Menlo Ventures: Chinese models take 89% of corporate open-source AI spend

Menlo Ventures' 2026 enterprise AI report quantifies a shift many builders sensed: when companies pay for open-source models, they overwhelmingly pick Chinese ones. Chinese models capture 89% of corporate open-source AI spend. Moonshot's Kimi leads with 61%, DeepSeek follows at 15% and Zhipu's GLM at 11%. Meta's Llama, once the default open model, sits at 7%.

The figures measure spending, not downloads, so they reflect production deployments where companies pay for hosting, fine-tuning or support. Kimi's dominance suggests enterprises are optimizing for capability per dollar on agentic and coding work, where Chinese labs have released aggressive open weights.

The context makes the number sharper. Meta has pivoted to the closed Muse family, shipping Muse Wonder and Muse Ember this week, effectively ceding the open-weights field. Mistral released open weights for its 1T-parameter Large 4, the main Western counterweight. Alibaba confirmed Qwen 4 is in development, and DeepSeek's Harness platform reports 1.54 million daily users.

Caveats: the report's methodology and sample size are not detailed in the source, and spend share can be skewed by a few large contracts. Policy is a wildcard, since governments are scrutinizing Chinese AI in sensitive sectors. Open-source builders noted Llama's collapse as the end of an era; watch whether Mistral Large 4 dents Kimi's lead.