A new NBER working paper drops a cold bucket of reality on the AI-eats-software narrative. Researchers Mert Demirer, Leon Musolff, and Liyuan Yang tracked over 100,000 GitHub developers from 2022 through 2026 in their study "Writing Code vs. Shipping Code," and the numbers are wild: sync agents boosted lines of code by 741% and pull requests by 65%. But shipped releases moved only 20%, and end-user downloads barely budged at all. The gain shrinks at every step from keystroke to user, which should make everyone pumping out AI coding benchmarks reconsider what they're actually measuring.

The Productivity Paradox

Here's the uncomfortable math: more code does not automatically equal more value delivered. When supply explodes faster than demand, you get exactly what we see in app stores today—more apps, flat downloads, and a growing graveyard of software with zero audience. This isn't a tooling problem. It's a market structure problem. The same AI that lets a startup ship a polished consumer app over a weekend also generates enormous amounts of code that never escapes the developer's laptop or gets abandoned in some forgotten repository.

Why Consumer Markets Stay Startup-Friendly

Consumer software has always favored the challenger. Low switching costs, attention-driven distribution, and novelty as competitive moat—AI doesn't change any of this. If anything, it sharpens the startup advantage because faster build cycles compress time-to-market without requiring massive capital or entrenched relationships. The race is loud, visible, and brutal. Everyone's watching.

Enterprise Is Where It Gets Interesting

Enterprise software operates under completely different physics. The constraint was never speed of building—it was integration, compliance, sales cycles, data residency requirements, audit trails, and the decade-plus trust accumulation that procurement teams demand. AI compresses exactly none of those. What it does compress is the cost side of backlog mathematics. Work that couldn't justify a developer-week three years ago can justify an agent-hour today. This is where the incumbent advantage flips hard. Every organization has a curated queue of deferred work: low-priority bugs, unfunded accessibility audits, internal tools everyone needs but no team owns. Those items never shipped because the ROI math didn't work—not because demand was missing. Now that math has moved. The fifteen-year-old enterprise vendor finally closing tickets they've been deferring since 2014? That's not disruption. That's amortization.

Key Takeaways

  • AI boosts code output massively (741%) but shipped value barely moves (20%)
  • Consumer markets stay favorable to startups—low friction, attention-based wins
  • Enterprise advantage goes to incumbents with existing backlogs and trust
  • This is a productivity shift, not a platform shift that creates new winners

The Bottom Line

The enterprise race will be slower, less photogenic, and probably worth way more than the consumer circus. When supply is up and demand is flat, the 741% code surge has to go somewhere—and it's landing in backlogs owned by vendors who've been waiting years for this cost collapse. Platform shifts create openings. Productivity shifts reward whoever already owns the category.