India’s digital transformation offers an important lesson: some of the most consequential innovations are not applications, but the infrastructure on which applications are built.
Digital identity (Aadhar), interoperable payments (UPI), and other elements of India’s Digital Public Infrastructure (DPI) have demonstrated how common technological rails can lower barriers to innovation and enable governments, businesses, and civil society to build solutions at population scale. As artificial intelligence enters education, India is applying similar principles.
(Sign up for THEdge, The Hindu’s weekly education newsletter.)
The recently released foundational AI models, as key components of the Bharat EduAI Stack as a sovereign Digital Public Infrastructure by Bodhan AI is an important step in that direction.
BodhanAI, India’s Centre of Excellence in AI for Education, led by Professor Mitesh Khapra and housed at IIT Madras under the aegis of the Ministry of Education, has released foundational AI models, developed in partnership with AI4Bharat, for 22 Indian languages. They provide capabilities spanning speech recognition, speech generation, machine translation, and optical character recognition. The models are available as open weights for others to build upon, alongside hosted APIs for organisations seeking production-ready services.
The significance is not simply another set of AI models. It is the aspiration to create a common foundational layer for AI innovation in Indian education.
A distinctly Indian pathway
India is not alone in treating AI capability as strategic infrastructure. Governments from China and South Korea to Japan and the UAE are investing in domestic AI capabilities, foundation models and local-language technologies. China, notably, is also developing shared AI capabilities specifically for education.
India’s emerging pathway, however, could be distinctive: applying its experience with DPI to AI by creating the Bharat EduAI Stack, a publicly supported, open foundational capabilities platform on which government, academia, start-ups, EdTech companies, and civil society can innovate.
Much of the conversation around AI in education focuses on applications: AI tutors/bots, teacher assistants, automated assessments and personalised learning.
Yet beneath them lie difficult technological problems. An AI tutor/bot must understand a child’s speech. Technology assessing written work must interpret handwriting. And for quality learning resources to travel across India’s linguistic landscape, translation must be accurate, pedagogically sound, contextual and affordable.
If every start-up, institution, non-profit, or State government must independently solve these underlying technological problems before addressing its actual educational challenge, the result is duplication, redundant costs and slower innovation.
Shared infrastructure changes that equation. An organisation building an oral-reading application, for example, can concentrate on reading pedagogy and learner feedback rather than first building speech-recognition infrastructure. This is when foundational AI models begin to function as infrastructure rather than merely as technology products.
Designing for the Indian learner
Much of today’s generative AI ecosystem has evolved around adults interacting with machines predominantly through typed text. That is not how millions of Indian children experience learning.
Children speak before they type. They read aloud. They write by hand. They move between languages. Their pronunciation, vocabulary, and linguistic context vary enormously across regions.
AI for Indian education therefore needs to be multilingual and multimodal by design, rather than adapted later.
Speech technologies could enable oral-reading assessments and conversational learning. Vision systems could help interpret handwritten responses. Translation could substantially reduce the cost of making quality educational resources available across languages.
These are not educational solutions by themselves. They are building blocks from which hundreds of solutions could emerge.
The organisation with the best idea for improving foundational literacy may not have the capital or computing resources to train sophisticated speech or language models. Shared foundational capabilities can reduce that disadvantage.
The decision to release model weights openly for the Bharat EduAI Stack is particularly relevant to India’s innovation ecosystem. Open models allow universities to research and improve them, start-ups to adapt them, and non-profits to build public-interest applications without being entirely dependent on proprietary platforms. Hosted infrastructure complements this: organisations with technical capacity can operate the models themselves, while others can consume the capabilities as services.
The principle is simple: Bharat EduAI Stack and its components should expand the number of organisations capable of innovating.
Affordability is infrastructure too
For public AI infrastructure, however, openness alone is not enough. It must also be inexpensive enough to disappear into the economics of the application.
BodhanAI’s hosted stack makes this tangible. An hour of Indian-language classroom audio can be transcribed for about ₹6; a 300-page textbook can be digitised for about ₹60. At these price points, capabilities that could otherwise become prohibitive recurring costs begin to look like utilities that can support population-scale applications.
This may prove as consequential as model performance. India’s digital public infrastructure succeeded partly because the marginal cost of using the rails became extraordinarily low. For AI to reach every government-school classroom, every teacher, every learner, and their homes across India, AI cannot remain a premium layer. It has to become affordable infrastructure.
Bodhan AI’s approach suggests a useful division of roles. Public institutions and academia can help build foundational capabilities where markets may underinvest — Indian-language models, datasets, benchmarks and common technological infrastructure.
Entrepreneurs, edtech companies, and civil-society organisations can then innovate and compete at the application layer.
The objective should not be for the State to determine the winning AI application, but to create conditions under which many organisations can build, test and compete.
A beginning, not a destination
The Bharat EduAI Stack release is ultimately a beginning, not a destination.
The real questions are whether a child can receive useful reading feedback in her own language; whether a teacher can understand where students are struggling; whether States can deploy AI without becoming locked into a single vendor; and whether Indian innovators can build affordable products for contexts global AI companies may never prioritise.
Answering these questions will require research, standards, responsible deployment and, above all, evidence of learning.
India’s earlier digital transformations have shown what can happen when foundational capabilities become shared rails rather than closed systems. BodhanAI creates an opportunity to explore whether the same principle can be extended to AI in education — designed around India’s languages, learners and classrooms.
The most important question about the Bharat EduAI Stack, therefore, is not how many models it has released. It is how many useful things India will now be able to build because these foundations exist.
(Bhanu Potta is the founding partner at ZingerLabs and the senior advisor at Birla AI Labs.)
