A Theory Ventures analyst projects per engineer AI bills at $363,000 by 2029, above a $224,000 engineer salary, inverting 'AI replaces humans' economics with model efficiency versus usage as the swing.
If Theory Ventures' Tom Tunguz is right about where AI compute bills are headed, the 2029 budget cycle at heavy AI adopters will not be a hiring decision. It will be a procurement decision. And the dominant "AI replaces the engineer" narrative is about to invert.
The number Tunguz keeps returning to is $89,000. That is the per-engineer AI bill at the top 1% of customers tracked by Ramp's June 2026 AI Index, a vendor data panel that reads card spend from mostly US small and mid-market businesses. The fully-loaded senior engineer salary it is benchmarked against is $224,000. The arithmetic is unambiguous: the AI tab already runs at 40% of the engineer, paid to a vendor that scales with usage rather than headcount.
The median customer on the Ramp panel spends $137 per engineer per year. The bimodality is the actual story: power users and laggards, with little mass in between. A company on the wrong side of that distribution is not getting cheaper software. It is paying a second payroll that hardly budges when headcount falls.
In an analysis published this month, Tunguz models three paths for the per-engineer AI bill by 2027, with a bear case at $106,000, a base case at $164,000, and a bull case at $258,000. By 2029, the same model puts the bear case at $106,000, the base case at $363,000, and the bull case at $596,000. The salary reference stays flat at $224,000. In the base case, the AI bill clears the paycheck by more than 60%.
The mechanism is concrete and runs against the dominant media narrative. Frontier capability requires frontier inference compute. Inference compute scales with usage. Cheaper tokens do not lower spend by default. They tend to raise it. This is what economists call Jevons paradox, and it is one of the central objections raised on Hacker News when Tunguz's piece landed: usage growth has eaten every efficiency gain shipped in the last three years, and there is little reason to assume the next three behave differently.
Epoch AI's revenue-per-employee dataset puts Anthropic at roughly $14 million per employee and OpenAI at roughly $6.5 million per employee. Tunguz estimates Anthropic is spending about 2.3 times its payroll on compute this year, around $10 billion on inference and training combined, against about $2 million in compute per employee and an all-in comp above $500,000. Both are analyst estimates from public reporting, not audited disclosures, and the ratios swing with headcount definitions and revenue windowing. The capital intensity is payroll-adjacent before any revenue lands. The frontier-lab business model is not yet "AI replaces the engineer." It is "AI is the engineer, and the engineer bills for electricity."
The line-crossing inverts the standard economic argument for AI. A company that crosses it is no longer substituting machines for people. It is substituting capex bills for headcount budgets, and the cash economics of headcount, a fixed line on the income statement, become the cash economics of cloud bills, a variable line that scales with seats, tokens, and decisions. That is a different corporate-finance problem, and the tools to manage it, including usage budgeting, model routing, and ROI plumbing, are still early.
If model efficiency improvements outpace usage growth, the per-engineer bill falls back below the salary line. The bear case Tunguz models assumes efficiency wins. The bull case assumes usage outruns everything. The empirical track record of the last three years has been closer to the bull case. The open-weights community and frontier labs are not standing still, and an efficiency surprise is the single variable that breaks the trajectory.
The number that matters for planning cycles right now is not the 2029 scenario. It is the $89,000 per engineer at the top 1%, and the fact that almost no Ramp customer sits in the middle of the distribution. Procurement leaders are not facing a slowly rising AI bill. They are facing a fork: build the discipline to manage a compute bill on track to clear the salary line, or watch it cross.