Compute & Cost
Inference Yield
The metric behind the hype: value per token spent. Not how much AI you use, but what comes out of it. Two companies can have identical API bills and completely different returns, depending on whether tokens produce drafts nobody reads or speed up decisions. Thinking in yield means prioritizing use cases, not celebrating usage statistics.
An example
Two systems handle complaints. The first costs 3 cents per transaction but delivers a result that passes without rework in only 60 percent of cases. The second costs 11 cents and manages 92 percent. Calculated per usable transaction: 5 cents versus 12, so the cheap system still looks better. Add the four minutes of rework a person needs for each of the 40 percent, and the ratio tips decisively.
A closer look
The metric sets the benefit achieved against the computing effort used: completed transactions per euro rather than tokens per euro. It forces you to put a number on the benefit in the first place.
Why it matters
The term is young and the calculation behind it old: what counts is not the price per request but the price per usable result, including human rework. In almost all the cases we have seen, rework is the larger item, and the only one that does not appear on the vendor's invoice. Anyone comparing two offers therefore needs two numbers from their own operation: pass-through rate without rework and minutes per rework case. Without them, you are comparing list prices.
Commonly misunderstood
Without this metric, utilization gets confused with value. Many requests can mean something is well received, or that users have to try it three times.