Enterprise AI Spending Hit $176.8 Billion, and Companies Spent 2.5 Times What They Budgeted, Because Agents Do Not Clock Out
The budget was written for seats. The bill arrived for work.
A seat licence is easy to budget: count people, multiply by a price, and the number stays still. An agent runs in parallel, keeps going after its owner logs off, and consumes more with every longer job. Price per user, and the bill still scales per task.
On November 3, Menlo Ventures released its fourth annual State of Enterprise AI report. It puts enterprise AI deployment spending at $176.8 billion by September 2026, up 5.4 times from $33 billion in 2025, with growth accelerating from 3.9 times the year before. Over three years, spending has grown roughly 100 times, from $1.7 billion in 2023 through $8.5 billion in 2024 and $33 billion in 2025.
Adoption barely moved. Spending per company did the rest.
The surge did not come from a wave of new buyers. The share of enterprises paying for AI rose only from 55 percent to 65 percent. The growth came from companies already in, spending far more than they intended.
On average, companies spent 4.7 times their 2025 level, against a plan of 1.9 times. That is roughly 2.5 times the budget. The median tells a calmer version: median spending reached 2.2 times last year against an expectation of 1.8 times. Menlo calls the gap between average and median a fat head: a small group of power users accounts for a disproportionate share of spending and grew far faster than everyone else.
A market growing through heavy users rather than new adopters has more headroom, because the long tail has barely started, and more budget risk, concentrated in the teams moving fastest.
Agents are now the main customer
Menlo’s clearest evidence comes from OpenRouter, the routing platform that gives developers access to hundreds of models through one interface. On OpenRouter, agents now generate 80 percent of all tokens, 130.9 trillion tokens a week, roughly seven times human-driven usage.
Agent traffic tripled in the week after Anthropic released Opus 4.5 in November 2025, the model Menlo credits with first letting AI move from answering questions in turns to executing multi-step tasks with limited oversight. It then overtook human traffic in early February, after the launches of Opus 4.6 and GPT-5.3-Codex. Since February, agent consumption has grown 37 times, against five times growth for human-driven usage.
That is why the old unit stopped working. Human usage tracks headcount and hours. Agent usage tracks work assigned. Menlo’s summary is the line to keep: spending is now tied to work performed rather than to users.
Where the $176.8 billion actually went
Menlo splits the money into using AI and building AI, and the split is itself a finding:
- Deployment, using AI: $176.8 billion, made up of infrastructure at $99.5 billion and applications at $77.3 billion, up from $19 billion last year.
- Development, building AI: $102.5 billion, made up of training compute at $87.2 billion and training data and services at $15.3 billion.
For every $1 spent building models, enterprises spend about $1.70 putting them to work. The buildout is underwritten by deployment demand, and agents are now driving that demand.
Inside applications, the money concentrates where agents can act, learn a workflow, and have a result checked:
- Coding: $34.0 billion.
- Chatbots, assistants and productivity tools: $25.0 billion.
- Together, those two are 76 percent of the $77.3 billion application market.
The labs take much of that directly. Their first-party applications total $40.0 billion: Claude Code and Codex account for $24.0 billion, or 70 percent of coding spend, and enterprise Claude, ChatGPT and Gemini make up $15.5 billion, or 62 percent of assistant spend.
Three labs take 69 cents of every dollar
Concentration rose sharply. Anthropic, OpenAI and Google together command $122.7 billion, 69 percent of all deployment dollars, up from 50 percent last year. Their spending grew 7.5 times, against 3.3 times for everyone else.
At the API layer it is closer to a two-horse race:
- Anthropic: 40 percent to 55 percent of enterprise LLM API spend, on the strength of Opus models at long-horizon tasks and coding.
- OpenAI: 27 percent to 31 percent, with a strong second half after its GPT-5.6 family in July and GPT-6 in September.
- Google: 21 percent down to 9 percent, with its remaining API spend skewing toward cheaper Flash models.
- Anthropic and OpenAI together now take 86 percent of API spend, up from 67 percent.
Behind them sits the puzzle. Grok 4.7 and open models such as Kimi K3, DeepSeek V4 and GLM-5.3 approach frontier performance at a fraction of the price, yet together they hold just under 5 percent of API spend. Buyers are paying up for capability rather than trading down for price.
The price war is real, and buyers are ignoring it at the top
Between the third quarters of 2025 and 2026, frontier models more than doubled their share of tokens, from 18 percent to 42 percent, and their share of spending rose from 35 percent to 63 percent. Open-source models also gained, from 11 percent to 19 percent of tokens. Smaller closed models fell from 29 percent to 13 percent of tokens.
Open source now handles about one token in five but takes only about one dollar in 25, 4 percent of spend against 19 percent of usage. Small closed models lost share despite price cuts of 80 percent or more at the low end, squeezed by open models below and frontier models above.
The most popular models tell the story in prices. A year ago they were Claude Sonnet 4.5 at $3 per million input tokens and $15 per million output tokens, and GPT-5 at $1.25 and $10. Today they are Claude Opus 5.5 at $4 and $20, and GPT-6 Astra at $10 and $60. For work an agent does unsupervised for hours, a failed run costs more than a pricey successful one.
This is mostly new money, not software money moved around
Earlier this year, fears that AI would replace traditional software triggered a selloff Menlo calls the SaaSpocalypse. Its data points the other way. New budgets funded 55 percent of incremental AI spend, existing software budgets funded 22 percent, and only 9 percent, $12.6 billion, moved to AI-native vendors. Menlo found limited evidence of companies replacing core CRM, HR or finance systems with AI. AI deployment spend now equals more than 30 percent of public cloud revenue, up from 6 percent a year ago. Who wins depends on the category: coding is 96 percent AI-native, while incumbents gain in IT at 91 percent, data science at 88 percent, and HR at 68 percent.
Read the report with its limits attached
Menlo is a participant as well as an observer, and it says so. Four limits belong in any board slide built from these numbers:
- Modeled totals, not audited accounts. The headline figures are Menlo estimates, built bottom-up from survey responses, public filings and third-party reports.
- A survey of nearly 400 people. The survey, run by Morning Consult from September 8 to 26, 2026, covered nearly 400 U.S. enterprise decision-makers responsible for more than $4 billion in AI spending. That is a serious sample for spending behaviour, and a U.S.-only one.
- Disclosed investments. Menlo invests in Anthropic and in OpenRouter, the two sources behind its share and token findings. The disclosure does not invalidate the data, but the firm benefits when both look strong.
- Spending is not return. The report measures what companies paid. It does not show that the $176.8 billion produced matching productivity or profit. An overrun can signal value found, or control missing, and the same 2.5 times figure can be either.
OpenRouter is one channel, favoured by developers and agent builders, not a census of enterprise AI. Its 80 percent agent share is a leading indicator, not a typical company’s current average.
What to do with this, this week
- Rebudget AI as consumption, not seats. Move the line to a monthly reviewed usage budget with a forecast range. If your plan assumes 1.9 times growth, know what 4.7 times would do to it.
- Cap the fat head first. Find the small group of teams and agents driving most of your tokens, and put per-team and per-agent spend limits, alerts and review on them before rolling controls out to everyone.
- Price finished work, not tokens. Track cost per accepted task: a merged pull request, a resolved ticket, a reviewed contract. That unit tells you whether a frontier model at $10 and $60 beats a cheaper model that needs two retries and a human rescue.
- Route by risk, deliberately. Keep long, unsupervised, high-stakes work on the frontier tier. Move verifiable, retryable work to open or small models, where usage is already 19 percent of tokens at 4 percent of spend.
The bottom line
Enterprise AI did not grow 5.4 times because ten points more companies bought it. It grew because the customer changed: most consumption now comes from agents doing assigned work, at 130.9 trillion tokens a week on one platform alone, and work expands to fill the capacity given to it. That is why budgets written for people missed by 2.5 times, and why three labs take 69 percent of the money. The question for 2027 is how much work a company will hand to machines it pays by the task, and whether it can prove the finished work was worth the bill.