Stanford MS&E435: From Megawatts to Molecules — the Complete Course in Nine Reviews
That wraps up Stanford MS&E435, “Economics of the AI Supercycle,” on the channel: nine lectures and nine reviews, from GPU economics and data-center construction to coding agents and drug discovery. The course is compelling because it does not try to pick the “best model.” Instead, it treats AI as a large industrial and economic system. The same questions recur at every layer: Where is the bottleneck now? Who pays the capital costs? What becomes a commodity? And who retains pricing power, data, and a closed feedback loop?
Taken together, the lectures show the bottleneck constantly moving: from chips to electricity and energized data centers, then to organizational context, human attention, evals, distribution, and laboratory experiments. The boundary moves with it between what a company can buy as a service and what it eventually has to control itself.
All nine reviews, in order:
1️⃣ The economics of the AI supercycle — the opening lecture is valuable as a map of the entire chips → infrastructure → models → applications stack and an inquiry into why most of the economics still sits at the bottom.
2️⃣ Inference as a production system — Sunny Madra and Brad Gerstner show why prefill and decode may call for different hardware, and why the cost of a verified outcome is more useful than the number of tokens burned.
3️⃣ The scarce megawatt of an AI factory — Crusoe’s Chase Lochmiller brings “cloud AI” back to earth: substations, cooling, construction, and the shortage of sites where GPUs can actually be switched on.
4️⃣ Ali Ghodsi: AGI is already here, but the company is not — the lecture is interesting for a Databricks case in which throughput improved materially only after redesigning the process, not replacing the model.
5️⃣ Sachin Katti and the human as the AI system’s bottleneck — a frontier-lab operator connects dynamic agentic workloads and heterogeneous infrastructure with the cost of human attention, verification, and responsibility.
6️⃣ Yash Patil: internal company knowledge as a learning loop — organizational expertise stops being abstract “context” and becomes evals, reward, post-training, and a reproducible learning cycle.
7️⃣ Who chooses the technology stack—the developer or the coding agent? — Guillermo Rauch shows how the agent becomes a new gatekeeper, while open source, documentation, and agentic ergonomics become distribution channels.
8️⃣ Baseten and how unit economics changes strategy — the story offers a useful symmetry: as they scale, applications take greater control of models while infrastructure platforms take greater control of compute capacity.
9️⃣ Chai and Claude: who captures the value in AI biotech? — the final lecture applies the same economic framework to drug discovery, where value depends not on a beautiful molecule but on a closed hypothesis → design → experiment → data loop and ownership of the result.
This is not a neutral textbook: the guests include investors, founders, and executives who sell the very layers of the stack they discuss. But once the mechanism is separated from the vendor pitch, the course becomes a rare end-to-end map of AI—from electrons to molecules.
If you do not have time for the entire playlist, pick the layer you are responsible for. If you do, I would go in order: that is when the movement of scarcity and value across the stack becomes especially clear.
#AI #Engineering #Infrastructure #Product #Strategy #Economics