AI server electricity per year: kWh for 2, 4 and 8 GPUs, power limits and cooling

How much electricity an AI server uses per year

An AI server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards draws about 1.8 kW at the wall at full load by our estimate, so in a year of 8,760 hours it uses at most about 15,800 kWh; with four cards the ceiling is about 29,800 kWh and with eight about 55,200 kWh. Most company servers do not run at full load day and night, so the real figure is lower and depends on how many hours a year the server works hard and what it draws while it waits.

This article shows how to get from those inputs to a number you can multiply by your own electricity tariff. It is written for the finance director or managing director who has to put the running cost of an own AI server into a budget, and for the person who plans the room. We give no prices and no tariffs; the kWh figures are our estimates or arithmetic, labelled as such. The energy per answer, a metric for engineers, is in our guide to GPU energy per token; here the question is the annual total.

The formula in plain words

Electricity is billed in kilowatt-hours (kWh): a machine that draws 1 kW for one hour uses 1 kWh. The annual energy of a server is therefore its power in kW multiplied by the hours it runs at that power, added up over the different states it is in.

For an AI server two states matter. Under load, when it answers questions, reads documents or runs a training job, it draws close to its full power. Idle, with the model loaded and no requests, it draws much less, but not nothing: the processors, memory, fans, drives and the cards themselves stay switched on. A year has 8,760 hours (365 × 24), so:

annual kWh = load power × load hours + idle power × (8,760 − load hours)

Company servers usually stay switched on between working days, so that the assistant answers when people arrive and updates and backups can run at night. So the idle hours count, and for a server used only in office hours they are most of the year.

Full-load power of 2, 4 and 8 cards

The cards are the largest load. NVIDIA rates the RTX PRO 6000 Server Edition at up to 600 W, configurable down to 300 W in its product brief SP-12355-001. The processors, memory, fans, drives and the losses of the power supplies come on top. Our estimate adds the cards, two AMD EPYC processors at their rated power and 15 per cent for the rest of the server, the method of our article on RTX PRO 6000 Server Edition power at 600, 450 and 300 W.

Our estimate: card power from NVIDIA's product brief SP-12355-001 (600, 450 and 300 W); two AMD EPYC 9135 for two cards, 9355 for four cards and for the eight RTX PRO 4500 Server Edition (165 W each), 9555 for eight RTX PRO 6000; plus 15 per cent for the rest of the server.

These are planning figures for full load. Dell's datasheet for its eight-card servers with RTX PRO 6000 Server Edition gives "~8 kW" per node, more than our 6.3 kW, so for an eight-card server the room and the feeds should be planned on the maker's figure. An eight-card server with H200 NVL cards has the same 600 W maximum per card, so the eight-card row gives its order of magnitude as well.

The annual ceiling at full load

Multiplying the full-load power by 8,760 hours gives the most a server can use in a year. A budget that uses this figure will not be exceeded by the server itself, but for most companies it is far too high.

Our arithmetic: full-load power from the table above (Dell's figure from its RTX PRO 6000 datasheet for PowerEdge XE7740 and XE7745) times 8,760 hours, rounded.

The ceiling is a fair figure for one case: a server that trains or fine-tunes models, or processes a queue of documents, around the clock. A training node that is busy every night and weekend comes close to it. An assistant for office staff does not.

Office hours, overnight jobs and idle time

For a company assistant, the load follows the working day. Take 250 working days of 10 hours as an example, 2,500 hours a year. If the server ran at full load in all of those hours, which is the upper bound, the load part of the year would be:

Our arithmetic with example values: 250 working days × 10 hours = 2,500 load hours; the remaining 6,260 hours idle; 100 W × 6,260 h = 626 kWh.

The last column is the open question. Over the 6,260 hours outside the working day, every 100 W the server draws while idle adds 626 kWh, and every 1 kW adds 6,260 kWh. We found no idle figure for servers with these cards in the datasheets, technical guides and reviews we read in October 2026; NVIDIA's brief gives the power limits and server makers give maximum figures. The idle draw depends on the processors, the memory, the fan settings and whether a model is loaded on the cards, so it has to be measured on the server itself.

Overnight work changes the split. If the same four-card server also reads a document queue for 6 hours on each working night, as in our guide to batch LLM inference on-premise, the load hours grow by 1,500 to 4,000 a year. At full load that is about 13,600 kWh for those hours, and the idle hours fall to 4,760, so each 100 W of idle draw adds 476 kWh instead of 626. Batch jobs at night use hours that would otherwise be idle, so the extra energy is the difference between load and idle draw over those hours.

Two cautions work in the other direction. An assistant used by a few hundred people does not hold the cards at full power through the whole working day; the board power follows the number of requests in flight. And a GPU waiting for requests still draws idle power, which our energy-per-token guide counts against the tokens produced in the busy hours.

We build these servers with two, four or eight cards and size the processors and power supplies per configuration; send us your planned users, working hours and any night jobs, and our team replies with a configuration and its power per feed.

Power limits: what 450 W and 300 W change

The RTX PRO 6000 Server Edition runs at up to 600 W by default with a power cable strapped for 600 W, and its limit can be lowered to 300 W, according to NVIDIA's brief. A lower limit reduces the full-load power and with it the annual ceiling. For four cards, 450 W brings the ceiling from about 29,800 to about 23,700 kWh a year, and 300 W to about 17,500 kWh.

The price of the cap is throughput, and it depends on the work. Our power article reports two published measurements. StorageReview measured 2 to 5 per cent less LLM throughput at short context and 4 to 9 per cent at long context at 450 W, with more output per rated watt. Its figures come from a Workstation Edition in a tower workstation, not from a Server Edition; StorageReview writes that the two editions profile identically in its testing. Delta Computer's compute sweep of the Server Edition kept about 86 per cent of its tensor result at 450 W and about 54 per cent at 300 W. For a chat assistant, whose speed depends mainly on memory bandwidth, 450 W costs little; for training and long-context work the cap makes jobs run longer, so part of the energy saved per hour is spent on extra hours. Whether a cap saves energy per job is measured as our energy-per-token guide describes.

The cap also matters for the building. A four-card server at 600 W, about 3.4 kW, is above what a 230 V 16 A feed carries at 80 per cent (2.94 kW); at 450 W, about 2.7 kW, it stays below it by our estimate. Our guide to GPU server power on 16 A, 32 A and three-phase feeds works through the feeds, and our article on power limits from 300 W to 600 W shows how the limit is set.

Cooling comes on top

All the electricity a server draws ends up as heat in the room, as our guide to a GPU server in an office or small server room explains. That heat has to be removed, and removing it costs electricity too.

In a data centre. The operator's power usage effectiveness (PUE) describes the overhead. Commission Delegated Regulation (EU) 2024/1364 defines it as the total energy consumption of the data centre divided by the energy consumption of its IT equipment, both in kWh. A PUE of 1.5, as an example value, means 0.5 kWh of cooling and distribution on top of every kWh the server uses. The same regulation requires operators of data centres with an installed IT power demand of at least 500 kW to report these figures each year. Whether and how the overhead appears on your invoice is in your hosting contract. Our guide to a private LLM in an EU data centre covers that option.

In your own server room. Here the air conditioner's electrical input is the overhead. Its datasheet lists the cooling capacity and the electrical power input; the ratio of the two is its efficiency ratio. A unit that removes 3.4 kW of heat with a ratio of 3, as an example value, draws about 1.1 kW while it does so. The unit has to run whenever the server runs, at night, at weekends and in winter, so check its operating range at low outdoor temperatures, as our guide to preparing a server room for a GPU server sets out. If the server runs on an uninterruptible power supply, count its heat as well, as the office guide notes.

Running costs: a worksheet for your budget

1. Choose the configuration and its full-load power from the table above: 1.8, 3.4 or 6.3 kW at 600 W, or the capped values.
2. Count the load hours per year: working hours of the assistant, plus night or weekend jobs. Our example is 2,500 hours for office use and 4,000 with nightly batch work.
3. Multiply full-load power by load hours. This is the upper bound of the load part.
4. Add the idle part: idle power times the remaining hours (8,760 minus load hours). Until the server is measured, show it as a line "idle draw, measured on delivery" or use the 626 kWh per 100 W step for 2,500 load hours.
5. Add cooling: multiply by the data centre's PUE, or add the air conditioner's electrical input for the hours it runs.
6. Multiply the total kWh by your tariff per kWh from your energy contract, with any grid or capacity charges your contract bills separately.

The full-load ceiling from the annual table is the safe upper value for a budget line; the worksheet gives the expected value once the hours are known.

Measuring on the delivered system

The final figure comes from the server you receive. Each server is assembled and tested before delivery, and the power estimate is confirmed by a measurement on the delivered system. Read the input power at the server's management controller (BMC) or at a metered power distribution unit (PDU), once idle with the model loaded and once under your real load, and multiply by your hours. The cards report their own share: nvidia-smi gives each GPU's "Average Power Draw", which NVIDIA defines as the average power draw for the entire board for the last second, in watts. Our guide to GPU server acceptance testing shows how to put such readings into your own acceptance test.

Repeat the measurement after a few weeks of real use. The share of hours under load is the number that decides the annual total, and it is easier to read from the server's own records than to guess before the project starts. We size the configuration, including the power limits and the power supplies, before we quote; tell us your feed per rack position and whether the server goes into your own room or a data centre, and we return the configuration with the power per feed.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two, four or eight RTX PRO 6000 Server Edition cards at 600 W or capped, and with RTX PRO 4500 Server Edition cards, assembled and tested before delivery, delivered across the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our guide to buying a GPU server for LLMs lists every configuration.

FAQ