Behind China’s AI gift, a constraint, Pg8

Key Highlights:

• China is promoting its Artificial Intelligence (AI) models as a global public good, offering open-weight models unlike the proprietary closed-source models from the U.S.
• This strategy is largely driven by China's significant compute constraints, particularly for AI inference, due to U.S. export restrictions on high-end chips.
• Since 2022, the U.S. administration has imposed export restrictions on China, limiting its access to advanced semiconductor chips.
• Chinese President Xi Jinping framed AI as "humanity’s collective wisdom" and a "historic opportunity" for international cooperation.
• Despite chip limitations, Chinese models like those from DeepSeek and Qwen are gaining popularity globally, even among U.S. firms, due to their cost-effectiveness.

China AI Strategy.jpg

Detailed Insights:

• Open-weight models allow users to download, customize, modify, and further train the AI models without subscriptions or terms of service.
• This approach ensures data remains within national borders, addressing data sovereignty concerns for countries like Papua New Guinea or Senegal.
• The U.S. export controls have made it difficult for China to acquire the vast number of chips needed for AI inference, which scales with user base.
• While China can accumulate enough chips for AI training through various means, serving models to a mass consumer base requires significantly more chips.
• Domestic chip production in China faces hurdles, such as low yields for 5nm chips without Extreme Ultraviolet (EUV) lithography tools, which are also under export restrictions.
• The inability to serve models at scale has led to issues like DeepSeek freezing new registrations due to traffic surges.
• American firms are increasingly using Chinese open-weight models, with their share reaching 60% on platforms like OpenRouter, due to comparable performance at lower costs.
• China's "openness" is thus a strategic necessity, allowing it to outsource the compute burden to users while still advancing its AI capabilities.

Key Concepts Involved:

• Open-weight models: AI models where the underlying parameters (weights) are publicly available for download and modification.
• Closed-source models: AI models where the underlying parameters are proprietary and not publicly shared, typically accessed via Application Programming Interfaces (APIs).
• AI Inference: The process of using a trained AI model to make predictions or generate outputs based on new data.
• AI Training: The process of teaching an AI model using large datasets to learn patterns and make decisions.
• Export Controls: Government regulations restricting the trade of certain goods or technologies to specific countries, often for national security reasons.