Descriptions:
Dylan Patel, CEO of semiconductor research firm SemiAnalysis, joins Dwarkesh Patel to deliver one of the most data-dense analyses available of the AI compute supply chain and what it means for the competitive positions of Anthropic, OpenAI, and the hyperscalers. The conversation begins with the headline figure: the Big Four — Amazon, Meta, Google, and Microsoft — have combined forecasted CapEx of $600 billion, but Patel breaks down exactly how much of that is actually deploying as compute this year versus committed to turbine deposits, data center construction, and power purchasing agreements for 2027–2029.
Patel estimates Anthropic and OpenAI are each currently operating at 2–2.5 gigawatts of compute and argues Anthropic needs to exceed 5 gigawatts by year-end to sustain its revenue growth trajectory — a target he describes as possible but very difficult. He contends Anthropic has fallen behind OpenAI in securing capacity due to a more conservative approach to signing large compute contracts, and is now being pushed toward lower-quality infrastructure providers. The discussion also covers why the performance gap between Hopper and Blackwell GPUs is approximately 20x for inference workloads like DeepSeek and Kimi K2.5 — a figure that dwarfs the raw FLOPS difference — because of NVLink bandwidth, HBM memory architecture, and advanced CoWoS packaging.
The episode closes with a detailed examination of physical limits on chip interconnects: how on-chip, within-rack, and cross-rack communication speeds differ by orders of magnitude, how wafer-scale integration (as attempted by Tesla’s Dojo) pushes these limits, and why architectural improvements in Nvidia’s Rubin on 3nm cannot simply be backported to older nodes. Essential viewing for anyone tracking AI infrastructure investment and model deployment economics.
📺 Source: Dwarkesh Patel · Published March 13, 2026
🏷️ Format: Interview
