This episode delves into the critical bottlenecks hindering the scaling of AI compute, primarily focusing on the semiconductor supply chain. Dylan Patel explains that while power and data center construction are being rapidly addressed, the long-term constraints lie in logic (TSMC's advanced nodes), memory (HBM scarcity), and ultimately, ASML's limited production of EUV lithography tools, which caps global AI chip output to around 200 gigawatts annually by 2030.
Summarized by Podsumo
Contrary to traditional depreciation, the *value of H100 GPUs is increasing* due to rapid AI model improvements (e.g., GPT-5.4 being more efficient and higher quality than GPT-4), extending their effective lifespan and driving higher rental prices.
A severe *HBM memory crunch* is forcing prices to quadruple, leading to a significant reduction in *smartphone and PC volumes* (projected to drop to 500-600 million units/year by 2025-2026), as memory is diverted to more profitable AI chips.
By 2028-2030, the *limited production of ASML's EUV machines* (max ~100 tools/year by 2030) will be the primary constraint on global AI compute, capping total annual chip production at roughly *200 gigawatts*.
Despite initial concerns, *power generation is not the ultimate bottleneck* for AI scaling. Diverse energy sources (gas turbines, fuel cells, solar+battery) and grid optimization can provide hundreds of gigawatts, albeit with higher costs and labor challenges.
If AGI timelines extend to *2035*, China's aggressive, vertically integrated approach to semiconductor manufacturing could allow it to *surpass the West* in mass flop production, especially if Western supply chains remain "not AI-pilled" enough to scale aggressively.
"An H100 is worth more today than it was three years ago."
— Dylan Patel
"The biggest bottleneck is compute and for that the longest lead time supply chains are not power or data centers. They're actually the semiconductor supply chain themselves."
— Dylan Patel
"ASML is, you know, maybe one of the most generous companies in the world, right? ... And yet they haven't taken price and margins up like crazy."
— Dylan Patel