Enterprise AI spending surged 5.4-fold in one year, yet the proportion of companies willing to pay increased only from 55% to 65%. Menlo Ventures has just released the “2026 Enterprise AI Report.” As of September this year, enterprise AI deployment spending reached $176.8 billion, up from just $33 billion in the same period last year. Several intriguing data points stand out: First, Agents have become the largest consumer of AI compute power. On the OpenRouter platform, 80% of token consumption comes from Agents. Since surpassing human direct usage in February this year, Agent traffic has grown 37-fold. Second, the model market is rapidly consolidating around leaders. Anthropic now accounts for 55% of enterprise LLM API spending, OpenAI holds 31%, and together they capture 86%. Google’s share has dropped from 21% last year to just 9%. Another striking figure: the enterprise AI applications market has reached $77.3 billion, with programming tools alone accounting for $34 billion. Claude Code and Codex together capture 70% of spending in the programming tools segment. I believe the most significant shift here is a change in the logic of AI commercialization. In the past, SaaS revenue primarily correlated with the number of users—how many employees a company had, and thus how many licenses it could sell, largely determined its revenue ceiling. But Agents are different. A single employee can run dozens of Agents simultaneously—Agents that work continuously, repeatedly invoke models, and consume tokens. Even without hiring a single new employee, a company’s AI spending could multiply several times over. This explains why the percentage of paying enterprises rose by only 10 percentage points while market size expanded 5.4-fold. More notably, 55% of new AI spending comes from newly allocated budgets, while only 22% comes from reallocated spending on existing software. At this stage, AI is creating a substantial new software market. However, this growth exhibits clear structural disparities. Programming tools have quickly reached a $34-billion market because code operates in a well-defined execution environment with verifiable outcomes—making it easy to measure whether an Agent performs well. In contrast, many Agents claiming to replace white-collar workers still struggle with autonomously completing tasks and validating results. Therefore, I’ll be paying closer attention to one key metric: how much real work enterprises are willing to delegate to Agents, and how much they’re willing to pay per completed task. As AI begins consuming AI itself, the correlation between model invocation volume and employee headcount will weaken further. This may well become the most important variable to study in the AI market over the coming years.