The Trillion Dollar AI Software Development Stack

• ~30M developers generate $3T annually; AI could double that output.
  • Current assistants lift productivity ~20%; best-in-class doubles it.
  • Equivalent to adding the GDP of France to global output.
• Three-phase AI dev loop: Plan → Code → Review.
  • AI drafts specs and surfaces gaps before touching code.
  • Agentic loops handle implementation; devs evaluate output.
• .cursor/rules = first NL knowledge repos built for AI, not humans.
• Legacy code migration (Fortran/COBOL → Java) is the top enterprise win.
• Cost at scale: Claude Opus reasoning mode ≈ $10K/year per dev workflow.
  • Multi-model platforms absorb cost; cheaper models still deliver ROI.
• Code won’t die — it runs ~100 billion times faster than raw LLM execution.
• Best startups win supercycles: Copilot’s lead challenged despite MSFT platform lock.
  • Opportunity for dozens of $1B cos and one potential trillion-dollar giant.

Guido Appenzeller and Yoko Li — a16z (October 2025) · 2025-10-09 · Read on a16z.com