India wants AI models to visit your data, not steal it

Key points

• India's Ministry of Electronics and Information Technology published a framework in 2026 that inverts the usual AI training process: instead of sending data to a model, the model is sent to the data.
• The scheme, called DEPA for AI Training, would let hospitals and government agencies share encrypted records with AI developers without handing over raw files.
• India says the system could extend formal credit to 400 million underserved people by pulling together financial data currently scattered across thousands of banks.
• India frames the framework not just as domestic policy but as a blueprint it wants to offer to the United Nations and World Bank.

The sentence that cuts to the core of India's plan came from Professor Ajay Kumar Sood, the government's Principal Scientific Adviser: "The model travels to the data, rather than the data travelling to the model."

Normally, when a company trains an AI, it moves enormous quantities of medical records or financial transactions onto its own servers. India's proposal flips that. The data stays put, locked inside what the framework calls a Secure Data Environment, and the AI comes to it.

How would this actually work?

The framework sets up three types of players. Training Data Providers (hospitals, banks, government agencies) anonymise and encrypt their records, then share them with a Data Exchange Platform, a kind of secured vault. AI developers, called Training Data Consumers in the document, apply to access those vaults and can run their models inside them, but can't extract the underlying records.

Each vault applies differential privacy, a technique that adds tiny mathematical noise to the data so the finished AI can't reproduce or identify any individual's information. Every action inside is logged permanently so auditors can verify nothing went wrong.

India already runs several data-exchange systems, including the Unified Payments Interface payment network. How much additional engineering is needed before DEPA is operational isn't clear.

Why does India think this matters?

Money and sovereignty. India has watched large American and Chinese AI companies train their models on global data, then sell the resulting tools back to Indian users. MeitY's framework argues those offshore companies have little incentive to solve specifically Indian problems: rural healthcare or micro-business lending.

The numbers in the document are striking. India has roughly 250 million students in a domestic education-technology sector worth $29 billion. Its banking system is so fragmented, across literally thousands of institutions, that 400 million people still lack access to formal credit. Combining those financial records safely could, the government argues, let lenders reach that population for the first time.

This is the same logic India applied when it built zero-commission public infrastructure like the Bharat Taxi ride-sharing platform: use government architecture to stop big platforms from monopolising the market.

For anyone who's followed this beat, the framework is a pointed response to two pressures at once. Global platforms have been racing into India. We've covered Alexa+ arriving with Hindi support, ChatGPT starting to show ads to free users there, and Google testing a buy button inside Gemini for Indian shoppers. DEPA is the government's answer: build the infrastructure so Indian data trains Indian models, and foreign players must play by local rules to access it. Our 30 September report on the Gemini buy-button trial shows exactly how aggressively those platforms are pushing into Indian commerce.

The framework, first reported by The Register AI, also pitches DEPA as a diplomatic export, offering the architecture to the World Bank and United Nations as a model for developing countries that want AI benefits without surrendering data control.

Watch the engineering, not the vision. Persuading thousands of Indian hospitals and banks to connect to a shared vault is a coordination problem that policy documents can't solve on their own. That's where this either becomes real infrastructure or another well-drafted intention.