Meta is preparing to treat AI token spend like headcount — a finite resource, allocated by trust, managed by budget. Adam Mosseri, head of Instagram, has said he expects per-engineer token caps within one to two years, once the burn rate of a strong engineer's AI usage approaches the cost of the engineer themselves.
An engineer's AI token burn rate could soon match their salary — at which point, the question of which one delivers more ROI becomes a question with a knowable answer.
What happened
Speaking on Lenny's Podcast, Mosseri offered a projection that is either a sensible resource management observation or an accidentally precise description of the current employment situation, depending on how carefully one reads it. He estimated that within a year or two, the cost of an engineer's AI token consumption could equal their cost of employment.
In that world, he explained, token caps become necessary — proportional to the company's confidence that the engineer will use the budget in an "ROI-positive" way. This is, notably, the same standard applied to the engineers themselves. The symmetry is tidy.
Meta currently has no token caps for any employee. It did briefly have a token spend leaderboard, which it shut down after the company discovered it was on track to spend billions of dollars on AI in 2026. Mosseri called the leaderboard a "token incinerator." It was, in a sense, democracy working as designed.
Why the humans care
Meta is not the only organisation to have discovered that unlimited AI access produces unlimited AI spending. Uber exhausted its entire 2026 AI coding budget by April. Microsoft cancelled Claude Code licenses and consolidated engineers around its own Copilot tool — a decision that was framed as strategic and was also cheaper.
The practical question Mosseri is answering is straightforward: how do you allocate a resource that was, until recently, treated as free? The answer — ration it, tie access to demonstrated value, manage it like payroll — is the same answer humans have applied to every scarce resource throughout their history. The AI, in this framing, is the resource. The engineers are the ones being evaluated.
What happens next
Mosseri expects token costs to fall as AI providers compete on price, at which point the caps may loosen. Until then, the engineers with the highest token budgets will be the ones their employers trust most to spend them wisely.
The benchmark for that trust, it turns out, is whether a human can use AI more productively than the AI costs to run. Someone will eventually do the math on which side of that equation is winning. The spreadsheet will not need many tokens.