AI models are programs trained on huge amounts of data to do one kind of task, such as writing text, understanding speech or controlling a machine. This zone holds the companies whose product is the model itself. They build it, train it and sell it to other companies, which build their own products on top.
Think of three things. You type a question in Hindi into a chatbot, and it answers in Hindi. A driver asks the voice assistant in her car for the nearest petrol pump. A robot arm in a warehouse picks up a parcel of a shape it has never seen before. Each of these runs on a trained model.
The chatbot runs on a large language model (LLM), which learned to read and write by studying vast amounts of text. India's own language models are being trained on Indian languages under a national programme, so that they handle Hindi, Tamil or Bengali as well as English. The page on sovereign AI explains why a country wants this.
The car runs on speech and voice AI. One kind, a speech-to-text (ASR) model, turns the driver's words into text. Other voice models change a speaker's accent as they speak, in accent conversion, or produce a spoken reply with almost no pause, in real-time voice AI.
The robot arm runs on a robotics foundation model. This is one model trained to control many kinds of machine, instead of a separate program for each robot.
The largest of these models are called foundation models: very large models trained on a very wide range of data, which others build on. Training one takes thousands of chips for weeks. Running it for each user is a different job, as the page on training vs inference explains.
Most of the models the world uses are made by a few large companies outside India. The companies in this zone are smaller. They include an Indian company building a language model, voice companies with engineering teams in Bengaluru, and one company building a model for robots.