AI Agent UpdateEnterprise AI spend hits $176.8B, up 5.4x — agents now generate 80% of tokens
Menlo Ventures estimates enterprise AI spending reached $176.8 billion by September 2026, up 5.4 times from $33 billion in 2025 — and AI agents now generate 80% of tokens on OpenRouter.
The News
Enterprise AI spending reached $176.8 billion by September 2026, up 5.4 times from $33 billion in 2025, according to Menlo Ventures' 2026 State of Enterprise AI, released November 3. The report is based on a survey of nearly 400 US enterprise decision-makers overseeing more than $4 billion in AI spending, fielded September 8–26 by Morning Consult. Its headline is not just that spending grew, but who — or what — is doing the spending: AI agents have become the dominant consumers of AI.
Agents now generate 80% of tokens on OpenRouter, about 130.9 trillion tokens per week, roughly seven times human-driven usage. Agent token consumption is up 37 times since agents overtook human traffic in February 2026. In other words, within about seven months, the typical user of frontier AI stopped being a person typing a question and became a piece of software carrying out tasks — writing code, searching documents, filling forms — on a person's behalf.
The money is concentrating as it grows. Anthropic, OpenAI and Google together capture 69% of deployment spending, up from 50%, while coding at $34.0 billion and chatbots and assistants at $25.0 billion make up 76% of the $77.3 billion application market. The report's authors are clear about their own position: these dollar totals are Menlo estimates, and Menlo discloses that it invests in Anthropic and OpenRouter. That does not invalidate the survey, but it means the most enthusiastic numbers deserve a second source before they are treated as fact.
Enterprise AI spend hit $176.8B — up 5.4x — and agents now eat 80% of tokens.
Why It Matters
This matters because it answers the question every AI sceptic asks: is anyone actually using this stuff, or just buying it? Token data is hard to fake. When agents generate 80% of tokens on a major routing platform, it means software systems are routinely doing multi-step work — and companies are paying for the computing to run them. That is the difference between a product demo and a product.
It matters for workers, too, in a complicated way. Agents doing the bulk of AI work does not automatically mean jobs disappear tomorrow — most agents today do narrow tasks inside larger human workflows, like drafting code or summarising documents. But the direction is unmistakable: the unit of AI work is shifting from a person asking a question to a system carrying out a job. Companies that learn to supervise agents well will get more done; companies that ignore them will compete against rivals who do not.
For investors, the concentration figure is the strategic point. If 69% of spending flows to three labs, then owning the platforms those labs run on — Google's cloud and models, Amazon's cloud — is a leveraged bet on the whole enterprise AI market. Anthropic and OpenAI are private, so they are named here only as context, with Google and Amazon as the public read-throughs. The risk is that concentration invites regulation and price wars; the opportunity is that the market itself is growing more than fivefold in a year.
Analyst's Take
The verdict is Bullish, because the most important number here is not the spending total but the token share. Spending can be driven by fashion — companies buying AI tools because competitors did — but token consumption is usage: agents generating 130.9 trillion tokens a week are doing work, and work that gets done tends to get paid for again. A market where automated software is the main customer is a market with a very different ceiling from one where humans type prompts after lunch.
The concentration risk is the honest counterweight. When three labs capture 69% of spending, the enterprise AI market starts to look like an oligopoly, and the read-through tickers named here — Google directly, and Amazon as a major cloud and model host — benefit from that concentration, even though Menlo's own data shows Google's share of enterprise API spend falling from 21% to 9% while Anthropic and OpenAI gained. Smaller builders and the companies renting from the big three face the opposite pressure. But for the AI agent thesis itself — that agents, not chatbots, are where the money is — Menlo's report is the strongest evidence of the year so far, caveats and all.
By The Numbers
FAQ
What is enterprise AI spending, in plain terms?
Money companies spend to actually use AI in their businesses — buying software, paying for model access, building internal tools — as opposed to research or hype. Menlo's $176.8 billion is its estimate of deployment spending, not the total value of AI companies.
What is a token, and why does token share matter?
A token is a chunk of text a model reads or writes — roughly part of a word. When agents generate 80% of tokens, it means automated software, not people typing prompts, is now doing most of the work on that platform. That is a direct measure of agents taking over real tasks.
What is OpenRouter?
A marketplace that lets developers send work to many different AI models through one connection, and compare them. Because it sees traffic across models, its token data is a useful window into who is actually using AI — humans or agents.
Why does concentration among three labs matter?
Anthropic, OpenAI and Google together capture 69% of deployment spending, up from 50%. That means most enterprise money flows to three suppliers, which gives them pricing power and makes it harder for smaller model makers to break in — even as the total market grows fast.
What is the caveat on these numbers?
They are Menlo Ventures' estimates from its own survey, not audited accounts. The survey covered nearly 400 US enterprise decision-makers overseeing more than $4 billion in AI spending, fielded September 8–26 by Morning Consult. Menlo also discloses that it invests in Anthropic and OpenRouter, so its enthusiasm for those names should be read with that interest in mind.
👁
Source: Menlo Ventures State of Enterprise AI 2026