The Economics of the
AI Supercycle
Which Layer Does the Money Ultimately Accrue To

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

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

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

The course has no assigned textbook, but it has a clear pre-read system—two versions of Apoorv's own "The Economics of Generative AI" (first published April 2024, and the April 2026 "two-year lookback") form the entire data baseline. The core argument: the current AI value chain is an "inverted pyramid"—the closer a layer is to the chip, the higher its gross margin.

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

Mobile Internet Analogy

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

Four Improvement Levers (How Application Layer Margins Go from 0–50% to the SaaS Era's 75–80%)

• ▸Better pricing / value alignment—usage-based pricing, redefining profit models
• ▸Custom chips lowering 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 margins. This translates "where value accrues" into a calculable "where the bottleneck lies."

Effective supply growth rate = min(chips, memory, energy, network)
Gap = Effective supply growth rate − Demand growth rate

Gap < 0 → Undersupply → Prices rise → Profits concentrate
Gap > 0 → Oversupply → Commoditization → Profits disperse

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

The course has no academic paper textbooks, but it has a very clear required reading system—Apoorv's own Tailwinds articles. In addition to the two already expanded upon in §01 and §02, there are three more corresponding to guest speakers in subsequent classes.

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

Core argument: inference is compounding growth, not linear growth. Jensen Huang said on the BG2 podcast that "inference is about to grow a billion-fold"—the next generation of AI products shifts 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, surging to 480 trillion in April 2025 (50x in 12 months), and then exceeding 1.3 quadrillion in October 2025 (another 3x in 3 months). Apoorv likens Baseten to "the Stripe of inference"—abstracting model deployment, autoscaling, routing, and performance tuning into a product; its 2025 revenue grew 10x, and its customer list is itself a roster of "winners" in the AI application layer (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, slide decks, charts; knowledge workers spend more and more time "finding answers." Apoorv's judgment is that Glean is not an "enterprise search tool," but rather an enterprise AI platform—this aligns with the problem domain of the Class #6 guest Applied Compute: how to unlock internal enterprise knowledge.

The State of Consumer AI Part 2 (March 2026) — Recommended Reading

Core takeaway in one sentence: consumer markets follow a power law distribution, and AI is no exception. Data shows ChatGPT holds about 70% market share, Gemini about 20%, with all others fighting for the remainder—the same historical pattern as Google taking 90%+ of search, Facebook taking social, and Apple taking mobile profits. This is two sides of the same coin as the application layer's "two-player game" (OpenAI + Anthropic capturing 75% of revenue).

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

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

The most significant story of the course: 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's IP, and acqui-hired founder Jonathan Ross, president Sunny Madra, and core engineers (the Groq company itself retains independent operations). Jensen Huang had publicly mocked ASICs, saying "even if they were free, they couldn't beat NVIDIA," only to spend $20 billion to acquire an ASIC company himself—one of the most dramatic tech acquisitions of 2025. Brad Gerstner's key quote in class: "Inference costs dropped 99% in two and a half years," which he defined as "a paradigm shift from pre-training to inference-time reasoning."

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, bringing modular data centers to well sites for Bitcoin mining; after 2022, it pivoted to AI compute, and starting in 2025, it took on institutional-grade AI data center projects. The core argument directly couples with the §02 bottleneck model: energy is the next true bottleneck of the AI supercycle—even if chip supply is completely unlocked, the grid can't deliver the power; "stranded energy" (oil field tail gas, geothermal, excess hydropower) represents a massive arbitrage opportunity.

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

At the time this guide was compiled, the video for this class had not yet been uploaded; based on Ali's recent public statements, the themes can be anticipated: "80% of databases on the Databricks platform are now built by AI agents, not humans"; and the "Service as Software" concept—traditional SaaS is "software as a service," while 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% gross margin ceiling could be broken, because the comparable benchmark becomes consulting fees and legal fees, rather than SaaS subscriptions.

Each week, a guest enters from a specific layer of the AI stack to discuss the real business they are building—this list is itself a roster of "who has gained a foothold in the AI supply chain."

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

Mainstream "AI economics" research has several distinct threads: Daron Acemoglu / Pascual Restrepo work on task automation and income distribution (labor economics); Erik Brynjolfsson / Tom Mitchell work on general-purpose technologies and the productivity J-curve (growth economics); Anton Korinek works on AGI transition paths (macroeconomics); Hal Varian / Avi Goldfarb work 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 are backed by Epoch AI, SemiAnalysis, and various earnings reports
• ▸Top-tier guests—one founder/CEO per week, not professors recounting second-hand
• ▸Clear reference system—consistently uses cloud stack history as a mirror for the current AI stack
• ▸Falsifiable—all predictions have timelines
• ▸Fully public—complete YouTube availability, essentially 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, raising "self-promotion" suspicions
• ▸Lacks labor/social dimensions—no discussion of unemployment, income distribution, or AI governance
• ▸Lacks counter-examples from open-source/China/Europe—defaults to all competition occurring among top US companies

The value of this course for AI industry observers lies not in teaching "what AI is," but in providing a reusable set of analytical tools—the three-layer stack, margin distribution, and bottleneck model—to ask the right questions: rather than vaguely asking "how big is the AI bubble," ask "which layer, how overpriced, and what's supporting it."

For readers focused on the Chinese AI ecosystem, the most valuable application 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 figures. China's bottleneck is in chips, not energy, and SaaS penetration was already low—concepts like "Service as Software" may follow entirely 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.