Key Moments
Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Building AI Factories
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
AI data centers are requiring unprecedented capital investment, rivaling national defense budgets, primarily driven by the need for vast amounts of energy and specialized infrastructure.
Key Insights
Hyperscalers are investing $650 billion in AI data centers, a sum comparable to national defense budgets and larger than investments in space exploration.
The core components required to produce AI are data, algorithms, compute (GPUs), energy, and data centers, with compute, energy, and data centers representing the most significant capital expenditures.
The development of AI infrastructure, exemplified by Crusoe's Abilene, Texas campus, involves constructing colossal data centers and power facilities, with a single gigawatt substation capable of powering an entire city like Denver.
The construction of a gigawatt AI data center campus can cost approximately $60 million per megawatt capitalized, with labor alone accounting for roughly $4.7 million per megawatt.
While GPU costs dominate IT CapEx at around $30 million per megawatt, networking, CPUs, storage, and in-room equipment also represent substantial investments, totaling about $40 million per megawatt for IT infrastructure.
Serving AI models directly as a managed service, rather than just renting out compute, can significantly increase revenue per megawatt, potentially shortening payback periods from four years to as little as two.
The unprecedented scale of AI infrastructure investment
The current AI supercycle is characterized by massive capital expenditure, with the top five hyperscalers investing billions into AI capabilities. This investment in data centers is so significant that it rivals, and in some cases surpasses, major national and global initiatives. For context, the CapEx spend is described as being 'bigger than space, bigger than our highway system, our Manhattan Project, second only to the US defense budget.' This staggering investment underscores the foundational role of physical infrastructure in the burgeoning AI economy. The talk highlights that while data and algorithms are crucial, the real bottleneck and largest cost lie in the compute, energy, and data center infrastructure required to power and house these advanced AI systems. This focus on tangible infrastructure is what companies like Crusoe are addressing.
Deconstructing the components of AI production
Producing AI, at its core, requires a specific combination of elements: data, algorithms, compute, energy, and data centers. Data is essential for training models, with companies like Scale AI and Merkur emerging to label and prepare vast datasets. Algorithms, developed in research labs, enable models to learn and process information. Compute, particularly high-performance computing through GPUs, is vital for parallel processing and tensor computations. Energy is the lifeblood of these GPUs, and data centers provide the physical environment to house and operate this complex infrastructure. While data and algorithms are key, the infrastructure aspects—compute, energy, and data centers—are where the substantial financial investments and operational challenges lie, forming the core focus for companies building AI capacity.
Creating digital labor through accelerated infrastructure growth
The current AI boom represents a paradigm shift in economic growth, akin to the Cobb-Douglas model which links GDP growth to labor, capital, and technology. For the first time in history, AI enables the creation of 'digital labor.' Unlike traditional human labor, which has a 20-year lead time and is constrained by birth rates, digital labor can be scaled rapidly through investment in data centers and GPUs. This acceleration of digital labor provides an unprecedented opportunity to boost GDP growth and improve quality of life by making labor more productive. The massive capital expenditures seen in the AI supercycle are driven by this fundamental ability to expand the digital workforce at an accelerated pace, transforming economic potential.
Crusoe's vertically integrated approach to AI infrastructure
Crusoe operates as a vertically integrated AI infrastructure business, aiming to unblock any impediments to building the necessary intelligence infrastructure. This strategy involves two main phases: energy development and data center construction. Recognizing the immense energy demands of AI, Crusoe prioritizes locating facilities in areas with abundant, low-cost energy resources. The data center itself is viewed as a building that requires power and cooling, but at scale, it becomes an intricate engineering marvel combining various disciplines—chemical, mechanical, electrical, and computer engineering. This holistic approach allows Crusoe to manage bottlenecks as they shift across the AI stack, from power and cooling to compute deployment.
The evolving bottleneck: From compute to energized data centers
Historically, bottlenecks in AI infrastructure have shifted. Initially, it was compute, then power memory, and more recently, labor. Today, the primary bottleneck is the availability of 'energized data centers'—that is, power shells where GPU clusters can be plugged in and operated. While access to chips has improved, finding locations with sufficient power and infrastructure remains the critical constraint. Crusoe's vertically integrated model is designed to navigate these moving bottlenecks. By controlling aspects of energy development, data center construction, and compute deployment, the company can adapt to the changing landscape of critical resources, ensuring a more robust supply chain for AI power.
Leveraging energy abundance in West Texas: The Abilene campus
Crusoe's strategy of an 'energy-first' approach led them to markets outside traditional data center hubs. Abilene, Texas, exemplifies this, chosen for its abundant wind and solar resources. Renewable energy developers had over-invested in the region, leading to negative power prices due to insufficient transmission infrastructure. Crusoe capitalized on this by building a massive AI computing campus, including a 1-gigawatt substation—the largest privately owned in the U.S., capable of powering a city the size of Denver. This site also features a 350 MW natural gas power plant to ensure stable energy supply for interconnected GPU clusters. The campus is a testament to turning an energy surplus into a compute advantage, with initial phase tenants including Oracle and OpenAI.
The immense cost and labor demands of building AI campuses
Constructing AI data center campuses involves staggering costs and intensive labor. The Abilene campus, for instance, requires a 5,000-car parking lot to accommodate approximately 9,000 on-site workers daily during construction. The total infrastructure for a gigawatt campus is estimated at around $20 billion, with labor alone contributing a significant portion, estimated at $4.7 million per megawatt capitalized. This translates to billions in wages for electricians, welders, plumbers, and construction workers, creating a substantial job market in smaller towns like Abilene (population ~120,000). Crusoe's vertically integrated approach, including on-site concrete batch plants and modular construction (Crusoe Spark), aims to mitigate these labor challenges and reduce costs by 30-50%.
Dissecting the economics of compute infrastructure
The IT CapEx for AI compute infrastructure is substantial, estimated at $40 million per megawatt. The largest component by far is GPUs, accounting for roughly $30 million per megawatt, highlighting NVIDIA's dominant position. Networking, essential for interconnecting GPUs in large clusters (e.g., using NVLink and InfiniBand), adds about $4 million per megawatt. CPUs and storage represent another $3 million per megawatt, with a recent shortage of CPUs noted due to the rise of agentic AI workflows. Other costs, including in-room equipment and shipping, bring the total IT infrastructure cost to approximately $40 million per megawatt. Combined with power plant and data center costs, the total upfront CapEx can reach about $60 million per megawatt for a gigawatt cluster.
Revenue, depreciation, and the path to profitability
With a projected CapEx of $60 million per megawatt and ongoing OPEX of approximately $1-2 million per megawatt annually, profitability hinges on revenue generation. Renting out compute power can yield roughly $15 million per megawatt annually, suggesting a potential payback period of around four years based on revenue alone. However, the true economic viability is influenced by depreciation, particularly the useful life of hardware like GPUs. Recent trends show that demand from agentic AI is increasing the value of existing hardware (like H100s), potentially extending its useful life beyond traditional depreciation schedules of 5-6 years. By offering managed services—hosting and serving AI models directly—Crusoe can enhance revenue per megawatt, potentially reaching $30 million and reducing payback periods to as little as two years, significantly improving the business model.
Innovations in electrical infrastructure and future trends
The expansion of AI drives significant innovation in the electrical infrastructure sector. Traditional components and companies in power distribution and transformation (e.g., Eaton, Schneider Electric) are well-positioned for near-term growth but may face long-term disruption. The demand for high-efficiency power delivery to racks (e.g., transitioning to 900V DC) and advancements in solid-state electronics and power transformers present opportunities for new innovators. Electrical engineers are encouraged to focus on power electronics and efficient voltage conversion. While data centers in space are a long-term prospect, facing challenges in thermals and maintenance, their potential to avoid terrestrial site preparation and permitting costs is significant, though likely not material for at least 5-10 years. The broader trend favors open-source solutions, with companies focused on developing advanced, modular data center solutions like Crusoe Spark to reduce costs and labor dependency.
Advice for students: Learning, growth, and leveraging AI
For students navigating their academic and career paths, the advice is to focus less on the exact subjects learned and more on the process of learning and continuous improvement. Embracing a philosophy of 'infinite growth loop,' where individuals are always a work in progress and strive to get better daily, leads to exponential compounding of skills and knowledge—the most valuable asset. The rapid adoption of AI will fundamentally change the workforce, providing individuals with access to capabilities equivalent to a million people. Therefore, students should focus on the 'how' of learning and working, leveraging AI tools to enhance their productivity and adapt to this evolving landscape, rather than solely on the 'what' of their studies.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Concepts
●People Referenced
AI Infrastructure Cost Breakdown (Per Megawatt)
Data extracted from this episode
| Component | Estimated Cost (USD) | Notes |
|---|---|---|
| Data Center & Power Plant | $20 Million | Includes power plant and building costs. |
| IT CapEx (Compute Infrastructure) | $40 Million | Includes GPUs, networking, CPUs, storage. |
| Total Upfront CapEx (per MW) | $60 Million | Sum of Data Center/Power Plant and IT CapEx. |
| Labor (Capitalized CapEx) | $4.7 Million per MW per year (during construction) | Represents investment in construction workforce. |
| Gas Turbine Manufacturing Cost | $1 Million (formerly) to $3 Million (current) | Significant price increase due to demand. |
Crusoe Cloud Revenue vs. Cost (Per Megawatt per Year)
Data extracted from this episode
| Category | Estimated Value (USD) | Notes |
|---|---|---|
| Annualized Revenue (Renting Chips) | $15 Million | Based on renting infrastructure like H100. |
| Ongoing OPEX | $1-2 Million | Includes power, insurance, on-site labor for repairs. |
| Optimistic Revenue (Managed Services) | $30 Million | Includes managed services for serving models and APIs. |
| Payback Period (Optimistic Case) | 2 Years | With managed services, significantly improving return on investment. |
Common Questions
AI production fundamentally requires data, algorithms (like backpropagation and neural networks), compute power (especially high-performance GPUs), energy to run the GPUs, and data centers to house the infrastructure.
Topics
Mentioned in this video
A company mentioned as an example of a data labeling business that has benefited from the AI boom by producing useful data.
The initial major project at the Abilene AI campus, involving Oracle and OpenAI utilizing Crusoe's infrastructure.
Has a company called Solar that manufactures gas turbines, mentioned in the context of the limited expansion of production capacity.
A manufacturer of gas turbines mentioned in the context of companies that could be impacted by innovation in the electrical stack.
Mentioned as a manufacturer of gas turbines.
A company in the electrical stack that could be challenged by new innovations if they do not adapt.
An AI infrastructure company focused on building and operating data centers, with a strategy centered around abundant, low-cost energy.
A manufacturer of gas turbines whose stock has performed well due to increased demand for power generation infrastructure.
A partner with Oracle and Crusoe on Project Stargate, operating out of the Abilene AI campus.
A tenant at the Abilene, Texas AI computing campus, collaborating with OpenAI on Project Stargate.
Mentioned as a data labeling company that benefits from the AI data demand.
A company in the electrical stack that could be challenged by new innovations if they do not adapt.
A partner of Crusoe that has launched the first H100 GPUs into space, exploring the feasibility of space-based data centers.
A manufacturer of gas turbines mentioned in the context of companies potentially impacted by innovations in the electrical stack.
An AI model mentioned as an example of technology that relies on data centers for operation.
An AI model mentioned as an example of technology that relies on data centers for operation and AI workloads.
A large language model whose operation and access to tokens are enabled by data center infrastructure.
More from Stanford Online
View all 79 summaries
42 minStanford CS153 Frontier Systems | Scale, AGI, and the Future of Everything
39 minAI in Healthcare Series: Inside the Rise of AI in Healthcare, Open Evidence and Cyber Risks
72 minStanford CS336 Language Modeling from Scratch | Spring 2026 | Guest Lecture: Dan Fu
60 minStanford CS547 HCI Seminar | Spring 2026 | The Modern Motivators of Play
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