Start with product risk
The AI backend decision should not start with a fashionable stack. If Postgres and SQL control are central, look at Supabase or Neon. If realtime agent state and a TypeScript workflow matter most, Convex can be faster. If the agent must run at the edge or execute code, Cloudflare becomes relevant.
What each layer solves
Supabase gives a Postgres-first platform, Neon gives database branching for agents, Convex gives a reactive backend, Cloudflare gives edge compute and sandboxing, and Vercel AI Gateway gives model routing. A strong product can combine several layers, but each needs a clear role.
The WebEdge rule
Choose the simplest stack that reliably performs the job. A prototype often needs one backend; production needs observability, access control, tests, backups and a migration strategy.