Agrawal — Stanford MS&E435 Spring 2026 Session 1: Economics of Generative AI (Stanford Online)
AI is the biggest technology supercycle since the PC, Internet, and Mobile. MS&E435 is a seminar that unpacks the economics at each layer of the AI stack. Each week, we learn from a practitioner with a unique vantage point across the stack.
Instructor: Apoorv Agrawal is a Stanford Adjunct Lecturer in Management Science and Engineering and a partner at Altimeter Capital.
— channel description, Stanford Online
TL;DR
A ~34:13 Stanford Online MS&E435 Session 1 lecture (published 2026-05-20; auto-generated English captions, 322 segments). Instructor: Apoorv Agrawal — Stanford Adjunct Lecturer in Management Science and Engineering and partner at Altimeter Capital (the investment firm). Prior career: 13–14 years ago started at Palantir (the wiki’s first Palantir anchor) writing Spark in government buildings; came back to Stanford for grad school.
The lecture is the opening session of a 9-week seminar on the economics of the AI stack — the seminar’s recurring guests will be practitioners with vantage points at each layer (semis to infrastructure to foundation models to applications and agents). The course is conducted under Chatham House Rules; the wiki ingest is limited to the publicly-released Session 1 instructor content.
The substantive contributions are three.
1. The triangle / inverted-triangle industry-structure framing (~7:00–11:00). Agrawal opens with the recurring class question — where’s the money? The current snapshot:
- Top of the value-capture pyramid (small triangle today): applications and agents layer
- Middle (medium triangle today): infrastructure / cloud / inference layer — “the most competitive part of the whole ecosystem… a lot of startups doing really well, but you’ve also got the hyperscalers who want a dominant say in that layer.”
- Bottom (largest triangle today): semiconductors layer (NVIDIA dominance)
“If you have a set of users using cursor or using AI applications, the incremental user of an AI application is not free. It’s not marginally free. It’s actually quite a bit more expensive to have AI users because turns out you’ve got to burn those GPUs.” This is structurally different from the prior software-eats-the-world supercycle, where marginal cost of running software was near zero and SaaS businesses ran at 80–90% gross margins. Agrawal flags this as a central puzzle the seminar will explore: when does the pyramid flip (i.e., when does the value-capture mass shift to the application layer)?
2. The AWS-historical-analog timing anchor (~9:38–10:30):
“AWS started in the year 2004. AWS had its first customer in Netflix in 2010 again, and ultimately Amazon shifted fully to AWS in 2012. 8 years from breaking ground, 8 years from the first capex investment cycle. I don’t know if any of you were around reading earnings reports 20 years ago, but the big debate was, hey, is Amazon going to go bankrupt? And that was the biggest question everybody had about the buildout of AWS.”
The 8-year-capex-cycle historical anchor is the central temporal frame Agrawal hands the seminar: the AI buildout we’re watching today is structurally analogous to AWS 2004–2012; the value-capture pattern won’t fully resolve for years; the right way to track it is capex cycle timing + which-layer-wins resolution. “Thankfully nobody at least yet is on the verge of bankruptcy, but these are large numbers.”
3. The are-you-a-feature-or-a-platform diagnostic (~12:01–13:00):
“For the speakers, we’re going to talk a lot about [profitability]. For the folks in the middle — the inference layer — this is the most competitive part of the whole ecosystem. There’s a lot of startups doing really well, they’re winning so far. But you’ve also got the hyperscalers who want a dominant say in that layer. So honestly the jury is still out, and the biggest question there is: are you a feature or a platform? A lot of new businesses on the infrastructure side feel like very good ideas — but if you ask yourself the question ‘hey why is this not a part of AWS,’ you are thinking about maybe it should be a part of AWS.”
This is Agrawal’s investor-altitude diagnostic for the AI-startup landscape: any infrastructure-layer startup must answer the am-I-a-feature-of-a-hyperscaler-or-an-independent-platform question. The diagnostic is reusable downstream as a 2026 which-startups-will-survive-the-hyperscaler-consolidation lens.
Salesforce / Palantir as old-economy worked examples (~12:34–13:18): Agrawal addresses the question of how to classify reinventing-themselves businesses with new AI SKUs (Salesforce Einstein, Palantir AIP): “the answer is yes they should [be at the apps layer]. The way I solve for that in this calculation is I get the model revenue. So if you were running Salesforce, you’re probably running either one of the big models or running inference. So their spend is captured in the app layer by way of the substrate.”
Google as a conglomerate worked example (~14:43–15:44): Agrawal explicitly disaggregates Google: “Any large conglomerate like Google deserves to be broken into business units. I would put the TPU business unit in semis. Their GCP unit is in the infrastructure layer. And the Gemini unit is at the apps layer.” This is the wiki’s first articulation of business-unit-level disaggregation as the right way to slot conglomerate companies into the AI stack pyramid.
Caveats. Stanford-conducted seminar; Agrawal flags that subsequent sessions (with practitioners across the stack) are under Chatham House Rules — the publicly-released Session 1 is the only ingestible material from the seminar. Session 1 is structured as a course introduction (logistics + framing + opening conceptual lenses); the substantive sessions with practitioners are not in this video. Agrawal is partner at Altimeter Capital — treat his framing as motivated by Altimeter’s investment thesis (sustained capex cycle, value will eventually flow to apps layer if you bet right).
Why this matters in the corpus
This source is the wiki’s first explicit AI stack value-capture framing at the venture-capital-altitude. The wiki holds related framings at adjacent altitudes:
- AI Index 2026 — supply-side measurements (compute, investment, capability).
- Anthropic Economic Index — demand-side deployment measurements.
- Stanford AI Club — VC-altitude how to win framing.
- Stanford GSB — macroeconomic-growth weak-links model.
- Agrawal (this source) — industry-structure / value-capture pyramid at the investor-altitude.
The AWS-historical-analog (8-year capex cycle from 2004 to 2012) is a reusable temporal anchor for downstream synthesis on when does the AI buildout produce sustainable profitability at each layer. The are-you-a-feature-or-a-platform diagnostic is a reusable filter for evaluating any AI-infrastructure startup against hyperscaler consolidation risk.
What was actually ingested
The full ~34:13 transcript was read end-to-end. Session 1 = course introduction (logistics + framing + opening conceptual lenses + Q&A); subsequent sessions are Chatham House Rules and not ingestible. The Q&A section adds substantive material on (i) timing-mismatch between semis capex and apps revenue (analogous to mobile super-cycle); (ii) Google as a conglomerate to be disaggregated by business unit; (iii) the stable-equilibrium / inverted-pyramid prediction.
Linked entities and concepts
Entities promoted by this source:
- Stanford Online — channel; already entity. Bumps source-count.
Dangling — single-source mention, deferred:
- Apoorv Agrawal — Stanford Adjunct Lecturer + Altimeter Capital partner. First wiki mention.
- Altimeter Capital — investment firm (public and private businesses). First wiki mention.
- Palantir — already a wiki reference (named throughout other sources); Agrawal’s career-origin venue. Promote on the next substantive Palantir-focused ingest.
- Salesforce — referenced via Einstein AI SKU; promote on substantive Salesforce-focused ingest.
Concept pages touched:
- foundation-models — adds the foundation-models-as-the-middle-layer-of-the-AI-stack industry-structure framing; the are-you-a-feature-or-a-platform hyperscaler-consolidation diagnostic.
- enterprise-ai-adoption — adds the incremental-user-of-an-AI-application-is-not-free gross-margin observation that distinguishes the current AI supercycle from the prior software-eats-the-world supercycle.
- strategic-foresight — adds the AWS-8-year-capex-cycle-historical-analog temporal anchor; the AI is the biggest technology supercycle since PC, Internet, Mobile framing.
Source quality
- Channel: Stanford Online — Stanford’s official online education channel.
- Format: ~34-minute seminar Session 1 = course intro + Q&A; ASR-cleaned auto-generated English captions.
- Empirical anchors: AWS historical timeline (verifiable from public Amazon disclosures); Google business-unit disaggregation framing.
- Bias / motive: Investor-vantage from a partner at an active AI-investment firm; treat as motivated by Altimeter’s investment thesis.
- Transcript provenance: youtube-transcript-skill (Playwright path); ASR-only auto-captions.