AI Compute Costs Now Rival Engineer Salaries at Frontier Labs, With 680× Gap to Median Companies
Summary
- • Anthropic spends 2.3× its payroll on compute — roughly $2M per employee vs ~$500k in salary in 2026
- • Top 1% of companies spend $89k/engineer/year on AI; the median company spends just $137 — a 680× gap
- • Goldman Sachs projects a 24× rise in token consumption by 2030 as agentic workflows displace chat-based usage
- • Bull scenario: AI bill per engineer could reach $596k/year by 2029, matching an entire median-SaaS employee's revenue contribution
Details
Anthropic compute ratio
Anthropic spends ~$2M compute per employee in 2026 — 2.3× its payroll — with ~$10B inference/training spend across ~5,000 staff.
AI spend gap (leaders vs median)
Top 1% of companies spend $89k/engineer/year on AI (40% of a $224k senior engineer salary); the median firm spends $137/year — a 680× gap (Ramp AI Index, June 2026).
AI lab revenue efficiency
Anthropic generates ~$14M revenue/employee; OpenAI ~$6.5M — the highest ratios in the Forbes Global 2000 (Epoch AI, 2026).
Goldman Sachs 24× projection
Agentic AI workloads are projected to drive a 24× rise in token consumption by 2030 versus current chat-dominated usage patterns.
Bear/Base/Bull cost scenarios
By 2029 AI bill per engineer: Bear $106k, Base $363k, Bull $596k. The Bull case matches a full median-SaaS employee's revenue contribution.
Token price deflation
GPT-4 class input pricing fell from $30 to under $3 per million tokens across three years — roughly 10× per year deflation.
Structural cost divergence
At AI-native firms, infrastructure costs already dominate payroll; for the median enterprise, AI is barely a rounding error — the gap is structural, not cyclical.
AI compute costs have crossed a threshold where frontier labs spend more on infrastructure than people, while most enterprises have barely started spending.
What This Means
AI compute costs have crossed a threshold where, at frontier labs, they dwarf traditional payroll — and the structure of the AI economy makes this divergence likely to persist or widen. For most enterprises, AI is still a rounding error ($137/year median), but top-1% companies are spending at a rate that could reach an engineer's full salary within two years. The wildcard is agentic workflows: if Goldman Sachs's 24× token consumption projection is even half right, AI will become a strategic cost center comparable to headcount for any serious AI user. Companies that don't model these scenarios now risk being blindsided when the transition accelerates.
Sources
- When AI Costs More Than the EngineerTomtunguz