The burn rate of an engineer could soon equal their salary

In plain terms

The head of Instagram estimates that, within one to two years, what a good engineer spends on AI tokens could cost about as much as their salary. Logical consequence: companies will start capping this budget, engineer by engineer. This isn't an HR anecdote - it's the end of an implicit assumption of our profession.

A cost variable reappears where we had forgotten it

We have been following for several weeks the return of the CFO in the AI loop. Adam Mosseri's announcement, on Lenny's podcast, marks the next step: rationing goes down to the individual. His wording is measured: « I think that you can imagine, at least in a year or two … that the burn rate of a strong engineer might be the same as their salary ». He doesn't talk about a punitive cap, but about a resource allocation, on the same level as the payroll or the GPUs - proportional to the demonstrated capacity of each to produce ROI.

The facts around it are more eloquent than the prediction. Meta closed its internal ranking of token usage when costs got out of hand - Mosseri himself qualifies the practice as « silly ». Uber exhausted its entire 2026 AI coding budget by April. Microsoft cut Claude Code licenses due to the expense. Meta, on the other hand, has not yet set a per-person cap.

Under the hood

For twenty years, software architecture has been based on a marginal cost of code production that is almost zero: the engineer's time was the only scarce resource, and we optimized that. An agentic loop that runs at night on a migration changes the equation - it consumes a resource billed per token, whose consumption grows with the size of the context, therefore with the technical debt of the repository. A poorly split monorepo no longer costs only human understanding time: it costs context at each agent iteration, on each task.

So what

The per-person cap is the wrong answer to a good question. It rations access without addressing the cause, and it creates a metric that engineers will learn to game - we've seen it with lines of code, we'll see it again with tokens.

The right answer is architectural, and it's encouraging: in a world where context is paid for, modularity, clear boundaries, and up-to-date documentation become measurable savings items, not abstract virtues that we plead in retrospect. Refactoring has just found its business case.

Mosseri also expects a drop in token prices due to competition. Likely. But the demand for an agent is elastic: if the token is worth ten times less, we will consume a hundred times more.