Key Moments
Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Economics of Generative AI
Stanford OnlineWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
NVIDIA's data center GPUs earn 75% gross margins, while AI applications struggle with profitability, often between 0-30%, highlighting the concentration of value in the semiconductor layer.
Key Insights
The AI ecosystem's economic structure, visualized as an inverted triangle, is dramatically different from cloud, internet, and mobile due to the high cost of running inference on GPUs for each user, unlike the near-zero marginal cost of traditional software.
NVIDIA's data center revenue boasts a gross margin of approximately 75%, significantly higher than the estimated 0-30% for application layer revenues in the generative AI space.
While consumer AI applications like ChatGPT have reached approximately 1 billion users, they monetize at roughly $10 per user per year, a fraction of Alphabet's ~$100/user/year or Meta's ~$70/user/year from established consumer franchises.
The infrastructure layer, often seen as the most competitive and unstable part of the AI ecosystem, is where a significant battle is brewing both horizontally and vertically across the stack.
Currently, about 60% of NVIDIA's GPUs are used for training and 40% for inference, a ratio expected to shift in favor of inference over time, though the exact timing is uncertain.
The potential for AI applications to significantly increase monetization beyond current levels (e.g., from $10/user/year to $100/user/year) may lie in more advanced ad models that leverage user intent and trust, similar to how mobile ads evolved.
The AI supercycle and course overview
The course explores the current generative AI supercycle, aiming to equip students with mental models to understand and navigate this rapidly evolving landscape. Instructor Apoorv Agrawal, a partner at Altimeter Capital, emphasizes that this is likely the biggest technology supercycle yet, comparable to the internet, mobile, and cloud eras. The course structure involves guest speakers from across the AI stack – from semiconductors to applications – with a focus on business economics. Grading is a simple 50/50 split between class attendance and a final assignment. Agrawal stresses the importance of being present and asking hard questions to speakers from companies like OpenAI and Anthropic, with the goal of understanding the fundamental 'laws of physics' governing AI businesses.
The inverted triangle: AI's unique economic structure
A central theme is the economic shape of the AI ecosystem, depicted as an inverted triangle, which starkly contrasts with previous technology revolutions like cloud, internet, and mobile. In traditional software and cloud models, software could be built and distributed at a near-zero marginal cost, leading to high gross margins. However, the AI model is different because the incremental cost of an AI user is significant due to the need to burn GPUs for inference. This has led to companies with billions in revenue still struggling with profitability. The chart shows a much smaller application layer compared to the massive investment in foundational layers like energy, chips, power, interconnects, and memory required for data centers that rent out compute power. Unlike previous cycles where the application layer dominated, AI's value is currently heavily concentrated in the infrastructure and semiconductor layers.
Where the money is: The dominance of the semiconductor layer
The most striking economic disparity lies in profitability. NVIDIA's data center revenue, for instance, commands an estimated gross margin of around 75%. This is a substantial lead compared to the application layer, where revenues are estimated to be between 0% and 30% gross margin, depending on the company. This concentration of value is attributed to a few dominant players, particularly NVIDIA, who effectively 'run the tables' on the semiconductor layer. This profitability gap means that from a financial perspective, the inverted triangle is even more pronounced. While the total AI ecosystem revenue has grown significantly, with approximately $350 billion added in the last two years, a disproportionate 75% of that growth has flowed directly to the semiconductor sector.
Challenges in the application and inference layers
The application layer in generative AI faces significant economic hurdles. While platforms like ChatGPT have achieved massive user adoption, reaching about 1 billion users, their monetization is currently low, averaging around $10 per user per year. This is a stark contrast to established consumer internet giants like Alphabet (Google) and Meta, which monetize their user bases at significantly higher rates ($100/year and $70/year, respectively). The core issue is that unlike traditional software, each AI user engagement requires substantial computational resources, making scalability and profitability challenging. The instructor posits that to bridge this monetization gap, AI applications might need to move beyond solely 'knowledge work' and explore new revenue streams, potentially including more advanced advertising models.
Competition and instability in the infrastructure layer
The infrastructure layer, which includes cloud providers and related services, is characterized by intense competition and high instability. This segment sees a rapid formation and acquisition of companies, making it a dynamic area. A key question for startups in this space is whether they are building a feature that could be absorbed by a larger platform like AWS or a distinct platform in its own right. The hyperscalers (AWS, Google Cloud, Azure) aim for dominance, complicating the landscape for independent infrastructure providers. Furthermore, the cyclical nature of capital expenditure (capex) in this layer, driven by the need to build out compute capacity for future revenue, adds another layer of complexity, similar to the 'laying down the railroads' phase seen in earlier tech cycles.
The future of AI monetization: Advertising's potential role
The discussion turns to how AI applications might increase their monetization, particularly aiming to scale from $10/user/year to $100/user/year. While subscriptions are a possibility, the instructor suggests that advertising could be a more significant revenue driver. Unlike current free-tier models, future AI advertisements could be highly effective due to a deeper understanding of user intent, seamless integration, and greater trust. This is analogous to how mobile advertising navigated early skepticism about fitting ads into a smaller screen, eventually finding effective formats. The ability to leverage logged-in user data for precise targeting and attribution is seen as a major unlock for the AI advertising model, potentially offering substantial 'alpha' for those who understand it best.
Vertical integration and market dynamics
The conversation touches upon vertical integration within the AI ecosystem. While companies like Google are highly integrated, running their own TPUs for their AI models and cloud services, the market may see a few winners emerge at each layer. Historically, dominant players in supercycles have sometimes been highly integrated (e.g., Google in the internet, Apple in mobile), but this hasn't always been the case (e.g., cloud heterogeneity). NVIDIA is also exploring deeper integration with efforts like DGX Cloud. For startups focusing on custom silicon (ASICs), the primary customers are likely to be the large hyperscalers, representing a small number of very large orders, a different customer acquisition dynamic compared to consumer or enterprise software.
The training vs. inference balance and long-term equilibrium
A critical factor in the AI economy is the balance between training and inference workloads. Currently, NVIDIA reports that roughly 40% of its GPU usage is for inference and 60% for training. This ratio is expected to shift towards inference over time, which would significantly impact the market. The instructor acknowledges that while AI is unlikely to be a fad, the stable equilibrium of the industry — how the inverted triangle might eventually evolve — is still uncertain. It could remain in its current form for longer than anticipated, or a breakthrough in ASIC development by a hyperscaler could lead to significant repricing of the semiconductor layer. Another indicator of potential shifts would be changes in hyperscaler capex guidance, signaling adjustments to the current equilibrium.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●People Referenced
AI Ecosystem Value Distribution vs. Cloud
Data extracted from this episode
| Layer | AI Ecosystem (Current) | Cloud Ecosystem (Mature) |
|---|---|---|
| Semis | Dominant Value | Smaller Value |
| Infrastructure | Competitive Intensity | Established Players (AWS, GCP, Azure) |
| Applications/Models | Smaller Value | Dominant Value |
Consumer Application User Scale and Monetization
Data extracted from this episode
| Category | Example Apps | User Scale (Billions) | Monetization (USD/User/Year Approx.) |
|---|---|---|---|
| Mandatory Utility | WhatsApp, Chrome, YouTube | 3+ | ~$100 |
| Social | Instagram, TikTok, Facebook | 1.5-2 | ~$70 |
| Niche | Amazon, Spotify, Twitter | N/A (Smaller Scale) | N/A (Variable) |
| AI (Current) | ChatGPT | ~1 | ~$10 |
| AI (Current) | Gemini | Has not overtaken niche | Not specified |
Common Questions
The Stanford MS&E435 course focuses on the economics of the AI supercycle, particularly generative AI. It aims to provide students with a deep understanding of the economic forces, business models, and investment opportunities within the rapidly evolving AI landscape.
Topics
Mentioned in this video
The instructor, Apur, started his career at this company 13-14 years ago, leading engineering teams and working with technologies like Spark.
The investment firm where the instructor, Apur, currently leads. It focuses on concentrated investing with public and private businesses.
A company whose representatives are expected to speak in the course, discussed in the context of AI models, profitability, and serving a billion-user franchise.
A key player in the AI supercycle, discussed for its dominant market share in compute, its role in the semis layer, and its data center revenues and gross margins.
Mentioned as the first customer of AWS in 2010, illustrating the timeline of cloud adoption.
The parent company of AWS, discussed in the context of its buildout and the financial debates surrounding it during the early days of AWS.
An incumbent platform mentioned as an example of an 'old economy' business reinventing itself with AI products like Einstein, whose AI spending is captured in the app layer.
Discussed in multiple contexts: its TPU business unit in semis, its GCP unit in infrastructure, and its Gemini unit in apps. Also highlighted as a major winner of the internet supercycle.
Identified as a major winner in the social media supercycle, with a significant market cap, but noted as not being as fully integrated as Apple or Google.
Considered the winner of the mobile supercycle with a high market cap and significant integration.
Mentioned as a tool students might spend time discussing or arguing with regarding assignments, and as a significant consumer AI application with a large user base.
Mentioned alongside ChatGPT as a tool students might argue with regarding assignments.
Amazon Web Services is used as a historical example of a cloud ecosystem that took 8 years from initial investment to full transition, and is discussed as a potential home for infrastructure-layer businesses.
AI product offered by Salesforce, mentioned as an example of how incumbent platforms are integrating AI.
Google Cloud Platform, considered part of the infrastructure layer in Google's business units.
Google's AI model, classified as part of the apps layer and discussed as a highly used consumer application, with questions about its success relative to distribution advantage.
Microsoft's cloud platform, mentioned as one of the three major players in the cloud market.
More from Stanford Online
View all 88 summaries
50 minStanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Applications, AI in Life Sciences
48 minStanford CS547 HCI Seminar | Spring 2026 | Toward Ontological Multiplicity in AI and Computing
40 minStanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Infrasctructure, Enterprise AI, SaaS
48 minStanford CS547 HCI Seminar | Spring 2026 | Just-in-Time Objectives for Specialized AI Interactions
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free