# China's strategy in AI: Influence vs Money | Lex Fridman Podcast

Source: https://www.youtube.com/watch?v=8DGIAvlJr4A
Recap page: https://rapidrecap.app/video/8DGIAvlJr4A
Generated: 2026-02-02T01:33:02.314+00:00

---
## Quick Overview

Chinese AI companies are rapidly releasing competitive open-weight models with strong performance, evidenced by the expectation that their models will soon overtake US models in usage downloads on platforms like Hugging Face, driven by different business incentives focused on open-source deployment versus API subscriptions.

**Key Points:**
- Open models from China are projected to overtake US models in usage downloads on Hugging Face in the near future, indicating a significant shift in global momentum.
- The Chinese approach favors open-weight models (like DeepSeek's V3/R1 and Kimi K2) often built by hedge funds or companies with different incentive structures than US counterparts.
- The speaker notes that DeepSeek's V3.2 model performs comparably to GPT-4 and surpasses GPT-3.5, achieving gold medals in the 2025 IMO and IOI.
- US tech companies often rely on API subscription models, creating a disincentive to release similar powerful open-weight models, leading to stalled growth in the US open ecosystem.
- Chinese companies are motivated to build open models to gain influence, build a tech ecosystem, and potentially drive broader adoption, even if initial revenue is not the primary goal.
- The speaker suggests that by 2026, many of the most notable new open model builders will likely be based in China, contrasting with the US focus on proprietary API access.
- DeepSeek's architecture utilizes Sparse Attention (DSA) and Scalable Reinforcement Learning frameworks, contributing to its high performance and efficiency.

![Screenshot at 00:42: The guest presents a slide titled 'Global Open Model Momentum,' illustrating that open models from China are set to overtake the US in Hugging Face downloads, marking a critical 'flip' in the AI landscape.](https://ss.rapidrecap.app/screens/8DGIAvlJr4A/00-00-42.jpg)

**Context:** This segment of the Lex Fridman Podcast features a discussion about the rapidly evolving landscape of open-source Large Language Models (LLMs), focusing specifically on the rising momentum of Chinese AI developers compared to their US counterparts. The discussion centers on the differences in business models—Chinese companies favoring open-weight releases for ecosystem influence versus US companies prioritizing closed, API-based monetization—and how this divergence is affecting the global distribution of powerful AI technology.

## Detailed Analysis

The conversation highlights a significant shift in the open-source AI landscape, where Chinese companies are aggressively releasing high-performing open-weight models, challenging the dominance previously held by US entities. The speaker points to data suggesting Chinese models will soon surpass US models in cumulative downloads on Hugging Face, describing this transition as 'The Flip' happening right now (00:43). This is attributed to differing business incentives: Chinese firms, like those behind DeepSeek and Zhipu AI (with models like Kimi K2), are releasing open-weight models to build influence and an ecosystem, often backed by entities like hedge funds (00:44, 03:46). In contrast, many major US tech companies are constrained by business models centered on API subscriptions (01:55), making them hesitant to open-source models that compete with their core revenue streams. The DeepSeek V3.2 model is specifically mentioned for achieving performance comparable to GPT-4 and surpassing GPT-3.5, even winning gold medals in 2025 IMO and IOI (03:56). The speaker predicts that by 2026, many leading open model builders will be Chinese, suggesting a strategic advantage gained through open releases, whereas US efforts have stalled in the open ecosystem. The discussion also touches upon the architectural breakthroughs in DeepSeek, such as Sparse Attention (DSA) and Scalable Reinforcement Learning, which contribute to efficiency and performance (03:56).

### Global Open Model Momentum

- Open models from China are set to overtake the US in near future usage on Hugging Face
- The American ecosystem has stalled in growth, indicating 'The Flip' is happening now
- Chinese companies are releasing open-weight models for influence, not solely money.

### Chinese Model Performance & Strategy

- DeepSeek V3.2 performs comparably to GPT-4 and surpasses GPT-3.5, winning 2025 IMO/IOI gold medals
- Chinese companies like Zhipu AI and DeepSeek are backed by entities like hedge funds, incentivizing open releases
- Many notable open model builders in 2026 are expected to be Chinese.

### US Business Model Conflict

- US tech companies are constrained by API subscription models, making them reluctant to release fully open, competitive models
- Many US companies won't pay for API subscriptions to Chinese models due to security concerns (01:29).

### Architectural Details

- DeepSeek utilizes DeepSeek Sparse Attention (DSA) for efficiency and Scalable Reinforcement Learning Frameworks
- Architectural diagrams comparing DeepSeek V3/R1 (671B, more heads, fewer experts) and Kimi K2 (1 Trillion, fewer heads, more experts) are shown (03:02).

![Screenshot at 00:02: Opening shot displaying Lex Fridman and an image of the Earth from space, setting the context for a high-level discussion.](https://ss.rapidrecap.app/screens/8DGIAvlJr4A/00-00-02.jpg)
![Screenshot at 00:22: A slide appears listing 'Recent AI Models \(Released in 2025+\)' categorized by developer, showing the vast landscape of current and near-future models.](https://ss.rapidrecap.app/screens/8DGIAvlJr4A/00-00-22.jpg)
![Screenshot at 01:42: A chart titled 'Global Open Model Momentum' illustrating the divergence where China's model download growth \(red line\) is projected to overtake the USA's \(blue line\), marking 'The Flip'.](https://ss.rapidrecap.app/screens/8DGIAvlJr4A/00-01-42.jpg)
![Screenshot at 03:02: A detailed technical diagram comparing the architectures of DeepSeek V3/R1 and Kimi K2, highlighting differences in MoE layers and parameter usage.](https://ss.rapidrecap.app/screens/8DGIAvlJr4A/00-03-02.jpg)
![Screenshot at 03:54: An abstract slide detailing the technical breakthroughs of DeepSeek-V3.2, including Sparse Attention \(DSA\) and Scalable Reinforcement Learning.](https://ss.rapidrecap.app/screens/8DGIAvlJr4A/00-03-54.jpg)