Stanford MS&E435 Economics of the AI Supercycle: Building AI Factories

Stanford MS&E435 Economics of the AI Supercycle: Building AI Factories

AI Infrastructure as Digital Labor

The current surge in AI adoption is manifesting physically as a massive investment in data centers, which Chase Lochmiller, CEO of Crusoe, describes as the "infrastructure of intelligence." This investment is one of the largest in history, comparable in scale to the Manhattan Project or the US highway system, and is second only to the US Defense budget.

From an economic perspective, the value of AI tokens lies in their ability to provide "digital labor." Historically, labor growth was limited by birth rates and long incubation periods (education and upbringing). AI allows for the digital acceleration of labor (delta L in the Cobb-Douglas production model), which can fundamentally transform the economy and accelerate GDP growth by providing an unprecedented increase in the available workforce through agents and bots.

The Anatomy of an AI Factory

Building a gigawatt-scale AI data center is a multidisciplinary engineering challenge involving chemical, mechanical, electrical, and computer science engineering. Lochmiller defines a data center at its most basic level as a building with power and cooling where computers can be plugged in, but at scale, it becomes a complex "AI factory."

The Energy-First Approach

Crusoe employs a vertically integrated strategy, prioritizing energy development to unblock infrastructure growth. This involves locating data centers in areas with abundant, low-cost energy resources rather than traditional data center hubs like Northern Virginia.

  • Abilene, Texas Site: Crusoe leverages West Texas's overinvestment in renewable energy. Due to production tax credits and limited transmission, power prices in Abilene were sometimes negative. Crusoe built a campus here featuring a 200 megawatt substation and a one gigawatt substation (the largest privately owned substation in the US), alongside a 350 megawatt natural gas power plant to ensure reliability.
  • Claude, Texas Site: This site focuses on wind energy, utilizing an "across the meter" strategy. Crusoe generates power on-site via wind (with plans for solar, batteries, and gas) and sells excess power back to the grid, while drawing from the grid during maintenance or low renewable production.

Cooling and Water Usage

Contrary to narratives that AI consumes excessive water, Lochmiller states that Crusoe's systems use recirculating chilled water loops. Once filled, the annual water consumption is comparable to that of a single-family home, making it sustainable for water-scarce regions like West Texas.

The Economics of AI Infrastructure

Standing up a gigawatt-scale compute cluster requires an estimated total capital expenditure (CapEx) of approximately $60 billion.

Infrastructure CapEx (Approx. $20 million per megawatt)

  • Labor: A significant bottleneck and cost, estimated at $4.7 million per megawatt. This includes a massive blue-collar workforce of electricians, welders, and plumbers.
  • Power Generation: Gas plants cost roughly $2-3 million per megawatt, with prices rising due to limited turbine manufacturer capacity (e.g., GE Vernova, Siemens).
  • Electrical Equipment: Includes power distribution centers, transformers, and switchgear to step down high-voltage power (e.g., 345 kV) to usable levels.
  • Mechanical Equipment: Includes air-cooled chillers and plumbing for recirculating water.

IT CapEx (Approx. $40 million per megawatt)

  • GPUs: The largest cost, accounting for $30 million per megawatt.
  • Networking: High-performance back-end networks (NVLink, InfiniBand, or RoCE) cost roughly $4 million per megawatt to interconnect GPUs into a coherent cluster.
  • CPUs and Storage: Approximately $3 million per megawatt. Lochmiller notes a recent shortage of CPUs driven by the rise of agentic workflows.
  • Other: Labor for deployment and shipping accounts for $1 million per megawatt.

Revenue, Payback, and Depreciation

For a gigawatt cluster costing $60 billion, the revenue model depends on the service layer provided:

  1. Chip Rental (Infrastructure as a Service): Generates roughly $15 million per megawatt annually. This results in a revenue-based payback period of approximately four years.
  2. Managed Services (Model as a Service): By hosting models and providing API endpoints for tokens, margins improve by an additional $5-15 million per megawatt, potentially reducing the payback period to two years.

The Depreciation Debate

While many public companies depreciate compute over five or six years, Lochmiller argues that the value of older chips (like the H100) can actually increase during demand spikes (e.g., the agentic AI boom), suggesting that the useful life of compute is tied to its value to the user rather than a fixed calendar date.

Future Outlook and Bottlenecks

The Electrical Stack Opportunity

Lochmiller identifies a major opportunity for innovation in the electrical stack. He suggests that traditional companies (e.g., Eaton, Schneider) may be at risk if they do not innovate toward solid-state electronics, solid-state transformers, and 900-volt DC architectures to move power more efficiently from high-voltage substations to the rack.

Space-Based Data Centers

Crusoe has partnered with Starcloud to launch H100s into space. While space data centers eliminate the need for concrete foundations, permitting, and fiber cabling (using optics instead), they face extreme challenges in thermal management and hardware maintenance, as GPUs cannot be easily replaced or reseated. Lochmiller views this as a long-term play (10+ years) dependent on the reduction of payload costs via vehicles like SpaceX's Starship.

Sources