Survey finds Chinese models take most US enterprise open-model spending

Menlo Ventures, a venture firm, published the fourth edition of its annual State of Enterprise AI report, estimating that Chinese models accounted for 89% of US enterprise spending on open-source models. By model family, the report puts Moonshot’s Kimi at 61% of open-source spend, DeepSeek at 15%, Zhipu’s GLM at 11% and Alibaba’s Qwen at 3%. Meta’s Llama, which it calls the most popular Western open-source model, stood at 7%.

The shares come from a survey rather than billing data. Morning Consult fielded it from 8 to 26 September among nearly 400 US enterprise decision-makers who, between them, are responsible for more than $4 billion in AI spending. Menlo combined the responses, in which people reported the share of their AI workloads on each model, weighted by spend, with a bottom-up model of token volumes and prices informed by public filings and third-party reports. The report says its market figures are Menlo’s estimates, treats OpenRouter token data as directional because that platform skews to individual developers and start-ups, and covers US enterprises only. It does not state a margin of error.

The headline needs the denominator. Open-source models were 4% of enterprise LLM spend in September, down from 6% a year earlier, though they handled 19% of tokens, up from 11%. So the 89% is a share of a small category whose portion of spend has fallen: open models carry about one token in five but about one dollar in 25 of model spending, according to the report.

One example in the report of Chinese open weights underpinning a Western product is that Cursor built its Composer 2 model on Kimi K2.5, with Fireworks providing inference infrastructure for its reinforcement-learning rollouts. The report marks Cursor as a Menlo portfolio company. Other 2026 entries record similar choices, such as Harvey building an in-house model on Kimi K3, and the finding sits beside the continuing argument over the security implications of Chinese open models.