The data behind the enterprise token budget

Seventy trillion tokens a month was the peak consumption rate for a single major tech firm earlier this year, a period of “tokenmaxxing” that saw internal leaderboards and unconstrained testing. That era of experimentation is now being replaced by a calculated budgeting phase. In a survey of over 50 large enterprises, including Fortune 500 firms in aerospace and pharmaceuticals, a clear split has emerged between the median user and the high-intensity engineering department.

The median corporate AI spend remains modest, often below $100 per employee annually. However, the top 10% of customers—largely tech-forward firms and those focused on software development—are spending upwards of $7,300 per employee, with the 99th percentile reaching $90,000. These high-end users are the primary revenue drivers for labs like Anthropic and OpenAI. To manage these costs, firms like Uber have moved to establish monthly limits, such as $1,500 per employee, requiring case-by-case approval for overages.

Coding remains the dominant driver of this expenditure, accounting for an estimated 70% of current annual recurring revenue across frontier AI providers. While some organizations are downgrading default models to save costs, the broader market for “Token-as-a-Service” is expanding. Providers like Together and Fireworks now represent over $4 billion in annual revenue, suggesting that even as individual companies impose budgets, the aggregate demand for frontier and open-source tokens continues to scale alongside new verticals in cybersecurity and white-collar knowledge work.