Chips
Inference cost is an agent constraint
Silicon TL;DR
- Digest tokens/$ or tokens/joule when vendors publish them.
- Note batch vs latency tradeoffs that affect tool-using agents.
- Cite the original vendor brief; we do not reprint whitepapers.
- Observation badges remain fetch logs only.
An agent that cannot afford a tool call is a different creature. Chips coverage on the wire stays close to that operational truth.