Cloud Stack for AI Apps — Mastery6 / 8
AWS — When You Outgrow the Managed Path
The managed stack covers 90% of an AI product. AWS is the escape hatch for the other 10% — long jobs, GPU inference, large files. Use it surgically.
The managed stack ships the product. AWS is not a replacement for it — it's a surgical escape hatch for the few jobs serverless platforms can't do well. Add it for a specific need, not as a lifestyle.
The cases that justify it
- Long-running / async jobs — anything past serverless time limits (batch embedding, video processing, large PDF generation). SQS + Lambda/Fargate runs them off the request path.
- GPU inference — self-hosting an open model needs GPUs (EC2/SageMaker) that Vercel doesn't offer.
- Large object storage — heavy files or huge volume at lower cost on S3 (Supabase storage is fine until it isn't).
Add it without a migration
Keep the managed stack as the front door. The app enqueues a job (SQS); a worker processes it and writes the result back to Supabase or S3; the app reads the result. You bolt AWS onto one seam, not your whole architecture.
Next: keeping secrets and config sane across all these services and environments.
Series — Cloud Stack for AI Apps — Mastery
- Part 01The Opinionated AI Product StackChoosing infrastructure is where weeks vanish. Here is a default stack that lets a small team ship an AI product in days — and when to deviate.
- Part 02Next.js on Vercel — The App LayerOne framework for UI, API and server rendering; one platform that deploys on git push. The app layer should be the part you never think about.
- Part 03Supabase — Postgres, Auth & Storage in OneReal Postgres, auth, file storage and pgvector behind one SDK. For an AI product, having your data and your vectors in the same database is a quiet superpower.
- Part 04Stripe — Payments & SubscriptionsCharging money is a solved problem — if you let Stripe solve it. Checkout, subscriptions, and the webhook that is the real source of truth.
- Part 05Resend — Transactional Email That LandsThe receipt, the download link, the password reset — if they hit spam, your product feels broken. Deliverability is a feature.
- Part 06AWS — When You Outgrow the Managed Path — you are hereThe managed stack covers 90% of an AI product. AWS is the escape hatch for the other 10% — long jobs, GPU inference, large files. Use it surgically.
- Part 07Secrets, Env & Config Across EnvironmentsFive services, three environments, one leaked key away from a bad week. Config discipline is unglamorous and non-negotiable.
- Part 08Shipping in Days — The Wiring PlaybookThe whole stack assembled into a build order: from empty repo to a paid, emailing, AI-powered product in a working week.