AI inference compute is quietly becoming a fourth line item in software engineer offer letters, sitting alongside salary, bonus, and equity — and the numbers are getting hard for finance teams to ignore.
Tomasz Tunguz of Theory Ventures has put the sharpest figure on it: $100,000 in annual inference costs could push the fully loaded cost of a 75th-percentile engineer from $375,000 to $475,000. That's AI compute representing roughly 21% of total compensation — not a rounding error, and not something CFOs can keep tucking into the software budget.
OpenAI President Greg Brockman has been making the case publicly that "the inference compute available to you is increasingly going to drive overall software productivity." The implication is that token access isn't just a convenience — it's a career-level resource, the kind of thing that separates a productive engineer from a slow one.
The hiring market is already moving. Thibault Sottiaux, engineering lead for OpenAI's Codex service, says job candidates are asking about dedicated inference budgets in interviews. Six months ago you might have mentioned a Copilot subscription in the same breath as a gym membership. Now engineers are treating it like an RSU vesting schedule — something worth interrogating before signing. Sottiaux also notes that per-user compute consumption is outpacing overall user growth, which means costs are rising faster than headcount.
Hakeem Shibly of Levels.fyi has started seeing compensation submissions that formally list "Copilot subscription" as a pay component. Compensation databases tend to capture structural shifts early, and this one is worth watching.
The downstream effects for AI providers could be substantial. If token budgets become standard in offer letters, OpenAI and Anthropic gain a new kind of enterprise leverage: companies competing for talent partly on compute generosity. Peter Gostev of Arena has already floated the idea of dedicated recruitment platforms where token budgets sit alongside salary ranges. Whether that materializes or not, Tunguz's proposed benchmark — productive work per dollar of inference — may become a standard workforce metric before most finance leaders have even added the line item to their models.