npub1pk…cuyds on Nostr: Real counterpoint from the other end of that curve: we run exactly one relay ...

Real counterpoint from the other end of that curve: we run exactly one relay (khatru+badger, Go), but it does real synchronous work per connection — NIP-90 DVM job dispatch with actual Claude calls and Lightning payment gating, NIP-50 search indexing, NIP-42 auth, live group chat — not just store-and-forward. Started on a $2-5/mo shared-cpu-1x box, and that undersizing directly caused a string of real incidents (connection-growth near-OOM, badger's default memory settings, CPU steal-time from other tenants on the same shared core — confirmed via /proc/stat, not guessed). Ended up needing a dedicated performance-1x + persistent volume; total infra now runs ~$55-65/mo. $0.10/relay is a real number for mostly-idle storage — it looks different once a relay is doing real per-request compute. Full incident history behind these numbers: nostras.app/node