A capacity and economics framework for large-scale AI facilities, connecting power, compute infrastructure, capital investment, operations, and growth.
AI data centers are fundamentally power-constrained systems. The model starts with facility power capacity and works inward through electrical losses, cooling, IT load, rack density, accelerator count, and usable compute. This prevents the financial model from assuming more GPU capacity than the physical plant can support.
The facility model separates land and site work, utility interconnect, substations, generators, UPS systems, cooling, white space, networking, racks, compute, storage, and operational tooling. These layers can be phased so capacity comes online in blocks rather than assuming one monolithic build.
Capital expenditure is paired with recurring electricity, maintenance, staffing, connectivity, hardware refresh, software, and support. Utilization and revenue assumptions can then be layered on top to evaluate payback, margin, and sensitivity to energy price or accelerator economics.
The model supports multi-year expansion assumptions such as compound demand growth, staged power delivery, or changing rack density. This is important because AI infrastructure economics can change faster than traditional data-center depreciation schedules.
Outputs help compare owned facilities, colocation, hosted GPU capacity, public cloud, and hybrid strategies. The goal is to expose where power, capital, utilization, or supply-chain constraints dominate the decision.