Jul 17, 2026, 2:19 PM ETLecture

Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Economics of Generative AI

The nine-week course involves approximately three hours of weekly commitment, comprising one hour of live class and one to two hours of reading.
Grading is split 50/50 between class attendance and a final assignment released at the end of the term.
The session utilizes Chatham House rules for guest speakers, who include leaders from OpenAI, Anthropic, and infrastructure firms; recording these sessions is prohibited.
The instructor, Apoorv, leads Altimeter, an investment firm focused on a concentrated public and private AI strategy.
The course objective is to equip students with mental models to determine where to invest in Series A rounds, what businesses to join, and the physical/economic laws governing the AI sector.

The current AI ecosystem is characterized as a "triangle" rather than an inverted pyramid, contrasting sharply with the software-heavy, high-margin shapes of the internet, mobile, and cloud eras.
AI infrastructure costs follow a "five-layer cake" model defined by Jensen Huang: energy, chips, power, interconnect, and memory.
Unlike legacy software which achieved 80-90% gross margins due to near-zero marginal distribution costs, AI inference carries significant incremental costs (GPU burn) per user.
The semiconductors layer currently holds the highest profitability, estimated at ~75% gross margins (data center revenue), whereas application layer margins are estimated between 0% and 30%.
The instructor argues the triangle may remain stable for a decade or longer due to the difficulty of optimizing the underlying substrate, unlike the rapid inversion seen in cloud computing (AWS took 8 years to fully shift from CapEx to revenue dominance).

NVIDIA maintains a dominant stranglehold on compute, with approximately 40% of fleet utilization currently dedicated to inference and 60% to training; this ratio is expected to shift toward inference over time.
A single vertical integration model is not yet established; the instructor notes Google (Internet), Apple (Mobile), and Meta (Social) as historically dominant integrated players, whereas the cloud era produced a heterogeneous oligopoly (AWS, GCP, Azure).
Consumer AI usage is currently dominated by free tiers (95% of users for ChatGPT), with significant uncertainty regarding monetization pathways (subscriptions vs. high-intent advertising).
Current consumer AI apps (ChatGPT, Gemini) are positioned in the "niche" category (similar to Spotify or Twitter) rather than the "mandatory utility" category (like WhatsApp) or "social" category (like TikTok).
ChatGPT has reached a scale of ~1 billion users, monetized at ~$10/user/year, representing a significant gap compared to Meta ($70/user/year) or Alphabet ($100/user/year).

The market structure may fundamentally shift if hyperscalers (Google TPU, Meta MTIA, etc.) achieve breakout success with specialized ASICs, potentially causing a major repricing of the semiconductor layer.
A key indicator of structural change is hyperscalers reducing CapEx guidance in earnings calls, suggesting the current heavy-investment equilibrium is no longer sustainable or productive.
The transition from training-heavy to inference-heavy workloads is a critical economic unlock, as inference workloads are more bursty and potentially less profitable to serve unless volume scales significantly.
Revenue growth in the AI stack has been heavily skewed toward semiconductors; despite a 10x increase in application growth over the last two years, ~75% of the $350 billion in added revenue went to the semis layer.
The instructor posits that future economic value will likely require moving beyond "knowledge work" to become a daily utility, potentially unlocked by high-intent advertising models where AI providers can leverage deep user intent data for superior attribution.