DECISION LAB
Inference Economics
Explore cost per million accepted output tokens for a defined serving scenario.
Keep the boundary visible.
JavaScript is required for local calculations.
No information is sent.
Formula and units
Accepted output tokens = GPUs × hours × 3,600 × measured output tokens/s/GPU × serving load × accepted-output share. Period cost = GPU-h × supplied hourly price + other costs + egress GB × egress price. US$/million accepted output tokens = period cost ÷ (accepted output tokens / 1,000,000).
How to read the result
This linear scenario does not simulate prompt processing, batching, queueing, latency, cache behaviour, multi-GPU scaling or retries. Include all serving costs in the supplied period boundary. When accepted output is zero, cost per million is undefined and the tool reports no useful output. Load and acceptance are distinct from measured GPU utilisation.
Show the method
Accepted output tokens = GPUs × hours × 3,600 × measured output tokens/s/GPU × serving load × accepted-output share. Period cost = GPU-h × supplied hourly price + other costs + egress GB × egress price. US$/million accepted output tokens = period cost ÷ (accepted output tokens / 1,000,000).
Recalculate after changing inputs. The download records inputs and results together. There is no live data connection and no automatic persistence. Reloading may discard your inputs.