Chinese open-source models account for 89% of the total expenditures incurred by open-source companies in this field, yet their revenue amounts to only 10% of what two American companies generate.

Viewed together, the two reports can illustrate the current situation of China's big models.

The first, from Menlo Ventures, surveyed nearly 400 U.S. corporate decision-makers in September. The conclusion is: Chinese models account for 89% of the money companies spend on open source models. Among them, Kimi’s family from Dark Side of the Moon accounts for 61%, DeepSeek 15%, GLM series 11%, and Llama from Meta only 7%.

But open source models alone account for only 4% of total enterprise model spending. Anthropic and OpenAI combined took 86% via API.

The second is from Epoch AI. According to its estimates, the combined annualized AI-related revenue of OpenAI and Anthropic is about 105 billion US dollars; the six Chinese companies of Byte, Alibaba Qwen, Zhipu, Dark Side of the Moon, DeepSeek, and MiniMax add up to about 11 billion, which is about 10.5% of the former. Among the six, the highest number of bytes is around 4 billion.

Why use more and earn less? Epoch gave a detail: On OpenRouter, 75% of GLM-5 calls come from third-party service providers, and 70% of DeepSeek V4 Pro is externally hosted. After the model was open sourced, most of the money was earned by the hosting and resale platform, and the model owner only got a small piece.

Other statistics show that China’s open source model’s share of global usage has increased from about 1.2% at the end of 2024 to nearly 30%. Influence is real, monetization is another matter.