01.What was actually launched
On Friday 30 October, MeitY Secretary S Krishnan unveiled the Data Empowerment and Protection Architecture (DEPA) for AI Training handbook at a roundtable in Bengaluru. The framework was prepared with the iSPIRT Foundation, the volunteer-led think tank behind parts of UPI and the India Stack, and the report calls it India's digital public infrastructure for AI. It also puts a number on the ambition: the potential to unlock 500 billion dollars in AI-driven value for India's GDP by 2030, as Business Standard reported from the launch.
The central idea is simple to state and hard to build: instead of moving data to researchers, you move the model to the data. A bank, a hospital or a public agency keeps its records where they are. A developer sends a model into a secure environment that runs only what an authorised digital contract permits, and gets back the trained model, or in some cases just a decision, such as on a loan application. According to the launch presentation, nobody can read the data inside that room, not even the company operating it.
02.The deadlock it is trying to break
The handbook names the problem a data deadlock. The data needed for the next generation of Indian AI sits split across domains, locked behind different institutional and regulatory walls, and the cost of sharing usually outweighs the benefit. Hospitals and banks hold some of the most useful datasets in the country, make little money from them, and face fines of up to 25 crore rupees under India's data protection law if that data leaks.
Stripping out names and ID numbers, the report argues, does not solve it. Simple anonymisation fails because of re-identification risk and the loss of utility that comes with it. The report ties the work to a credit gap of 20 to 25 trillion rupees for micro, small and medium enterprises, an estimate it attributes to an expert committee constituted by the Reserve Bank of India. It warns that without such a framework, Indian startups and researchers fall behind while the market concentrates around a few large players.
03.How the plumbing is meant to work
The architecture rests on four technical pillars: a Data Exchange Platform, machine-readable electronic contracts, a secure data environment and differential privacy. Organisations that hold data, called Training Data Providers, register datasets with metadata and cryptographic identifiers. Startups, researchers and enterprises, called Training Data Consumers, request access for a specific use case. A new class of intermediaries called Data Collabs aggregates data across sources and sets the sharing rules.
Training then happens inside the secure environment, with differential privacy used so individual records cannot be pulled back out of an aggregate. Trained models face independent checks for bias, fairness, security, safety and explainability before they are cleared for pilot deployment, and oversight is meant to continue through monitoring and periodic reviews. Governance would sit with a consortium of data providers, consumers, technology providers, domain experts and civil society, with a government authority able to certify models and suspend non-compliant activity. Day-to-day operations are expected to sit with Samyog, a new industry group powered by iSPIRT that would register participants, supply standard contracts and handle disputes.
04.The pilots already running
Two early pilots were demonstrated at the launch, the most concrete evidence that this is not only a slide deck. In the first, a bank testing the approach since March 2026 worked out which customers of a fintech partner qualified for its credit cards before asking anyone for consent, then approached only those with a real offer. Mint reported the pilot handled more than two million requests in two weeks. In the second, hospitals and a health-tech startup are training a tumour-detection model on MRI scans that the startup never sees.
A senior official framed the approach as a democratic third way, between the United States, where companies buy data through commercial deals, and China, where the state pools it.
05.What is still missing
The framework is a draft and is not legally binding, and officials did not say if or when it might be. Consultations are expected to begin in the coming weeks, and bringing it under a legal framework could take years. The handbook says only that payment methods for data owners will be explored, and it does not specify who would be liable if data leaked from a sealed room. Samyog's own site does not describe the group as government-appointed, which matters for how independent the referee looks.
There are technical doubts too. Jibu Elias, India lead for responsible computing at the Mozilla Foundation, told Mint that while the framework could be vital for startups building domain-specific AI in healthcare, education and critical communications, the crucial part is building a trust framework first and validating the quality of vast tranches of sensitive data, both hard to establish. Observers expect binding rules through sectoral regulators rather than a new law, the way Account Aggregator took five years from RBI rule-making to launch. Much also depends on the Digital Personal Data Protection Act, whose main rules take effect in May 2027.
For Indian builders, clean-room training is becoming the expected way to work with regulated data, so learn how these contracts and environments behave before a client asks. For a bank, hospital or government department, joining is a governance decision rather than an IT one, since the liability question is still open. Treat the 500 billion dollar figure as the framework's own projection, not a forecast, because the mechanism behind it has been tested only in two pilots.
Frequently asked questions
What is DEPA for AI Training?
It is a draft framework from MeitY, prepared with the iSPIRT Foundation, that lets startups, researchers and companies train AI models on data held by hospitals, banks and public agencies without the raw data leaving those institutions. Training runs inside a secure environment and the developer receives only the trained model.
Is DEPA for AI Training a law yet?
No. It is a draft handbook that is not legally binding. Consultations are expected in the coming weeks, and stakeholders said bringing it under a legal framework could take years, with binding rules more likely to come from sectoral regulators than from a new law.
How is it different from the existing DEPA or Account Aggregator?
The existing framework handles consented sharing of individual data, as Account Aggregator does in finance. DEPA Training targets bulk, de-identified data used to build models in the first place, which is why it needs secure environments, electronic contracts and differential privacy.