@Asianometry & Dylan Patel — How the semiconductor industry actually works

Dylan Patel and Jon Y break down China’s semiconductor catch-up, export control failures, and why OpenAI needs $50–100B for its next cluster.

• SMIC’s Shanghai fab has ~25–35k 7nm wafers/month capacity; at 50–80 chips/wafer, that’s millions of potential GPUs if China goes scale-pilled.
• Liang Mong Song left TSMC for Samsung, poached a wave of Taiwanese engineers, won Apple’s business, then moved to SMIC and rapidly closed the process gap.
• China received over 1 million H20 and Hopper GPUs in the year after October 2023 sanctions; leading labs train on 100k-GPU clusters, China has none that large yet.
• Export controls are structurally broken: China was 48% of ASML revenue in recent quarters, and domestic SMIC capacity can now build chips better than what the West is legally allowed to ship China.
• If Xi centralizes compute, a 10GW data center near the Three Gorges Dam is feasible in months — China adds as much power as half of Europe per year and could hide the load behind aluminum smelters.
• OpenAI cannot fund planned 2025 cluster builds without raising $50–100B; Microsoft is absorbing the credit risk because Nadella treats underinvesting as the worse Pascal’s Wager outcome.
• Private AI investment is ~$55–60B/year; the dot-com bubble peaked at $150B/year in private capital, and prior bubbles were each smaller than the next.
• DRAM density improvement stalled around 2012 while logic continued scaling; Dylan flags memory architecture as the most consequential underinvested layer in the stack.

2024-10-02 · Watch on YouTube