#22 Dylan Patel: China’s Robotics Dominance; AI Infrastructure Breakdown

In this talk

1. 5:04

   How the AI infrastructure stack fits together

   Patel describes the stack from AI compute needs down to accelerators, data centers, power, cooling, networking, chips, and manufacturing. His team tracks components such as transformers, optical fiber, packaging, and fab equipment, using satellite images, regulatory filings, and industry expertise.

2. 10:38

   Why progress feels faster

   Patel says progress is accelerating across models and infrastructure, though physical goods move more slowly than software. He points to math performance rising from under 4% to 70% on FrontierMath in six months, alongside rapid gains on software engineering benchmarks.

3. 13:56

   Connecting research across the stack

   SemiAnalysis assigns researchers to specific layers, including models, power delivery, accelerators, packaging, and fabs. Patel says connecting those layers helps explain problems such as GPU reliability, which can involve cloud software and manufacturing processes. Consulting projects also deepen the research and data products.

4. 29:48

   China’s lead in robot manufacturing

   Patel says China deployed more factory robots in 2023 than the US, Japan, South Korea, and Germany combined, which together reached about one-third of China’s installations. He says Chinese robots made up half of China’s market in 2023, with the share expected to reach 60% to 70% in 2025. He argues that Chinese production of components such as actuators and power supplies leaves Western companies dependent on supply chains they do not control.

5. 36:25

   The case for industrial investment

   Patel says building domestic capacity for materials such as neodymium magnets could take a few billion dollars, but refining is dirty and may require sustained support. He contrasts Western companies’ tendency to retain cost savings in margins with Chinese firms’ cost-plus approach, which passes lower costs to customers. On robotics training, he says researchers and companies are making progress on limited real-world data.