Economics of the AI Supercycle
Where Does the Money Actually Accrue

MS&E 435 "Economics of the AI Supercycle" belongs to Stanford's Management Science & Engineering department, a 1-unit lightweight seminar, held Spring 2026 every Thursday 4:30–5:20PM in Hewlett 201 (4/2–5/28, 9 weeks total), with all lectures publicly uploaded to YouTube channel @MSE435EconomicsofAI, official site mse435.stanford.edu.

The instructor Apoorv Agrawal is a partner at Altimeter Capital, an engineer by training (Palantir Forward Deployed Engineer), who led Altimeter's largest single investment in history—entering OpenAI at a $150 billion valuation (now valued at $500 billion), and co-led Glean's Series E, early-invested in Baseten. He has long written industry memos on Substack under "Tailwinds," and these articles form the entire course's pre-read system—no textbook, only first-hand observations written by a VC.

The entire course pursues one question: Where does value accrue in the AI supercycle? As of April 2026, Apoorv's answer is:

The course has no assigned textbook, but a clear pre-read system—Apoorv's own "The Economics of Generative AI" in two versions (first published April 2024, April 2026 "looking back after two years") constitutes the entire data baseline. Core argument: the current AI value chain is an "inverted pyramid"—the closer to the chip, the higher the margin.

In two years the overall ecosystem grew 5x (~$90B → ~$435B), with the application layer growing fastest (12x vs semiconductor 4x), but in absolute incremental terms, NVIDIA alone is still 3x larger than the entire application layer—the semiconductor layer's margin share only dropped from 87% to 79%, declining ~4 percentage points per year. At this rate, for the application layer to reach the margin share of a mature cloud stack (apps ~70% / semis ~6%), it would take "well over a decade."

Mobile Internet Analogy

Apoorv uses the decade-long value flow of the mobile internet as reference: First leg semiconductors (Qualcomm, ARM capturing margins) → Second leg infrastructure (iOS/Android/app stores) → Third leg applications (Uber, Instagram, TikTok pulling profits to the top layer). Cloud computing followed the same pattern: first build data centers (2004–2010 AWS), then the SaaS explosion. AI is now at the "end of the first leg"; Apoorv's 2024 prediction was "enter the third leg within a decade"—in 2026 he still believes the flip will happen, just acknowledges the timeline is longer than originally predicted.

Four Improvement Levers (how application layer margins go from 0–50% toward the SaaS era's 75–80%)

• ▸Better pricing / value alignment—usage-based pricing, re-architecting profit models
• ▸Self-designed chips to lower TCO—Google TPU, Meta MTIA, Microsoft Maia
• ▸Model architecture optimization—state-space models, JEPA, MoE, etc.
• ▸Model cost reduction—batching, distillation, quantization

The course homepage features an interactive tool that applies Liebig's law of the minimum to the compute stack: whichever layer is scarcest has pricing power, and that layer captures the margin. This transforms "where does value accrue" into the calculable "where is the bottleneck."

Effective supply growth = min(chips, memory, energy, network)
Gap = effective supply growth − demand growth

Gap < 0 → undersupply → prices rise → profits concentrate
Gap > 0 → oversupply → commoditization → profits disperse

Using default values, energy is the current bottleneck (15%/yr), compared to demand growth of 310%/yr, a gap of −295%/yr. The tool provides four projection scenarios: A NVIDIA solves chip supply (bottleneck shifts to energy, i.e., the actual 2025–2026 path); B AI Winter (demand pulled to 100%/yr, gap narrows); C Energy breakthrough (small nuclear/modular reactors pulled to 50%/yr, bottleneck shifts to memory bandwidth); D All balanced at ~60% (no single bottleneck, full-stack margin compression). Understanding this explains why NVIDIA can currently command 73% margins—it sits precisely at the scarcest position.

The course has no paper-style textbook, but a very clear required reading system—Apoorv's own Tailwinds articles. Beyond the two already expanded in §01 and §02, three more correspond to guest speakers in subsequent classes.

Why we invested in Baseten (January 2026) — corresponds to Class #7

Core argument: inference is compound growth, not linear growth. Jensen Huang said on the BG2 podcast that "inference is about to grow a billion times"—next-generation AI products shift from "single request-single response" to loops (retrieve→reason→tool call→rerank→synthesize), where each user interaction triggers multiple inference events. Google's monthly token volume corroborates this curve: 9.7 trillion tokens/month in April 2024, growing to 480 trillion in April 2025 (50x in 12 months), then exceeding 1.3 quadrillion in October 2025 (another 3x in 3 months). Apoorv analogizes Baseten to "Stripe for inference"—abstracting model deployment, autoscaling, routing, and performance optimization into a product; 2025 revenue grew 10x, and the customer list itself is a who's who of AI application layer "winners" (Cursor, Notion, Clay, Gamma, etc.).

Glean: Putting AI to work, at work (September 2024) — corresponds to Class #6

Core question: 90% of enterprise knowledge is scattered across unstructured data—documents, code, screenshots, emails, slides, charts; knowledge workers spend increasing time "finding answers." Apoorv's judgment is that Glean is not an "enterprise search tool," but an enterprise AI platform—this aligns with Class #6 guest Applied Compute's problem domain: how to unlock enterprise internal knowledge.

The State of Consumer AI Part 2 (March 2026) — recommended reading

One-sentence core: consumer markets follow power-law distributions, and AI is no exception. Data shows ChatGPT holds ~70% market share, Gemini ~20%, and all others combined fight for the remainder—the same historical pattern as Google taking 90%+ of search, Facebook taking social, Apple taking mobile profits. This is two sides of the same coin as the application layer "two-player game" (OpenAI + Anthropic capturing 75% of revenue).

As of compilation time (2026-04-26), Class #1–#3 videos are online, Class #4 has not yet been uploaded. Below are the core facts and takeaways from held sessions.

Class #2 (4/9) Silicon: The GPU Economy — Sunny Madra (Groq→NVIDIA VP) + Brad Gerstner (Altimeter CEO)

The course's most blockbuster story: in September 2025, Groq raised $750M at a $6.9B valuation; on Christmas Eve 2025, NVIDIA announced it would spend $20 billion to acquire a non-exclusive license to Groq IP, and acqui-hire founder Jonathan Ross, president Sunny Madra, and core engineers (Groq the company remains independently operated). Jensen Huang had publicly mocked ASICs saying "even if free, they can't beat NVIDIA," yet he spent $20B to acqui-hire an ASIC company—one of 2025's most dramatic tech acquisitions. Brad Gerstner's soundbite in class: "inference costs dropped 99% in two and a half years," which he defined as "the paradigm shift from pre-training to inference-time inference."

Class #3 (4/16) Energy & Data Centers — Chase Lochmiller (Crusoe founder/CEO)

Crusoe started in 2018 using oil field associated gas (natural gas that would otherwise be flared) to generate power, hauling modular data centers to well sites for Bitcoin mining; after 2022 pivoted to AI compute, and from 2025 began undertaking institutional-grade AI data center projects. Core argument directly couples with §02 bottleneck model: energy is the next real bottleneck of the AI supercycle—even if chip supply is fully unblocked, the grid can't deliver enough power; "stranded energy" (oil field tail gas, geothermal, surplus hydro) represents a massive arbitrage opportunity.

Class #4 (4/23, preview) Enterprise AI & Service as a Software — Ali Ghodsi (Databricks founder/CEO)

At compilation time this video had not been uploaded; based on Ali's recent public statements, themes can be previewed: on the Databricks platform "80% of databases are now built by AI agents, not humans"; and the "Service as Software" concept—traditional SaaS is "software as a service," the new paradigm is "service as software," directly delivering service outcomes and competing with human outsourcing. If this paradigm holds, the application layer's 33% margin ceiling could be broken, because the comparable becomes consulting fees and legal fees, not SaaS subscriptions.

Each week a guest enters from one layer of the AI stack, talking about the real business they're building—this list itself is a roster of "who has a foothold in the AI supply chain."

Class #5 (4/30) is most likely a frontier lab capstone case (OpenAI or Anthropic); Class #9 (5/28) is student group "long/short" investment thesis presentations—using the entire course's framework to force students to make capitalization judgments, rather than停留在概念层面.

Mainstream "AI economics" research has several distinct threads: Daron Acemoglu / Pascual Restrepo on task automation and income distribution (labor economics); Erik Brynjolfsson / Tom Mitchell on general-purpose technology and productivity J-curve (growth economics); Anton Korinek on AGI transition paths (macroeconomics); Hal Varian / Avi Goldfarb on information economics and platform theory (industrial organization). MS&E 435 takes a completely different route—VC-perspective stack-layer value capture analysis, closer to Michael Porter's value chain analysis + Bill Gurley / Benedict Evans-style industry teardowns.

Strengths

• ▸Solid data—all arguments backed by Epoch AI, SemiAnalysis, and company earnings reports
• ▸Top-tier guests—one founder/CEO per week, not professor retellings
• ▸Clear reference system—consistently uses cloud stack history as mirror for AI stack present
• ▸Falsifiable—all predictions have timelines
• ▸Fully public—YouTube completely open, equivalent to a free industry research report

Limitations

• ▸Entirely US-centric perspective—the Chinese AI ecosystem barely appears
• ▸Capital perspective carries inherent bias—Altimeter portfolio companies (OpenAI, Glean, Baseten) appear frequently, with "self-advertising"嫌疑
• ▸Missing labor/social dimensions—no discussion of unemployment, income distribution, AI governance
• ▸Missing open-source/China/Europe counterexamples—defaults to all competition occurring among US head companies

The value of this course for AI industry observers lies not in teaching "what AI is," but in providing a reusable set of judgment tools—three-layer stack, margin distribution, bottleneck model—to ask the right questions: not vaguely asking "is the AI bubble big," but asking "which layer, how expensive, sustained by what."

For readers focused on the Chinese AI ecosystem, the most valuable use is "transfer analysis": use the same framework to re-estimate the Chinese side's three-layer stack revenue, margins, and bottleneck distribution, rather than directly applying US numbers. China's bottleneck is chips not energy, and SaaS penetration was already low—concepts like "Service as Software" may follow completely different paths in China. These are questions every AI practitioner should write their own answers to, rather than taking ready-made conclusions from this guide.