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

Key Concepts

• Data Center Economy: The physical infrastructure (power, cooling, buildings) required to support AI compute.
• Digital Labor: The ability to use AI agents to perform tasks, effectively scaling the labor force digitally rather than through traditional population growth.
• Vertically Integrated Infrastructure: A business model (Crusoe) that controls multiple layers of the stack—from energy generation to data center construction and managed compute services.
• Across-the-Meter Energy: A strategy of co-locating data centers with on-site power generation (wind, solar, gas) to create energy abundance and grid stability.
• Compute Bottlenecks: The shifting constraints in AI growth, moving from chip availability to power, cooling, and skilled labor.
• Cobb-Douglas Model: An economic production function used to explain how AI-driven technology increases labor productivity and GDP growth.

---

1. The Data Center Economy and AI Infrastructure

Chase Lochmiller, CEO of Crusoe, defines the current AI boom as a physical infrastructure challenge. The massive CapEx spend by hyperscalers is driven by the need to create "digital labor." Unlike traditional labor, which requires a 20-year incubation period, digital labor can be scaled instantly through the deployment of GPUs and data centers.

• The AI Equation: AI = Data + Algorithms + Compute (GPUs) + Energy + Data Centers.
• The Economic Premise: By accelerating the growth of digital labor, companies are fundamentally up-leveling GDP and economic productivity.

2. The "Energy-First" Framework

Crusoe’s strategy is to work backward from energy. Instead of building in saturated hubs like Northern Virginia, they target areas with abundant, low-cost, stranded energy.

• Case Study: Abilene, Texas:
  • The Opportunity: West Texas had an over-investment in renewable energy (wind/solar) with insufficient transmission capacity, leading to negative power prices.
  • The Solution: Crusoe built a 2.1-gigawatt campus (the size of two Denvers) to consume this stranded power.
  • Infrastructure: The site includes a 350MW natural gas plant for firming power and a 1GW substation—the largest privately owned substation in the U.S.
  • Labor Impact: The project employs 9,000 construction workers, significantly impacting the local economy of 120,000 people.

3. Cost Breakdown and Technical Specifications

Lochmiller provides a detailed breakdown of the capital expenditure (CapEx) required to build modern AI infrastructure, normalized per megawatt (MW).

• Data Center & Power Plant: ~$20 million per MW.
  • Key Components: Power distribution centers (stepping down 345kV to 480V/415V), air-cooled chillers, and massive plumbing systems.
  • Water Usage: Contrary to popular belief, these systems are closed-loop; once filled, they consume minimal water annually.
• IT CapEx (Compute): ~$40 million per MW.
  • GPU Spend: ~$30 million/MW.
  • Networking: ~$4 million/MW (InfiniBand/RDMA over Converged Ethernet).
  • CPUs/Storage: ~$3 million/MW (essential for orchestrating agentic workflows).
• Total Investment: Approximately $60 million per MW.

4. Revenue and Payback Periods

• Revenue Model: Renting out compute capacity generates roughly $15 million/MW annually.
• Payback Period: Pure infrastructure rental yields a ~4-year payback. By adding a Managed Services layer (hosting models and serving tokens), revenue can double to $30 million/MW, shortening the payback period to ~2 years.
• Depreciation: While standard accounting uses 5–6 years, the high demand for H100s and Blackwell chips suggests these assets may remain valuable longer than traditional hardware.

5. Future Outlook and Industry Perspectives

• Bottlenecks: The current primary bottleneck is "powered shells"—finding locations with sufficient power and cooling to plug in chips.
• Innovation in Electrical Stack: Lochmiller argues that traditional electrical equipment companies (e.g., Eaton, Schneider) are currently critical but face long-term disruption if they do not innovate in power electronics (e.g., 900V DC architectures).
• Space Data Centers: While technically feasible, space-based data centers face significant hurdles regarding thermal management and the inability to perform physical maintenance (reseating GPUs). He views this as a 10+ year horizon.
• Advice to Students: Focus on the "how" of learning rather than the "what." Cultivate grit and the ability to leverage AI tools to maintain an "infinite growth loop" in personal development.

Synthesis

The AI infrastructure boom is a massive, multi-disciplinary engineering challenge that requires the integration of power generation, mechanical cooling, and high-performance networking. The shift toward "energy-first" development and vertically integrated infrastructure is essential to overcoming the physical constraints of the grid. As AI agents become more prevalent, the demand for compute will continue to drive innovation, potentially disrupting legacy electrical industries and creating new economic models for digital labor.