Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Economics of Generative AI

ELI5/TLDR

A Stanford lecture on where the money in AI actually lives. Right now the picture is upside-down compared to past tech cycles — chip makers like Nvidia are eating roughly three-quarters of the profit, while the apps people use barely make money. The instructor — who runs Altimeter, an investment firm — argues this triangle won’t flip for years, and that the next big unlock for app-layer revenue will be ads inside chatbots, not subscriptions.

The Full Story

The course is taught by Apoorv, a Stanford alum who runs Altimeter Capital, an investment firm with concentrated public and private positions in the AI stack. This is the first session of a nine-week MBA-flavored course built around guest speakers from across the supply chain — semis, infra, energy, models, applications. The structure is deliberately conversational. The grade is half attendance, half one final assignment. Most of the substance will come from outsiders; this opening hour is the instructor laying down the central question that will frame every conversation to follow.

The triangle that won’t flip

The lecture’s organizing image is a chart of the AI ecosystem drawn as an inverted pyramid — semis at the top (huge revenue, huge margin), infrastructure in the middle, and applications at the narrow bottom. The cloud ecosystem, by contrast, looks the opposite way: a thick app layer sitting on a thinner infrastructure base.

Why the difference? The room offered the obvious guesses — AI is early, Nvidia has a near-monopoly. The instructor accepted both, then added a third: the physics of inference. Software ate the world because the marginal cost of serving one more user was near zero, which is how SaaS companies ran 80-90% gross margins. AI breaks that pattern. Every extra user of Cursor or ChatGPT burns GPU time. That’s why businesses doing billions in revenue at “AI scale” still aren’t profitable.

“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.”

He then anchored the timing with a historical example. AWS broke ground in 2004, didn’t land Netflix as a customer until 2010, and Amazon itself only fully shifted onto AWS in 2012 — eight years of capex before the model paid off, during which the standing question on Wall Street was whether Amazon would go bankrupt. The lesson: triangle inversions take a long time. He thinks AI might stay this way for longer than the decade cloud took, because the substrate (chips, power, interconnects, memory — “a five-layer cake” as Jensen Huang calls it) is harder.

Where profitability actually concentrates

A student asked about profitability directly. The instructor pulled up the number: Nvidia’s data center business runs at roughly 75% gross margin. Application-layer businesses sit somewhere between 0% and 30%, depending on who you ask. When you redraw the triangle by profit instead of revenue, the concentration gets even more extreme. One company is running the tables on the layer that prints money.

The infrastructure middle is the most contested zone. It has the highest metabolic rate — most startups forming, most acquisitions, most deaths. The framing question he wants students to apply to any infra startup: is this a feature or a platform? If the answer to “why isn’t this just part of AWS?” is uncomfortable, the company is probably a feature.

What could flip the triangle

Two specific catalysts to watch:

A breakout ASIC program. Google’s TPU, Meta’s MTIA, Amazon’s Trainium, plus whatever OpenAI and Microsoft are quietly building. If any of these crosses some threshold of independence from Nvidia, the semis layer gets repriced.
Hyperscaler capex guidance. If the four big spenders stop guiding to ever-larger capex numbers on earnings calls, that itself signals the current equilibrium isn’t working. He repeatedly nudged the class to actually listen to those calls — four times a year, public CEOs telling you what’s worrying them, mostly for free.

On the training-versus-inference question: Nvidia recently disclosed that roughly 40% of its installed fleet is being used for inference and 60% for training. He expects inference share to rise — agents running 24/7 will push it harder than humans ever could — but couldn’t put a date on the crossover. The shape of the workloads is also different: training is predictable and high-utilization for short bursts; inference is bursty, follows the sun, and inexplicably dips around Thanksgiving and Christmas.

The $300 billion question for chip startups

A student asked the natural follow-up: if every hyperscaler is building its own silicon, who do the ASIC startups sell to? The answer was deflating in its arithmetic. There’s about $300 billion of data-center chip revenue to fight over. Half of it comes from a handful of hyperscalers. So the customer base for a new chip company is “a very small number of very large orders.” The first question any chip founder should answer is which of the five hyperscalers they’re going to sell to first. The long tail of enterprises mostly just buys from the clouds anyway.

Vertical integration as the lottery ticket

He walked through the winners of each prior super-cycle and how vertically integrated they were:

Internet — Google, ~$3 trillion, near-99% search share, integrated from servers up through ads.
Mobile — Apple, ~$2.5 trillion, integrated from silicon up through services.
Social — Meta, ~$2 trillion, less integrated. “Maybe they lost a trillion because they didn’t fully go down to the servers.”
Cloud — heterogeneous, no single winner, three oligopolies (AWS, GCP, Azure), none fully integrated.

The implication for AI: the most vertically integrated player has historically captured the most. Today, Google is the only AI player that owns silicon (TPUs), infra (GCP), models (Gemini), and distribution (Search, Android, Chrome). That stack might matter more than people currently price in.

Why apps stay small

The second big slide compared this picture to two years ago. The ecosystem has grown roughly 5x in total revenue — about $350 billion added — but the shape hasn’t moved. Around 75% of that new revenue went straight to semis. Apps have grown more than 10x in absolute terms and still barely register as a slice.

The most striking framing was on consumer AI. He plotted ChatGPT and Gemini against the three tiers of consumer franchises:

Mandatory utilities (3B+ users): WhatsApp, Chrome, YouTube
Social (1.5–2B users): Instagram, TikTok, Facebook
Niche (under 1B): Amazon, Spotify, Twitter

ChatGPT has just crossed into the niche tier. Gemini hasn’t. As a comparison: Alphabet has roughly 4 billion users monetized at about $100/user/year. Meta has 3.5 billion at about $70. ChatGPT has roughly 1 billion at about $10.

“How do we get the billion up to 4 billion? I’m not sure knowledge work is the answer. I think we’d have to go beyond knowledge work.”

Knowledge work, his argument went, is something only some people do. You actively ask the bot a question. You don’t passively scroll it. To become a mandatory utility, AI has to find a non-active mode — messaging, inbox, ambient assistance — not just Q&A.

The ads bet

The other multiplier is monetization. Going from $10 to $100 per user per year. He doesn’t think subscriptions get you there. He thinks ads will.

The Demis Hassabis announcement that Gemini won’t run ads got a polite nod and a “we’ll see.” His view: the ad format inside a chatbot will eventually be invented, the way ads on phones eventually got invented despite the 2012 Facebook short reports that said phones had no screen real estate for them. And the pricing will be better than anything that came before — you’re logged in, the AI knows your intent, attribution is clean, the trust signal is high.

“There’s a lot of alpha in understanding the ad model really well.”

He left the form factor open. He doesn’t know what an ad inside an intimate chatbot conversation looks like without breaking the thing. But he thinks it gets figured out, and that it’ll be the headline of the next year or two.

Key Takeaways

AI revenue split today: roughly 75% goes to semis, the rest split between infra and apps. Two years ago, same shape, despite 5x growth.
Nvidia data-center margin sits at ~75%. App-layer margins range 0-30%.
AWS took eight years from first capex to internal adoption. AI’s triangle inversion will probably take longer than a decade.
Nvidia’s installed fleet is roughly 40% inference, 60% training. Inference share will grow but the crossover date is unclear.
ASIC startups face a customer base of “a very small number of very large orders” — basically five hyperscalers.
Winners of prior super-cycles correlate with vertical integration depth. Google is the most vertically integrated AI player today.
ChatGPT: ~1 billion users, ~$10/user/year. Alphabet: ~4 billion users, ~$100/user/year. The gap is the bull case and the bear case at once.
The path to closing the gap likely runs through ads in chatbots, not subscriptions. Form factor undiscovered.
The two real signals to watch on whether the current equilibrium breaks: breakout ASIC success at a hyperscaler, or hyperscalers backing off capex guidance.
Recurring instruction to the class: listen to all four hyperscaler earnings calls per year. It’s free CEO disclosure.

Claude’s Take

This is an opening lecture, so it’s setup more than payoff — most of the meat will come from the guest speakers in later weeks. But the framing is unusually clean for a first session, and the instructor’s day job (running a concentrated AI-stack fund) shows in how specific the numbers are. The 75% Nvidia margin, the 40/60 inference-training split, the $10-vs-$100 ARPU comparison — these aren’t textbook figures, they’re the kind you only have at hand if you’re modeling them weekly.

The strongest claim is the historical analogy: triangle inversions take a decade or more, and AI’s underlying physics make it likely to take longer than cloud did. That feels right and is well-grounded. The weakest claim, or at least the most under-defended, is the ads-will-save-the-app-layer argument. He acknowledges he doesn’t know what chatbot ads look like, then asserts they’ll be huge because phone ads worked despite skeptics. That’s a vibes argument dressed as a forecast. Plausible, but he didn’t earn it in this session.

What’s missing: any serious discussion of energy as the binding constraint, despite naming it as the first layer of Jensen’s cake. Also no discussion of open-source models compressing model-layer pricing, which is arguably already happening and would be relevant to the “are app margins ever going up” question.

8/10. Substantive, well-structured, and grounded in real numbers. Not the highest-density 35 minutes I’ve watched on this stack, but a clean conceptual scaffold and worth coming back to as the course unfolds.