A Menlo Ventures study, as reported by the South China Morning Post on November 6, estimated that Chinese AI systems accounted for 89% of global enterprise spending on open-source models. That figure describes spending within this model category, not the entire enterprise AI market. Separate research on OpenRouter shows why spending associated with a model does not necessarily become revenue for its developer.
A model’s share of spending and its developer’s share of that spending are different questions. Keep the denominator in view.
What the reported 89% measures
The estimate concerns global enterprise spending on open-source models. The same South China Morning Post account said proprietary U.S. models still command most overall AI revenue, a broader market measure than the 89% figure.
The report attributed these shares to Menlo Ventures:
| Model family | Share of spending in the reported category |
| Moonshot AI’s Kimi | 61% |
| DeepSeek | 15% |
| Z.ai’s GLM series | 11% |
| Meta Platforms’ Llama series | 7% |
The three listed Chinese model families add up to 87%, while the reported total for Chinese systems is 89%. The South China Morning Post account does not explain the two-percentage-point difference.
Why model spending may not reach its developer
A separate Epoch AI analysis of AI company revenue and OpenRouter spending, published October 6, estimated that DeepSeek received about 8% of customer spending on its models accessed through OpenRouter. That estimate applies to spending through this platform, not the global enterprise category measured in the Menlo Ventures report.
Epoch AI’s annualized OpenRouter estimates used eight sampled days between August 27 and September 19, 2026. The analysis estimated daily spending by multiplying each host’s input and output token volumes by their respective prices, then annualized the results; its estimates can vary with the sampled days.
GLM-5.3 Flash drew more third-party hosts
Epoch AI reported that Z.ai released the weights for GLM-5.3 Flash and listed the model on OpenRouter on August 26, 2026. Within a day, 12 third-party providers offered it there. After 20 days, the number had risen to 26, and eight charged less than Z.ai.
Over that 20-day period, Z.ai’s share of GLM-5.3 Flash tokens served through OpenRouter fell from 88% to 22%. Its daily token volume fell 70% while overall use of the model rose 17. Token share, total model use and customer spending measure different things; together, the figures illustrate how access through multiple hosts can separate a model’s use from the developer’s share of activity on one platform.
AI companies have several revenue paths
Epoch AI identifies consumer apps, access to models through APIs or subscriptions, enterprise and government solutions, licensing, and indirect income from complementary businesses such as cloud services, computing and advertising. For model-focused companies, API access is a main revenue channel. Large groups such as Alibaba and ByteDance can also benefit from complementary businesses when their AI models attract users.
