🚀 The Free Ship-It Stack (2026)

Free model tokens are only half a product. This guide is the other half — the database, vector store, embeddings, hosting, storage, auth, and cron you need to actually ship an AI app for $0 — with a hard focus on which "free" tiers you can legally use commercially (several popular ones quietly can't).

Companion to the main Awesome Free Compute list, the Free Token-Maxxing Guide (LLMs/tokens), the Free Audio Generation Guide, and the Media Post-Production Guide. Compiled from a multi-model research sweep (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro) and reconciled against official pricing pages. All quotas are June 2026 snapshots — verify before relying.

Contents

• TL;DR — the commercial-safe $0 stack
• ⚠️ 2026 reality check (read this first)
• 1. Databases
• 2. Vector search and RAG
• 3. Embeddings
• 4. Hosting and serverless
• 5. Object storage
• 6. Auth, email, cron
• 7. Reference architectures
• The free-tier cliffs
• Sources

TL;DR — the commercial-safe $0 stack

One-liner: the cleanest $0 commercial stack is all-Cloudflare (Pages + Workers + D1 + R2) or all-Supabase (Postgres + pgvector + Auth + Storage) — both minimize the number of free tiers you can outgrow, and neither has a commercial-use trap.

⚠️ 2026 reality check (read this first)

The free-tier landscape shifted hard in 2024–2026. The internet is full of stale guides; these are the traps that actually matter:

• 🚫 Vercel Hobby is personal / non-commercial ONLY. The moment your project is commercial you must move to Pro ($20/mo) — this is a ToS rule, not a usage limit. For a commercial $0 app, deploy on Cloudflare Pages instead.
• 🚫 GitHub Pages forbids commercial use — "no business, e-commerce, or commercial SaaS." Fine for docs/landing pages, not your product.
• 💸 No longer free (perpetual): PlanetScale (Hobby removed), Fly.io (free allowances gone Oct 2024), Railway ($5 trial only), Xata (credit then PAYG). Don't plan a $0 stack around these.
• 📅 Deno Deploy Classic (1M req/mo) sunsets July 20, 2026 — no new signups; migrate to the new Deno Deploy.
• 📉 Netlify moved to a credit model (~300 credits ≈ 15 GB bandwidth/mo) — far tighter than the old 100 GB.
• 🕵️ Gemini free-tier embeddings (and chat) train on your inputs. Google's free tier uses your content to improve products; the paid tier doesn't. For anything proprietary, use Voyage (200M free tokens) or local embeddings.
• 😴 "Free" often means "sleeps": Neon scale-to-zero (5-min cold start), Render spin-down (15 min), Supabase pause (1 week idle), Qdrant/Weaviate idle-suspend (1–4 weeks), HF Spaces sleep (~48 h). Fine for side-projects; budget an "always-on" upgrade for latency-sensitive production.

1. Databases

Pick: Neon if you want Postgres + pgvector with the fewest moving parts; Supabase if you also want bundled Auth + Storage; Cloudflare D1 if you're going all-Cloudflare and want zero egress fees.

2. Vector search and RAG

Pick: pgvector on your existing Postgres for 90% of apps. Reach for Qdrant or Upstash Vector only when you need a dedicated ANN service; LanceDB (embedded) when you want zero infra.

3. Embeddings

Pick: local (Ollama) for unlimited + private, or Voyage for a hosted, commercial-clean key. Avoid Gemini's free tier for proprietary data (it's used for training).

4. Hosting and serverless

Pick: Cloudflare Pages + Workers is the $0 commercial hero — unlimited static, 100k req/day, zero egress. Use Vercel Hobby only for personal/non-commercial projects (the DX is great, but commercial = $20/mo).

5. Object storage

Pick: Cloudflare R2 — free egress means no surprise bandwidth bills, and it pairs natively with Pages/Workers.

6. Auth, email, cron

Pick: Supabase Auth if you're on Supabase, else Auth0 (25k MAU, no branding). Resend for email (mind 100/day). QStash or CF Cron for scheduled jobs.

7. Reference architectures

🅰️ Commercial-safe $0 (recommended):

Cloudflare Pages + Workers   ← frontend + edge API ($0 egress, commercial OK)
        │
        ├── Neon / Supabase Postgres + pgvector   ← data + RAG (one service)
        ├── Voyage (200M free) or local Ollama     ← embeddings (no data-sharing)
        ├── Cloudflare R2                          ← file/asset storage (free egress)
        ├── Supabase Auth (50k) / Auth0 (25k)      ← login
        ├── Resend (3k/mo)                         ← transactional email
        └── Upstash QStash / CF Cron               ← jobs & queues

🅱️ Simplest one-vendor: - Supabase-only: Postgres + pgvector + Auth + Storage + Edge Functions. - Cloudflare-only: Pages + Workers + D1 + R2 + Workers AI.

🅲 Best DX, non-commercial / hobby: Next.js on Vercel Hobby + Neon + Gemini-free embeddings + R2 + Resend. Smoothest to build — but the instant it's commercial, Vercel forces Pro ($20/mo) and GitHub Pages becomes disallowed.

The free-tier cliffs

When (and why) each free tier forces a paid upgrade:

• Vercel Hobby → Pro $20/mo — the instant the project is commercial (ToS, not usage).
• GitHub Pages → move off — any business / e-commerce / commercial SaaS is prohibited.
• Supabase → Pro $25/mo — >500 MB, >50k MAU, backups/PITR, or to stop the 1-week pause.
• Neon → PAYG — >0.5 GB, >100 CU-hrs/mo, or to disable scale-to-zero (kill cold starts).
• Render → ~$7/mo — any production app (free spins down after 15 min; "not for production").
• Resend → $20/mo — >3,000/mo or the 100/day cap; or >1 domain.
• Clerk → $25/mo — to remove branding, add MFA, or change session length.
• Vector DBs — outgrow Qdrant (1 GB, suspends 1 wk), Upstash Vector (200M, 10k q/day), Weaviate (suspends 7 d), or Pinecone (1M read / 2M write units) → paid.
• Latency-sensitive production — pay to make Neon / Render / Qdrant / Weaviate / HF Spaces "always-on" (kill the idle sleep).
• Already paid-only (no $0 path): PlanetScale, Fly.io, Railway, Xata.

Sources

Primary pricing/docs pages fetched & verified June 2026 (representative):

• Databases: Supabase pricing · Neon pricing · Turso pricing · Cloudflare D1 pricing · MongoDB Atlas M0 · Upstash Redis · PlanetScale · Xata
• Vector: Qdrant Cloud · Pinecone limits · Weaviate Cloud · Upstash Vector · Chroma Cloud · LanceDB
• Embeddings: Voyage pricing · Gemini pricing · Cohere rate limits · Nomic · Jina
• Hosting: Vercel Hobby · Netlify pricing · Cloudflare Workers / Pages · Render free · Fly.io pricing · Railway · Deno Deploy · GitHub Pages limits · HF Spaces
• Storage: Cloudflare R2 · Backblaze B2
• Auth / email / cron: Supabase · Clerk · Auth0 · Firebase · Resend · Upstash QStash

Some vendor pages are JS-gated; a few figures (MongoDB M0 512 MB, Zilliz/Weaviate exact vector caps, Jina/Mistral quotas, new Deno Deploy tier) are widely-documented but should be reconfirmed in-browser. Quotas change often — treat every number as a June 2026 snapshot.

Part of Awesome Free Compute 2026. PRs welcome — correct a stale quota, add a tier, flag a commercial-use trap.

📝 Spotted a stale quota or a license that changed? This guide is open source — edit it on GitHub.