IIT-Madras announces research projects on AI, seeks to study feasibility of small data centres

IIT-Madras has announced a feasibility study for small AI data centres, called “pico centres”. These centres would support inference rather than the large-scale training of language models. The proposal matters because it seeks a practical, lower-resource AI setup for different groups of users across India.

The centres would use frugal engineering methods and operate as a federated, distributed network. Bodhan AI, IIT-Madras’s Centre of Excellence for AI in Education, is studying whether such centres could consume less electricity and fewer resources. One example is personalised AI tutoring for school students.

This remains a feasibility study, not a completed rollout. IIT-M Director V. Kamakoti announced the plan at an AI Measurement and Governance conclave. If the model proves workable, similar centres could support local applications in agriculture, space, manufacturing, health and law.

An AI data centre is a facility containing computing systems, storage and networking equipment that run AI workloads. The source distinguishes between large data centres needed for training large language models and smaller centres intended for inference. Inference means using an already-trained model to produce an output, such as an answer, prediction or recommendation.

During training, a model processes many examples and adjusts its internal parameters to learn statistical patterns in language or other data. During inference, it receives new input and calculates an output using those learned parameters. A personalised tutor, for example, could generate a response to a student’s question.

The article focuses on making inference more resource-efficient. It does not provide technical specifications for the proposed pico centres. It says they could consume less electricity and resources than arrangements designed for large-scale model training.

The article identifies India’s 25 crore school students as a possible audience for personalised AI tutors. In Indian numbering, 1 crore equals 10 million, so 25 crore is approximately 250 million students. This is a potential user base, not a confirmed number of students who will receive the service.

Bodhan AI is studying frugal AI data centres that could provide such tutoring. The proposed pico centres would handle inference, meaning they would use trained AI systems to respond to students’ questions or learning needs. Their intended advantage is lower electricity and resource consumption.

The scale makes the feasibility question important. Serving millions of students would require an approach designed for many users and groups. The article does not state how many centres would be needed, when deployment might begin, or whether the proposed system will be adopted nationwide.

If small AI data centres work as intended, organisations could access specialised AI inference closer to the people and sectors they serve. That could support personalised education, agricultural advice, health-related tools, legal assistance, manufacturing applications and space-related work. The central idea is right-sized AI rather than one identical system for every need.

For example, a centre supporting education could provide personalised AI tutors. A centre serving another group could be designed around agriculture or manufacturing. The proposed mechanism is a federated and distributed network built with frugal engineering, intended to consume less electricity and fewer resources.

The article describes this as a feasibility study, so these benefits are possible implications, not established results. IIT-Madras is also pursuing responsible-use case studies, AI evaluation and trusted AI Commons. Together, these efforts suggest that practical deployment would need both efficiency and accountability.

A federated network could be preferable when different communities need different AI services. Small centres could be designed for particular groups or subjects, such as students, farmers, hospitals or manufacturers. This matches IIT-Madras’s idea of a “right-sized AI approach” for India and its emphasis on frugal engineering.

The key mechanism is distribution. Instead of sending every inference task to a few very large data centres, multiple pico centres could provide services across a wider network. The article says this model could consume less electricity and resources. It also describes the approach as federated, distributed and workable.

However, the article does not establish that small centres will always be cheaper, faster or more reliable. IIT-Madras has announced only a feasibility study. Its value will depend on whether the centres can deliver useful services, support intended user groups and meet responsible-AI expectations.

These three initiatives address different parts of responsible AI. The evaluation platform would measure AI tools across more than 50 metrics and 100 sub-metrics in multiple languages. That can make comparisons more systematic and expose strengths or weaknesses that a single score might miss.

The 500 planned videos would show case studies from across India on using AI responsibly and effectively. Real examples can help organisations understand how responsible use works in practice. Trusted AI Commons, which IIT-Madras said it would begin developing, is intended to enhance AI accountability and responsibility.

Together, measurement, practical guidance and a trust-focused resource could create a stronger basis for oversight. They would not automatically make every AI system safe or fair. The article does not specify the platform’s metrics, the Commons’ governance structure or how organisations will act on evaluation results.

During training, a large language model processes many examples of text and learns statistical relationships among words, phrases and broader patterns. It does not store a simple list of ready-made answers. Instead, its parameters are adjusted so the model becomes better at predicting what text is likely to follow given earlier text.

During inference, the model receives a prompt and calculates probabilities for possible next tokens, such as words or word fragments. It selects a token according to its decoding method, adds it to the context and repeats the process. This produces a response one step at a time. The result reflects learned patterns, not human-like understanding.

The article only says that large data centres are needed for large-scale LLM training and that smaller centres could handle inference. It does not describe model architecture, training data, token prediction or decoding. Those details come from established background knowledge about LLMs.