Run a tech biz
without a burn narrative.
For founders and tiny teams shipping a product — SaaS, civic tech, ops tooling, tech-adjacent services. The goal is not “zero forever.” It’sknown costs, commercial-safe licenses, and AI that doesn’t melt the card.
Pair this with the directory tabsStartup leanandAI & models— limits live there.
- 01
Stay commercial-OK
Edu licenses (JetBrains Edu, Figma Education, Autodesk, MATLAB Campus-Wide) are for learning. Production for a company — even a tiny one — needs seats and ToS that allow commercial use. Vercel Hobby is the same trap.
- 02
Buy outcomes, not platforms
Default to Cloudflare + Neon/Turso + Clerk + Resend + R2 + PostHog. Add a hyperscaler only when you have a concrete reason (GPU batch, compliance region, Activate credits that expire).
- 03
AI is a line item — meter it
Route drafts through cheap/fast models (Groq, Gemini free quota, Workers AI, OpenRouter free models). Keep Claude/GPT on the hard path. Put Langfuse or Helicone in front before the first customer demo so spend is visible.
- 04
Claim startup credits once you are real
AWS Activate, Google for Startups, Microsoft for Startups — apply when you have incorporation or an accelerator letter. Burn credits on load tests and GPU jobs, not idle VMs.
- 05
Ops stack that doesn't need a COO
Cal.com for bookings, Linear for issues, Chatwoot/Crisp for support, Lemon Squeezy or Paddle when you sell globally and hate VAT paperwork. Notion Free until SAML is forced on you.
Swap pieces freely. Keep the commercial chip green.
- Edge / web
- Cloudflare Pages + Workers (commercial OK)
- API
- Koyeb micro or Workers — skip always-on beef until traffic
- DB
- Neon or Turso free → paid when storage wakes you up
- Auth
- Clerk / Kinde free MAU
- Resend free → Pro when deliverability matters
- Files
- R2 — zero egress for user downloads
- Jobs
- Inngest free steps
- Analytics
- PostHog free events
- LLM gateway
- OpenRouter + spend cap · Helicone/Langfuse
- Default model
- Gemini / Groq / Workers AI for bulk
- Frontier model
- Claude or GPT only on tool-critical steps
- Payments
- Lemon Squeezy / Paddle (MoR) or Stripe when ready
Cache & dedupe
Identical prompts should not hit the API twice. Helicone caching or your own KV on R2/Workers buys real money back.
Structured outputs > chat soup
JSON schemas and tool calls fail cheaper than free-form prose you re-parse with another model.
Eval before scale
A 50-row golden set in Langfuse beats vibes. Don't raise temperature and hope.
Human in the loop where liability lives
Tech-adjacent (health, finance, legal-ish ops): AI drafts, humans approve. Cheaper than one incident.
Pre-revenue prototype
~$0–30/mo- ·Forever-free SaaS only (see directory filters: commercial ok)
- ·Student Pack if you still qualify — don't put edu tools in prod
- ·Oracle Always Free or Koyeb micro if you need a long process
First paying users
~$30–200/mo- ·Turn on spend caps everywhere (OpenAI, Anthropic, OpenRouter, cloud)
- ·Graduate Resend / Clerk / Neon when free limits page you at night
- ·Apply Activate / Google for Startups if incorporated
Team of 3–10
intentional- ·Linear + shared PostHog + one observability bill (Sentry)
- ·Drop self-host chores that aren't your moat
- ·Revisit Fly/AWS only with a region or compliance story
Things that look free and then eat the company.
- Intercom day-0Seat pricing arrives before PMF. Use Chatwoot/Crisp until tickets hurt.
- Vercel Hobby prodNon-commercial. Cloudflare Pages/Workers or paid Vercel if you charge money.
- Edu licenses in prodJetBrains Edu, Figma Education, Autodesk Edu — school only. Buy seats.
- Frontier model defaultClaude/GPT on every request. Local/cheap first; frontier on hard steps only.
- Idle Activate VMsCredits expire. Burn on load tests and GPU jobs, not always-on toys.