Table of Contents
🌏 中文版
This article covers Stanford CS221, Autumn 2025, Lecture 19, dated 2025-12-01. The official schedule calls it AI Supply Chains, while the executable lecture repository links to a deck titled Economics of AI. I downloaded and read the complete 70-page Google Slides PDF. Its cover lists Rishi Bommasani as deck author / credited presenter. That credit does not establish who presented the class, so this article does not call him the in-room speaker or extend Percy Liang’s course-instructor role into a Lecture 19 speaking attribution. The primary written artifact is Economics of AI Google Slides (deck snapshot checked 2026-08-22; the slides do not print a market date for every valuation), with the offering and schedule fixed by the official course site.
Material boundary: the schedule and repository use different names for the same 2025-12-01 lecture. I label claims directly stated in the PDF as “Deck fact,” research or forecasts cited by the deck as “Deck citation,” and connective reasoning as “My interpretation.” I did not fill gaps with another term or a lecture transcript. Market valuations, survey numbers, and forecast graphics are treated as snapshots in this deck, not as live statistics in 2026.
The lecture's problem
The 70-page deck does not open by asking which architecture comes next. It asks three connected questions: how might AI change the economy as a whole, which upstream resources and organizational choices make models possible, and through what mechanisms could model capability become growth? That is the overlap between the schedule’s AI Supply Chains label and the slides’ Economics of AI label.
The agenda has a deliberate order: technology and organizations, current supply chains, and growth economics. If we jump from “a more capable model” to “higher GDP,” we skip adoption, resource access, value capture, and the way free services enter national accounts.
[Deck fact] The opening slides show the health of the world economy, the future of work, and inequality, then state that overall economic impact depends heavily on non-AI organizations. Technology alone does not say who wins or loses. Even model developers still choose when to release a model, how to price it, which products to build, and whom to partner with. The deck first presents “Google = Anthropic = OpenAI” as an over-simple conclusion and then rejects it on the next “reality” slide: their choices and contexts differ, so model capability alone is not a ranking of outcomes.
How the method develops
Put technology back inside organizations.
[Deck fact] The first section puts technology and organizations under one lens. Technology does not automatically create economic impact; organizations must rearrange decisions, workflows, responsibility, and investment. The slides make the choices concrete: when to release a model, how to price it, what products to build, and whom to partner with. These are not marketing details added after training; they shape how value is captured and distributed.
[My interpretation] The model and serving stack form a lower layer; people, process, permissions, evaluation, and customer relationships form an upper layer. A lower-layer upgrade may produce no productivity gain without workflow and accountability changes.
Current supply chains: compute.
The second section begins with compute, while the deck explicitly warns that compute is more than chips. It includes chip design and manufacturing, extreme-ultraviolet lithography equipment, packaging, data centers, electricity, networking, cooling, and compute services supplied by major clouds. The deck does not turn these into a complete cost model; it uses them to show that model production is a chain of interdependent bottlenecks.
[Deck fact; snapshot checked 2026-08-22; underlying market date not printed] A critical-companies slide gives three examples and valuation snapshots: ASML is described as the global monopoly in extreme-ultraviolet lithography and labeled at about $400 billion; TSMC is described as the top manufacturer of advanced chips and labeled at $1.5 trillion; Nvidia is described as having dominant market share in the chip market and labeled at $4.4 trillion. The official deck is the source for those valuation and dominance descriptions. They are not current prices checked here and cannot establish what share of all AI value accrues to each company.
[Deck fact] The next slide asks why this matters and names supply-chain resilience, who accrues value from AI growth, and US-China geopolitics. TSMC (Taiwan) and chip export controls appear under geopolitics. The deck does not offer a complete policy conclusion or claim that one country controls the entire chain; it puts geographic concentration, export restrictions, and value capture on the same risk map.
[My interpretation] Engineering review should ask whether a cloud region, GPU, packaging, or power failure has a substitute. Compute cost is not only per-token price; capacity, utilization, location, and latency matter too.
Data: acquisition method determines price and rights.
The deck divides data acquisition methods into three sources. Data within the firm includes synthetic data such as RLAIF and usage data such as ChatGPT interactions and Gmail. Data that is public includes public datasets such as SQuAD and The Pile, plus web crawling. Data that third parties have includes new data such as Scale and Mercor, and old data such as the New York Times and Reddit.
[Deck fact] The point is not that “public” means zero cost. It is that pricing depends on the acquisition method. Synthetic data within a firm requires compute; usage data may be obtained by paying for services; public datasets are marked free on the slide; web crawling is marked currently free; new third-party data is associated with paying for labor; old third-party data is a negotiated large-scale transaction. “Free” here is a simplified label for direct acquisition price in the deck’s taxonomy, not a promise of perpetual free access or zero compliance cost.
The deck then makes transaction scale concrete: $1.5 billion for 500,000 works = $3,000 per work. [Deck fact; snapshot checked 2026-08-22] The official deck supports only the claim that this is one market example printed on the slide. It does not identify the parties, work types, rights, term, territory, or transaction date. It cannot support a general price for copyrighted data or a universal per-work rate.
The classification raises at least three questions. First, will we run out of data? That requires looking at the current mix, data-generation rates, and total data, not merely counting how much text remains online. Second, compliance includes copyright, piracy, CSAM, and privacy; these are constraints inside the data chain, not a checklist added after training. Third, competition changes when certain data is available exclusively: data is not only an input but can become a barrier to entry.
[My interpretation] A data inventory should record provenance, acquisition, permitted use, update, and withdrawal together. Dataset size alone misses rights, freshness, exclusivity, and cleaning cost.
Distribution: the downstream channel reshapes the chain.
The third current-supply-chain element is distribution. The deck does not reduce distribution to an app store or API gateway. It asks how distribution shapes downstream supply chains, listing vertical control, pricing, data flows (privacy and security), and adaptability (fine-tuning, distillation, and related options).
[Deck fact] The same model distributed through a proprietary product, a cloud platform, a partner, or open weights creates different downstream relationships. Vertical control affects who can change the interface and routing. Pricing affects who can afford use. Data flows change the privacy and security boundary. Distribution also affects whether downstream users can fine-tune or distill. The deck does not declare one channel universally best.
[My interpretation] The chain is also a feedback loop: distribution changes feedback data, pricing changes demand and capacity planning, and vertical integration raises control while potentially reducing substitutability.
Growth economics: from micro decisions to aggregate growth.
The third section separates micro from macro: microeconomics studies individual decisions, while macroeconomics studies the whole economy. The deck says computer science is predominantly micro-oriented, whereas this section asks about aggregate macro outcomes. Its scope is deliberately narrow: AI means frontier AI as currently understood; it excludes Waymo, industrial robots, and Facebook recommendation systems. Economics here means aggregate outcomes; it excludes distribution outcomes, economic policy, student and leisure uses, and misuse.
The deck uses a Technological Richter Scale, impact examples, and labeled forecasts from LEAP and Murphy et al. to show a range from utopia to doom. [Deck citation; snapshot checked 2026-08-22] The official deck presents these as an attributed forecast landscape, not a single number guaranteed by Bommasani. [My interpretation] The useful reading is to ask what each forecast assumes about adoption speed, capital, automatable tasks, reinvestment, and complementary innovation. The deck frames the magnitude question; it does not provide a plug-in forecast calculator for a company.
Is AI a general-purpose technology?
The slides return to normal technologies and generic normal technologies, cite Bresnahan and Trajtenberg’s GPT research, and ask whether AI is a general-purpose technology (GPT) and, if so, whether it follows GPT principles. They also cite Bommasani and collaborators (2021): pervasiveness, improvements over time, and innovation complementarities. These are the deck’s explicit literature links, not conclusions to attribute to Bommasani personally.
[Deck fact] Pervasiveness means that the technology can spread across sectors. Improvements over time means that capability continues to improve. Innovation complementarities mean that downstream applications and organizational changes reinforce a general-purpose technology. The deck’s technological example is downstream applications such as coding tools; its organizational example is workflow redesign such as verifying AI outputs. It also presents the productivity J-curve of GPTs, citing Brynjolfsson et al. (2018): complementary investment and organizational adjustment may come before productivity appears in measured statistics.
GDP, free services, and three growth pathways.
GDP is the central measure in the latter part of the deck. It shows GDP trends over roughly 150 years and cites Chad Jones (2023). It also revisits the observation associated with Solow and Brynjolfsson: we can see the computer age everywhere except in productivity statistics. The deck offers a concrete reason: the Internet clearly changes life, but many Internet technologies—Google Search/Maps and Facebook are examples—are free, while GDP is mainly about payments.
[Deck fact; snapshot checked 2026-08-22] To address the gap, the slides introduce GDP-B and consumer-surplus measurement through choice experiments, citing Brynjolfsson et al.’s GDP-B research. A 2025 survey slide lists 40% of adults using genAI, willingness to pay of $98 per month, and $97 billion in annual consumer surplus. These are the cited study/example as presented by the deck, not a new estimate here. Willingness to pay, consumer surplus, revenue, and GDP growth are not interchangeable; the study year and deck-snapshot date must remain distinct.
[Material gap] The PDF slide does not include the full sample, question wording, uncertainty intervals, or every calculation step. I therefore use the values as a cited example of how free AI services may be missing from GDP, not as estimates that automatically generalize across countries, years, or populations.
The deck then gives three ways AI might affect the economy: make one sector super productive, act as an injection of cheap labor, and generate new ideas.
The first pathway does not mean that a highly productive sector raises GDP proportionally. Using the illumination sector as an example, the deck says prices may fall in that sector even if consumption rises, so the GDP increase may be small. It then invokes Baumol’s cost disease: other sectors match the productive sector’s wages, while “good/cheap software, expensive healthcare” describes relatively costly sectors that still dominate GDP share. [My interpretation] This is not a denial of productivity; it is a reminder to track prices, quantities, sector shares, and wages together.
The second pathway treats AI as a labor injection. The deck writes the Cobb-Douglas production function:
Y = A · K^α · L^(1−α)
Here Y is output (GDP or production), A is total factor productivity, K is capital, L is labor, and α is the capital share with 0 < α < 1. [Deck fact] If AI injects labor, the GDP curve may shift upward for a one-time change; if labor grows persistently and capital keeps up, the exponential growth rate may change. The deck also says that labor in developed economies has plateaued, citing population stagnation and labor-share stagnation. What matters is not only how much labor is added, but which tasks change.
The third pathway is AI generating ideas. The deck uses the Romer growth intuition that ideas are at the heart of growth; unlike ordinary goods, ideas are nonrival and reusable. Linear algebra is its example. If AI changes R&D or science, the most likely effect is on the exponential rate of GDP growth; the deck points to Ben Jones (2025) for more specific models. [Deck citation] This is a growth mechanism and research direction, not a verified estimate of how much AI has already accelerated science.
Together, the three pathways produce a better checklist than a single automation percentage. Did one sector become cheaper? Did the amount or scope of usable labor expand? Or did the generation, verification, and diffusion of new ideas change? Each pathway requires different assumptions about capital, prices, labor, research, and organizational complements.
Connection to the course
The first half of CS221 trains us to describe AI systems with states, actions, objectives, data, and uncertainty. Lecture 19 applies the same discipline outside the model. The model is no longer just a callable function; it is a decision node embedded in compute, data, distribution, organizational workflows, and markets. This is not a second lesson on search, learning, MDPs, games, Bayesian networks, or logic. It asks under what resource and institutional conditions those tools become usable products.
[My interpretation] The most transferable engineering artifact from this lecture is not a new Python API but a dependency map. For model selection, write five columns: compute dependency, data provenance, distribution position, organization/workflow change, and the economic metric to observe. This prevents a local benchmark from becoming the definition of system success.
Extension
Turn the deck into four exercises: draw model and organization layers; label compute, data, and distribution edges with substitutes, fixed/marginal cost, and single points of failure; record rights, freshness, labor, and compliance per data method; classify an impact hypothesis as sector productivity, labor injection, or ideas, separating GDP, consumer surplus, revenue, and cost.
Keep “what the deck says” separate from “what we want to test.” The ASML, TSMC, and Nvidia descriptions and valuations are deck snapshots. The data-acquisition table is an analytical framework. The 2025 genAI survey numbers are a cited survey example. The links among GPTs, the J-curve, Cobb-Douglas, and ideas are research-backed concepts assembled by the deck. My extension must not turn them into a complete current industry census.
Open gaps include compute cost/capacity decomposition, contract context for the $1.5B transaction, empirical distribution-channel comparison, and a reproducible macro model. These remain follow-ups, not conclusions supplied by common sense. First label a number as deck fact, deck citation, or interpretation; then ask which bottleneck and complement a market conclusion omits.
References
Loading...