Enterprise AI spending reached $176.8 billion in September 2026, up from $33 billion a year earlier. On November 3, Menlo Ventures published the fourth edition of its State of Enterprise AI report, and its verdict was direct: growth did not come from new companies signing up, but from existing users spending much more, while agents have replaced people as the main source of work.
The report draws on a Morning Consult survey of nearly 400 US enterprise decision-makers, conducted from September 8 to 26, who oversee more than $4 billion in AI spending. The headline total is Menlo's own bottom-up estimate, built from the survey, public filings and third-party reports. The share of companies using AI rose only from 55% to 65%, yet average spending reached 4.7 times last year's level, far above the planned 1.9 times. In other words, budgets were pushed up by actual usage, not by new customers.
Where the money went: coding leads, and three vendors take nearly 70%
Application spending reached about $77 billion, up from $19 billion last year. Coding accounted for $34 billion and AI assistants and productivity tools for $25 billion, together 76% of application spending. Anthropic, OpenAI and Google together took 69% of deployment spending, up from 50% last year. Looking only at large-model API spending, Anthropic's share rose from 40% to 55%, OpenAI's from 27% to 31%, and Google's fell from 21% to 9%. Concentration is rising, and so is the cost of choosing the wrong vendor.
Agents take over usage: 80% of tokens are no longer sent by people
On OpenRouter, a model-routing platform widely used by developers, agents generated 80% of tokens, roughly seven times human usage. Agent usage overtook human usage in February 2026 and has since grown 37-fold, while human usage grew only five-fold over the same period. The report calls this shift "the customers are machines now": purchasing, calls and spending decisions are increasingly triggered by agents, and the old way of budgeting per employee is breaking down.
How to read the numbers: estimates have a methodology, so do not overstate the conclusion
Menlo states that the dollar totals are estimates, not audited figures, and discloses that it invests in Anthropic and OpenRouter, so readers should treat the share figures with some caution. The report counts spending only; it does not prove how much productivity the $176.8 billion bought. For enterprises, the more practical conclusions are three: set usage and cost caps for agents first, check the real return in high-frequency uses such as coding, and do not bet everything on a single vendor.
