Tokens may be the wrong unit for compute value

Tokens are easy to count, which makes them attractive as a meter. But they are not necessarily the thing a user values, and they do not represent the same amount of work across models, hardware, or tasks.

If distributed compute becomes a market, we will need a unit that connects resource consumption to service quality. Time, memory, energy, latency, reliability, and model capability all matter.

I do not know the right answer yet. I am fairly sure that one token should not be assumed to mean one unit of economic value.

The useful meter should tell us something about the service delivered, not only the text processed along the way.