What did the wall meters measure?
On October 20, 2026, Big Iron published a test of a 27B Qwen model running in Ollama on an RTX 4070 Ti SUPER and an external RTX 3060. A Tapo P115 smart plug monitored each enclosure. Counting both plugs matters here: the external card alone drew about 169 W under load. Original experiment, meter readings, and script.
| Measurement | Reported result |
|---|---|
| Idle power with model loaded | 143.3 W baseline from idle windows before and after the run |
| Power during generation | 452.6 W from median plug readings sampled every 5 seconds |
| Generation time | 151.9 seconds of decode time reported by Ollama |
| Generated output | 5,600 tokens, about 36.9 tokens per second |
How do 5,600 tokens translate to kWh per million?
Convert watts to kilowatts and seconds to hours. During decoding, the estimated energy was 0.4526 kW × (151.9 / 3,600 hours) ≈ 0.0191 kWh. Multiply by 1,000,000 / 5,600 to get 3.41 kWh per million output tokens.
To find the extra electricity that generation used, subtract the 143.3 W idle baseline first. Repeating the calculation with roughly 309 W gives 2.33 kWh per million. Both results extrapolate from the same short run. They use sampled power and decode time, so model startup, prompt processing, and idle hours between jobs need separate measurements for a monthly bill.
How much would that add to my electricity bill?
Multiply kWh by your own tariff. At $0.30 per kWh, the whole-machine figure rises to about $1.02 per million output tokens. The energy stays at 3.41 kWh. If you keep a dedicated server running all day, also count the hours it waits for work. Ten idle hours at the measured baseline add 1.43 kWh, even if the machine serves zero tokens in that time.
To compare with a cloud API, price the same prompts and answers on both sides. Add hardware and operating costs on the local side, and input and output charges on the API side. Compare electricity-only and fully loaded GPU costs.
What would the emissions be on the UK grid?
For a UK scenario, the government’s 2026 methodology gives 0.14396 kg CO₂e per kWh consumed for data year 2025, including grid losses. At that factor, 3.41 kWh corresponds to 0.49 kg CO₂e per million output tokens. Counting only the 2.33 kWh above idle gives 0.34 kg CO₂e. These estimates cover operational electricity. The UK is an example location for the calculation. UK government electricity conversion factors, 2026 methodology.
To reproduce the first estimate, open the calculator’s self-hosted section. For a month with one million output tokens, enter 3.41 kWh, set PUE to 1 for the wall-meter boundary, and enter 143.96 g CO₂e/kWh. Read the operational CO₂e result. For your own deployment, substitute your monthly metered kWh, regional factor, and any facility overhead outside the meter. Open Ganivra’s self-hosted AI carbon calculator.
How can I measure my own local model?
- Plug the host and any external GPUs into energy meters. Load the model. Record idle watts before and after the run.
- Run a representative batch. Record elapsed time, output tokens, and the change in the meters’ kWh counters over the same interval. Use a longer batch if the meter rounds small readings to zero.
- Divide the measured kWh by generated tokens and multiply by one million. For energy above idle, first subtract idle kilowatts multiplied by elapsed hours.
Save the model, quantization, prompt length, and concurrency with the result. Those details let you repeat the test after a model change and see whether the same task uses less electricity.