The $137 Engineer

Here are two numbers from the same week. The median software company spends one hundred and thirty-seven dollars per engineer per year on AI. Anthropic spends roughly two million dollars of compute per employee per year. Both are real, both were reported by people trying to be careful, and the distance between them is the whole story of where this is going.

The figures come from Tomasz Tunguz, a venture capitalist at Theory Ventures, in a June 29 analysis with the deliberately flat title "When AI Costs More Than the Engineer."¹ It is not a manifesto. It is a spreadsheet with a thesis, drawing on Goldman Sachs, the Ramp AI Index, Levels.fyi, and Fortune, and its thesis is a line on a chart that has not happened yet.

The three numbers

Consider the floor first: the median company spends $137 per engineer per year on AI. That is a rounding error, a couple of seats of a coding assistant, the cost of a nice dinner. If you only looked at the median, you would conclude that AI is a cheap productivity tweak most firms have barely started paying for.

The frontier of adopters looks different. The top one percent of companies spend $89,000 per engineer per year on AI, which Tunguz puts at forty percent of a fully loaded two-hundred-twenty-four-thousand-dollar senior-engineer salary. At the leading edge, the tooling already costs almost half of what the human costs, and not the human's output but the human.

The lab itself is the outlier. Anthropic, by Tunguz's arithmetic, spends something like $2 million of compute per employee per year: roughly five thousand employees set against about ten billion dollars of 2026 inference and training spend. Put that next to all-in compensation of around five hundred thousand, and the company that sells the model spends roughly four times as much running the machines as it pays the people who build them.

At the leading edge, the tooling already costs almost half of what the engineer costs. Not the engineer's output. The engineer.

The line that has not happened yet

A snapshot of $137 versus $2 million is just an inequality, and inequalities are easy to shrug off as the gap between a hobbyist and a hyperscaler. The reason to keep reading is the slope. Tunguz's base case projects AI spend reaching about 140 percent of a fully loaded engineer salary by 2029, roughly three hundred sixty-three thousand dollars a year, per engineer, on AI. The driver he leans on is a Goldman Sachs projection of a 24-fold rise in token consumption by 2030.

In the base case, in three years, the average company's AI bill for a developer crosses the developer's own salary and keeps climbing. The tool costs more than the person it was bought to help. The title is not a metaphor; it is a date.

Two honest caveats, because this is the kind of number that wants to be overstated. First, these are Tunguz's modeled scenarios, base and bull cases, not measured facts. A projection is an argument about the future dressed as a chart, and token prices have historically fallen as fast as consumption has risen, which could bend the curve back down. Second, the per-engineer framing counts spend against headcount, not against output; if one engineer with the tooling does the work of three, the economics look different than the raw ratio suggests. Tunguz's own point is not that the line is destiny. It is that almost nobody is pricing for it.

Why the floor is the tell

The most revealing number is not the two million. It is the $137. The frontier labs and the top one percent are spending like they believe the curve. The median company is spending like AI is a browser extension. That gap is not a disagreement about capability; both groups have access to the same models. It is a disagreement about what the meter is going to do, and only one side has run the math.

Everything we have been writing about this summer is what the top of this curve feels like from the inside: the classifier that reroutes your request to a cheaper model, the reasoning budget that quietly shrank, the token leaderboard a company built and then panicked over, the vendor telling its own staff to "demand efficiency." The $137 company has not felt any of it yet. Tunguz's chart is a note passed from the people who have to the people who haven't, and the note says: this gets expensive faster than you think, and the number on your invoice today is the smallest it will ever be.