Power is the line item that most often makes a colocation quote look larger than expected, and the reason is almost always the same: the rack was sized by reading labels. A power supply's nameplate rating is the maximum the supply can deliver, not what the equipment behind it consumes. Real draw is usually a fraction of it, and the gap compounds every time you add a device.
Step one: measure the draw
If the equipment is already running, this is the easy part. A rack PDU with per-outlet metering, or a clamp meter on the feed, gives you a number that is true for your workload rather than for a specification sheet. Note that draw varies with load: an idle database server and the same server at full query volume are different numbers, so measure under representative load, and take the peak rather than the average.
If the equipment is not running yet, plan from published power figures and assume you are overestimating. Then verify with a meter in the first week and adjust the plan while it is still cheap to change.
Step two: derate the circuit
A circuit should not be run at its rated maximum continuously. The long-standing practice is to keep continuous load at or below 80% of the circuit rating, which leaves headroom for inrush current at power-on and keeps the breaker from tripping on a hot afternoon. This is the number to compare your measured draw against.
Take a common single-phase 208V, 20A circuit. The arithmetic, which you can check against your own equipment:
- 208 volts × 20 amps = 4,160 VA at the circuit rating.
- 80% of that is 3,328 VA of continuous load the circuit should carry.
So the question for that circuit is not "is my rack under 4,160 watts?" but "is my measured peak under about 3,300?" A full rack on a single such circuit therefore has a real ceiling of roughly 3.3 kW of usable, continuous load — which is the figure to carry into the planning conversation, not the nameplate total.
Step three: plan redundancy, not just capacity
Servers with two power supplies are not drawing half their power from each one in the sense that matters here. Under normal operation the supplies share the load, but the design intent is that either feed alone can carry the whole server if the other fails or is taken down for maintenance. So when you size the two feeds, size each one for the full load — not for half of it.
That is why a rack with dual-corded equipment is specified as two independent circuits each capable of the entire draw, on separate power paths. Sizing each feed at 50% and calling it redundant produces a rack that survives a power supply failure and fails on a circuit maintenance window, which is not redundancy. Ask explicitly which feeds are on separate paths; two circuits from the same panel are one point of failure wearing two labels.
A note on power-supply efficiency: the draw at the wall is higher than the power the components inside use, because the supply itself loses some in conversion. That loss is already included in anything you measure with a meter, which is one more reason to prefer measuring over calculating. If you must calculate, divide the expected internal load by the supply's efficiency rating rather than treating them as equal.
Step four: leave room for the growth you actually expect
The most common mistake in the other direction is buying for a growth curve that never arrives, and paying for empty units for three years. The correction is not to plan for zero growth; it is to separate the two questions. Size the circuit for the workload plus a modest margin, because that is the part that is disruptive to change later. Size the commitment — the term and the cabinet allocation — for where you are now, because moving up a plan is a change to an agreement rather than a new deployment.
Two other numbers belong in the same conversation. Heat is the same figure as power, since essentially all the electricity a server draws leaves the rack as heat — dense compute needs to be somewhere that can actually remove it. And remote hands is priced by the hour in most facilities, so a small task stays a small charge; the question worth asking is not whether hands are available but what they cost at 3am.
Published rates remove most of the guesswork from the last part. You can work the arithmetic yourself before you speak to anyone, which is the only way to know whether a quote is reasonable.