Apoorv Agrawal · 00:33 "Where is the money in AI?"
In past tech supercycles — the internet 25 years ago, mobile 20 years ago, the cloud 10 years ago — the side that spent the capex eventually recouped from users in an upright-triangle economic structure, with profit flowing upward. In AI, that triangle is inverted: the semiconductor layer (NVIDIA and others) runs gross margins around 75%, while the application layer is overcrowded at 0–30%. This is the theme running through the entire course.
The instructor is Apoorv Agrawal — a partner at Altimeter Capital and the investor who led the firm's large position in OpenAI. Stanford GSB graduate, former Palantir engineer. Across the nine-week seminar, he invites one practitioner guest from each layer — semiconductors, power, models, inference, and applications — and applies the same question to each: "Who is dominant in your layer, and for how long? What are the price-compression vectors?"
Marc Andreessen's "Software is eating the world" (2011) worked because software's marginal cost was near zero, which let gross margins sit at 80–90%. AI applications are different — every additional user consumes more GPU. It is now normal to see businesses with billions in revenue still running at a loss. The course starts from the recognition that the underlying physical laws are different.
Another key idea is the timing mismatch between capex and revenue. Semiconductor construction is a 5–6 year forward investment; application revenue is now. So the lower half of the triangle moves with the periodicity of "laying railroad track." In past mobile supercycles, capex-heavy companies saw their market caps balloon in the early phase and then settle to cruising speed. Apoorv's framing: "we are in that early phase right now."
Key Observations
"Five years from now, everyone is going to ask you — 'did you see this coming?'" (04:36)
Apoorv addresses the students directly on why they should take this course. "The moment ChatGPT came out, when the plates were forming and the clay was about to harden — you want to be able to say you were there." Over the next nine weeks, the leaders of the companies that show up as guests will become primary sources, in hindsight, for "what were we thinking at that moment." The investor's sense of time: he packs the entire raison d'être of the course into the first five minutes.
AWS — 8 years from groundbreaking to full migration (09:48)
AWS started in 2004, landed Netflix as its first customer in 2010, and Amazon itself didn't fully migrate to AWS until 2012. Eight years from groundbreaking. In earnings calls of that era, the biggest debate was "is Amazon going bankrupt?" During a stretch where capex is heavy and the payback is invisible, the market gets nervous. The same pattern may repeat against today's framing — five hyperscalers spending a combined $650 billion. A reference point on the time axis.
NVIDIA fleet — 40% inference, 60% training (18:38)
One of the most-watched numbers in NVIDIA's earnings reports. Inference workloads are harder to forecast than training: bursty, concentrated when humans are awake, with troughs around Christmas and Thanksgiving. The framing: "if agents end up consuming tokens 24/365, it becomes even harder to predict." Inference is the next battleground for expansion, and ASIC / custom-chip players going after NVIDIA become visible from the numbers.
Video Outline
- (00:00) Self-introduction and course logistics (grading, Chatham House Rule)
- (03:47) Why this course is necessary — the value of being there when you can later say "I saw it coming"
- (05:50) The biggest question in generative AI — are the models generating economic value?
- (06:50) Comparison with past tech revolutions (internet, mobile, cloud)
- (09:00) Why AI's economic model differs from the past — marginal cost lives in the GPU
- (09:48) The history of AWS — 8 years from groundbreaking, the era of "Amazon is going bankrupt"
- (10:35) The theme across the whole course — who is dominant in each layer, what are the price-compression vectors
- (13:32) Capex-revenue timing mismatch — 5 years for semis, now for apps
- (14:14) The railroad-laying metaphor — periodicity and phase recognition
- (18:38) NVIDIA fleet — 40% inference, 60% training
- (20:04) Conclusion slide — semi layer 75%, app layer 0–30%
- (22:00) Winners of past cycles — Google ($3T), Apple ($2.5T), Meta ($2T)
Sources
Class #1 | MS&E 435: Economics of the AI Supercycle, Stanford University Spring '26, Apoorv Agrawal