AI Compute Already Costs Anthropic 2.3× Its Payroll — and the Gap Is Growing

New analysis by venture capitalist Tom Tunguz, drawing on the Ramp AI Index from June 2026, maps AI compute spending across the software industry and finds a striking concentration. At the frontier, Anthropic now spends approximately $2 million per employee per year on compute — against average fully-loaded compensation of over $500,000 — a ratio of 2.3x payroll. The company generates $14 million in revenue per employee, which makes the math work. But across the broader SaaS landscape, the picture is fragmented: the top 1% of companies spend $89,000 per engineer per year on AI, representing 40% of a typical engineering salary. The median SaaS company, however, spends just $137 per engineer annually — a 680x gap between leaders and the market center.

Tunguz models three scenarios for how compute costs evolve by 2029. In the bear case, continued model price declines of roughly 10x per year and competition from open-weight models keep AI spend at 40–41% of engineering salary. The base case lands at 70–140% of salary. The bull case — powered by a Goldman Sachs projection that agentic workflows will consume 24 times more tokens by 2030 — pushes AI spend to $596,000 per engineer, or 230% of annual compensation. At that point, compute crosses the salary line and becomes the dominant cost of employing an engineer. The practical implication is a structural shift in how companies measure team output: headcount becomes a secondary variable, and token-to-result efficiency becomes the primary one.