Menlo's $177 billion enterprise AI tally is big money measured with a moving yardstick

Menlo Ventures published its annual State of Enterprise AI report on Tuesday with a headline made for board decks: enterprise spending on putting AI to work reached an estimated $176.8 billion on an annualized basis, 5.4 times the year before, and Anthropic now takes 55% of enterprise spending on large-language-model APIs. The numbers deserve a close read before they become conventional wisdom, because the most quotable findings rest on a modeled estimate, a reworked taxonomy, usage data from a platform that skews toward startups, and an author with money in the company that comes out on top.

What Menlo actually measured

The survey behind the report is modest: nearly 400 US enterprise decision-makers, fielded by Morning Consult from September 8 to 26, who are responsible for more than $4 billion in AI spending. The $176.8 billion is not a sum of invoices. Menlo describes it as the output of a bottom-up model informed by those responses, public filings and third-party reports, and says revenue is shown as an annualized run rate. It is best read as a snapshot of the spending pace around the September fieldwork, extrapolated from a sample worth a few percent of the total, not a count of dollars spent in calendar 2026.

Menlo splits the total into $99.5 billion of infrastructure (model inference, inference platforms and agent tooling) and $77.3 billion of applications, where coding at $34 billion and chatbots and workplace assistants at $25 billion account for roughly three quarters. A separate $102.5 billion goes to building models, most of it training compute. The three frontier labs, Anthropic, OpenAI and Google, collect an estimated $122.7 billion, or 69% of deployment dollars, up from about half.

The yardstick has also moved. Last December's edition, based on 495 respondents surveyed in November 2025, put 2025 enterprise generative AI spending at $37 billion, a figure that counted model training and said it excluded chips, inference and model serving. This year's edition restates 2025 as $33 billion of deployment spending after reclassifying categories; data-labeling vendors such as Mercor, for example, were moved from deployment to development. Menlo says it reclassified prior years to keep its own comparison consistent, but a reader holding last year's report cannot line the two editions up, and the growth rate depends partly on where the lines were redrawn.

The budget overshoot is an average, not a typical company

The most striking claim is that companies budgeted for roughly 1.9 times their 2025 AI spending and ended up at 4.7 times, about two and a half times plan. Those are averages. Menlo's own medians are far tamer: the typical respondent planned for 1.8 times and landed at 2.2 times, an overrun of around a fifth. The dramatic number is being pulled up by a minority of very heavy spenders, which fits Menlo's other finding that enterprise adoption rose only from 55% to 65%. As Menlo principal Derek Xiao put it to Yahoo Finance: "If you look at per-company spend, that's really what drove the inflection."

Menlo attributes the blowouts to agents that work on their own, run in parallel and keep going after people log off, so spending scales with work rather than headcount, and to a small "fat head" of power users. It cites OpenRouter data showing agents now generate about 80% of tokens, roughly seven times human-driven use, with agent traffic up 37-fold since it overtook human traffic in February. The report itself notes that OpenRouter's users lean toward individual developers and startups rather than large enterprises, and Menlo discloses it is an investor in OpenRouter. Tokens are also not dollars: by Menlo's estimate, open-source models account for about 19% of usage but only 4% of spend. For finance chiefs the lesson is narrower than the headline: most companies overran modestly, and Menlo's own explanation points to a concentrated group of heavy agent users rather than a market-wide forecasting failure.

Market share depends on what you count

Menlo estimates Anthropic's share of enterprise LLM API spending rose from 40% to 55% over the year, OpenAI's from 27% to 31%, and Google's fell from 21% to 9%. Claude Code and OpenAI's Codex alone account for an estimated $24 billion, or 70% of coding spend. These shares are derived from survey-reported production usage weighted by AI spend, not from vendor revenue. Menlo marks Anthropic as a portfolio company, and Yahoo Finance's coverage carried a disclosure that Menlo is an Anthropic investor. Disclosure is not evidence of bias, but the lab rankings are an investor's estimate, not an audited count.

Independent transaction data paints a closer race. Ramp's AI Index, built from corporate card and bill-pay records at more than 85,000 US businesses, found 55.9% of them paying for AI in September, with 43.8% paying Anthropic and 40.6% paying OpenAI. Ramp counts paying customers rather than dollars, and its base includes many smaller firms, so both data sets can be right at once: Anthropic could lead narrowly on the number of businesses and widely on API dollars because coding agents consume so many tokens. Still, a three-point lead and a 24-point lead are very different stories, and which one gets quoted depends on who is selling what.

A Gartner forecast reported by Techstrong.ai in September shows how elastic the category is. It put total worldwide AI spending for 2026 at $2.67 trillion, dominated by $1.48 trillion of infrastructure, with generative AI models specifically at just $28.3 billion, up from $13 billion in 2025. Gartner's buckets do not map onto Menlo's, and that is the point: there is no agreed boundary for the enterprise AI market, and each firm draws it where its thesis lives.

Spending is not the same as value

A survey released the same day offers a check on the euphoria. Accenture's first CIO Pulse, covering 300 CIOs and senior technology leaders at companies with more than $1 billion in revenue and fielded in August and September, found 64% scaling enterprise AI but only 7% at what it calls optimization at scale. Fifty-four percent named governance, risk and responsible-AI concerns as a leading barrier, and 51% said they lack a trusted, governed data foundation. Seventy percent mainly use AI for productivity copilots, against 37% running end-to-end agent orchestration. Accenture sells AI transformation services, so it has its own reasons to stress the gap, but its findings are consistent with a market where money is moving much faster than measurable returns.

Taken together, the direction of Menlo's report is credible: spending is concentrating among the frontier labs and in coding tools, and agent workloads are making AI bills harder to forecast. The precise magnitudes are another matter. The $177 billion is a model output, the budget overrun is a skewed average, and the lab horse race comes from an interested party. As the frontier labs move toward public markets and audited disclosures, there will finally be firmer numbers to test these estimates against. Until then, enterprises setting 2027 budgets would do better to watch their own median, not Menlo's mean.