Meta told staff in a memo on Tuesday that it will start limiting how many AI tokens employees can burn through, weeks after it spent most of the year telling those same employees to use AI more. The Information reviewed the memo, which says internal AI use alone is on track to cost the company billions of dollars in 2026.
What the memo actually says
Two things, mostly. Meta is building an internal platform to track AI usage and spending in real time, and it wants engineers leaning on its own coding assistant, MetaCode (formerly Devmate), instead of outside tools. The structured version of this, with budgets and allocation decisions, arrives in 2027. Until then employees and teams have, by Meta's own admission, almost no visibility into what they are spending.
The tracking lives in a centralized dashboard the company calls AI Gateway. New controls roll out to staff in the coming weeks.
The part where the incentive ate itself
Here is the context the memo does not include. Back in November, Meta's head of people Janelle Gale told employees that "AI-driven impact" would become a core expectation in 2026, baked into performance reviews. The message landed: use AI, or explain why you didn't.
Then in April an employee-built leaderboard called Claudeonomics, named after Anthropic's Claude, started ranking the top 250 token users with titles like "Token Legend." Over one 30-day stretch the dashboard logged more than 60 trillion tokens across Meta's workforce. The Pragmatic Engineer pegged the raw figure at 60.2 trillion tokens, which at Anthropic's list prices would run around $900 million a month. Meta almost certainly pays far less than list, but even a steep enterprise discount leaves a nine-figure number sitting there.
And some of that was theater. Reporting described employees leaving AI agents running idle for hours, producing nothing, just to climb the board. The leaderboard came down within days of the coverage. Meta says the employee took it down voluntarily.
Tokenmaxxing, then
The Silicon Valley term for this is "tokenmaxxing," treating consumption as a stand-in for output. It is an input metric wearing an output metric's clothes. Meta CTO Andrew Bosworth reportedly told staff that an engineer spending their salary on tokens could hit 10x productivity with no upper limit, which is the kind of claim that sounds great in an all-hands and terrible on a finance review.
So you can read Tuesday's memo as a course correction. Reward token usage, watch token usage explode, then build the dashboard you should have had first. Meta is not alone here. Uber reportedly burned through its entire planned 2026 AI coding budget in the first four months.
The new controls and the MetaCode push start reaching employees over the next few weeks, with the full budgeting framework due in 2027.
