Developers became 180% more productive with AI. Yet the number of apps people actually use didn’t increase. Both metrics were measured across 100,000 developers, and together they form the most uncomfortable tech data point of 2026. The study comes from the NBER. Researchers cross-referenced actual activity on GitHub with each developer’s AI usage data, isolating the impact of each generation of tools. The surge in code output is staggering: → Code autocomplete: +40% → Conversational agents: +140% → Autonomous agents: +180% Almost three times more code coming from the same person. It looks like the greatest productivity revolution the world has ever seen. But then the researchers asked the question almost no one wanted to ask: How much of that output actually became a product in someone’s hands? When you track what happens to that 180% gain, it drops to +50% when measured by number of projects. When you get to what’s actually launched, only +30% remains. Almost all the gain vanishes along the way. And it gets worse. Look at the Financial Times chart. The line for app launches skyrocketed. Meanwhile, the number of apps with real user adoption stayed flat, and user ratings began to decline. AI has flooded the world with software nobody opens. The greatest production capacity in history is being spent creating things that die on launch day. Why does this happen? Code has always been the easy part of this story. The hard work comes after: reviewing, ensuring it works, deploying it, and convincing someone to use it. AI brutally accelerated just one link in this chain—and exposed just how much everything else still depends on people. What this means for builders going forward is direct: AI has made writing code so cheap it’s become a commodity. Anyone can produce it now—and quickly. What will be worth its weight in gold over the next decade is “taste” and judgment: knowing what deserves to exist before you spend your first hour building it.