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

Infrastructure captures disproportionate value

Unlike cloud, mobile, and internet ecosystems that formed pyramid-shaped value distributions with large application layers, generative AI currently exhibits an inverted triangle where semiconductors and data centers command the majority of revenue.

High marginal costs destroy software margins

Traditional software achieved 80-90% gross margins because marginal distribution costs approached zero, but AI applications face significant per-user GPU inference costs that prevent profitability even at billion-dollar revenue scales.

AWS endured eight years of investment before returns

Amazon Web Services required eight years from initial 2004 capex investment to full 2012 adoption, surviving bankruptcy speculation that mirrors current hyperscaler AI spending, suggesting this cycle requires similar patience.

Semiconductor timelines mismatch application revenue

Chip buildouts follow 5-6 year cycles while application revenue manifests immediately, creating temporary infrastructure valuation inflation similar to railroad laying phases before transport value accrues.

Custom silicon could break Nvidia's grip

Successful hyperscaler ASIC programs like Google's TPU or Meta's MTIA achieving breakout performance would trigger massive repricing of the semiconductor layer and shift economic power toward the application stack.

Training-to-inference ratios indicate maturity

Nvidia's current fleet utilization is approximately 60% training and 40% inference; a sustained shift toward majority inference workloads would signal maturation toward utility-like economics capable of flipping the revenue triangle.

Hyperscaler guidance signals equilibrium viability

Reductions in quarterly capex guidance from major cloud providers would indicate the current economic model is unsustainable, making earnings calls essential monitoring points for sector health.