Free tool · no sign-up

AI energy and carbon calculator: kWh and CO2e per million tokens

Estimate operational electricity and carbon from GPU power, throughput, utilization, PUE and grid intensity. The defaults are planning assumptions; replace them with your own measurements.

Workload and facility

Advanced: batch mode and custom carbon intensity

Batch mode only changes the utilization assumption. It does not assume a lower carbon grid. If you schedule batch work into a cleaner hour, enter that hour's intensity in the custom field.

Rent a GPU

Throughput is an input, not a benchmark. Use a number you measured on the same GPU, model, quantization and context length. The default 1,000 tokens/sec/GPU is a planning value, not a claim about any model.

Result real-time

Enter a throughput to see energy per million tokens.
kWh / 1M tokens—
gCO2e / 1M tokens—
Monthly energy—
Monthly CO2e—
Monthly electricity cost—
Facility power—

Operational electricity only: GPU TDP × utilization × PUE. It excludes embodied carbon, networking, storage, water and training. Arithmetic, not a measurement.

Energy per million tokens is a division problem

The core formula is small enough to check by hand:

facility_power_W = gpu_tdp_W × gpu_count × utilization × PUE
kWh_per_1M_tokens = facility_power_W ÷ (tokens_per_second × 3.6)

The 3.6 is unit conversion: one watt for one hour is 3.6 kJ, and one kWh is 3.6 MJ. With one 700 W GPU at 45% utilization and PUE 1.52, facility power is 479 W. At 1,000 tokens per second, that is 479 ÷ 3,600 = 0.133 kWh per million tokens. Double the throughput and the energy per token halves; double the PUE and it doubles.

This is why “kWh per 1M tokens” is not a property of a model. It is a property of the serving stack at a moment in time. A 70B model on an idle H100 and a 7B model on a saturated L4 can have the same energy per token if their throughput-to-power ratios match. The model matters because it changes throughput and memory pressure, not because it has an intrinsic energy label.

Thirty GPUs at four throughput levels

The table below applies the formula to each GPU's published TDP at 45% utilization and PUE 1.52. The throughput columns are assumptions. They let you compare the shape of the trade-off, not claim that a GPU reaches any particular speed on your model.

kWh per 1 million tokens at 45% utilization and PUE 1.52. GPU TDP from vendor datasheets; throughput is an input, not a measured benchmark.
GPUMemory GBBandwidth GB/sTDP W250 tok/s500 tok/s1,000 tok/s2,000 tok/s
NVIDIA H100 SXM803,3507000.5320.2660.1330.067
NVIDIA H100 PCIe802,0003500.2660.1330.0670.033
NVIDIA H200 SXM1414,8007000.5320.2660.1330.067
NVIDIA B200 SXM1928,00010000.7600.3800.1900.095
NVIDIA A100 SXM 80GB802,0394000.3040.1520.0760.038
NVIDIA A100 PCIe 80GB801,9353000.2280.1140.0570.028
NVIDIA A100 SXM 40GB401,5554000.3040.1520.0760.038
NVIDIA A30249331650.1250.0630.0310.016
NVIDIA A10246001500.1140.0570.0280.014
NVIDIA A10G246001500.1140.0570.0280.014
NVIDIA L40488643000.2280.1140.0570.028
NVIDIA L40S488643500.2660.1330.0670.033
NVIDIA L424300720.0550.0270.0140.007
NVIDIA T416320700.0530.0270.0130.007
NVIDIA V100 SXM2 32GB329003000.2280.1140.0570.028
NVIDIA V100 PCIe 32GB329002500.1900.0950.0480.024
NVIDIA RTX 4090241,0084500.3420.1710.0860.043
NVIDIA RTX 4080 Super167363200.2430.1220.0610.030
NVIDIA RTX 3090249363500.2660.1330.0670.033
NVIDIA RTX 3080107603200.2430.1220.0610.030
NVIDIA RTX 5090321,7925750.4370.2180.1090.055
NVIDIA RTX 5080169603600.2740.1370.0680.034
NVIDIA RTX 5070 Ti168963000.2280.1140.0570.028
NVIDIA RTX A6000487683000.2280.1140.0570.028
NVIDIA RTX 6000 Ada489603000.2280.1140.0570.028
NVIDIA RTX 5000 Ada325762500.1900.0950.0480.024
AMD Instinct MI300X1925,3007500.5700.2850.1420.071
AMD Instinct MI250X1283,2005600.4260.2130.1060.053
AMD Instinct MI210641,6003000.2280.1140.0570.028
Intel Gaudi 31283,7006000.4560.2280.1140.057

At 250 tokens/sec/GPU, a 700 W H100 uses about 0.53 kWh per million tokens; at 2,000 tokens/sec/GPU it uses about 0.07 kWh. The 7.6× spread is throughput, not hardware efficiency.

Model size changes the GPU count, not the formula

A model has to fit in GPU memory before it can serve a token. At 4-bit quantization, weights take roughly 0.5 GB per billion parameters, plus KV cache and runtime overhead. That gives a minimum GPU count, and the count multiplies facility power and throughput together. The table uses an 80 GB GPU and a 90% memory budget; it is a sizing check, not a throughput claim.

Minimum 80 GB GPUs for 4-bit weights and the resulting facility power at 45% utilization and PUE 1.52. KV cache and activation memory are not included; add headroom for long contexts.
Model4-bit weightsMinimum GPUsGPUsFacility kWM tokens/hour at 1,000 tok/s/GPUkWh / 1M tokens
7B≈4 GB1 × 24 GB10.4793.60.133
13B≈7 GB1 × 24 GB10.4793.60.133
32B≈18 GB1 × 24 GB10.4793.60.133
70B≈39 GB1 × 80 GB10.4793.60.133
120B≈66 GB1 × 80 GB10.4793.60.133
405B≈223 GB3 × 80 GB31.43610.80.133

Energy per token stays flat because power and throughput both scale with GPU count. The 405B row draws three times the power and serves three times the tokens at the same per-GPU rate. Real tensor-parallel serving adds interconnect power and rarely scales perfectly; use a measurement when you have one.

PUE: the facility multiplier

PUE is total facility power divided by IT power. A PUE of 1.52 means every 1 W of GPU power pulls 1.52 W from the grid, with the extra going to cooling, power distribution and losses. The Uptime Institute's 2026 survey reports an industry-wide annual average of 1.52; facilities commissioned in the previous five years average 1.48, and large facilities of 20 MW or more average 1.44.

Use the number your facility publishes, and check whether it is annual, monthly or instantaneous. A design PUE is not an operating PUE. For a colocation quote, ask for the contracted PUE and the measurement boundary. For a cloud GPU, the provider may publish a region-level PUE; if it does not, 1.5 is a defensible planning default, not a fact about that provider.

Grid carbon intensity: annual averages, not hourly truth

The carbon table uses Ember's latest annual intensity values via Our World in Data. The spread is large: Sweden and Norway are under 40 gCO2e/kWh, France is about 41, the United States about 384, India about 670. At 0.133 kWh per million tokens, the same workload emits roughly 5 g in Sweden and 89 g in India.

Annual average carbon intensity, latest available year (2025). The gCO2e column uses 0.133 kWh per 1M tokens, the 700 W / 1,000 tok/s / 45% / PUE 1.52 case.
RegiongCO2e/kWhgCO2e / 1M tokenskgCO2e / 1B tokens
Sweden3550.005
Norway2840.004
France4160.006
Brazil110150.015
Spain154200.020
Canada191250.025
United Kingdom217290.029
Netherlands254340.034
Ireland257340.034
Italy285380.038
Germany330440.044
United States384510.051
South Korea417550.055
Japan477630.063
Singapore497660.066
Australia525700.070
China525700.070
Poland589780.078
India670890.089
World458610.061

Source: Ember via Our World in Data, “Carbon intensity of electricity”, retrieved 20 November 2026. Values are national annual averages; subnational grids and hourly marginal intensity can differ substantially.

Batch versus real-time

Real-time chat keeps a GPU resident and often runs at low utilization between requests. Batch work can queue enough requests to keep the same GPU busy, which raises utilization and improves tokens per kWh, provided the GPU does not hit a power or thermal limit. The calculator's batch mode changes utilization from 45% to 85%; it does not change the carbon intensity unless you enter a different value.

There are two separate savings. The first is utilization: the fixed power of an idle GPU is spread over more tokens. The second is carbon-aware scheduling: running deferrable work when the grid is cleaner. The second only helps if the workload can wait and the grid signal is real. A batch job that must finish in five minutes cannot wait for a windy hour.

Do not double-count. If you raise utilization from 45% to 85%, facility power rises by 89%, but throughput may rise by more than 89% because the GPU was previously idle. Measure both before and after; the ratio is the only honest efficiency claim.

Methodology and limits

The calculator uses published GPU TDP, a user-supplied throughput, a user-supplied utilization and PUE, and a regional annual carbon intensity. It does not model GPU frequency/voltage curves, memory power, CPU host power, networking, storage or cooling water. It does not include embodied emissions from manufacturing. It does not assume that any GPU achieves a particular tokens-per-second figure on any model.

For a procurement estimate, replace the default throughput with a measurement from your own stack at your p95 context length. Then replace the default PUE with the facility's reported value. The remaining uncertainty is the grid; use the annual average for reporting and an hourly signal for scheduling. Related: renting GPU capacity, local AI deployment, and what actually drives inference cost per million tokens.

FAQ

How many kWh does 1 million tokens use?

It depends on throughput and facility overhead, not on the model name. At 700 W per GPU, 45% utilization and a PUE of 1.52, a GPU drawing 479 W and serving 1,000 tokens per second uses about 0.13 kWh per million tokens. At 250 tokens per second the same GPU uses about 0.53 kWh. Enter your own throughput; the calculator does not assume a benchmark.

What PUE should I use?

Use the PUE your facility reports. The Uptime Institute's 2026 survey puts the industry-wide annual average at 1.52, with newer and larger facilities often between 1.3 and 1.48. A hyperscale facility may be lower; a small on-premises server room may be well above 1.5. PUE multiplies the GPU power, so a 0.2 difference changes the result by about 13%.

Does batching reduce energy per token?

It can, but not automatically. Batching raises GPU utilization and spreads the fixed cost of weights and idle power across more tokens. If the GPU is already saturated, adding batch does not reduce the energy per token; it increases throughput at roughly the same power. The saving comes from running the same hardware closer to its power-performance sweet spot, not from the word “batch”.

Where does the carbon intensity data come from?

The regional values are Ember's latest annual carbon-intensity figures, published through Our World in Data and retrieved on 20 November 2026. They are annual averages for 2025, not hourly marginal emissions. A grid can be much cleaner or dirtier at a particular hour, so use the custom-intensity field if you are scheduling batch work.

Does this include embodied carbon or water use?

No. The calculator covers operational electricity for the GPUs plus the facility overhead captured by PUE. It does not include GPU manufacturing, server networking, storage, cooling water, or the energy used to train the model. Those are real costs, but they need a different dataset and a different boundary.

Sources and verification

Primary sources were opened and checked on 20 November 2026. Product terms and datasets change; the linked source is authoritative.

Skip the hardware and the PUE spreadsheet

Plugsky serves hosted models through one OpenAI-compatible API. You get published plan caps instead of a GPU bill, and the free tier starts at 100 requests and 50K tokens per day.

See plans → Browse models