1. Convert watts
Divide the measured wall draw or the board rating by 1,000. Eight 700 W GPUs are 5.6 kW of board power before CPUs, memory, fans and networking.
GPU infrastructure energy guide
GPU Server Power Consumption: Calculate Electricity Cost
Estimate the kWh and electricity bill for a GPU server by combining board power, host components, operating hours, utilization, and cooling overhead. The practical answer depends on the whole server, not the GPU label alone.
Cloud rate checkpoints updated from Sep 10, 2026. Electricity examples use an explicit $0.15/kWh reference tariff so you can substitute your local rate.
Short answer
What determines GPU server power consumption?
Start with the power you can measure at the wall. Add the GPU boards, CPUs, memory, storage, fans, network cards, and other host components. Then convert watts to kilowatts and multiply by the hours the server is actually running. A data-center estimate should also apply PUE, which captures cooling and facility overhead.
Monthly electricity cost = server kW x hours x utilization x PUE x tariff
For example, a 1.05 kW host running 300 hours at 80% utilization with PUE 1.3 uses 327.6 kWh. At $0.15/kWh, the energy line is about $49.14. That example is a planning calculation, not a quote: replace every assumption with a meter reading or a local utility rate.
GPU power consumption by model
These rows use GPU Cost data for board power and the lowest observed cloud rate. The 730-hour kWh column is a deliberately conservative board-only checkpoint at continuous rated power; it excludes the rest of the server and does not claim that every workload draws its TDP. The $0.15 column is a transparent reference tariff, not a universal utility price.
| GPU | VRAM | Board power | 730-hour board energy | Board cost at $0.15/kWh | Lowest cloud rate |
|---|---|---|---|---|---|
| H100 | 80 GB | 700 W class | 511.0 kWh | $76.65 | $1.47/GPU-h |
| H200 | 141 GB | 700 W class | 511.0 kWh | $76.65 | $2.30/GPU-h |
| A100 SXM | 80 GB | 400 W class | 292.0 kWh | $43.80 | $1.30/GPU-h |
| RTX 4090 | 24 GB | 450 W class | 328.5 kWh | $49.27 | $0.149/GPU-h |
| MI300X | 192 GB | 750 W class | 547.5 kWh | $82.13 | $2.39/GPU-h |
For a full-server bill, add host draw and multiply by utilization and PUE. The cloud rate is included for context only; it already bundles provider infrastructure and cannot be converted directly into watts.
How to calculate GPU electricity cost
Use separate lines for the GPU and the host so a change in utilization is visible. A simple monthly model is:
energy (kWh) = [GPU kW + host kW] x active hours x utilization x PUEelectricity cost = energy (kWh) x tariff ($/kWh)2. Choose hours
Use actual active hours for a project. A 300-hour training month and a 730-hour always-on month answer different cost questions.
3. Apply utilization
Utilization describes how much of the active window is doing useful work. Include idle, queue and warm-up time when the server remains powered.
4. Add facility overhead
PUE above 1.0 accounts for cooling, power distribution and other facility loads. A workstation at home may use a simpler wall-meter reading instead.
Worked example: one H100 SXM at 700 W plus 350 W of host draw is 1.05 kW. At 300 hours, 80% utilization, PUE 1.3 and $0.15/kWh, the estimate is 1.05 x 300 x .8 x 1.3 = 327.6 kWh, or $49.14. If the meter reads 1.35 kW at the wall, use that measurement instead of the component estimate.
Why GPU electricity estimates vary
GPU load
Training can hold a board near its power limit for long periods. Inference may alternate between bursts and idle, so a nameplate wattage can overstate the monthly average.
Host components
Two systems with the same GPU can have different CPUs, DIMMs, drives, fans and network cards. Those watts are paid every hour the server is on.
Cooling and PUE
Air conditioning, pumps, chillers and power conversion add facility energy. Apply a measured PUE for a data center or keep the wall meter as the final household answer.
Tariff and demand
Commercial bills can include time-of-use rates, demand charges and taxes. A per-kWh estimate is useful for comparison, but it is not a complete utility invoice.
Idle and standby time
A queued job, warm worker or paused VM can keep fans, memory and storage powered. Track active, idle and stopped hours separately.
Peak and transient draw
Monthly energy and electrical capacity are different. Use a power limit, PDU reading and vendor guidance for breakers, PSUs, rack density and transient events.
Treat TDP as a design and thermal reference, not as a promise about every workload. For a budget, record both an average watts value for energy and a peak value for power delivery. This keeps the electricity estimate useful without hiding the capacity constraint.
Measure a real GPU server before you budget
- Measure at the wall. A PDU, UPS or calibrated plug meter captures the host, not only the GPU board.
- Log several workload states. Record idle, warm-up, inference, training and shutdown values instead of one screenshot.
- Separate energy from capacity. Use average watts for kWh and peak watts for PSU, rack and breaker planning.
- Record the tariff window. Note currency, residential or commercial class, time-of-use period, taxes and demand charges.
- Recalculate useful work. A slower GPU with lower watts can lose if it takes twice as long; compare completed jobs, not watts alone.
How power cost changes a buy-versus-rent decision
Cloud GPU rental hides facility electricity inside the hourly price, which is useful when demand is uncertain. An owned server exposes the full energy line but can be cheaper when utilization is high and predictable. Compare the same useful workload: include setup time, storage, transfer, maintenance, idle hours, cooling, support and the cost of replacing an unavailable machine.
Buying is easier to justify when
- Measured utilization is high for several years.
- You already operate power, cooling, monitoring and spares.
- Data locality or latency makes local infrastructure necessary.
- The workload is stable enough to keep the GPU productive.
Renting is safer when
- Usage is bursty, seasonal or still being tested.
- You need a different GPU generation quickly.
- Facility upgrades would dominate the energy budget.
- Interruptions, egress or storage can be priced per project.
GPU server power consumption FAQ
How much power does an H100 server use?
An H100 SXM is a 700 W class accelerator, while an H100 PCIe is lower in the local data set. A complete server also includes CPUs, memory, storage, fans and networking, so use a wall meter or a documented system figure for the final estimate.
How much does it cost to run a GPU 24/7?
Multiply average wall watts divided by 1,000 by 730 hours and your tariff. At 1 kW and $0.15/kWh, the energy-only checkpoint is $109.50 per 30-day month before any demand charge or facility adjustment.
Is GPU TDP the same as electricity consumption?
No. TDP is a thermal and design reference for sustained load. Actual draw changes with workload, clocks, power limits, utilization and the host platform. Use TDP for a conservative capacity check and measured average watts for the monthly bill.
What is PUE in a GPU power calculation?
Power Usage Effectiveness is total facility energy divided by IT equipment energy. A PUE of 1.3 means each 1 kWh delivered to the server is associated with about 0.3 kWh of additional facility overhead in that model.
Does lower GPU power always mean lower cost per job?
No. A lower-watt GPU may take longer, use more idle hours or require a second card because of memory limits. Compare energy per completed job, throughput, memory fit and the time value of the result.
Where can I check GPU specifications?
Use the NVIDIA DGX H100/H200 system documentation for system composition and power-planning context. For a tariff benchmark, consult the U.S. Energy Information Administration electricity tables and replace the value with your own utility rate.
Method and limits
GPU Cost combines local D1 GPU specifications with provider pricing snapshots. Board power rows are not a substitute for a measured server draw, and cloud rates can change by region, availability, contract and spot conditions. Use this page to make assumptions explicit before a quote, facility design or procurement decision.
