In estimated US enterprise production API spending on open models, Chinese-developed models account for 89%, including Kimi’s 61%, according to Menlo Ventures’ November 3 report.
The big change
- From September 2025 to September 2026, open models rose from 11% to 19% of all model tokens.
- Their share of model spending fell from 6% to 4% over that period.
- Buyers must assess price, token mix and deployment costs, not just usage volume.
Why volume and spending diverge
A token is a unit of model input or output. More tokens do not necessarily produce a proportionate bill: prices and the mixture of requests differ. Menlo estimates average open-model token prices fell 3.3-fold; that does not guarantee savings.
Fireworks, a Menlo portfolio company, publishes separate serverless rates for fresh input, cached input and output, and offers reserved capacity. A buyer’s bill depends on the model, usage and arrangement.
Open weights can be run on a buyer’s infrastructure, subject to the license and deployment requirements. Moonshot publishes Kimi K2.5 weights under a modified MIT license. Running weights still requires computing and operational resources.
What the estimate can show
Nearly 400 US enterprise AI decision-makers overseeing over $4 billion were surveyed September 8–26. Menlo weighted reported production workload shares by spending and modeled token economics using public disclosures and third-party data. API spending excludes a comprehensive accounting of self-hosting costs.
Menlo invests in AI businesses including Anthropic, Fireworks and OpenRouter. Those ties are context, not proof of error.
