Inference unit
The provider or infrastructure consumption used to produce model output, such as tokens, requests, compute time, accelerator time, or provisioned capacity.
Why it matters
Billing units are necessary inputs to a forecast, but they must be translated into the venture's useful-task unit.
Measure it with
- Input and output units
- Requests
- Provisioned and idle capacity
- Retries
Do not assume
Two providers define, meter, bundle, or price the same unit in the same way.