to back it up. That cost is real, and it is part of what you are paying for.
For AI training, the redundancy is worth it. An unplanned outage mid-run can mean hours of lost compute and a checkpoint rollback. For inference, availability is the product.
Questions to take into a call with an operator
- Is the kW figure in this quote IT load or facility load?
- What is the usable power limit per cabinet, and how does that differ from the contracted allocation?
- Is billing metered or committed, and from what date does the commitment start?
- What is the site's PUE, and is it measured or designed?
- What redundancy configuration covers the power path to my cabinets?
- What is the maximum kW per cabinet the site can support today, and what cooling infrastructure does that require?
For a broader look at what colocation, which is space and power for hardware you own, commits you to beyond power, see What a colocation contract commits you to and Space and power: what a colocation site provides.
Sources
Per-server power figures for H100, H200 and B200 are planning figures widely used in data center design. DGX B300 power figures (14 kW typical, 15 kW maximum) and GB300 NVL72 rack power (142 kW maximum, 132 to 140 kW in operation) are from NVIDIA and Supermicro reference documentation. PUE is defined by the Green Grid consortium. Tier III design standards are defined by the Uptime Institute.
Share your requirements. We'll canvas the market.
Tell us what you need (for colocation space and power: market, kW or MW, cooling and timing; for reserved GPU capacity: region, GPU type, cluster size and timing) and we'll canvas colocation and neocloud operators in parallel. Shortlist in 48 hours. We don't sell hardware. We broker colocation space and power, and reserved GPU compute capacity.
Earlier conversations get better terms. When you engage early, we have time to negotiate with vendors before you need to commit. We're paid only when a deal closes, and the fee structure is agreed in writing before any introduction.