Menlo Ventures published its fourth annual State of Enterprise AI report on November 3, and one figure has travelled fastest in Chinese-language coverage: 89% of what enterprises spend on open-source models goes to Chinese models.
According to the breakdown reported in the press, Moonshot AI's Kimi accounts for 61%, DeepSeek for 15%, Zhipu AI's GLM family for 11%, and Meta's Llama for just 7%. A year ago, Llama was still the most common default for US companies deploying open models.
First, look at the denominator
That denominator is small. Another figure in the report says open-source models make up only 4% of enterprise model spend, yet 19% of usage. Little money, high volume: companies mostly run open models on price-sensitive, high-call-volume workloads.
On the closed-source side, Anthropic and OpenAI together capture 86% of enterprise LLM API spend through their APIs, up from 67% a year earlier. Anthropic rose from 40% to 55%, OpenAI from 27% to 31%, and Google fell from 21% to 9%. The three leading labs combined account for 69% of all deployment spend, roughly $122.7 billion.
The report estimates enterprise AI deployment spend at about $176.8 billion as of September this year, versus $33 billion in 2025. The data comes from a survey of about 400 US enterprise decision-makers who together oversee more than $4 billion in AI budgets.
A few caveats on scope:
- All respondents are in the United States; some coverage describes the figure as "global enterprise spend," which is a looser wording than the data supports.
- The Kimi 61%, DeepSeek 15% and GLM 11% split comes from South China Morning Post and Cailian Press accounts; we could not find the original table in the report's English summary.
- All of these are Menlo's survey results and estimates, not audited market statistics.
What it means for Chinese model makers
Take Kimi, the largest share: 61% of 4% works out to roughly 2.4% of total US enterprise model spend. That is small on a bill dominated by closed-source APIs, but it shows one thing: when US companies pick an open model, Chinese models are already first in line.
For Chinese vendors, the limits of this route are also clear. Open weights get self-deployed by enterprises or hosted by third parties, so most of the money flows to cloud providers and inference hosts, and the API revenue that model developers collect directly is limited. A considerable part of the "share" attributed to Kimi, DeepSeek and GLM in this report may be inference fees that companies pay to hosting providers. Whether Menlo's methodology separates that layer is not stated in the public material.
There is also a time lag. The survey was run in September; since then several Chinese models have shipped new versions, and Llama has not stood still either. Those changes will only show up in the next report. The report also does not break out how many respondents are government contractors or work in sectors such as finance where procurement is restricted, and those are exactly the customers most likely to drop Chinese models for compliance reasons.
Two other numbers from the report
The first is agent traffic. Citing OpenRouter data, the report says agents now account for 80% of tokens on that routing service, about 130.9 trillion per week, roughly seven times the volume driven directly by humans. Agent traffic overtook human traffic in early February this year.
The second is coding. Enterprise AI application spend is about $77.3 billion, of which coding accounts for $34 billion, and Claude Code and Codex together make up $24 billion, about 70% of coding spend. Companies are concentrating their money on a handful of coding tools, and open models have almost no presence there; their home turf remains back-end batch processing and cost-sensitive calls.
Sources: Menlo Ventures State of Enterprise AI annual report, South China Morning Post, CocoLoop, Cailian Press; the report summary was used to verify the open-source spend share, API shares and survey sample, while the model-level shares follow the media accounts.