Lecture
Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Economics of Generative AI
- The course is scheduled to run for approximately nine weeks, with weekly commitments estimated at no more than three hours (one hour of sessions plus one to two hours of readings and discussions), featuring guest speakers in every session starting next week and an optional post-class dinner.
- Approximately 95% of current AI consumer users are on free tiers, driving a focus on future monetization through subscriptions and advertising, with the latter expected to yield higher pricing and attribution due to intent understanding.
- The AI ecosystem currently exhibits a triangular shape with the semiconductor layer generating about $350 billion in added revenue over the last two years, despite application layer growth exceeding 10x, while gross margins for semiconductor data center revenues are estimated at 75% and application layer revenues between 0% and 30%.
- The current ecosystem structure is projected to persist for about a decade or longer, with a potential inversion to a cloud software-like shape possible in five, ten, or fifteen years, or never, depending on substrate development and hyperscaler CapEx guidance.
- One or two unlocks, such as breakout success in ASIC programs (e.g., Google TPU, Meta MTIA), or a shift in hyperscaler capital expenditure messaging, may eventually trigger an ecosystem flip where startup revenue becomes dominated by large hyperscaler orders comprising about 50% of the $300 billion market.
- Google is suspected to be fully vertically integrated on the right side of the AI ecosystem, while the timing of the semiconductor build-out (five to six years) may create a cyclical mismatch with immediate application revenue generation.
- Consumer AI usage is expected to scale toward the level of social applications like Facebook or TikTok rather than utility-based platforms like WhatsApp, with incumbent platforms such as Salesforce and Palantir expected to capture revenue within the application layer.
- DeepMind is not planning to utilize advertising or subscriptions as revenue models, whereas the broader market expects an ad model to be a significant economic unlock this year, potentially following the trajectory of mobile advertising.
- The proportion of GPU usage for inference is predicted to increase over time, though the specific timing remains uncertain, and the speaker views AI as unlikely to be a fad or unsuccessful endeavor.