India’s government has launched the handbook for its Data Empowerment and Protection Architecture for AI Training, an attempt to let Indian data train AI models without the data itself changing hands. The Ministry of Electronics and Information Technology unveiled the framework at a roundtable in Bengaluru, with secretary S. Krishnan saying India’s AI ambitions must be matched by the ability to use data responsibly and securely.
The design inverts the usual training pipeline. Principal Scientific Adviser Ajay Kumar Sood summarised it in the line the framework is built around: the model travels to the data, rather than the data travelling to the model. Data holders such as hospitals, banks and government agencies, termed Training Data Providers, place anonymised and encrypted data with Data Exchange Platforms. Model builders, termed Training Data Consumers, train inside hardware based secure data environments where the data stays encrypted and isolated during processing, the providers’ usage rules are enforced, differential privacy techniques are applied, and every activity is immutably logged for audit.
The stated prize is the data that is currently unusable. MeitY argues the framework could unlock fragmented health data for diagnostics and drug discovery, extend formal credit to 400 million underserved individuals and small businesses by enabling secure sharing across India’s thousands of banks, and support personalised learning for 250 million students. Those are the framework’s own projections, not results. The Press Information Bureau launch drew HDFC Bank, Infosys co-founder Kris Gopalakrishnan and the iSPIRT Foundation, which built much of India’s digital public infrastructure, into the announcement.
How much of DEPA exists beyond the handbook is unclear. India already operates several data exchanges, and the framework leans on the same digital public infrastructure playbook as the Unified Payments Interface: state built rails intended to keep Big Tech from dominating the local market, and to be offered to friendly nations as a diplomatic export. Whether model builders accept training inside someone else’s secure enclave, rather than on their own clusters, is the adoption question the launch does not answer.