Infrastructure for Modern Artificial Intelligence: A Systems Taxonomy of Large Language Model Training and Serving

Large language models (LLMs) have turned artificial intelligence infrastructure from a supporting concern into a primary determinant of model capability, development velocity, reliability, and cost. A modern AI platform is not simply a collection of accelerators: it is a vertically integrated system spanning power and cooling, compute and memory, scale-up and scale-out networks, storage and data pipelines, compilers and communication libraries, distributed training, post-training, inference serving, cluster control, and lifecycle governance. This paper develops a systems-oriented taxonomy of that stack. It distinguishes three planes—resources, execution, and lifecycle control—and analyzes how the requirements of pretraining, post-training, and online inference differ. Particular attention is given to distributed parallelism, state sharding, checkpointing, data quality, key–value cache management, continuous batching, prefill–decode disaggregation, observability, and multi-tenant scheduling. The analysis shows that LLM infrastructure is governed by shifting bottlenecks: training alternates among arithmetic, communication, memory, and input/output limits, whereas online generation is frequently constrained by memory capacity, bandwidth, queueing, and tail latency. The paper concludes with a measurement framework and research agenda centered on composable parallelism, heterogeneous and disaggregated resources, reliable long-running jobs, efficient reasoning and agent workloads, security, and energy-aware operation. The resulting taxonomy provides a common vocabulary for studying AI infrastructure as a complete sociotechnical system rather than as isolated hardware or software components.