Power and cooling for a GPU rack: what to check before the servers arrive

The gap nobody budgets for

The hardware conversation is usually about the cards. The delivery conversation, three weeks later, is about a circuit breaker. Between those two moments sits a set of physical facts that no one owns: how much current the row can actually supply, where the hot air goes, and whether the floor can carry the weight.

None of this is difficult. It is simply nobody’s job by default: the vendor sells the server, the integrator configures it, and the facilities team finds out on the day the pallet arrives. This is the checklist we walk through before a quote goes out, and it is short enough to do yourself.

What the numbers actually are

Start with the honest comparison. The rack your existing servers live in was designed for a different era.

Densities per rack; the middle band is where most first AI deployments land

Now the systems themselves. These are maximum system draws, not card TDP, the number that matters when you size a circuit:

Put those together and the arithmetic gets uncomfortable quickly. Four DGX H100 systems are roughly 41 kW in one rack, more than three legacy racks combined. A single 400 V three-phase 32 A feed delivers about 22 kW, and we plan on 80% of that as headroom (the US NEC makes 80% mandatory for continuous loads; IEC breakers are rated for continuous current at their nominal value, but the margin is still good practice). Two feeds, minimum, and they need to come from somewhere.

Airflow: the arithmetic that decides the room

Every watt that goes in comes out as heat, and the air has to carry it away. The working figure is about 160 CFM per kW at an 11 °C rise across the server. That gives you a number you can check against your actual floor:

CFM figures assume a ~11 °C (20 °F) temperature rise; ~800 CFM per grate is a conservative planning figure at typical plenum pressures (ASHRAE cites up to ~1,900 CFM per tile on best-of-breed floors)

Two details ruin more deployments than the totals do. First, a raised floor delivers air per tile, and you cannot put six tiles in front of one rack; there is no room. Second, without hot/cold aisle containment the hot exhaust loops back around the rack and re-enters the intake; the room gets hotter without cooling anything. Above roughly 20 kW per rack, containment stops being an optimisation and becomes the thing that makes it work at all.

ASHRAE’s recommended inlet range for class A1 equipment is 18–27 °C. If your measured inlet is already at the top of that band with the current load, the new rack has no headroom before it even arrives.

Where air stops and liquid starts

There is no hard cutoff, but there is a practical sequence, and knowing which step you are on saves a lot of argument:

The rear-door heat exchanger is the step most people have not considered and the one that most often solves the problem. It is a radiator that replaces the rear door of the rack: hot air passes through it and leaves the rack at close to room temperature, so the room itself never sees the load. It needs a water loop to the rack, typically at 14 to 30 °C and above dew point rather than chilled, which is a real building project, but it is a smaller one than rebuilding the room, and it does not touch the servers.

Direct-to-chip liquid cooling is a different commitment. It changes the servers, the rack, the facility loop and the maintenance procedure, and it is worth it when the density genuinely demands it, not before.

The four checks that stop a delivery

These are the ones we have actually seen block a switch-on.

1. The circuit, in amps, not in kilowatts. Ask the facilities team for the rating of the circuit feeding the rack position and the current already drawn on it. A 32 A three-phase feed at 400 V is ~22 kW, and you plan at 80% of that as headroom. Two 10 kW systems on one such feed are already at the limit.

2. The socket type on the PDU. A C13 outlet carries 10 A, about 2.3 kW at 230 V. A GPU server with dual 2,700 W supplies needs C19 outlets, and a PDU full of C13s will not do. This is a cheap part to get wrong and a slow one to fix, because the replacement PDU has a lead time.

3. Floor loading. A fully populated GPU rack easily exceeds 1,200 kg. Many raised floors are rated well below that per tile, and the route from the loading bay to the rack position has to carry it too. Check the point load, not just the total.

4. The cable that came in the box. On cards with the 12V-2×6 connector (formerly 12VHPWR) the sense pins set the power limit, and some bundled server cables are configured for 450 W rather than 600 W. The card runs, reports Max Power Limit 450 W in nvidia-smi, and quietly delivers less than you paid for. The 600 W cable is a separate part number; put it in the bill of materials.

Passive cards need a server, not a workstation

Server-edition cards (the L40S, the RTX PRO 6000 Server Edition, the H100 and H200) have no fans of their own. They are designed to sit in front-to-back chassis airflow and they depend on it completely. NVIDIA’s own forum carries a telling case: an L40S installed in a Dell Precision workstation, idling at 100 °C. Nothing was faulty. The card simply had no airflow, and outside a server chassis that cannot be fixed.

The practical rule: passive card, server chassis, and the chassis fan profile set for the card. If the machine is going under a desk, it needs an actively cooled card: the Workstation or Max-Q edition.

A worked example

A team wants to run a 70B model for about forty internal users and asks for four GPU servers.

Four servers at roughly 6.5 kW each is 26 kW in one rack. That rules out the existing row, which is designed for 12 kW per rack. Airflow at 26 kW is about 4,200 CFM: five or six tiles, which will not fit in front of one rack, so containment is required. Two three-phase feeds are needed rather than one, and the rack weight lands near 550 kg with 4U PCIe servers, or well past a tonne with DGX-class systems.

None of that makes the project impossible. It makes it a project with a facilities line in the budget, discovered in week one instead of week twelve. The alternative that often wins on this maths: two racks at 13 kW in the existing row, which needs no building work at all and costs less overall than upgrading one position.

What we supply

Eurokommerz supplies the GPU servers, Nutanix NX nodes and professional NVIDIA cards, and we size the power and cooling with the quote rather than after it: feed ratings, PDU outlet types, airflow figures and rack weight, in writing. Deployment engineering comes from our partner Vixen.UNO under the same contract. Climate and refrigeration equipment is a business we have run under our own EU licence since 2014, so the cooling side is not a subject we outsource.

FAQ