Model training and inference are the two jobs of running an AI model. Training teaches the model from data, once or a few times, on a large cluster of machines. Inference, also called serving, means running the finished model to answer each user's request, often millions of times a day. This zone holds the platforms for both jobs.
Training needs many GPUs working together. A GPU is a processor that does thousands of simple calculations at the same time, which is the kind of arithmetic AI models need. The page on AI accelerators explains these chips. A large model can need thousands of GPUs for weeks. A distributed computing framework splits that work across all the machines and restarts any piece that fails. The GPU servers themselves are wired together in AI clusters, which other companies build.
Serving has different problems. Requests arrive unevenly, more by day than at night, and each user expects an answer within a second or two. A model deployment platform puts a trained model on servers, adds servers when requests rise and removes them when requests fall. It usually runs the model in containers, packaged copies of the software that can be started on any server.
Not every company has data scientists to build its own models. Consider an insurer in Chennai that wants to predict which motor claims are likely to be fraudulent. Its analysts use automated machine learning (AutoML), which tries many kinds of model on the table of past claims and picks the one that predicts best. The insurer then serves the chosen model from its own cloud account.
Three other products in this zone change where the model runs. Edge AI runs a small model on a phone, a camera or a machine, without a data centre. A sovereign AI cloud runs models on GPUs that one country controls. A quantum computing cloud offers quantum computers beside AI tools. It is the one corner of this zone in which the hardware is not a GPU.