Skip to content

Chapter Summary: Infrastructure for AI-First Operations

Key Takeaways

  1. Start boring, add complexity when it hurts: The Day 1 stack—Vercel, Supabase, direct API calls—costs under $500/month and gets you to production. Linear launched with three integrations. Notion hit one million users on seed funding. Infrastructure sophistication should lag revenue, not lead it.

  2. Postgres handles more than you think: PostgreSQL with pgvector achieves 471 queries per second at 99% recall—11.4x better than dedicated vector databases on the same benchmark. Instacart pushed to 1 billion embeddings before reconsidering. The 5 million vector threshold is where most teams actually need specialized storage.

  3. Match autonomy to reversibility: The three-tier access model—free, supervised, forbidden—isn't about limiting AI. It's about proportioning risk. The Postmark MCP attack compromised 1,500 users because tools lacked central oversight. 73% of enterprises experienced AI security breaches in 2024-2025.

  4. Agents need identity, not just API keys: Non-human identities outnumber humans 50:1 in enterprise environments. Gartner predicts 25% of breaches will trace to AI agent abuse by 2028. OAuth 2.1 delegation chains ensure permissions flow down, never up—and every action traces to a human.

  5. Build what differentiates, buy everything else: Linear built a custom sync engine because real-time collaboration IS their product. They buy managed PostgreSQL because databases aren't their moat. Self-hosting beats SaaS only around 80 million queries monthly. If you can't answer "yes" to two of three questions—differentiator, team capacity, scale certainty—buy.

Next: Building with AI

← Previous: Build vs Buy for Infrastructure | Chapter Overview