Companies overspent their AI budgets 2.5x as enterprise AI hit $177B

The average company planned to roughly double its AI budget this year. It ended up spending about 4.7 times what it spent in 2025. That gap between plan and reality is the most telling number in the 2026 State of Enterprise AI report that Menlo Ventures published on November 3, 2026, and it explains almost everything else in it: AI agents do not scale with headcount, they scale with work, and work turns out to be a lot more expensive than seats.

Menlo estimates that enterprise spending on deploying generative AI reached $176.8 billion by September 2026, up 5.4 times from $33 billion a year earlier. It is the firm's fourth annual edition of the report, and Menlo also put out an accompanying press release on GlobeNewswire. Below is what the numbers say, where the money goes, and what you should keep in mind before you quote them.

A market that grew 100 times in three years

Menlo's series starts at $1.7 billion in 2023. It grew to $8.5 billion in 2024, a 5x jump, then to $33 billion in 2025, 3.9x, and now to $176.8 billion. That is roughly a hundredfold increase in three years, and the growth rate went up again this year instead of slowing down. More than $140 billion of new annual spending appeared in twelve months, which the authors point out is more than four times the size of the whole market a year ago.

To put that in proportion, Menlo says AI deployment spend now equals more than 30% of public cloud revenue, up from about 6% a year ago. It also estimates that Anthropic and OpenAI alone account for more than a quarter of the revenue of AWS, Azure and Google Cloud, much of it for training.

The report separates using AI from building it. Deployment, which is what the $176.8 billion figure measures, splits into $99.5 billion for infrastructure (model inference, inference platforms, agent infrastructure) and $77.3 billion for applications. Development, meaning training compute plus training data and evaluation services, adds another $102.5 billion. In Menlo's words, every dollar spent building models now generates about $1.70 of spending on putting them to work.

Agents broke the budget

The reason for the overshoot is not that more companies adopted AI. Adoption among enterprises only moved from 55% to 65%, according to the survey. What changed is how much each adopter uses. Average spend per enterprise was 4.7x last year's level against a planned 1.9x, about 2.5 times what was budgeted. The median company also overshot, but more modestly, reaching 2.2x against a planned 1.8x. The big gap between average and median means a relatively small group of heavy users is responsible for a large share of the money.

Menlo's explanation is agents. An agent keeps working after the person who started it goes to lunch, it can run in parallel with other agents, and it burns tokens for every step it takes. On OpenRouter, a gateway that routes traffic to hundreds of models, agents now produce about 80% of all tokens, roughly 130.9 trillion per week and about seven times human driven usage. Agent traffic overtook human traffic in early February and has grown 37x since then, compared with 5x for human usage. Menlo itself flags that OpenRouter skews toward individual developers and startups, so it is a signal, not a census.

The survey data backs the shift on the builder side too. 74% of AI builders now use agents to manage workflows and execute tasks, against 27% of startup deployments in 2025, and 32% already run several agents that work together. That pulls new infrastructure along with it: the share of builders using code sandboxes rose from 11% to 40%, search APIs from 5% to 30% and browser infrastructure from 4% to 14%.

What this means for you: if you run AI tools in a team, per seat thinking is the wrong mental model for cost. Usage based agent work is where the bill grows, and it grows fastest for the people who are best at using it.

Anthropic takes 55% of enterprise API spend

The headline that will travel furthest is the model market. Anthropic and OpenAI together now capture 86% of enterprise LLM API spend, up from 67% a year ago. Within that, Anthropic grew its share from 40% to 55%, which Menlo attributes to its Opus models' strength on long running tasks and coding. OpenAI also gained, from 27% to 31%, helped by a strong second half with its GPT 5.6 and GPT 6 releases. Google fell furthest, from 21% to 9%, as its API spend shifted toward cheaper Flash models.

There is an important caveat here, and it is printed in the report itself: Menlo Ventures is an investor in Anthropic, and it lists Anthropic among its portfolio companies. The share numbers come from survey respondents reporting which models carry their production workloads, weighted by AI spend, not from billing data. That does not make them wrong, but it means you should read the 55% as a well informed estimate from an interested party, not as an audited market share.

Frontier labs as a group also pulled away. Anthropic, OpenAI and Google together command an estimated $122.7 billion, or 69% of all enterprise deployment dollars, up from 50%. Their spending grew 7.5x while everyone else grew 3.3x. The labs' own apps are a big part of that: Claude Code and Codex together account for $24 billion, or about 70% of the $34 billion coding market, up from roughly $1 billion last year.

Open source: one in five tokens, one in 25 dollars

The most interesting tension in the report is open source. Its share of enterprise tokens rose from 11% to 19% in a year, and its volume grew 22.5x to an estimated 2.9 quadrillion tokens per month. Yet its share of spending fell from 6% to 4%. The reason is price: the average open source token got 3.3 times cheaper and now costs about $0.13 per million, while the blended price of closed model tokens fell only 24%, to about $0.70.

Chinese models dominate that slice. According to Menlo, they make up 89% of open source spend: Kimi alone has 61%, DeepSeek 15%, GLM 11% and Qwen 3%. Meta's Llama is the biggest Western open model at 7%. Near frontier models such as xAI's Grok 4.7, Moonshot's Kimi K3, DeepSeek's V4 and Z AI's GLM 5.3 together still hold just under 5% of LLM API spend, despite getting close to frontier quality at a fraction of the price.

That fits stories we have covered on djcroman. Self hosting an open model can cut costs dramatically for teams with the hardware and skills, and real companies are doing it. But Menlo's data suggests that, in aggregate, buyers still pay up for the most capable model. A year ago the most popular enterprise models were Claude Sonnet 4.5 and GPT 5. Today, according to the report, they are Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, and GPT 6 Astra at $10 and $60. The real price war is happening at the low end, where labs have cut small model prices by 80% or more and small closed models still lost token share from 29% to 13%.

Post training, the thing that lets companies shape open weights for their own tasks, is still rare. Only 10% of AI builders do it, up from 6%, and 91% of those rely on outside partners for infrastructure, data or the training itself.

What this means for you: if your workload is high volume and well defined, the cost gap between open and closed models has become huge. If your workload is long, messy agent work where quality decides whether the agent finishes the task at all, most companies are still deciding that the expensive model is the cheaper choice.

Coding is the killer app, and it is not close

On the application side, the market reached $77.3 billion, up from $19 billion. Two categories take 76% of it. Coding tools reached $34 billion, up 7.7x, and enterprise chatbots and assistants $25 billion, up 3x. Healthcare, the largest industry vertical, is a distant third at $3.7 billion.

Menlo's explanation for why coding wins is useful beyond this report. Agents do well where three conditions hold: they have an environment to act in, such as a repository and connected tools; they have complete context, such as code, documentation and worked examples; and their results can be verified, such as by compilers and tests. Where those conditions hold, spending follows. Customer support, with documented playbooks and clear resolutions, grew 4.3x to $2.7 billion. Human resources, where work involves more subjective judgment, grew only 2.5x to $0.9 billion, and marketing just 1.8x to $1.2 billion.

The report also says enterprise code activity is up 5x year over year, and that the labs' coding agents have not crowded out independent tools: the rest of the coding market still grew 2.9x to about $10 billion, led by companies such as Cursor, Cognition and Factory and app builders like Lovable. AI native providers hold 96% of the coding category, while incumbents such as ServiceNow, Databricks, Snowflake and Workday keep their ground in IT, data science and HR, where AI has to fit into an existing system of record.

Is AI eating software budgets?

One of the year's big fears was the so called SaaSpocalypse: the idea that AI would replace traditional software and drain existing budgets. Menlo's survey points the other way, at least for now. 55% of the incremental AI spend came from new budgets. 22% came from software budgets, but most of that, 13 points, stayed with incumbents converting existing licenses to AI products, for example moving from Microsoft Office to Copilot. Only 9%, about $12.6 billion, moved to AI native vendors. Menlo estimates that this costs the more than $600 billion application SaaS market about two points of growth, from 13% to 11%.

It also found limited evidence of companies ripping out core systems such as CRM or HR software. The authors' conclusion is that AI is expanding the software market while modestly slowing traditional SaaS. That is a reasonable reading of the data, but remember who is writing it: a venture firm whose portfolio benefits from AI spending continuing to grow.

How the numbers were made

Before you put any of these figures in a slide, it helps to know where they come from. The survey covered nearly 400 US enterprise decision makers responsible for more than $4 billion in AI spending and was fielded by Morning Consult between September 8 and 26, 2026. Market sizes are Menlo's own estimates from a bottom up model that combines the survey with public filings and third party reports, shown as annualized run rates. Earlier years were reclassified to match this year's categories. Model market shares are based on what respondents report about their production API usage, weighted by spend, and token prices come from a model triangulated against sources such as OpenRouter, Ramp and Vercel.

So the $176.8 billion is not a number anyone counted. It is an estimate built by people with good data access and a clear interest in the outcome. The broad direction is hard to argue with, since the frontier labs' own growth points the same way, but the precise shares deserve a little more caution.

The bottom line

Three things stand out. First, enterprise AI has stopped being a pilot budget: average spend per company went up nearly fivefold, far beyond plan, and agents are the reason. Second, the money is concentrating on the most capable models and on coding, where agents can check their own work. Third, open models are winning usage much faster than they are winning revenue, which makes them the cheapest way to run AI and, for now, a much smaller business than their token counts suggest.

If you are planning next year's AI budget, the practical lesson from this report is simple: budget for work, not for seats, and measure what your agents actually finish, not how many people have a license.

Sources: [Menlo Ventures, 2026: The State of Enterprise AI](https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/) (November 3, 2026), [Menlo Ventures press release on GlobeNewswire](https://www.globenewswire.com/news-release/2026/11/03/3394279/0/en/menlo-ventures-2026-state-of-enterprise-ai-report-enterprise-ai-spending-hits-177b-up-5-4x-in-one-year-as-agents-take-over.html). Disclosure from the report: Menlo Ventures is an investor in Anthropic, OpenRouter and several other companies named here.