Stanford MS&E435 Economics of the AI Supercycle — Annotated Lecture Notes & Key Takeaways

Course Information

Institution: Stanford

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

Instructor Bio: This session is led by **Apoorv Agrawal**, Adjunct Lecturer in Management Science and Engineering at Stanford University and Partner at Altimeter Capital. Apoorv Agrawal holds a Bachelor’s degree in Computer Science from the National University of Singapore and an MBA from Stanford University. At Altimeter Capital, he leads the firm’s investments in artificial intelligence and enterprise software, and has backed defining companies of the AI supercycle including OpenAI, Glean, Revolut, and Baseten. He brings deep operational and investment perspective across the full AI value stack, from semiconductors and infrastructure to end-user applications.

Course Description: As the inaugural session of the course, this lecture establishes the core framework of the AI supercycle and the economics of generative AI. It breaks down the layered AI value chain — semiconductors, infrastructure, foundation models, and end-user applications — and introduces the central thesis of the *inverted value triangle*: that value capture in the current AI boom is heavily concentrated at the bottom infrastructure layer, but will gradually shift upward to applications over the course of the supercycle. It also outlines course logistics, grading structure, guest speaker lineup, and learning objectives.

Course Origins and the Instructor's Stake in the Question

 

Apur (the instructor) opened by framing this nine-week course around a deceptively simple question: where does the money actually flow in generative AI? He is not a typical tenured academic. His background runs from India to Singapore, then to a data-engineering career at Palantir, grad school at Stanford, and now a leadership role at Altimeter — a concentrated investment firm with both public and private arms. That investing lens shapes everything in this lecture. He is not trying to teach AI as a technical subject; he is trying to teach the economics of the stack as an allocation problem.
 

One personal detail worth noting: he became a father six months ago and joked that it is "the biggest investment I will make," with "guaranteed negative IR" financially but the best possible return in non-financial terms. It is a throwaway line, but it subtly reinforces his framework — every decision in this course is about where capital goes and what kind of return (financial or otherwise) it produces.
 

The course itself is deliberately light-touch: no more than three hours a week including readings, one hour of class, an hour or two of readings, and guest speakers every session from the second class onward. The TA is Chloe Fang. Grading is 50% attendance and 50% a final assignment released at the end. There is also a quiz on day one — partly for fun, with a prize for the winner, and students are allowed to use AI tools like Claude as long as they give human players five seconds to answer first. The message is clear: this is a conversational, low-friction course, but the content is meant to be operationally serious.
 

Key takeaway:​ The instructor's core motivation is that he could not find a course, when he was a student, that went deep on where value accrues in the AI stack — only courses about how the technology works. That gap is what this class is trying to fill.
 

---

The Central Visual: An Inverted Triangle

 

What the chart shows

 

The anchor of the entire lecture is a chart comparing the AI ecosystem to earlier technology revolutions — internet (roughly 25 years ago), mobile (20 years ago), and cloud (10 years ago). The right-hand side shows the AI value stack; the left shows the same analysis from about two years earlier.
 

The striking observation is structural. In a mature software or cloud ecosystem, value is distributed broadly: many application companies, healthy infrastructure providers, and so on. The AI version looks dramatically different — it resembles an inverted triangle (or wedge) that is wide at the bottom (semiconductors, energy, data centers) and narrow at the top (applications).
 

Apur calls this the "punchline" of the course. The big question it raises: the hyperscalers and Nvidia are investing enormous sums into capital expenditure — building data centers layer by layer. He referenced Jensen Huang's description of it as a "five-layer cake": energy, chips, power, interconnects, and memory. All of that produces a data center that can be rented by the hour or by the token to train and serve models. The question then becomes: are the models being built on top actually creating commensurate economic value?
 

Why the shape differs — student theories and instructor responses

 

Several students offered hypotheses, and Apur accepted all of them as partially correct:

1. It is early.​ The ecosystem has not had time to mature. This is the simplest explanation, and he agrees it is a factor.
2. Nvidia's dominance.​ One student noted that Nvidia holds a near-monopoly and can extract leverage. Apur deferred the detailed discussion to a later session with Nvidia speakers but confirmed the point — Nvidia has a "stranglehold" — and referenced a chart in the readings showing their compute market share.
3. Physics of the problem.​ Unlike traditional software, the marginal cost of an additional AI user is not close to zero. Running inference "burns GPUs," and that cost has to be paid every time someone uses an application like Cursor.

 

Apur added a fourth, historical dimension: in previous cycles, it took years for the triangle to "flip." His example was AWS — founded around 2004, first major customer Netflix around 2010, Amazon fully shifted onto AWS by 2012. That is roughly eight years from the first capex cycle to a mature value distribution. He also reminded students that 20 years ago, the prevailing question about Amazon was whether the AWS buildout would drive the company bankrupt.
 

Key takeaway:​ The inverted triangle is not necessarily a sign that AI is broken; it may just be a very early-stage version of a pattern that eventually normalizes. But the timeline is genuinely uncertain — and that uncertainty is where both risk and opportunity live.
 

---

Where Profitability Actually Sits

 

The semis layer dominates

When a student asked where profitability is concentrated, Apur gave a direct answer: the most profitable part of the stack is semiconductors, "by a long shot." He estimated Nvidia's data center business at roughly 75% gross margin (with a caveat not to quote him exactly), compared to application-layer companies operating somewhere between 0% and 30% depending on whose numbers you use.
 

He also noted that on a profitability basis, the concentration is even more extreme than the revenue chart suggests. The "triangle" becomes even more lopsided when you measure margins rather than top-line size.
 

A few implications he drew out:

• The inference layer is the most competitive.​ Sitting between the chip makers and the applications, this middle layer has the highest "metabolic rate" — companies being formed and acquired rapidly. Both startups and hyperscalers want control of it, so the jury is still out on who wins. A recurring question for founders: "Are you a feature or a platform?" If AWS could absorb your product tomorrow, that is a serious vulnerability.
• Incumbent platforms complicate the picture.​ A student asked whether companies like Salesforce (with Einstein) or Palantir (with AIP) should count toward the application-layer revenue. Apur said yes — their AI-related spending is captured in the app layer, but it is very difficult to extract from public disclosures. He handles this by tracking model and inference revenue separately where possible.
• Hyperscalers span multiple layers.​ Google, for example, has TPUs (semis), GCP (infrastructure), and Gemini (applications). He treats each business unit independently when placing it on the chart. Gemini, he noted, is currently the second-most-used consumer application, though how much of that comes from distribution advantage versus product quality remains an open question.

 

The timing mismatch

 

One of the more thoughtful exchanges concerned the "semis buildout" being driven by five-to-six-year investment cycles, while application revenue is only beginning to materialize. Apur agreed this creates a cyclical dynamic in the lower half of the chart. His analogy: laying down railroads. The capex-heavy portion behaves cyclically, and a basket of capex-heavy names should be expected to go through boom-and-bust phases — something similar happened in the early innings of mobile.
 

He also addressed training versus inference. Nvidia's fleet was roughly 40% inference / 60% training at the time he checked, and he expects that mix to shift toward inference over time. But he explicitly declined to predict when, noting that training workloads remain substantial and have a very different shape (predictable, high utilization over short bursts) from inference (bursty, tied to human waking hours, with usage dipping around Thanksgiving and Christmas).
 

What I found confusing:​ The distinction between "the triangle is early" and "the triangle is structurally permanent" was never fully resolved. Apur clearly leans toward "it will flip eventually," but he also said it might stay inverted longer than he anticipates — possibly because getting the substrate right is intrinsically hard. That tension is real, and he is unusually honest about not knowing.
 

---

What It Would Take for the Triangle to Flip

 

Signals to watch

A student asked a sharp question: if the triangle is currently inverted, what would an unsuccessful technology look like — would it also be a triangle? How do you tell whether it will invert or stay stuck?
 

Apur reframed this as the question of the industry's "stable equilibrium." He is confident AI is not a fad and not an unsuccessful endeavor. His best guess is that the current shape could persist longer than expected — perhaps a decade, based on the cloud comparison — because the substrate layer is so difficult to get right.
 

But he identified two concrete catalysts that could trigger a repricing:
 

1. A breakout ASIC program.​ If a hyperscaler's custom silicon — Google's TPU, Meta's MTIA, or efforts from Amazon, OpenAI, Microsoft, or labs not yet public — achieves major success, that could be the biggest single shift in the semiconductor layer.
2. Capex guidance changes.​ Hyperscalers report earnings four times a year, and Apur recommends listening to every call. If they stop guiding to large capex numbers, that would imply the current equilibrium is not working.

 

He also noted there is roughly $300 billion of revenue "to fight about" in this ecosystem, about half of which comes from the big hyperscalers (per Jensen Huang's disclosures). So if you were starting a chip company today, your customer base would be a very small number of enormous orders — a fundamentally different shape from a consumer or enterprise software business.
 

Key takeaway:​ The triangle flipping is not a matter of hope; it is tied to observable events (custom silicon success, changes in capex guidance, inference overtaking training). Those are the metrics worth tracking over the quarter.
 

---

The Consumer Monetization Problem

 

How big can AI apps get?

The final major segment tackled consumer AI — currently the largest market for AI outside of coding. The data point: ChatGPT has roughly a billion users monetized at about $10 per user per year. Compare that to Alphabet (~4 billion users at ~$100/year) and Meta (~3.5 billion at ~$70/year).
 

Apur walked through three tiers of consumer products:

• Mandatory-scale (~3 billion users):​ WhatsApp, Chrome — products you cannot live without.
• Social-scale (~1.5–2 billion):​ Instagram, TikTok, Facebook — not mandatory, but powerful network effects.
• Niche / transactional:​ Amazon (shopping), Spotify (music), Twitter (news/debate) — task-specific, not universal.

 

Where do ChatGPT and Gemini fit? ChatGPT has just overtaken the "niche" category; Gemini has not yet. Both are heading toward "social-scale," but Apur expressed doubt they will reach the "mandatory utility" tier — at least not through knowledge work alone.
 

His reasoning is worth sitting with: the number of people in the world who actively ask technology questions is not the entire online population. There are roughly 8 billion people on the planet, 4 billion online, but knowledge work is a subset of that. ChatGPT is not (yet) where you message friends, check email, or get a dopamine hit — it is where you go to do active work. That limits its natural ceiling.
 

The two big unanswered questions

 

This produces two central questions for the course:

1. How do you get from 1 billion to 4 billion users?​ Knowledge work alone may not be the answer. AI probably needs to expand beyond it.
2. How do you get from $10 per user per year to$100?​ Neither ads nor subscriptions alone may be sufficient.

 

On the ads question, Apur was intriguingly specific. He suspects AI-native advertising will be more valuable than current display ads because it will understand user intent, have strong attribution, and benefit from logged-in trust. He drew a parallel to the Facebook IPO, when short-sellers argued mobile ads would fail because "there is no space on a phone" — and then the industry found the space. The same debate, he argues, is playing out now: no one wants ads interrupting a personal AI conversation, yet the economics will likely force a solution.
 

A limitation worth flagging:​ This is genuinely speculative. Apur explicitly said "I couldn't tell you what it's going to be like" when describing AI-native ads. It is a well-reasoned hypothesis, not evidence. If the ad model does not materialize, the entire application-layer revenue thesis looks much weaker.
 

---

Course Structure and What to Expect

 

The nine-week schedule runs the full stack, from semiconductors through infrastructure and energy to models, applications, and agents. Confirmed or referenced speakers include people from OpenAI, Anthropic, and Nvidia. The format is guest-heavy, with Chatham House rules (don't record or share what speakers say), and optional dinners afterward.
 

Recurring questions Apur plans to pose to every speaker:

• Are you a dominant force, and how long will you stay one?
• What competitive threats worry you most?
• What are the pricing-compression vectors for your business?
• For model labs: is your user base profitable? Will ads or subscriptions be the larger revenue source?
• For infrastructure companies: are you a feature or a platform?

 

What I would have liked more of:​ concrete numbers or a worked example of the triangle's revenue split. Apur references figures throughout — $300 billion, 75% margins,$10/user/year — but these are scattered across the lecture rather than compiled into a single, clear table. That is probably intentional given the first-day format, but it does make the picture harder to hold in your head.
 

---

Where the Argument Feels Strong — and Where It Does Not

 

Strongest point:​ The inverted triangle is a genuinely useful visual, and the historical comparison to cloud is persuasive. AWS took ~8 years to reach maturity, and the idea that AI is simply earlier in its cycle than people realize is a disciplined counter to the "AI is a bubble" narrative. The instructor is also refreshingly explicit about what he does not know.
 

Weakest point:​ The claim that the triangle will eventually flip rests on analogy rather than a mechanism. Cloud flipped because software businesses could distribute at near-zero marginal cost; AI applications, by contrast, have persistent GPU costs. That is a structural difference, not just a timing difference. Apur acknowledges the GPU-cost problem but does not fully grapple with whether it prevents the flip rather than merely delaying it.
 

There is also an implicit assumption that advertising will unlock the missing revenue. That may be right, but it is essentially a bet on consumer behavior — and one that has not been tested at scale in an AI-native interface.
 

---

Bottom Line

 

If you take one thing from this first session, it is this: the money in AI is currently at the bottom of the stack, not the top.​ Chips and infrastructure are extracting most of the value, while application companies — despite enormous user growth — struggle with unit economics because every user costs money to serve. The central drama of the course will be watching whether (and when) that imbalance corrects, and what specific events — custom silicon, capex guidance shifts, inference dominance, or new ad formats — signal the change.
 

For students, the practical value is a set of "mental models" for evaluating any AI business: what layer is it on, who holds pricing power, what are the unit economics, and what forces could compress them? Whether you end up starting a company or funding one, those questions are the point of the class.
 

Content Disclaimer:

This article is for general reference only and does not constitute professional R&D guidance, production process advice or quality certification. All material performance data has specific test premises; readers should verify parameters against actual equipment and working conditions.