a16z Podcast H20s to China + 15% with Chris Miller and Lennart Heim
218 snips
Aug 14, 2025 
This discussion features Chris Miller, an expert on semiconductor strategy and author of "Chip War," alongside Lennart Heim from RAND, a technology policy specialist. They dissect the recent U.S. ban on Nvidia's H20 AI chips to China and the surprising reversal that allows sales after implementing a 15% export fee. The conversation highlights the geopolitical ramifications for AI competition, the importance of high-bandwidth memory, and the complexities of semiconductor export regulations amidst escalating tensions between the U.S. and China.
AI Snips
Chapters
Books
Transcript
Episode notes
H20 Is A Memory-Optimized Export-Compliant GPU
- The H20 was designed to maximize allowed memory bandwidth while staying within export-control flops limits.
- It became a high-demand, export-control-compliant GPU that prompted fast policy action in April 2025.
Compute Leadership Shapes Geopolitical Influence
- Leading in AI depends on access to advanced compute produced by a small set of firms.
- Who controls that compute matters strategically for geopolitical influence and future AI leadership.
Flops Vs Memory Bandwidth: Different Risks
- GPUs have orthogonal specs: flops (training) versus memory bandwidth (deployment/inference).
- The H20 is weak on flops but strong on memory bandwidth, making it valuable for inference and deployment use cases.
