Local LLM hosting isn't always cheaper: cached-token rate decides the break-even point
paq85 · reddit · 2026-10-17
Reddit user paq85 released a browser-based LLM cost calculator comparing self-hosted GPUs (RTX 4070 Ti Super / 5090 / 5090 eco presets at 500–900W) against cloud API pricing ($0.25/$1.20/$0.02 per 1M input/output/cached tokens). Key finding: the cached-input rate is the variable that moves the break-even point most — most calculators guess at it. On long, repeated prompts (60k in / 50k out / 2k requests), electricity stays nearly flat while API costs keep climbing, pushing the crossover further out than expected.
More from coding & agent
- Over 20,000 agent harnesses so far in 2026, one hit 95,000+ in 2 days — jasonkneen · 2026-10-17
- Voice AI product lead on where Jev-style classifier models actually fit in phone bots — VladimirSamukov · 2026-10-17
- Cutting AI inference bills from $250K to $7.6K/month: Eugene Cheah's tokenomics 101 — AI Engineer · 2026-10-17
- Tailscale uses Cyrillic 'ж' in Go code to flag variables you shouldn't touch — mitsuhiko · 2026-10-17
- After an agent deleted their CRM leads, dev ships a trusted merge layer for Postgres — Antique-Willow-5841 · 2026-10-17
- Nuwa.skill (34.8k stars) releases promo animation fully generated by Opus 5.5 — AlchainHust · 2026-10-17