AI’s Midas problem: When code is free, features are worthless

Teams that pushed Claude Code, Cursor, or Copilot into their workflows wanted speed. What they’re running into instead is a basic economic problem: if output gets cheap enough, more output stops being a moat.

That’s the argument behind The New Stack’s Midas comparison. In the myth, Midas gets everything he touches turned to gold, then discovers he can’t eat or drink it. The article says AI can do something similar to software: turn what used to be scarce craft into abundant, copyable commodity work.

The piece points to signs that the glow is already fading. Microsoft and Meta have reportedly cut internal spending on Anthropic’s Claude, while Uber and Accenture have capped employee spending on AI tools. McKinsey’s own 2026 research is cited as finding that record AI spending hasn’t moved earnings for 94% of enterprises. That’s not a small miss. That’s a lot of expensive enthusiasm looking very ordinary.

The bigger point is less about code generation itself than about what companies think they’re buying. If software is just a translation layer from user requests into code, then AI can flatten the value of that work fast. The article argues teams need to stop acting like feature factories and start understanding the jobs users are actually trying to do, because that’s where software can still be worth more than the tokens used to make it.

DORA’s research gets pulled in to back that up. Its 2024 report says a user-centric approach improves product performance, regardless of delivery speed. So the answer isn’t to make more code. It’s to make software that solves something users couldn’t have easily asked for in the first place.