The report's figures
Chinese artificial intelligence models absorb 89% of global enterprise spending on open-source models. The estimate comes from a report by U.S. venture capital firm Menlo Ventures, picked up by the South China Morning Post and republished by IT之家. It is a share close to the entire segment, leaving Western models only marginal space.
The internal ranking is led by Kimi, the model family developed by Moonshot AI (月之暗面), with 61% of spending. It is followed by DeepSeek with 15% and Zhipu's GLM (智谱) with 11%. The top three positions are therefore all occupied by Chinese projects, with a clear gap between the leader and the pursuers.
The gap with the West
The Western model with the highest share is Meta Platforms' Llama family, which according to the report accounts for only 7% of enterprise spending on open source. It is the same figure that, in the order presented by the source, is set against Kimi's 61%: the leading Chinese model alone is worth almost nine times the leading Western model.
The figure should be read with caution because the available material does not clarify some methodological aspects. It does not specify, for example, how Menlo Ventures built the sample of companies, what period the survey covers, or how spending is measured. Without these elements, it is not possible to establish whether the 89% reflects the entire market or a specific segment of users.
What remains to be verified
The report is cited via two journalistic intermediaries: the South China Morning Post summary and the IT之家 summary. The original Menlo Ventures text is not available in the material provided, so the percentages should be attributed to the venture capital firm as reported by the two outlets, not independently verified.
A second element to clarify concerns the definition of “open-source model” adopted in the report: the material does not specify whether it covers open weights, permissive licenses, or both categories. This distinction affects the comparability of the data, especially for Llama, which Meta distributes under its own licenses rather than a canonical open-source license.
Finally, the source does not indicate whether spending is expressed in absolute value or as a share of a basket, nor whether it includes inference costs, licenses, or related services. In the absence of these details, the comparison between 61%, 15%, 11%, and 7% describes a hierarchy internal to the sample analyzed, but does not allow estimating the overall size of the market.
