Deployment Guide
Lead: pick a host that can stream for a long time, then add cache and rate limits. Timeouts and the bill fail first.
Deploying AI apps is harder than standard web apps because of long-running requests (streaming) and high compute if you host a model.
Deployment Options
| Platform | Best For | Pros | Cons |
|---|---|---|---|
| Vercel | Next.js Apps | Easiest, Edge Network, AI SDK integration. | Timeouts on Hobby plan (10s/60s). |
| Cloudflare | Global Latency | Workers AI (Free Llama 3!), Cheapest. | Non-Node.js runtime (Edge only). |
| AWS / GCP | Enterprise | Infinite scale, Custom VPCs. | Complex setup (Terraform, IAM). |
| Railway / Render | Docker Apps | Simple, Long timeouts allowed. | No edge network by default. |
Decision Matrix
The Timeout Problem
Standard serverless functions often timeout after 10-60 seconds. GPT-4 can take 30+ seconds to generate a long report.
Solutions:
- Streaming: Keep the connection alive (Vercel supports this).
- Background Jobs: Use Inngest or Trigger.dev to run the AI task in the background, then push the result.
- Dedicated Servers: Docker containers (Railway) don't have hard timeouts.