Correction · 29 May 2026
The energy comparison has been withdrawn
It rested on a batch size nobody has published, so it was not a figure we could defend.
Specifics
The numbers, and where they come from
| Quantity | Value | Basis |
|---|---|---|
| Withdrawn figure | 2 to 6 ×10⁵ | Was modelled |
| Dominant unknown | counterparty batch size | Unpublished |
| Latency figure | ~7,700× | Measured, unaffected |
| Cost figure | parity at 10⁶ turns | Modelled |
What was claimed
And why it does not survive
Earlier material carried an energy advantage of 2 to 6 ×10⁵. The range came from a model whose dominant term is the batch size used by the counterparty at inference, and no counterparty publishes that number. A figure whose largest input is a guess about someone else is not a claim, it is an assertion.
The value is struck rather than deleted. Anyone who saw it should be able to find out that it was withdrawn, when, and why, which is not possible if the number simply disappears.
Why it matters more than the number
Keeping the axes apart
Latency, energy and cost are three separate arguments with three different bases. Latency is measured and architectural. Cost is modelled with stated assumptions. Energy rested on figures the counterparty has not published. Combining them into one line about being faster, cheaper and greener loses all three, because the weakest basis contaminates the strongest.
The latency figure is measured on an eighty-word reply and is unaffected by this withdrawal. It stands on its own and it does not need the energy claim beside it.
Reinstatement condition
The comparison returns only if a counterparty publishes batch size, or if we can measure end-to-end energy on hardware we control against a workload we can describe. Until then it stays struck.