Almost every customer looking at a colocation quote for the first time asks the same question: if I connect my servers to two power supplies, do I use twice as much power? The answer is no — and the reason why explains at once how power in a data centre is put together.
Why there are two feeds
In the data centre, power reaches your rack over two independent paths: the A feed and the B feed. They come from separate distribution boards, hang off separate backup power systems and enter your rack independently of one another, each on its own PDU.
That is not a luxury; it is what keeps the system maintainable — and redundancy is safety. A distribution board needs work now and again, and a UPS needs testing. With two paths one half of the chain can be taken out while your equipment simply carries on running on the other. The same applies when something breaks unexpectedly: one path gone is then an event in the logs, not an outage.
Two power supplies, one consumption
A server with two power supplies does not draw twice as much power. The two supplies share the load: if the machine uses 400 watts, roughly 200 watts of that comes from each feed. Still 400 in total.
If the A feed drops out, the B feed takes over completely — those 400 watts then go over a single path in their entirety. Nothing changes for the server; it notices nothing at all, which is exactly the point.
What your meter counts is therefore the same, whether you hang off one feed or two. Redundancy costs no extra kilowatt-hours.
Where something is lost after all
There is one place where a dual setup does cost you money, and it is not on the meter but in the efficiency of the power supplies themselves.
A power supply is not equally efficient at every load level. The sweet spot sits roughly between 50 and 75 per cent of what it was built for; below that efficiency drops, and under 20 per cent it falls below 80 per cent for everything except the highest certification.
In a redundant setup each supply carries only half. A server drawing a third of its total supply capacity therefore leaves both supplies running at a sixth — precisely the region where efficiency falls away. What is lost there disappears as heat, and you pay for that heat twice: once on your meter and once in the cooling.
That is why the 80 PLUS programme explicitly measures redundant supplies at 10 per cent load as well. Titanium still has to reach 90 per cent there; the lighter certifications are allowed to fall well below it.
The practical conclusion is not "take a single feed then", but: do not size your power supplies more generously than necessary. A heavier supply "just to be safe" costs you efficiency in every hour of the year.
But capacity is reserved twice over
Here is the nuance that keeps the original question from being entirely unreasonable. If the B feed has to be able to carry your full load on its own, that capacity has to be there — all of the time, even though it normally only does half.
That headroom is held for you and cannot be sold to anyone else. That is what you pay for with redundancy: not for extra consumption, but for reserved capacity that has to be there the moment the other side fails.
That is why a colocation quote has two line items for power:
- A capacity charge per kW — the capacity held for you, whether you draw on it fully or not.
- Your actual consumption per kWh — what your meter really counts.
Connected load is not the same as consumption
The next pitfall: customers estimate their consumption from what is on the nameplate. That figure is the maximum draw of the power supply, not what the machine does.
A server that can draw 750 watts on paper often sits around 200 to 300 watts in practice — and only peaks higher under heavy load. Reserve capacity on the basis of nameplates and you are paying for air every single month.
Conversely, reserving too little is not an option. Your connected load is a hard limit with a fuse behind it, and that fuse does not negotiate. Go over it and the circuit trips — and in a dual setup that means: one side first, and as soon as the other feed takes over the whole load, often that one too.
Rules of thumb that work in practice:
- Measure, do not estimate. Are the machines already running somewhere? Then read the meter for a week under normal load.
- Count peak load, not peak rating. What happens during the nightly backup or the month-end close?
- Include planned expansion. Extra disks and a heavier CPU add up, and adjusting your connected load is not a matter of a quick phone call.
Power at Xyphen IT
With colocation we supply an A and a B feed on separate PDUs as standard, and we help you work out the connected load that matches your actual consumption — not the sum of your nameplates. If you do not know your consumption precisely yet, we can measure it for a period first and set the agreement afterwards.
That works the same way as with bandwidth: you fix a level and we settle any excess afterwards. Exactly how that billing works is set out in 95th percentile billing explained.
Curious what your setup needs? Request a proposal or get in touch — we will work it out for you.