The AI SaaS Tech Stack We Ship in 2026 (and What We Killed)
The exact frameworks, models, infra, and tools we use to ship AI-native SaaS in under 30 days — with the reasoning behind each choice.
Every 6 months we re-run our stack. Half the tools we used in 2024 aren't in production anymore — either the ecosystem moved on or we found something that costs 5× less and does the job better. Here's the stack we ship with in 2026 across every AI SaaS build.
Frontend
- Next.js 15 or TanStack Start. Next.js when the client wants Vercel and mature RSC. TanStack Start when we care about edge-first, minimal magic, and typed routing.
- Tailwind v4 + shadcn/ui. Still the fastest way to a design system that doesn't look like Bootstrap.
- Framer Motion for animation. TanStack Query for data.
Backend
- Server functions in the same repo. Whether Next server actions or TanStack server fns, one deploy, one auth boundary, less glue.
- Cloudflare Workers for edge routes and webhooks. Fast cold starts, cheap egress.
- Postgres via Supabase for everything OLTP. Row-level security, realtime, storage, auth — one buy.
AI layer
| Job | 2026 default | Why |
|---|---|---|
| Reasoning | Claude Sonnet, GPT-5-mini | Best cost/quality for product features. |
| Cheap classification | GPT-4.1-mini, Haiku | 10× cheaper than reasoning models, same accuracy for narrow tasks. |
| Embeddings | text-embedding-3-large or Voyage-3 | Voyage wins on retrieval quality; OpenAI wins on ecosystem. |
| Speech | Deepgram Nova-3, ElevenLabs | Latency + quality still beat OpenAI for real-time. |
| Image gen | Nano Banana, Ideogram, GPT-image | Depends on brand/type — no single winner. |
Vector search
Postgres + pgvector for anything under 5M vectors. Turbopuffer or Pinecone once you cross that line. We wrote up the tradeoffs in detail.
Payments & auth
- Stripe for US-first, Paddle when the buyer is global and you want MoR.
- Supabase Auth unless you need enterprise SSO from day one — then WorkOS.
Observability
- PostHog for product analytics + session replay.
- Sentry for errors.
- Langfuse or Helicone for LLM tracing, evals, cost tracking.
Background work
Trigger.dev or Inngest for durable jobs and long-running AI tasks. Both beat rolling your own queue in a serverless world.
What we killed this year
- LangChain — over-abstracted. We call SDKs directly.
- Firebase — Postgres is cheaper and doesn't lock you in.
- Redis for queues — Trigger.dev is more reliable at our scale.
- Custom vector DBs for <5M docs — pgvector is fine and one less service.
Cost benchmark (typical AI SaaS at 500 MAU)
| Line item | Monthly |
|---|---|
| Vercel / Cloudflare | $20–$150 |
| Supabase | $25–$250 |
| LLM APIs | $200–$2,500 |
| Analytics + monitoring | $0–$100 |
| Domain, email, misc | $50 |
| Total | ~$300–$3,000 |
Every stack we ship boils down to: one repo, one deploy, one auth boundary, one Postgres, one API-first LLM provider. If your architecture doc has more than 8 boxes at seed stage, you're going to spend more time integrating than building.
FAQ
Frequently asked questions
What is the best tech stack for an AI SaaS in 2026?+
Next.js or TanStack Start on the frontend, Postgres via Supabase for data and auth, Cloudflare Workers for edge routes, Claude or GPT via SDKs (not LangChain), pgvector for retrieval under 5M docs, and Trigger.dev for background AI jobs. That's ~90% of production AI SaaS in 2026.
Should I use LangChain in production?+
For most product teams, no. LangChain adds abstractions that make prompts, retries, and streaming harder to debug. Call the OpenAI/Anthropic SDKs directly and add your own orchestration where you need it. LlamaIndex is fine for pure RAG pipelines.
Is Supabase enough for an AI SaaS?+
Yes, until 10k paid users or hard multi-region requirements. Postgres, auth, storage, realtime, and pgvector in one product removes 4–5 services from your architecture. Migrate off when you hit those limits, not before.
How much does the infra cost to run an AI SaaS?+
For a product at 500 monthly active users, plan on $300–$3,000/month total. LLM API spend is usually the biggest line item — cache aggressively and choose the cheapest model that passes your evals.
Building something similar?
Let's talk in 30 minutes.