All AI model training starts with raw data from somewhere. When moving terabytes or even petabytes of data to your server, high-capacity storage with fast sequential write performance keeps things moving.

Nobody likes dirty data. During this phase (sometimes called preprocessing or extract-transform-load), raw data is cleaned and organized into tokens for use during training. In storage speak, this is mostly sequential read activity.

Your nascent model is exposed to training tokens in random order, developing a set of parameters that’ll drive later outputs. Expect heavy random read activity here while the GPUs work overtime. Frequent checkpoints rely on sequential write throughput.  

Your shiny new AI model is deployed and processes new inputs to generate responses. Low-latency storage enables real-time inference for that “living in the future” feeling.

Save your work! Not only is it increasingly important for compliance and audit reasons, but all those inputs and outputs can be used to re-train your model later. High capacities are key here.