Menlo Ventures Report: 89% of US Enterprise Spending on Open-Source Models Goes to Chinese Models

US enterprises' open-source model bills: 89% paid to Chinese AI

Guancha.cn Download the client | Exclusive first look

November 06, 2026 20:33:52 From Shanghai

Updated November 06, 2026 20:28:24

► Article by Guancha.cn Lai Jiaqi, edited by Lü Dong

The "2026 State of Enterprise AI Report" released on November 3 local time by the well-known US venture capital firm Menlo Ventures shows that Chinese models account for as much as 89% of US enterprises' spending on open-source AI models.

Among them, Moonshot AI's Kimi ranks first with a 61% share, followed in order by DeepSeek, Zhipu GLM, and Alibaba's Tongyi Qianwen (Qwen), with shares of 15%, 11%, and 3% respectively. The most-used Western open-source model in the report's statistics, Meta's Llama, accounts for only 7%.

The core data of this report comes from a survey commissioned by Menlo Ventures and conducted by Morning Consult. The survey was conducted between September 8 and September 26 this year and asked 400 US enterprise decision-makers, who are responsible for a total of more than $4 billion in AI spending.

The results show that in 2026, enterprise spending on AI applications surged from $33 billion a year earlier to $177 billion. Of this, spending on the three AI models from Anthropic, OpenAI, and Google was $122.7 billion, with their share rising from 50% last year to 69%.

The growth of open-source models is equally striking, with their share rising from 11% to 19%. In terms of usage, from September 2025 to September 2026, token usage of open-source models grew 22.5 times, reaching about 2.9 quadrillion tokens per month, a growth rate clearly exceeding the 12.1 times of closed-source models.

The sharp increase in open-source model usage is partly due to their price advantage. The average token price of open-source models fell 3.3 times to about $0.13 per million tokens; by comparison, the average price of closed-source models is about $0.70 per million tokens.

Another reason, the report mentions, is that open-weight models give developers a degree of control that closed-source models find hard to provide: enterprises can perform post-training on this basis to adapt models to their own business, thereby reducing costs and lessening dependence on leading model companies.

However, this market is still at an early stage. In 2026, the proportion of AI developers adopting post-training technology rose only from 6% in 2025 to 10%; and among developers who have already carried out post-training, 91% rely on external partners.

The report specifically cites a case: the US AI coding company Cursor is based on Chinese open-source models and uses Fireworks to provide inference infrastructure, supporting model calls during the reinforcement learning process.

In addition, the report believes that the rapid development of agents is changing the way enterprises use AI models and has also become an important factor driving demand growth.

Unlike traditional chatbots, agents can independently perform complex multi-step tasks, process work in parallel, and continue running with little human intervention. As agents take on more and more practical work, the frequency of enterprises' model calls and token consumption also increase substantially.

Data shows that on the OpenRouter platform, agents currently already contribute 80% of token usage, equivalent to about 7 times the usage of human users. Since agent traffic surpassed human users in February this year, it has grown 37 times. This means that the main users of AI models are gradually shifting from people asking AI questions to machines using AI to autonomously perform tasks.

This trend has also brought new market opportunities for open-source models. The operation of agents generates a large amount of token consumption, and the advantage of open-source models in inference costs makes them better suited to undertake large-scale, high-frequency model call tasks. Especially against the backdrop of enterprises hoping to control AI operating costs and reduce dependence on a small number of leading model suppliers, Chinese open-source models are expected to further expand their application scope.

Source | Guancha.cn

"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of "Dafeng Hao", a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."