Noah Smith
· Noahpinion
· June 02, 2026 at 09:31
· ⏱ 8 min read
| Read on Substack ↗
Summary
Despite massive spending on AI coding agents (tokenmaxxing), there is little evidence that the tokens translate into shipped products or measurable economic output. This gap suggests the software industry may be in an experimentation phase akin to early electricity or the internet, but it also raises questions about whether AI's productivity gains will materialize in mature software sectors or instead create entirely new applications.
•Anthropic is IPOing at ~$1 trillion valuation with $45B annualized revenue and 130% quarterly growth, driven by its Claude Code coding agent.
•Only 18% of company spending on AI coding tokens translates into shipped products reaching real users, per EntelligenceAI data on 2,000+ firms.
•Microsoft began canceling Claude Code licenses, and Salesforce redesigned employee targets to measure real output instead of AI input.
•Uber COO Andrew Macdonald stated it is not yet possible to draw a direct link between raw AI usage and useful products shipped.
•Companies like Meta encouraged employees to use as many tokens as possible (e.g., a leaderboard), but observers report little visible new features or app improvements.
•John Loeber described the trend as 'spinning their wheels and shipping absolutely jack shit,' with engineers churning unfinished marginalia.