Stanford CS221 ｜ Autumn 2025 ｜ Lecture 19： AI Supply Chains — Annotated Lecture Notes & Key Takeaways

Course Information

Institution: Stanford

Original Course: Stanford CS221 | Autumn 2025 | Lecture 19: AI Supply Chains

Instructor Bio: This lecture is delivered by Percy Liang, Associate Professor of Computer Science at Stanford University and core faculty of the Stanford Institute for Human-Centered Artificial Intelligence (HAI). Percy Liang leads the Stanford Natural Language Processing Group and the Center for Research on Foundation Models (CRFM). His research spans the theoretical foundations and practical systems of artificial intelligence, including machine learning, natural language processing, AI alignment, and rigorous model evaluation. He received his PhD in Computer Science from the University of California, Berkeley and his BA in Mathematics from Harvard University. His work has been recognized with the NSF CAREER Award, Google Faculty Research Award, and multiple best paper awards at top-tier AI conferences. He has taught CS 221 at Stanford for over a decade, shaping foundational AI education for thousands of students.

Course Description: This lecture examines the global ecosystem and layered supply chains that underpin modern AI systems. It maps the full AI value chain from semiconductor hardware and cloud infrastructure to foundation models and end-user applications. The lecture analyzes geopolitical dynamics, resource constraints, and economic factors shaping AI development, and discusses how supply chain resilience, accessibility, and concentration impact the future trajectory of AI technology.

Percy opens this lecture by reminding everyone that the course has been steadily moving attention from what happens inside the model to what happens around it — the upstream resources that make AI possible and the downstream places where it gets used. The guest for the day is Rishi Bommasani, who finished his PhD at Stanford and is now a senior research scholar at HAI, best known for leading the paper that coined the term "foundation models" and for working on AI policy in both the US and the EU. His subject here is the economics of AI, and specifically AI supply chains: who supplies what to whom, where the bottlenecks sit, and what any of that implies for growth.
 

A note before anything else: the transcript I worked from is auto-generated, and proper nouns are rendered phonetically. So "Eric Nolson" is presumably Erik Brynjolfsson, "Duron Asamoglu" is Daron Acemoglu, "Bob Salo" is Robert Solow, "Tim Breahan and Martin Chhattenberg" map onto Timothy Bresnahan and Manuel Trajtenberg, and "Arvin Sash" appears to be Arvind Narayanan. I'll flag the ones that matter as they come up rather than silently "fixing" them.
 

Why you need organizations, not just algorithms

 

Bommasani frames the whole lecture around a puzzle that most CS students, himself included as a Cornell undergrad, never got taught. You can spend years learning how to build AI systems and never think about how those systems land in an economy. His claim is that understanding the technology's trajectory is not enough, and he gives three reasons.
 

The first is that AI is a crosscutting technology, so most of its cumulative effect will arrive through the decisions of firms in sectors that have nothing to do with tech — whether a hospital, a bank, or a logistics company adopts it, and how. The second is distributional: knowing that AI matters tells you nothing about which firms capture the value. The third is that firms make a lot of choices that have nothing to do with capability.
 

That third point gets the sharpest illustration in the lecture. He puts up benchmark results — he deliberately doesn't say which benchmark — showing Google, Anthropic, and OpenAI clustered at roughly the same score. If you stopped there, you'd conclude these companies are substitutes for one another in the economy. His objection is that even granting the models are equally capable, the three firms differ in when they release models, how they price them, what downstream products they build and vertically integrate into, and which partners they work with. Capability parity does not imply economic equivalence (10:36).
 

From there he sets the two questions he'll keep returning to. One is macro: the seven most valuable AI companies already make up over a third of the entire S&P 500 (3:28), so how contingent is the whole economy on continued AI progress? The other is micro, and aimed directly at a room full of people about to enter the job market. He shows a plot built on ADP payroll data — ADP being the world's largest payroll processor — because payments turn out to be one of the main instruments for studying the economy empirically. The line that matters slopes down sharply after 2022: junior-level software development hiring dropped off fast (5:25).
 

The second individual-level study is even more interesting, and slightly counterintuitive. In a call-center case study where some workers began using a generative AI tool in 2023, the productivity gains were largest for the most junior staff and much smaller for veterans (6:08). Tenure, in that setting, seems to be worth more than the tool. I'd have guessed the opposite — that experienced workers would know better how to exploit a new assistant — and the lecturer doesn't really explain the mechanism, which left me wanting a concrete walkthrough of what the juniors were doing differently.
 

His synthesis is that we need a dual lens: AI as a class of technology and AI as a set of organizations, held in mind at the same time.
 

Walking the AI supply chain: compute, data, distribution

 

The supply chain can be described two ways at once. Technologically, it runs from datasets and GPUs through foundation models to the systems built on top. Organizationally, the same chain runs through news companies, cloud providers, model labs, and coding-tool vendors. He wants you to hold both descriptions simultaneously, with the technical assets mediating the commercial relationships. He then picks three regions he thinks are especially legible from both sides.
 

Compute is a story about monopolies

 

The CS-Student mental model — Nvidia GPUs sitting in data centers — isn't wrong, but it's too short. He shows a deliberately simplified semiconductor map, then narrows to three firms, each of which dominates one layer and is among the most valuable companies in its region of the world.
 

ASML, a Dutch company, builds the lithography equipment needed to fabricate advanced chips and is essentially the only game in town at that level (15:26). TSMC, in Taiwan, does the actual manufacturing and is therefore a critical dependency for Nvidia, which designs the chips (16:11). Three layers, three near-monopolies.
 

Two consequences follow. The first is resilience: when a single firm is wholly responsible for one layer, it becomes a bottleneck for every layer downstream, and a lot of value accrues to it precisely because it's indispensable. The second is geopolitics, which he calls a third layer of abstraction stacked on top of the technology and the firms. TSMC's location in Taiwan puts it at the center of US–China competition in DC conversations; Nvidia sits at the center of export-control debates about which chips can be shipped to China (18:37).
 

He then moves one step downstream to clouds, where compute serves two purposes: training and inference. Training is concentrated in Amazon, Microsoft, and Google, with most frontier labs holding partnerships or dependencies on one of the three. Inference looks more heterogeneous, with a lot of younger specialized providers competing. And he flags a part of the chain he explicitly doesn't have time for: the data center itself, plus the land, electricity, and water it needs — a piece that's becoming a bottleneck in its own right (20:33).
 

Data looks nothing like compute

 

The contrast is the point. Compute is capital-intensive and concentrated; data is cheap to produce and therefore far more heterogeneous (21:02). He organizes data acquisition into six channels, and the taxonomy is genuinely useful:
 

• Self-generated data​ — synthetic data used in post-training, produced via reinforcement learning and similar methods. He sets this aside, since the firm makes it itself.
• Usage data​ — what users hand over while using ChatGPT, or Gmail, or Facebook. It's already inside the firm before anyone starts training.
• Nominally public data​ — open datasets like SQuAD or The Pile, plus the crawlable web, where there may be a nominal owner but no negotiation.
• Commissioned annotation​ — new labels bought from firms like Scale or Mechanical Turk for post-training.
• Licensed proprietary data​ — Reddit's posts, the New York Times' articles, sold or licensed to labs.
• Web crawling​ — which he treats separately because it's where things are moving fastest.

 

Web crawling is where the supply chain visibly tightened. Plotting website policies over time, restrictions on crawlers have risen sharply, the share of sites with no restrictions at all has fallen, and the share blocking crawling entirely has grown (25:19). The asymmetry is the interesting part: when a site's robots.txt singles out a specific crawler by name, OpenAI's crawlers are blocked more often than Anthropic's (25:51). If data becomes a differentiator between models, that asymmetry is not a neutral fact.
 

Then there's pricing, and this is the part I found most novel. The six channels have almost nothing in common economically. With synthetic data you're really paying compute costs. With usage data you're paying nothing in money — the price is embedded in your terms of service. With annotation you're buying worker-hours. With licensing you're making one large transaction. Different cost structures imply different asymmetries in who can afford which data, and therefore in who can build what.
 

Law enters as a pricing mechanism, not just a constraint. He cites the Anthropic copyright settlement — $1.5 billion across roughly 500,000 works, which works out to about $3,000 per work (28:29) — and points out that a settlement is, among other things, a revealed price. Copyright and piracy law, CSAM prohibitions that ban possession outright, GDPR-style privacy rules: all of these mean you have to understand the provenance of your data, not just the final artifact, to know whether you're compliant.
 

Why care about any of this? Three answers. The technical one is whether we're going to run out of data — and he notes that technologists often hold strong opinions about this while not actually knowing the current mixture, the generation rates, or the existing stock. The legal one is compliance. And the competition one, which he promises to return to: if some developers can acquire data that others can't, does that convert directly into better models (30:54)?
 

Distribution decides what downstream can exist

 

The last region is how models get out into the world, and he lays it out as a spectrum rather than an open/closed binary. At one extreme, a developer releases nothing externally — models stay siloed and only products and services are exposed, which is entirely normal in software. In the middle sit API access and other restricted arrangements. The line he draws for "open" is whether the model weights are distributed at all (32:47). Past that line, further choices remain: use restrictions in the license, whether you also release training data and code, and how far toward conventional open source you go.
 

The consequences ripple downstream in three ways he can see empirically. Closed distribution buys vertical control and lets a firm integrate into downstream markets, but it also means ceding opportunities it isn't personally investing in. Closed models also price higher on average, because open weights create a serving market — release them and many companies can host inference, and the competition shows up in price (34:52). And openness unlocks application types that closed access can't support: running locally gives firms a clearer picture of where their data goes, which matters enormously in regulated sectors, and having the weights permits fine-tuning and distillation (36:15).
 

Is AI a general purpose technology, and what would that predict?

 

The back half of the lecture stops describing and starts forecasting, and Bommasani pre-warns the room that this will feel foreign to CS people because it's macro: less precision about individual decisions, more reasoning about aggregates. He also cabins the scope — frontier LLMs only, no autonomous driving; aggregate trends only, no distributional questions; growth only, so nothing about how students use AI, since that never appears in workplace statistics.
 

Before his own framing he surveys other people's. On the informal end, there's Nate Silver's "technological Richter scale," which assigns technologies to decades, centuries, and so on up an exponential ladder — credit cards for a decade, electricity for a century. He also mentions a recent Forecasting Research Institute exercise where AI professors, economists, and superforecasters mostly landed on "technology of the century." On the formal end, Acemoglu's "The Simple Macroeconomics of AI" gives a headline estimate that most readers read as modest — important, visible in GDP, not game-changing (42:51). A Princeton computer scientist's "AI as Normal Technology" comes out slightly more bullish; Dario Amodei's description of a "country of geniuses in the data center" is considerably more bullish still; and then the sign flips entirely depending on whether you're hearing Sam Altman on curing cancer or someone worrying about extinction.
 

The framework he wants instead comes from the economics of general purpose technologies. Bresnahan and Trajtenberg give three criteria, and he and his co-authors claimed in the foundation models paper that foundation models satisfy all three (46:17):

1. Pervasiveness​ — not "used a lot" in the colloquial sense, but adopted across many economic sectors. The evidence he offers is the Anthropic Economic Index showing non-trivial usage spread across downstream sectors.
2. Improvement over time​ — two sub-conditions: capability rising and price falling. Both appear to hold, with task-complexity benchmarks climbing and the dollar cost of inference at a given capability level falling.
3. Innovation complementarities​ — the ability to spawn complementary innovation, which he calls the heart of the matter.

 

The third is where he's most careful, and most interesting. Electricity mattered not because electricity existed but because of everything built on top of it: light bulbs, motors, night shifts, and the wholesale redesign of how factories and organizations were arranged. Apply that lens to AI and you get two candidate signals — people are building things on top of models (coding tools like Cursor), and organizations are reshaping workflows, including an entirely new class of work around verifying AI outputs (50:29). But he explicitly says we lack the quantitative evidence that economists would want, and that this is the condition he's least sure about.
 

Knowing whether AI qualifies matters because GPTs have a documented shape: a J-curve. Productivity dips first, while organizations learn how to use the thing, and only then climbs. He offers this as one explanation for a tension he thinks everyone in the room feels — people in the Bay Area believe models are extraordinarily capable, yet the measured economic effects so far are muted relative to that belief (51:27).
 

Three belief-conditioned views of growth, and the problem with GDP as a yardstick

 

Before the three views, a detour on measurement. Long-run world GDP growth is one of the most famous plots in growth economics, and the more famous zoom-in shows GDP per capita rising at a strikingly predictable ~2% a year (52:48). That's the benchmark any AI claim has to beat.
 

The problem is that GDP may simply be the wrong instrument. Robert Solow's line about seeing the computer age everywhere except in the productivity statistics anchors the point (53:42). Bommasani's explanation: many internet-era technologies are free, cheap, or ad-subsidized, so if using Google Search costs you nothing, nothing lands on the ledger and GDP registers nothing despite something consequential having happened. He adds that recent growth may actually be running below the 2% trend.
 

The alternative he likes is GDP-B, from Brynjolfsson and colleagues, which tries to measure consumer surplus directly through choice experiments — how much would you have to be paid to give a good up. In a survey run earlier this year, roughly 40% of people use generative AI tools frequently in daily life, and their average willingness-to-accept for giving them up was $98 a month (56:44). Scale that across the US population and you get on the order of $100 billion a year in consumer surplus that GDP accounting misses. He then sets all of this aside and says: assume GDP is the metric, and ask what different beliefs about AI imply.
 

View one: one sector becomes wildly more productive.​ Say software. The counterintuitive result is that GDP growth is much smaller than the productivity gain. The sector's prices fall as it gets efficient, so even with latent demand and higher consumption, its share of the economy shrinks. David Autor's example is illumination — candles to light bulbs raised lighting productivity enormously, crashed the price, destroyed lighting jobs, and left lighting a far smaller slice of the economy (1:00:30). Then there's a second-order effect, Baumol's cost disease: other sectors must offer comparable wages to compete for workers, so they get more expensive without getting more productive. His punchy version: great software, more expensive healthcare.
 

View two: AI is cheap labor.​ He sketches a production function — output as a function of capital, labor, and total factor productivity — mostly to get the vocabulary, and notes that not all economists accept this functional form. The practical point is constant returns to scale: K and L grow best together. If AI inflates L but K doesn't keep pace, you can't translate the new labor into useful output, and GDP doesn't follow (1:04:04). What makes this view attractive is the baseline: in developed economies, population growth has stagnated and labor force participation has plateaued or is falling, so a new labor supply arrives at exactly the moment the old one stopped growing.
 

View three: AI produces ideas.​ This is the most aggressive forecast, and it comes from Paul Romer's insight that ideas are non-rival — once linear algebra exists, everyone can reuse it, unlike a computer, where helping a second person requires building a second machine (1:06:42). If AI's real contribution is generating ideas, then the growth function could bend past exponential, and the thing to watch isn't AI's effect on existing tasks at all but its effect on R&D and science. He points to work by Ben Jones on exactly that.
 

Crucially, he presents these as conditional — pick a belief, and you can now write out what follows — without telling us which one he holds.
 

Where the lecture felt thinnest

 

The weakest link is the J-curve, and Bommasani is honest about it. Asked in Q&A for empirical instances of past J-curves, he says the paper is largely theoretical under a set of assumptions, that it matches intuition and some observations, and that he isn't aware of a clean measured instance where a specific technology's productivity actually traced the J (1:12:31). That's a load-bearing claim being supported mostly by plausibility. If the J-curve is your main explanation for why measured effects lag capability hype, "no clean demonstration exists" is a significant caveat, not a footnote.
 

The consumer surplus estimate is the second soft spot. Multiplying 40% usage by $98/month by the US population reads as neat, but it rests on stated willingness-to-accept in a survey — a number people have every incentive to inflate when nobody is actually paying. It's a useful示 indication that GDP misses something, not a measurement of how much.
 

Third, the lecture's centerpiece move — that capability parity doesn't imply economic equivalence — is stronger than it needs the benchmark evidence to be. He says "even if" the models are equally capable, but the whole argument only needs the firms to differ in release timing, pricing, integration, and partnerships, which they plainly do regardless of benchmark scores.
 

And the three growth views, while elegantly structured, leave you without a recommendation. Pressed directly, he says he'd lean "century" over "decade," that he's most confident in the crosscutting labor-substitution story, and that the augmentation-versus-automation distinction matters enormously for workers while being nearly irrelevant to GDP (1:09:22). That last point is worth sitting with: the question that matters most to the students in that room is explicitly outside the frame the lecture chose.
 

What I'd take away

 

The durable idea from this lecture is that AI supply chains are not a neutral technical description but a map of power. Three firms sit on three chip layers; crawlers get blocked at different rates depending on whose they are; a copyright settlement turns out to be a price list; the choice to release weights changes what applications can exist at all. None of that is visible from inside the model.
 

The second durable idea is that the macro question is genuinely open, and the reason it's open is partly instrumental — the yardstick we have may not register the thing we're measuring. If you believe AI is a general purpose technology, history says expect a trough before the gains, and expect the gains to come from reorganizations nobody has built yet rather than from the models themselves.
 

Two threads he flagged but didn't develop are, to me, the most interesting loose ends: whether exclusive data access really converts into better models, and whether the binding constraint on scaling is power generation or power transmission — he guesses transmission in the US, and says plainly he doesn't understand it well enough to say more (1:11:21).
 

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.