Sequoia Warning: US Won the Closed-Source Model, Could Lose the Open Foundation

According to Dynamics Beating monitoring, Sequoia Capital partner Dean Meyer and Konstantine Buhler stated that the United States leads in proprietary models, while Western companies are increasingly relying on open-weight models from China. Qwen, Kimi, GLM, and DeepSeek are being used as product bases, training teachers, and synthetic data sources.

The ATOM report shows that Qwen's share in monthly new fine-tuning and model adaptation has increased from 1% in January 2024 to 69% in February 2026. Thinking Machines' Inkling also used open-weight models like Kimi K2.5 to generate synthetic data in the early stages of post-training. However, this accounts for only a small portion of the training power.

The article argues that the issue lies in the "distillation" rule. Distillation is using the output of a strong model to train another model. OpenAI and Anthropic restrict clients from using model outputs to train competing products, while U.S. companies can legally learn from Chinese open models.

The two suggest that U.S. forward labs sell controlled, delayed, and auditable training rights to qualified companies. Otherwise, the U.S. may continue to lead in proprietary models but hand over the open-model base to China.

China has recently discussed restricting overseas access to some advanced models, but a final policy has not yet been established. The article warns that even though existing models can still be downloaded and used, Western companies may gradually fall behind as they cannot access subsequent versions.