You don't need more startup advice. You need sourced answers.

Every "AI eats everything" post is a hot take with no receipts. The Startup AI Playbook is the opposite: 111 real sources — Sequoia, Y Combinator, NVIDIA, Anthropic, Carta, OpenAI, WEF, McKinsey — synthesized into the concrete 2026–2027 strategy for deciding what to build, why, and how to win.

Decide what to build — with data

• The autopilot thesis. For every $1 spent on software, ~$6 is spent on services. The winning 2026 play is to sell the work (books closed, NDAs drafted, claims filed) — not the tool. Nine verticals are mapped with labor TAMs and the companies already winning each one.
• The 21 agent-startup categories. Greg Isenberg's framework for the economy where agents — not humans — are the new customers.
• The honest counter-case. A playbook that never argues against its own thesis is marketing. This one tells you exactly when "sell the work" breaks — liability, services economics, incumbent trust — so you don't bet wrong.

Where money actually flows

• The 5-layer AI stack, and NVIDIA's $500B AI-factory financing.
• Chamath's margin-migration lens: physical scarcity today → customer proximity tomorrow.
• Sovereign AI ("not your weights, not your product") — when to own your intelligence, and when renting is right.

Get recommended by AI (GEO)

The Cal.com reverse-engineering — 47 comparison posts, 23 niche directories, and the "citable listicle" architecture that made it the #1 ChatGPT-recommended scheduler — plus how to score your own site on is-agentic.com.

Model reasoning — the actual, current picture

Not the stale "LLMs can't reason" line. You get both the 2024 GSM-Symbolic fragility finding and the 2026 evidence (OpenAI's unit-distance disproof, "Astra"'s ten Lean-certified proofs), plus the abduction gap and why verification bandwidth — not intelligence — is now the binding constraint.

It ships its receipts

• 111-source bibliography + 46 verbatim primary-source exports.
• 11 research PDFs bundled offline — Sequoia's thesis, the GSM-Symbolic paper, Anthropic's 81k-interview study, Carta's solo-founder report, and more.
• A working idea evaluator (evaluate-idea.py) scoring any idea across 14 dimensions.
• 7 OKF v0.2 templates to turn the frameworks into documents.
• 25 evaluation prompts — the exact QA harness used to test this playbook, shipped for transparency.

For founders, indie builders, and strategists who want the sourced version of the AI-startup conversation — not another thread of vibes.