The AI Startup MVP stack
For solo founders or 2-3 person teams shipping an LLM-powered product end-to-end in under 4 weeks.
You're building a Cursor-or-die AI app. The non-negotiables: Anthropic or OpenAI for inference, a vector store, fast hosting with edge functions, and just enough auth + email to ship paid plans. Everything else is procrastination dressed as productivity. This stack ships you to first paying customer in 28 days, not 6 months.
Foundation models
Start with Anthropic Sonnet — best output quality per dollar in late 2025/2026. Add OpenAI as a fallback for tool-calling depth and to hedge availability. Don't self-host.
AI infra
You'll do RAG before you finish the first page of the app. Pinecone is the lazy default; LangChain is the framework if you want agent loops without writing them yourself.
Hosting + database
Vercel + Neon is the path of least resistance for Next.js. Skip the AWS detour — you'll lose two weeks to IAM and gain nothing until you hit 1M MAU.
Auth + payments
Don't build auth. Don't build payments. Pay the $25-49/mo and ship.
Email + observability
Resend for email because it has React Email built in. Sentry for errors. Axiom for logs. Skip Datadog until you have an SRE.
Pay for the AI Startup MVP stack with the right card
At ~$165/mo USD spend (the midpoint of this stack's burn range), the right card recovers 1.5–4% in forex savings + rewards. Default retail cards lose 3.5% of every USD subscription to forex.
EXPRESS
Amex Hilton Honors
No-fx US card — pay USD with USD, zero markup; 3% rewards on engineering spend
IDFC FIRST Wow
0% forex markup vs 3.5% on retail Indian cards
Tradeoffs & pitfalls
The honest stuff. No vendor blog will tell you this.
- 01Vendor lock-in is real but cheap to escape. Keep your DB schema portable, your prompts in version control, and your auth user-IDs in a column you own.
- 02You will outgrow Pinecone's serverless tier before you outgrow Anthropic's rate limits. Plan accordingly.
- 03Cursor + Copilot are not a duplicate spend at the indie scale — different surfaces (multi-file refactors vs in-line autocomplete). Many devs run both.