Keynote, Lecture, Conference Presentation, Fireside Chat
Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Building AI Factories
- AI Infrastructure Investment Scale: Hyperscalers' combined AI CapEx is projected to surpass the U.S. Highway System and the Manhattan Project, ranking second only to the U.S. defense budget.
- Core Economic Thesis: Chase Lock Miller defines the current boom as the creation of "digital labor," allowing the economy to accelerate GDP growth without the 20-year biological lag associated with increasing human labor supply.
- Bottleneck Shift: The primary constraint has evolved from chip availability to securing "powered shells" (energized data centers) and skilled labor; chip supply has softened, while access to reliable power and construction trades is now the critical friction point.
- Abilene, Texas Deployment:
- Crusoe is constructing one of the world's largest AI computing campuses with an aggregate capacity of 2.1 gigawatts.
- Tenants include Oracle and OpenAI (Project Stargate), with future expansion planned for Microsoft.
- The site features a 1-gigawatt substation, the largest privately owned in the U.S., capable of powering a city the size of Denver.
- A 350-megawatt natural gas plant was co-located to firm renewable energy and ensure operational continuity.
- On-site staff (~3,500 currently, expanding to ~2,000 permanent operations) exceeds the local town population of 12,000.
- Cost Structure Breakdown:
- Total CapEx: Approximately $20 million per megawatt for infrastructure/power plant and $40 million per megawatt for IT/compute, totaling ~$60 million per megawatt.
- Labor Intensity: Labor accounts for $4.7 million per megawatt, representing a $4.7 billion investment for a 1-gigawatt site during construction, with a critical shortage of skilled electricians and plumbs driving wage inflation.
- Power Equipment Inflation: Gas turbine costs have tripled from ~$1 million/MW to ~$3 million/MW due to limited manufacturing capacity among major suppliers (GE Vernova, Siemens, Mitsubishi).
- OpEx: Ongoing operational expenses are estimated at ~$1–2 million per megawatt annually, covering power, insurance, and maintenance.
- Compute Economics & Revenue:
- IT CapEx Allocation: Of the $40 million/MW IT spend, ~$30 million is allocated to GPUs (NVIDIA), ~$4 million to high-performance networking (NVLink, InfiniBand), ~$3 million to CPUs/storage, and ~$3 million to rack fit-out.
- Asset Appreciation: Contrary to fears of rapid obsolescence, H100 spot pricing has surged above initial launch prices due to demand for agentic workflows; Blackwell pricing follows a similar trajectory.
- Revenue Potential: Renting compute infrastructure generates ~$15 million per megawatt annually, implying a 4-year payback period; adding managed services (hosting models/APIs) can double revenue to ~$30 million/MW, achieving a 2-year payback.
- Technological Innovation:
- Crusoe Spark: A modular, self-contained data center unit (500kW air-cooled or 2MW liquid-cooled) designed to reduce infrastructure and labor costs by 30–50% through centralized manufacturing.
- Electrical Stack Evolution: High-voltage distribution (765kV lines) requires innovative solid-state transformers and DC conversion to 900V, suggesting long-term disruption risks for legacy electrical hardware manufacturers who fail to innovate.
- Future Outlook & Risks:
- Space Data Centers: While partnerships exist with Star Cloud for space-based H100 deployment, Miller views this as non-material for 5–10 years due to logistical challenges in thermal management, component repair, and payload costs.
- Open Source Trend: Open-source AI models are expected to gain significant market share, potentially pressuring closed-source model providers.
- Advice for Students: Focus on the "how" (learning processes, grit, and leveraging AI tools) rather than specific technical knowledge, emphasizing the value of continuous self-improvement over static skill sets.