Home › Guides › Hardware, power and cost

PUE and the Real Electricity Cost of Running GPU Servers

Updated October 21, 2026 · By the Xingyu Zhitong editors

The power bill for a GPU server is easy to estimate badly. The usual shortcut multiplies the GPU's rated wattage by hours and a tariff, and ignores everything that makes the real figure larger: the rest of the server, the cooling that removes the heat, power-conversion losses, and the fact that GPUs rarely draw exactly their rated maximum. This article walks through a more careful estimate and explains PUE, the metric used to capture facility overhead.

What PUE is, and what it is not

Power Usage Effectiveness is defined by The Green Grid as total facility energy divided by IT equipment energy. A PUE of 1.0 would mean every watt entering the building reaches servers, storage and network gear; anything above 1.0 is overhead such as cooling, power distribution losses and lighting. A PUE of 1.5 means that for every 1 kWh the IT equipment uses, the facility consumes 1.5 kWh in total.

Uptime Institute's 2024 Global Data Center Survey reported an industry-average PUE of 1.56, essentially flat for several years, though newer and larger facilities achieve noticeably lower values. Be careful comparing PUE figures from different sources: some are per-site averages, others are weighted by capacity, and those methods give different results (Uptime has noted this distinction itself). A hyperscale operator advertising a very low PUE and a small colocation room with air conditioning built for office loads live in different worlds.

PUE is also not a measure of how efficient the computing is. It says nothing about how many tokens you generate per kWh. A facility with a great PUE hosting idle servers still wastes energy; a facility with a mediocre PUE running well-utilized servers may deliver more work per kWh than a near-empty efficient one. Use PUE to scale the overhead, not to judge the workload.

Step 1: Estimate the IT load

Start from component power, not the GPU alone. For a typical eight-GPU server:

• GPUs: NVIDIA lists H100 SXM at up to 700 W (configurable), so eight GPUs reach up to 5.6 kW. NVIDIA's DGX B200 page lists about 14.3 kW maximum for the whole eight-GPU Blackwell system, which gives a sense of how much a complete system draws beyond the GPUs themselves.
• CPUs, memory, NICs, storage, fans: often another 20 to 40 percent on top of GPU power in dense servers. Check the system vendor's published maximum or typical figures rather than guessing.
• Network switches and other rack gear if you want the complete picture.

Then apply a load factor. Rated maximum is a ceiling. Inference workloads at modest batch sizes can draw well below it, while sustained training can sit near it. If you can, measure the server under your real workload (for NVIDIA GPUs, the standard tools report board power) and use the average. The planning guide on PSU sizing explains why you must keep the ceiling for sizing power supplies and circuits even when the average is lower.

Step 2: Apply PUE and hours

Facility energy per year equals IT load in kW × PUE × hours. A year has 8,760 hours. Multiply by your price per kWh. The tariff is the other big uncertainty; the U.S. Energy Information Administration publishes average retail electricity prices by sector and state, and commercial or industrial rates differ from residential ones. Colocation providers often bill differently again: a fixed fee per kW of committed power, sometimes with PUE already built in, sometimes billing metered electricity on top.

Worked example (example assumptions throughout)

Assume a hypothetical eight-GPU server with a measured average IT draw of 8 kW, hosted in a facility with PUE 1.4, at an example tariff of $0.12 per kWh, running continuously.

• Facility power: 8 kW × 1.4 = 11.2 kW.
• Energy per day: 11.2 × 24 = 268.8 kWh.
• Energy per year: 11.2 × 8,760 = 98,112 kWh.
• Cost per year: 98,112 × $0.12 = $11,773, or about $981 per month.

Now see how the assumptions move the answer. With PUE 1.56 (the Uptime survey average) instead of 1.4, the annual figure becomes 8 × 1.56 × 8,760 × 0.12 = $13,119. With a poorly cooled room at PUE 2.0 it is $16,819. Using a measured average draw of 8 kW rather than, say, a 10 kW worst-case estimate cuts the bill by 20 percent. Doubling the tariff to $0.24 doubles every number. The sensitivity ranking is usually tariff first, then load factor, then PUE, which is why negotiating or locating by electricity price often beats shaving PUE. You can plug your own values into the HGX electricity cost calculator.

Per-GPU and per-token views

For comparing against cloud rental, convert to cost per GPU-hour: $11,773 per year ÷ (8 GPUs × 8,760 h) ≈ $0.17 per GPU-hour in this example. Electricity is therefore often a minor line next to hardware depreciation; in the break-even method in renting versus buying, capital and overhead dominate at typical tariffs. It becomes material in high-tariff regions, in always-on deployments over many years, and for consumer-grade deployments with poor facility efficiency.

Cost per token follows by dividing hourly cost by tokens served per hour. If a node serves 2,000 tokens per second at steady state (an example number), that is 7.2 million tokens per hour; and electricity costs $11,773 ÷ 8,760 h ≈ $1.34 per node-hour, then electricity costs roughly 1.34 ÷ 7.2 ≈ $0.19 per million tokens. The server's throughput, covered in the tokens-per-second guide, matters as much as its watts.

Home and office deployments

Smaller setups do not have a published PUE, but the same logic applies with an effective overhead you can estimate. A single workstation with an RTX 4090 (NVIDIA lists 450 W total graphics power for the reference card) plus a CPU and other components may draw 600 to 800 W under combined load; that is an assumption to verify with a wall-socket power meter, which costs little and removes the guesswork. At 700 W for 8 hours a day and a residential price of $0.15 per kWh (example), the yearly cost is 0.7 × 8 × 365 × 0.15 ≈ $307. Air conditioning in summer adds to this: if the room needs cooling, each kWh of heat may need a fraction of a kWh more to remove, depending on the air conditioner's efficiency. In winter, the heat may offset some heating costs, which is a legitimate effect but one that depends on your heating source and not worth counting unless it is large.

For a home user deciding between a local GPU and a rented one for occasional jobs, the electricity line is almost always dwarfed by the purchase price, which is why the break-even method focuses on utilization. Check the supply side too: a high-end PC on a shared household circuit competes with kettles and heaters, so see the PSU sizing article for how to size the supply for peaks.

Reading a colocation quote

Colocation contracts for GPU servers are usually priced per kilowatt of committed power per month, with a rack or space fee on top. Ask these questions before comparing quotes: Is the kW figure the contracted maximum or the metered average? Is cooling included, or billed through a PUE multiplier on metered usage? What happens if you exceed the commitment: a surcharge, a throttle, or a tripped breaker? Does the facility support the density you need per rack, or will you have to leave half the rack empty? A cheap per-kW price can turn expensive if you can only populate half the rack. Convert every quote to cost per used kW-month at your expected draw so offers are comparable.

Where overhead hides

1. Cooling. Almost every watt consumed by a GPU becomes heat that must be removed. Air cooling at high density needs large airflow and cold-aisle management; liquid cooling can reduce fan and chiller energy. The trade-offs are discussed in air versus liquid cooling.
2. Power conversion. UPS, transformers and PDUs lose a few percent each. Efficient designs minimize conversion stages.
3. Partial load. Fans and some components do not scale to zero when GPUs idle. An idle eight-GPU server still draws a meaningful baseline, so turning machines off or consolidating workloads has real value if your usage is bursty.
4. Demand charges. Commercial tariffs may bill on peak kW in addition to energy. A cluster that starts all jobs at once can set a peak that costs money all month.

Practical checklist

• Get the server vendor's maximum and typical power figures in writing.
• Measure real draw during a representative workload for at least an hour.
• Ask your facility or landlord whether PUE is included in the per-kW rate or applied separately, and whether the quoted tariff is energy-only or includes demand charges.
• Look up the commercial rate for your location in the EIA data rather than using a residential bill.
• Run the estimate at three PUE values (for example 1.2, 1.5 and 2.0) to see how much the facility matters.
• Check that circuits and breakers can carry the continuous maximum. Electrical codes treat continuous loads conservatively; consult a licensed electrician for installation.

This article is general planning information. Local electrical codes, utility tariffs and tax rules differ, and the numbers above are examples, not quotes.

Sources

• The Green Grid (originator of the PUE metric)
• Uptime Institute Global Data Center Survey 2024
• U.S. EIA Electric Power Monthly, average retail price of electricity
• NVIDIA H100 product page
• NVIDIA DGX B200 page

Figures are estimates based on published model configurations and vendor specifications. Real memory use, speed and power draw depend on the software stack, drivers, settings and workload, so test on your own hardware before you buy. Spotted a mistake? Email huhuwen09@gmail.com with the page address and we will check and correct it.