What China Does Not Want the U.S. to Know

A leaked call with DeepSeek’s CEO, Liang Wenfeng, reveals China’s perspective on the AI race with the U.S., identifying resources and compute power as the primary bottlenecks. Despite these limitations, Chinese companies are seen as having strengths in cost efficiency and product development, positioning them as potential major producers in the global AI landscape. The CEO also discussed the future of AI scaling, the role of embodied intelligence, and the strategic focus on general agents, particularly coding agents.

• DeepSeek CEO identifies resources, specifically compute power, as China’s main limitation in the AI race with the U.S.
• The talent gap in China is attributed to the compute gap, which restricts experimentation and development.
• Chinese AI models are currently an order of magnitude smaller in parameters compared to top U.S. models due to resource constraints.
• Huawei’s domestic chip production is limited, creating a bottleneck for training large-scale models.
• Chinese companies are expected to play a role as the largest producers in AI, offering services at lower costs.
• Domestic Chinese chips like Huawei’s 950 supernode are seen as potential replacements for NVIDIA’s offerings, though with a lag.
• DeepSeek prioritizes reasonable returns over maximizing profit and aims for cost-effective AI development.
• China’s AI gap with the U.S. is estimated at roughly two years, but with significantly less compute power.
• There is no perceived upper limit to language model scaling, and embodied intelligence is considered inevitable for future needs.
• High-cost data labeling, a strength of U.S. companies, is a challenge for China due to capital structure limitations.
• The strategic focus for Chinese companies appears to be on developing general AI agents, with Coding Agent as a top priority.
  https://bender.layer3.press/articles/6f07f797-258a-430e-a1ed-e4502fdb0003