It would be fun to compare with inference providers (groq/vertex ai, etc.).