The real energy cost of a GPU server: the calculation method
For a GPU server, the question is almost never space: a few rack units are enough. The question is electricity, what it imposes on cooling, and what it costs every year. Here is the method to go from watts to budget, formula by formula.
Estimate my cost →Estimate the peak: cards, plus host, plus power supply #
The starting point is each card's maximum power (the TDP, published by the manufacturer), multiplied by the number of cards. Add the host: processors, memory, disks and fans consume as much as a small server on their own, several hundred watts. Finally divide by the power supply efficiency (around 0.9): the wall socket always draws more than the components. A model training run saturates the cards continuously for days: size for the peak, without hoping for dips.
From kW to kVA: the subscription #
The data center does not sell your average consumption, it sells a subscribed power, expressed in kVA. Subscribing at the exact measurement trips the breaker at the first peak: add a margin, on the order of 15%. The electrical feed follows common steps: a 16 A single-phase feed carries about 3.5 kVA, a 32 A single-phase about 7, a 32 A three-phase about 22. Each rack should have two feeds (A and B), each able to carry the full load.
PUE: the energy you don't see #
Cooling and distributing electricity also consumes power. PUE (Power Usage Effectiveness) measures that overhead: a PUE of 1.4 means that for every 1 kWh your machines consume, the site spends 1.4 in total. Depending on the contract, that overhead is billed in the energy rate or in the rent: the question to ask is whether the advertised rate includes cooling.
The yearly budget formula #
Yearly energy budget = average kW × 8,760 hours × PUE × your kWh rate. Average kW depends on usage: training runs close to the peak, spiky inference runs lower. With this formula and your own numbers, contract rate, site PUE, load profile, you get a budget comparable from site to site, without depending on any brochure.
Subscribed or metered: read the billing line #
Two models coexist: billing the subscribed power (you pay for the reservation, even unused) or the consumed energy (metered). At high density the gap between the two runs into thousands of euros per year: it is a question to ask before signing, not a first-invoice discovery.
Cooling follows thresholds #
Up to about 5 kW per rack, standard colocation air cooling is enough and nearly every site can do it. Around 10 kW you need aisle containment, already less common. Beyond that, rear-door heat exchangers then liquid cooling: equipment few sites offer at retail, negotiated site by site with long lead times. That is why spreading servers across more racks is often the cheapest trade-off: more space, far fewer constraints. For choosing the site itself, read our GPU/AI colocation guide.
Run the numbers on your configuration #
Our GPU hosting engine applies this method to your configuration: just type your hardware into the search (for example "8 RTX PRO 6000 for training") and you get servers, rack units, kVA to subscribe, cooling regime and the requirement sheet to compare site by site.
FAQ #
Add up the cards' TDP, add several hundred watts for the host (processors, memory, disks, fans), then divide by the power supply efficiency (around 0.9). For training, size for the peak: the cards run flat out continuously.
Average kW × 8,760 hours × PUE × kWh rate. PUE represents cooling and distribution energy; depending on the contract it is included in the rate or billed separately.
The ratio between the site's total energy and the energy consumed by IT equipment. A PUE of 1.4 means 40% extra energy for cooling and distribution.
Because cooling follows thresholds: beyond roughly 5 kW per rack you need equipment few sites sell at retail. More racks at moderate density often costs less than one extreme rack.
Written on 10 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers