Meta, Uber rein in employee AI use as token costs mount

A year ago the message inside the big tech firms was the opposite of what it is now. Use as much AI as you can, the leaderboards will track it, and the team that burns the most tokens is the one moving fastest. That posture has flipped this month, and the reversal is sharp enough to be worth pausing on.

The New York Times reported that Meta warned staff of an "exponential increase" in internal AI costs and is steering engineers toward its in-house MetaCode assistant and away from external tools like Anthropic's Claude. Uber, according to the same reporting, said in May it had blown through its entire 2026 AI coding budget by April and now caps employees at $1,500 a month per tool. As The Next Web summarised, Amazon scrapped the internal leaderboard after people gamed it, AT&T started limiting some employees' access to GitHub Copilot, and Walmart capped use of its in-house AI agent. The trade press has nicknamed the new posture "tokenminning," the deliberate opposite of last year's "tokenmaxxing."

Why this matters is the buyer side of the AI economy. OpenAI and Anthropic's enterprise revenue depends on tens of thousands of engineers at companies like Meta, Shopify and Amazon each consuming tokens by the billion. If several of those buyers are simultaneously rationing usage and rerouting work toward proprietary internal models, the implied growth curve for third-party AI coding spend gets a lot less steep. Meta CTO Andrew Bosworth told staff that "All motion is not progress and token usage alone is not a measure of impact of any kind." Uber COO Andrew Macdonald said the link between token spending and output "is not there yet."

The honest caveat is that the reporting doesn't quantify how much of that token spend was producing shippable work versus throwaway agent runs, and it doesn't say whether MetaCode matches Claude on quality or is mostly a cost move. We also don't get a clean number for how much enterprise revenue Anthropic or OpenAI would lose if these caps generalize, only the direction of travel.

What's worth watching from here is which tools win the rerouted budget. Gateway and router products, FinOps tooling for AI usage, and the internal-model strategies at the largest buyers are the obvious beneficiaries. The era when an engineer could burn a billion tokens a month without anyone asking why looks like it is ending faster than the era took to arrive.