28 Oct 2026 · Hardening & Ops
Serverless vs a VPS for AI SaaS, Real Cost at 25,000 MAU
Why Vercel duration caps and streaming egress hurt an AI product at 25k MAU, and what a Hetzner Docker box costs for the same workload.
The architecture and cost post is the overview: serverless vs a 5 euro VPS, and when a starter kit beats wiring it yourself. This post is the same question at a specific size. 25,000 monthly active users, streaming chat, some of them hitting a multi-step agent.
AgenticPacks OSS is the app you can run on either host: Agentic-Packs/agenticpacks-oss. The VPS path with Caddy and Terraform is Pro.
The problem
Vercel is a good place to start a Next.js UI. An AI SaaS stops fitting when:
- A LangGraph run needs more than the function max duration (60s on many plans, 300s if you pay).
- Token streams count as outbound data, every chunk, every user.
- Cold starts make the first message after idle feel broken.
- Prisma wants a connection pool. Short-lived isolates fight that.
You still pay OpenAI or Anthropic either way. The question is the glue: function time, egress, hosted Postgres, and concurrency.
A back-of-envelope at 25k MAU
Assumptions, not a quote. 25,000 MAU. 20% use chat in a given month. Those users average 30 streamed completions. Average 2,000 completion tokens visible in the UI. Some share of requests also run a 60 to 180 second graph.
| Line | Serverless (Vercel Pro + hosted Postgres) | Self-hosted (Hetzner CX23 or CX33 + Compose) |
|---|---|---|
| App compute | Function GB-seconds on every token and every graph | Flat VPS, about 5 to 15 USD |
| Stream egress | Billed outbound, easy to reach hundreds of USD | Included in the VPS uplink until you are huge |
| Database | Neon or Supabase Pro, tens to low hundreds | Postgres on the same node or a second CX |
| Timeouts | Graphs die or you pay for 300s duration | Process can sit on SSE for minutes |
| Total glue (order of magnitude) | often 300 to 900+ USD / month | often 20 to 80 USD / month |
Model tokens are extra on both sides. Do not compare a VPS to a model invoice. Compare glue to glue.
If your graph regularly needs 3 minutes, serverless is not a pricing debate. It is a hard cap. See SSE drops behind proxies and LangGraph streaming.
What "leave Vercel" actually means
You do not have to move DNS on day one.
- Run the same Next.js app with
output: "standalone"in Docker. Guide: Docker Compose and Caddy. - Put Postgres on the Compose network. Stop using a serverless driver that assumes a short isolate.
- Point a hostname at the VPS. Caddy takes TLS. AgenticPacks OSS has no Caddy. Pro does.
- Keep the Vercel project as a preview if you want. Production traffic goes to the VPS.
Auth.js AUTH_URL must change to the new origin. Stripe webhook URL too. Both are runtime secrets. See Docker env vars and Stripe webhooks.
When to stay on serverless
- The product is a UI with short completions, not a graph.
- You have no one who can SSH.
- 25k MAU is a slide, not traffic.
If you are already hitting duration errors, staying is more expensive than a CX23 and a weekend.
Tired of a 60 second wall in the middle of an agent run?
The MIT app runs on Vercel or Docker: Agentic-Packs/agenticpacks-oss.
The Hetzner Terraform and Caddy path is the VPS deploy path in Pro.
25k is not 250k
At 25k MAU one well-sized node plus backups is the usual answer. At much larger scale you split Postgres, add a second app replica, and put a load balancer in front. That is still cheaper than unbounded function time on streaming. It is not a reason to start on Kubernetes this week.
Where to Go From Here
You do not have to wait for the Vercel invoice to prove the architecture.
Option A: Free open-source baseline (MIT)
- GitHub: github.com/Agentic-Packs/agenticpacks-oss
- What is included: standalone Docker, Compose, Postgres, Auth.js, streaming chat, single-plan Stripe. No Caddy. Deploy on Vercel or a VPS you set up.
- Cost: free, MIT.
Option B: Ship with AgenticPacks Pro ($149 one-time)
If production should be the VPS path:
- Automated Hetzner provisioning: Terraform, firewall, Caddy SSL, Compose.
- Production multi-agent core: LangGraph, pgvector, Redis, long-lived Node.
- Usage & Unit Economics: Stripe, team pools, token ledger.