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

Course Overview and Instructor Background

• The course instructor, Apoorv, will be leading the class for the next 9 weeks, and the course is designed to be no more than 3 hours a week, including class, readings, and other activities 10s.
• Apoorv's background includes starting his career at Palantir, leading engineering teams, and now leading Altimeter, an investment firm with a public and private business, and he recently became a father, which he considers the biggest investment of his life 2m6s.
• The course format will include guest speakers every class, starting from the next class, and students are welcome to join optional dinners with the speakers, but are asked not to record the speakers' conversations 4m30s.

Course Structure and Grading

• The grading for the course is 50/50, with 50% based on attendance and 50% based on an assignment that will be released at the end of the course, and the instructor encourages students to be involved and ask questions 6m20s.
• The course schedule includes a variety of speakers from impressive businesses across the stack, from semis to infrastructure, and students will have the opportunity to ask questions and learn from the speakers 8m40s.

AI Super Cycle and Industry Context

• The instructor believes that the course will provide students with a deep understanding of the AI super cycle and its implications, and he hopes that students will achieve a good understanding of the topic and be able to see the opportunities and challenges it presents 12m10s.
• The current state of AI is compared to the early days of the internet, mobile, and cloud technology, with many students likely to start or fund AI companies, and it is essential to understand where to invest and what questions to ask when evaluating AI businesses 10s.
• The biggest question in generative AI is whether the models being built are creating economic value, and this is related to the significant investments being made in building data centers and training models 2m6s.
• The ecosystem of AI is different from that of the cloud, with the AI ecosystem resembling a triangle, and possible reasons for this difference include the early stage of the AI cycle and the dominance of companies like Nvidia in the market 4m42s.
• The economic model of AI is distinct from traditional software businesses, as the incremental cost of adding new users is not close to zero due to the need for significant computational resources, such as GPUs, which can lead to lower profit margins for AI companies 8m40s.

Nvidia's Role and AI Ecosystem Dynamics

• The market share of Nvidia in the compute market is significant, and their dominance may be a factor in the differences between the AI and cloud ecosystems, with the possibility of Nvidia's monopoly allowing them to charge high prices for their products 10m20s.
• The readings provided include charts on the cloud ecosystem and the market share of Nvidia, and the shape of the AI ecosystem is expected to change as the technology continues to evolve 6m15s.
• The current state of the AI industry is still in its early stages, with Nvidia being a dominant force, and the physics of the problem are very different from how inference is run, which is where the industry is right now 10s.

Historical Comparisons and Industry Evolution

• Analyzing the history of internet, mobile, and cloud technologies can provide insight into the development of the AI industry, such as the example of AWS, which started in 2004, had its first customer in 2010, and Amazon shifted fully to AWS in 2012, taking eight years from the initial investment 1m30s.
• The course will explore the central theme of the AI industry, including the dominance of companies like Nvidia, the profitability of companies like Anthropic and OpenAI, and the competitiveness of the inference layer, with questions about pricing compression, ASICs, and revenue sources 3m20s.
• The inference layer is the most competitive part of the ecosystem, with many startups doing well, but also facing competition from hyperscalers like AWS, and the question of whether these startups are features or platforms 5m10s.

Course Speakers and Industry Themes

• Incumbent platforms like Salesforce and Palantir are reinventing themselves with AI-powered products, such as Einstein and AIP, and should be included in the analysis of the industry's pyramid structure, with their spend captured in the app layer by way of the substrate 7m40s.
• The course will have speakers from various companies, including Nvidia, Anthropic, and OpenAI, to discuss topics like dominance, profitability, and pricing compression, and to explore the forces that are driving the AI industry forward 2m50s.
• The build-out of the semis layer typically occurs over a 5-year or 6-year period, but application revenue is generated immediately, which can create a timing mismatch and lead to cyclical phases of capex cycles, similar to what happened in the mobile super cycle 10s.

AI Industry Cycles and Market Dynamics

• A chart in the readings illustrates what happened in the mobile super cycle, where the first inning had inflated market caps for capex-heavy businesses, and a similar phenomenon is expected to occur in the current AI supercycle 1m42s.
• Google is a large conglomerate with multiple business units, including the TPU business unit in the semis layer, the GCP unit in the infrastructure layer, and the Gemini unit in the apps layer, with Gemini being one of the most used consumer applications 2m6s.
• The question of whether the AI supercycle will be successful or not is unlikely to be a fad, and the stable equilibrium of the industry is still being debated, with the possibility that it might stay in its current shape for longer than anticipated, potentially for about a decade 4m30s.
• The stable equilibrium of the chart could be influenced by unlocks, such as a breakthrough in ASIC programs at hyperscalers like Google's TPU or Meta's MTIA, which could lead to a repricing of the layer, or changes in hyperscaler capex guidance and earnings calls 6m15s.
• Another factor to consider is the element of training versus inference, where the current equilibrium might flip if inference becomes larger than training, which could be a key factor in determining the success of the AI supercycle 8m30s.

Hyperscalers and Training vs. Inference

• The hyperscalers, such as Google and AWS, may stop spending on training due to not seeing the desired performance, and instead focus on inference, with Nvidia's share of inference in their fleet being around 40% 10s.
• The training workload is very predictable with high utilization for a short period, whereas the inference workload is bursty and harder to predict, with usage typically being high when humans are awake 2m6s.
• The most profitable part of the stack is the semis layer, with Nvidia's data center revenues earning the most margin of around 75%, while the application layer revenues are estimated to be between 0 and 30% 4m6s.

Profitability and Revenue Distribution in AI Stack

• The machine learning industry is still in its early stages, with some comparing it to being only 10 years into a long-term development cycle, and the investment cycle in cloud is a key element of this 6m6s.
• The ASIC infra startups will likely sell to the big hyperscalers, such as Google, AWS, and Nvidia, which account for around half of the $300 billion revenue, with the customer base being a small number of very large orders 8m6s.
• When starting a chip company, the primary consideration should be which of the major hyperscalers to sell to first, as they will be the main customers, and the long tail of other enterprises may not be a reliable source of revenue 10m6s.
• It takes multiple years for the revenue and earnings in each layer to play out, and the industry may be moving towards a fully vertically integrated model, with no single company having achieved this yet 12m6s.

Industry Leaders and Market Structures

• The biggest winners of previous internet super cycles include Google in the internet space with a market cap of around $3 trillion and near 99% market share in search, Apple in the mobile space with a market cap of around $2.5 trillion, and Meta in the social space with a market cap of around $2 trillion, although Meta is not as fully integrated as Google 10s.
• The cloud space is fairly heterogeneous with no single dominant player, but rather a mix of oligopolies including AWS, GCP, and Azure, and companies like Nvidia are trying to build their own cloud ecosystems, such as DGX Cloud 2m6s.

Course Engagement and Competitive Elements

• A quiz is being conducted with a prize for the winner, who will be determined by being both right and fast in their answers, with the option to use cloud computing to aid in their responses, but human players will be given a 5-second head start 5m42s.

Value Distribution and Competitive Layers

• The value in the AI space is accruing in a particular manner, with around 75% of the $350 billion in revenue added in the last 2 years going to semiconductors, primarily Nvidia, and the apps segment is dominated by two companies that make up around 90% of it 10m0s.
• The infra segment is the most competitive part of the AI ecosystem, with the biggest battle brewing both sideways and across the stack, and companies are trying to gain an advantage through various means, including building their own cloud ecosystems and developing new technologies 12m10s.
• The current state of the AI industry is characterized by a high metabolic rate, with many companies being formed and acquired, making it a competitive and unstable environment 10s.

Future Projections and Industry Maturity

• The question of how long it will take for the AI industry to reach a state similar to cloud software is being considered, with possibilities ranging from 5 to 15 years or never 1m42s.
• Consumer AI, the largest market for AI outside of coding, has high usage, with around 95% of ChatGPT users being free, and companies like DeepMind are exploring alternative revenue models, such as subscriptions 2m6s.

Consumer AI and Monetization Models

• The monetization engine of AI businesses is being explored, with considerations of whether subscription or ad-based models will be more successful, and comparisons are being made to large consumer franchises outside of AI 2m6s.
• Consumer products can be categorized into three groups: mandatory products with 3 billion users, social products with 1.5-2 billion users, and niche products, and the question is being asked where ChatGPT and Gemini will fall in these categories 3m30s.
• ChatGPT has just overtaken the niche app category, and Gemini has not yet reached this level, with considerations of whether these AI applications will reach the scale of mandatory or social apps 6m30s.
• The potential for knowledge work to become a universal activity is being questioned, with ChatGPT currently being a place where users must actively engage in work, such as asking questions, rather than a platform for passive activities like messaging or email 8m40s.
• The number of people in the world who are asking active questions of technology is not the entirety of the population that's online, with about 4 billion people out of 8 billion on the planet being online 10s.

Consumer App Economics and Market Positioning

• The economics of consumer applications show that Alphabet has about 4 billion users who are monetized at about $100 a user a year, while Meta has about 3 and 1/2 billion users who are monetized at about $70 a user a year, and the leading AI provider ChatGPT has about a billion users who are monetized at about $10 a user a year 42s.
• To increase the number of users and monetization, it's necessary to go beyond knowledge work and find new ways to serve ads, as the current models may not be sufficient, with the potential for ChatGPT and cloud to serve ads with better pricing due to understanding user intent and having good attribution and trust 1m30s.
• The ad model is expected to be a big headline this year, with a lot of potential for growth and alpha in understanding the ad model really well, similar to the situation 10 years ago at the Facebook IPO when people doubted the effectiveness of ads on phones 2m6s.
• The current debate is about finding a way to serve ads that are not shocking or intrusive, with the goal of finding a big unlock for the economic model, which will be explored in more detail in later sessions 3m30s.