Anyone wanting to bring a GPU server to a data centre usually starts by counting: a rack has 42 units, the server is 4U, so ten of them fit. That is not how it works, and the difference lies not in the units but in power and heat. What you need to know about A and B feeds and what you actually use applies here in full, only the numbers are an order of magnitude larger. This article works through the sums and ends with what you need to supply up front to get a quote that is right.
Units are not the limit
An ordinary server draws a few hundred watts in practice. A rack full of them rarely reaches the connected load. With GPU servers that flips. A modern GPU draws 350 to 700 W per card, data centre cards up to about 1000 W, and a server with eight heavy cards, counting CPUs, memory and power supplies, quickly sits around 10 kW. Two such servers draw more than a whole rack of web servers.
In the industry, racks of 40 to 120 kW are now common, but those are liquid-cooled and stand in halls built for it. In a regular colocation hall the limit is what the feeds deliver and what the cooling removes. The empty units underneath are not waste; they are the distance you need.
Worked example: three-phase 32 A per feed
A heavy rack supply in colocation is three-phase 32 A per feed. At 230 V per phase that is 3 × 32 × 230 ≈ 22 kW. That sounds like a lot, until you factor in redundancy.
With full A+B redundancy each feed has to carry your entire load on its own, because that is what happens when the other one drops out. Your usable budget is therefore 22 kW, not 44. That accommodates:
- Two servers of around 10 kW — 20 kW together, with a few kilowatts left for switches, storage and out-of-band.
- Or more servers with fewer cards — four servers with four cards each come out at the same power, but spread the heat better across the rack.
Two other pitfalls from the A/B article weigh more heavily here. The nameplate of a GPU server states the maximum of its power supplies, often well above what the machine ever draws; so measure, or work from the power of the cards plus a margin for the rest. And under inference load the peaks are sharper than on a web server: a batch that starts pulls full power within a second. The fuse behind your connected load does not bend.
Heat is power
Every watt your rack draws comes back out as heat. A 22 kW rack is a 22 kW heater, and the hall's cooling has to remove that heat. Next to power, that is the real limit, and it is harder to see because there is no fuse on it; you notice it through temperature warnings and cards throttling.
What counts in practice:
- Front-to-back airflow — the cold aisle supplies air at the front, the hot aisle removes it at the back. A server that breathes through its sides does not fit that pattern.
- Blanking panels in empty units — without panels the server pulls its own hot exhaust back in.
- Spread across the rack — two 10 kW heaters directly above each other make a hot spot; space between them helps.
- Depth and weight — GPU servers are deeper and far heavier than standard servers. Not every rack and not every floor is rated for them.
Whether air cooling is enough depends on the power per rack and on the hall. That differs per location, which is why the cooling question belongs in the conversation up front and not at delivery.
What you supply up front
A quote for GPU colocation is only as good as the information it is based on. Supply, per server:
- Power — nameplate and the expected actual draw under load. If the machine is already running somewhere, read it off for a week.
- Power supplies — number, rating and plug type, so the PDUs have the right outlets.
- Dimensions — height in units, depth, weight, and whether rails are included.
- Airflow — front to back or otherwise.
- Network — number and speed of the ports, and whether you want out-of-band exposed separately.
- Load pattern — continuous, office hours or peaks. That partly determines how your connected load and consumption relate.
- Growth — what gets added within a year. Raising a connected load is not a matter of one extra cable.
With that list we can say which feed fits, how many servers go in one rack and where cooling draws the line.
GPU colocation at Xyphen IT
We offer GPU colocation in BIT-2C in Ede, up to 96 A per rack, for example as three-phase 32 A per feed, always with an A and a B feed on separate PDUs. It is the same hall and the same network as our BIT colocation, so you connect to our EVPN fabric with a private VRF or a cross-connect to your other equipment. If you would rather not manage your own cards, AI inference as a service is the alternative; a combination, your own storage in colocation with inference on our cards, is in practice the most common. How power is billed is on the colocation page.
Have the list above to hand? Request a proposal or get in touch and we will work it out.