Menlo Ventures says enterprise AI spending reached $176.8 billion by September 2026, up 5.4 times from $33 billion a year earlier. Its sharpest finding for model buyers: within enterprise LLM API spending, Anthropic rose from 40% to 55%, OpenAI from 27% to 31%, and Google fell from 21% to 9%. The numbers come from Menlo's 2026 State of Enterprise AI report, announced in a November 3 press release. Menlo is an investor in Anthropic and OpenRouter, a fact it discloses in the release. This post explains the findings and where to be careful.
TL;DR: the headline numbers and how solid they are
| Finding | Menlo's figure | How to treat it |
|---|---|---|
| Enterprise AI spend | $176.8B by Sept 2026, up from $33B | Menlo estimate from a bottom-up model |
| Enterprises paying for AI | 55% to 65% | Survey of nearly 400 U.S. decision-makers |
| LLM API spend share, Anthropic | 40% to 55% | Survey-based share of spend |
| LLM API spend share, OpenAI | 27% to 31% | Survey-based |
| LLM API spend share, Google | 21% to 9% | Survey-based |
| Frontier labs share of deployment spend | 69%, up from 50% | Anthropic, OpenAI, Google together |
| Agent share of OpenRouter tokens | 80%, about 7x human use | One channel only |
| Open-source share | About 1 in 5 tokens, about 1 in 25 dollars | Menlo estimate |
| AI application spend | About $77B, coding $34B | Coding plus assistants is 76% of it |
What did Menlo actually report?
The press release lists six key findings. In Menlo's words and numbers:
- Spend more than quintupled. From $33 billion in 2025 to $176.8 billion by September 2026, after 3.9x growth the year before. The share of enterprises paying for AI rose only from 55% to 65%, so growth is mostly existing buyers spending more, not a wave of new ones.
- Budgets were beaten. Companies budgeted nearly twice their 2025 spend, then spent 2.5 times their budget, according to Derek Xiao of Menlo. For every dollar spent training models this year, enterprises spent $1.70 putting AI to work.
- Agents dominate tokens. On OpenRouter, agents generate 80% of tokens. Since overtaking human traffic in February, agent consumption grew 37x against 5x for human-driven use.
- New budgets, not replacement. New budgets funded 55% of incremental AI spend and existing software budgets 22%. Menlo found little evidence that enterprises are replacing core systems such as Salesforce or Workday.
- Applications are a growth engine. About $77 billion, up from $19 billion. Coding ($34B) and assistants ($25B) make up 76%. Healthcare is the largest vertical at $3.7B.
- AI-native versus incumbents. AI-native companies hold 96% of coding spend. Incumbents lead IT (91%), data science (88%) and HR (68%).
Why did Google fall from 21% to 9%?
A pie chart with one green wedge pulled out, showing how enterprise LLM API spend share shifted toward Anthropic in the Menlo Ventures report
Menlo does not give a cause in the release, beyond a quote from partner Deedy Das: "Every major model release is an opportunity to shift market share." We can say what the numbers show and what they do not.
They show that, in Menlo's survey, enterprise buyers allocate a far smaller share of LLM API dollars to Google than a year ago. They do not show Google's revenue falling. Total spend grew more than fivefold, so a smaller share of a much bigger market can still be a larger dollar figure. The percentages are also share of API spend, not usage and not total cloud revenue. Google sells through Vertex, Workspace and consumer products that fall outside "enterprise LLM API spend."
Other surveys can disagree on ordering, depending on what they measure: spend, production use, card payments or tokens, and many firms use several vendors at once. Our coverage of the a16z consumer spending report shows how different consumer money looks from enterprise money. Do not read one share figure as "the" market share.
How independent is the report?
Menlo's release carries an asterisk next to Anthropic and OpenRouter: "Menlo Ventures is an investor in Anthropic and OpenRouter." Two of its headline claims rest on those firms: Anthropic's share, and the 80% agent token figure from OpenRouter. That does not make the data wrong. It is a reason to read it as an informed estimate from an interested party.
Method matters too. The report "combines a survey of nearly 400 U.S. enterprise decision-makers overseeing more than $4 billion in AI spending with a bottom-up analysis of the market." Market size figures are "Menlo Ventures estimates built from a bottom-up model informed by survey responses, public filings, and third-party reports." Token figures "draw on OpenRouter data and Menlo Ventures estimates of model mix." We have read the press release, not the full methodology section, and we cannot audit the model. U.S.-only respondents also exclude Europe and Asia, where Google, Mistral, Qwen and others may hold different shares.
For a cross-check on Anthropic's own numbers, see our posts on the amended S-1 with Q3 revenue and losses and on the reported IPO valuation range. Company filings are firmer evidence than a survey share.
What do the agent and open-source numbers mean for builders?
Agents are the buyers. Menlo's line: "enterprise spending is no longer constrained by the number of employees using it, but by how much work it can get done." If that holds, token budgets scale with workload, not seats. That fits what we reported on GitHub's Octoverse numbers on agents and pull requests and on OpenRouter's token run-rate claim. Remember the caveat: OpenRouter users skew toward developers who pick models freely, so 80% may overstate enterprise-wide agent use.
Open models win tokens, not dollars. About one in five tokens but one in 25 dollars is a large gap. It means open-weight models are mostly used for cheap, high-volume work while frontier models take the expensive jobs. Our post on Gartner's view of open-weight models and agent steps and the Vercel AI Gateway token share show other measurements with other answers. The common thread is routing: send routine steps to cheaper models and reserve frontier models for hard ones.
Coding is the biggest application. $34 billion in coding, 96% to AI-native providers, is a reminder that the market for developer tools is where the money is today. If you build in that space, you compete with firms that grew up around it.
How should a finance or procurement team read the share figures?
Share of spend is a ratio, so it moves for two reasons: one vendor sells more, or another sells less relative to the whole. In a market that grew 5.4 times, nearly every vendor can grow in dollars while shares still shift sharply. A team reading this report should ask three questions before using a share number in a plan. What is the denominator, and does it include the products you buy? Which respondents were asked, and do they look like your company? And how fast do the numbers go stale? Menlo's own quote concedes that each major release can shift share, so a figure from a survey that closed around September may already lag the models released since. Use it as a prompt for your own vendor review, not as the result of one.
It is also worth asking what the report leaves out. The release does not say how much of the 2.5x budget overrun came from agent workloads versus seat-based tools, and it does not split Anthropic's share by use case. Without that, "coding drives Anthropic's lead" is a reasonable guess, not a finding.
What should you do with this?
- Do not pick a vendor from a share chart. Test your own tasks on two or three models and compare cost per solved task.
- Design for multi-model. A router or gateway lets you switch when a new release shifts quality or price. Menlo's own quote says share can move with each release.
- Budget for overruns. Menlo says companies spent 2.5x what they planned. Set per-agent spend caps.
- Ask who paid for the study. When a report comes from an investor in one of the vendors, find a second source before you cite it.
- Read the full methodology at Menlo's report page if you plan to quote a number in a board deck.
What is still unclear?
- How Menlo defines "enterprise" and whether Google's consumer-facing and Workspace AI revenue are counted.
- How share changes if you measure tokens, requests or contracts instead of dollars.
- Whether the 400 respondents represent firms outside the U.S.
- How much of Anthropic's lead is coding agents specifically, which the release does not break out by vendor.
Bottom line
Menlo's report gives a clear direction: more money, more agents, and a concentration of spend among three frontier labs, with Anthropic well ahead in its survey. It is also an estimate from a firm with a stake in one of the winners. Use the shape of the trend, check the dollar claims against filings and your own invoices, and keep your stack portable.
Figures are accurate as of November 3, 2026, based on Menlo Ventures' press release. We have not audited the full report or its methodology.
Related reading
- Anthropic's amended S-1: Q3 revenue and net loss
- Anthropic IPO valuation: claimed vs verified
- OpenRouter token run-rate claim
- Gartner on open-weight models and agent steps
- Open models at 78% of Vercel AI Gateway tokens
- GitHub Octoverse 2026: agents and pull requests
- a16z top 100 consumer AI spending
- Menlo Ventures 2026 State of Enterprise AI