Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Applications, Applied AI

Baseten enables companies to build profitable and defensible AI businesses by shifting their reliance from expensive, centralized frontier models toward customized, open-source models that the companies own themselves.

• Inference economics — Roughly 95% of AI spending currently goes to frontier labs, but moving to post-trained, open-source models helps companies improve profit margins and retain control over their proprietary user data [06:37].
• Market defensibility — Relying solely on external frontier models creates a dangerous dependency, effectively handing over trade secrets to entities that could eventually use that data to compete directly with their own customers [11:07].
• Compute scarcity — GPU access is an extreme seller's market where demand far outstrips supply, causing rental costs to double and lead times to extend over a year [33:55].
  • Rental prices are becoming untenable, pushing the company to transition from renting to owning its hardware infrastructure [34:41].
• Infrastructure strategy — Baseten acts as an abstraction layer that manages reliability, performance, and multi-cloud access, essentially making disparate GPUs fungible for customers who lack internal hardware teams [30:06].
• Future infrastructure — To maintain service levels amid projected growth, the company estimates it will require 150,000 advanced chip equivalents within the next two years to meet demand [35:49].
• Hardware dominance — Nvidia currently holds a massive advantage due to the deep integration of its CUDA software ecosystem, making it the primary choice for rapid development, though heterogeneous computing architectures are likely the long-term future [27:41].

Discussion Questions

• How does the technical separation of "prefill" and "decode" stages in model processing influence the move toward heterogeneous hardware architectures?
• What are the functional differences between modular data centers and traditional virtual machines regarding industrialization and hardware efficiency?