Fireside Chat, Interview
@Asianometry & Dylan Patel — How the semiconductor industry actually works
Dwarkesh PatelDylan Patel, Jon Y, Jane Street, Stripe, Xi, Liang Mong Song, China, Huawei, Taiwan, US, John Y
- OpenAI is projected to raise between $50 billion and $100 billion by the end of this year or early next year to fund infrastructure clusters, with expectations to deploy multiple 100,000-GPU clusters next year, including some reaching 300,000 to 500,000 GPU equivalents.
- XAI is anticipated to secure over $30 billion to support similar massive cluster ambitions, while the broader AI investment bubble is forecast to surpass the dot-com bubble, potentially reaching $150 billion annually in private capital with near-term estimates of $55 billion to $60 billion this year.
- Capital flow is driven by a "Pascal's Wager" mentality among CEOs, who prioritize avoiding being left behind over immediate ROI, though a failure of GPT-5 or subsequent models to demonstrate exceptional capabilities could render current multi-hundred billion dollar expenditures void.
- China is predicted to centralize its compute resources to aggregate 600,000 Ascend 910B chips, potentially enabling a 1e27 model by 2026 and a 1e30 model before the U.S. by 2028 or 2029, possibly within a single gigawatt data center.
- U.S. power infrastructure faces significant bottlenecks regarding transformers, substations, and generation in 2026 and 2027, necessitating a revitalization of the power industry with production capacity scaling on six-month to three-year timelines.
- By 2028, 60% to 80% of TSMC's 2-nanometer capacity is expected to be dedicated to AI applications, with total deliverable flops (including training and inference) reaching 1e30 by 2028 or 2029, representing 100,000 times more compute than GPT-4.
- The cost of intelligence is expected to decline significantly due to efficiency gains, AI-driven chip design optimizations capable of 100x gains, and Moore's Law equivalents halving transistor costs every two years, allowing for 10x year-over-year compute growth.
- China's AI models may architecturally diverge from U.S. counterparts by utilizing more video/image recognition and state-space models due to hardware constraints, while the U.S. is expected to maintain leadership in leading-edge nodes and advanced packaging capacity.
- Current centralized trends in the U.S. are expected to intensify as "decentralized" lab efforts shift toward larger clusters to achieve necessary scale, potentially mirroring China's consolidation strategy.
- The semiconductor industry may reach a point where the next process node is only economically viable if justified by AI demand, as H100 ownership costs are dominated by the GPU itself (75-80%), making power delivery capability more critical than power costs for data center decisions.