For a couple of years, the main problem AI firms had was convincing people — or, more specifically, people in charge of companies — to pay for their products. Artificial-intelligence applications were gaining popularity, but they were expensive and not yet easy to deploy in clearly and measurably productive ways. Lots of companies wanted to use AI, insofar as they understood it as a theoretical source of efficiency and a tech megatrend that they didn’t want to miss, but early AI pilots and half-baked enterprise tools didn’t make a strong case for spending more.

In late 2025, though, the case became much stronger: AI coding improved massively as tools like Claude Code and OpenAI’s Codex went from coding assistance tools to code-writing tools. This elevated so-called vibe-coding from a fascinating curiosity into a plausible way of working. Here, for companies developing and maintaining software, was a clear and actionable use case, a place to spend real money with the expectation of specific returns, and a way to satisfy a strategic need to be doing something — anything! — not to fall behind. A few months later, some companies have a different problem with their AI deployments: Now they’re spending too much.

This has all been good news for Anthropic, which became the standard-bearer for AI coding, surpassed OpenAI’s valuation, and reportedly logged its first profitable quarter this year, far ahead of projections. If the trend continues, the company will ride enterprise AI-coding adoption to a massive IPO this year. (The enormous surge in spending with the frontier AI companies has also quieted talk of a bubble, which AI executives themselves were fretting about as recently as December.)

But this period of rapid adoption, for all its surface-level obviousness — you can just talk through software development now, and the code writes itself — led to some strange behaviors in the broader tech industry. Top-down AI usage mandates became common practice, and companies including Meta and Amazon created internal leaderboards to rank and incentivize AI usage. This resulted in some well-publicized episodes of “tokenmaxxing,” where employees at these companies blew through billions of tokens — the basic unit of information that a model receives and generates — to unclear ends, throwing AI agents at pointless tasks, using the most expensive models to do simple work, and defaulting to AI for work where other tools might suffice. (Why check your weather app when you can send Claude to check for you? Better yet: Why not whip up a whole new weather app?)

BREAKING: CEO discovers tokens cost money pic.twitter.com/4rhm6ayAJU— Alberta Tech (@albertadevs) June 2,........

© Daily Intelligencer