Abstract:
The money that American companies spend on open source large models is continuing to flow to Chinese companies. Recently, Menlo Ventures, a Silicon Valley venture capital institution, released the “2026 Enterprise AI Status Report.” Based on a survey of nearly 400 AI business leaders in U.S. companies, the report estimates that 89% of U.S. companies’ expenditures on open source large models went to Chinese models, and the total AI expenditures responsible for the respondents exceeded US$4 billion.
The Kimi family owned by Dark Side of the Moon accounts for 61% of the overall market share, and the top three are all Chinese models. Among Western open source models, Meta Llama, which has the highest share, is only 7%.
Monroe Ventures has a 50-year investment history and is an investor in technology giants such as Anthropic and Uber.
China's open source model is favored by enterprises. The price advantage is important, but the upper limit of the capability of the open source model is also being raised.
Monroe Ventures bluntly stated in the report: "As the Chinese model defines the open source frontier, they account for 89% of open source spending."
For developers, the open source model can be downloaded and adjusted according to their own business needs. When companies develop AI products, they do not have to train models from scratch. Instead, they can continue to develop based on existing models, thus saving time, computing power and financial investment. Opening model weights also allows companies to adapt to their own tasks through subsequent training, reducing dependence on a few closed-source model manufacturers.
In the past year, the average token price of the open source model has dropped by two-thirds, from US$0.45 to US$0.13 per million, but usage has increased 22.5 times at the same time, with monthly token calls reaching 2.9 trillion. The price drop coincides with the surge in usage, which also forms a seemingly contrasting data: open source models already account for 19% of the total platform tokens, but only account for 4% of enterprise model expenditures.
In other words, the open source model is not only cheap, but is being used on a large scale.
Comments