Top AI Adopters Spent $90K Per Employee on Tokens in May 2026 — At 14% Monthly Growth, Engineer Salaries Are Next
The most aggressive AI adopters in the US are approaching a threshold that would have seemed absurd twelve months ago: spending more per employee on AI tokens than on the employee. New enterprise data, published through a16z and attributed to Hebbia CEO George Sivulka, puts the top 1% of AI-spending companies at $90,000 per employee annually on AI compute in May 2026. The monthly growth rate is 14.1%.
At that compounded rate, the same figure crosses $192,000 before the end of 2026. The average US software engineer earns approximately $192,000 in total compensation. Token spend is on a trajectory to match labour cost at the frontier of corporate AI adoption.
What $90K Per Employee Looks Like
The figure is per-employee AI spend, not total company spend. For a 50-person technical team at a frontier AI-adopting company, that translates to $4.5 million annually on tokens — an operating expense line that did not exist in 2024 and now rivals the team’s salary budget.
The spending is concentrated. The top 1% of companies in the dataset drive the headline number; median AI-adopting companies spend significantly less. That concentration reflects a pattern where a small set of firms — primarily enterprise software companies, coding-intensive startups, and professional services organisations running agentic workflows at scale — are running frontier models at near-continuous utilisation.
The Employment Counter-Narrative
The analysis runs against the job-displacement narrative that dominates AI coverage. Companies in the top quintile of AI spending added 10.2% more employees over a 24-month period. Companies that spent modestly on AI showed no statistically significant headcount change. Entry-level job openings grew 12% at the heaviest AI adopters.
The mechanism is familiar from prior technology cycles: AI raised the productivity ceiling high enough that demand for the output expanded faster than efficiency gains eliminated roles. A team that ships three times as much code does not need to shrink by two-thirds — it can take on three times as many projects. The marginal cost of scope falls; scope expands.
Meta’s Adam Mosseri said publicly he expects token budgets to become a standard management discipline within one to two years — the AI equivalent of travel and expense limits. At $90,000 per employee annually and 14% monthly growth, the companies at the frontier are already there.
The Price-Volume Tension
There is a structural resolution to runaway token spend: model prices are falling faster than consumption is growing for most companies.
GPT-5.6 Sol is 54% more token-efficient than GPT-5.5 xHigh on agentic coding. DeepSeek V4 Flash runs at $0.14 per million tokens. The economics of AI infrastructure are a race between falling per-token prices and rising per-company utilisation. For most of the enterprise market, falling prices are winning.
For the top 1% running frontier models at maximum effort on the most demanding tasks, the dynamics are different. They need the best model for the hardest tasks, and the best model’s price is set by whoever holds the Intelligence Index top spot this month. Claude Fable 5 at $1.57 per task on LiveBench Max Effort costs 2.7 times as much as GPT-5.6 Sol at $0.59. For an organisation with $90,000 in per-employee token spend, that difference in model routing is the difference between a manageable cost line and a budget crisis.
The companies spending $90,000 per employee are not doing so on commodity tasks. They are running frontier models on work where the quality differential justifies the price. As that cohort grows, the question is whether per-task efficiency gains arrive fast enough to prevent the crossover point from arriving exactly when the analysis predicts.