Stanford's School of Engineering launched a course this spring called Economics of the AI Super-Cycle (MS&E 435), taught by Apoorv Agrawal, a partner at Altimeter Capital — a major shareholder in NVIDIA and investor in OpenAI and Anthropic. Each week, a core industry player joins the class to dissect a different layer of the AI economy, from Databricks' CEO to Crusoe's CEO to Altimeter founder Brad Gerstner himself. Star watched the first three sessions and built an 8-layer panoramic map of the AI industry stack, which he and Ruby walk through together in this episode.
The conversation covers every layer of the AI value chain — energy, manufacturing, storage, chips, interconnects, compute services, data, model orchestration, applications, and physical execution — explaining why each layer exists, which companies represent it, and where the moats are. Along the way, Star and Ruby unpack the macro case for why AI is the largest infrastructure buildout in the world right now, with capex already exceeding $600 billion and AI revenue growing faster than spending — a key reason this cycle looks nothing like the dot-com bubble.
One of the most illuminating threads in the episode is a concrete follow-the-money exercise: where does your $20 monthly AI subscription actually go? From the model company down to NVIDIA, TSMC, ASML, and the memory giants, the profit distribution chain reveals why Jensen Huang is always smiling at his keynotes. Whether you're building a startup, planning a career move, or thinking about long-term investments, this episode is a strong foundation for understanding how money flows through the AI era.
In this episode:
- What Stanford's MS&E 435 course is, who teaches it, and why the guest lineup covers the entire AI stack
- Why AI capex has already surpassed $600 billion — second only to the U.S. defense budget — and why this cycle isn't a bubble
- Anthropic's 10x revenue growth three years running, and what it signals about the pace of AI adoption
- How value capture flipped: in the cloud era, applications took the lion's share; in the AI era, NVIDIA captures ~80% of the pie
- Star's self-built 8-layer AI industry map: energy → manufacturing → storage → chips → interconnects → compute services → data → models → applications
- The "1,000 researchers" analogy that explains GPU training, SRAM, HBM, NVLink, optical modules, and why every storage layer exists
- Where your $20 Claude subscription actually goes — a full profit-distribution teardown from model company to NVIDIA to TSMC to ASML to the memory trio
- Three dimensions for evaluating AI companies: total addressable opportunity × profit capture ability × indispensability
- The Crusoe CEO's framework for going long and short in AI: first-principles thinking and why hyped energy stocks deserve extra scrutiny