Local inference is not totalizing, and can be cheaper per user or community

Claim (restated 2026-10-26 per James: the point is that local hosting is not totalizing and can be cheaper per user or community, not that it is more efficient in aggregate; applied to inference by analogy with hosting-cost-claim, to confirm with James): local or municipal-scale LLM inference is not totalizing and can realistically be cheaper and simpler for each user or community than inference from mega-cloud providers. Aggregate efficiency (batching, utilisation) can favour hyperscale without refuting this; what tests it is per-user or per-community cost at realistic load, and lock-in.

~~Was:~~ Claim: local or municipal-scale LLM inference is only marginally less efficient than centralised cloud inference, and much cheaper and simpler to realise. His inference testbed exists to measure this. Evidence welcome, for and against: batching economics, utilisation, model-size thresholds where the claim breaks.

Contested: collect evidence for and against on this page; do not treat as established.

Provenance: James, in conversation with Claude, Sept 2026 (unverified). Disposition: contested.

Related: convivial-computation, inference-testbed, hosting-cost-claim

Part of thread: thread-convivial-computation