# AI Factory **Domain:** Artificial Intelligence / High-Performance Computing / Facilities **Doc Type:** Canonical Bridge Concept **Maturity:** Developed ## Definition An **AI factory** is a co-designed cyber-physical system that turns data, energy and model-development work into trained models, inference services and derived machine capabilities. Compute, networking, storage, power, cooling, orchestration and operations are parts of one production architecture. The term has been strongly popularized by particular vendors, but the underlying category is not vendor-specific. Open and heterogeneous systems can implement the same facility-level relationship. ## Historical Lineage Scientific computing joined specialized processors to parallel storage and high-performance networks. Cloud data centers added fleet automation and elastic service. Frontier model workloads intensify collective communication, accelerator power density, checkpointing and data motion until the boundary between facility and computer becomes difficult to maintain. ## Governing Variable The AI factory governs useful model throughput under constraints of power, heat, bandwidth, memory, storage, reliability and scheduling. ## Material and Institutional Manifestation Accelerator systems, scale-up links such as [[wiki/UALink|UALink]], scale-out networks such as [[wiki/Ultra Ethernet|Ultra Ethernet]], [[wiki/Optical Fabric|optical fabrics]], storage, [[wiki/Rack-Scale Computing|rack-scale computing]], high-density power and liquid cooling compose the plant. [[wiki/Open Compute Project|Open Compute Project]] standards demonstrate how mechanical, electrical and thermal interfaces can be opened above the component level. ## Computational Analogue The factory is the computational object. A scheduler that ignores topology wastes accelerators; a model pipeline that ignores checkpoints cannot survive routine failure; a rack design that ignores heat cannot preserve information processing. ## Relation to Governance [[articles/Cyber-Physical Lessons from Imperial Beekeeping|Cyber-Physical Lessons from Imperial Beekeeping]] treats the facility as computational metabolism. The higher-order question is who can allocate its capacity, define its objectives and connect its outputs to operational systems beyond the site. ## Constitutional Question and Failure Modes Failures include resource concentration, opaque allocation, grid and water conflict, vendor lock-in, common-mode software error and optimization of model throughput at the expense of resilience or public infrastructure. Standards improve composability but can also concentrate power in the bodies that define them. ## Related Ontology [[wiki/Data Center|Data Center]] · [[wiki/Rack-Scale Computing|Rack-Scale Computing]] · [[wiki/Thermal Architecture|Thermal Architecture]] · [[wiki/Optical Fabric|Optical Fabric]] · [[wiki/Facility Digital Twin|Facility Digital Twin]]