Published: October 31, 2026
StartupFeed Quick Take
- MeitY unveiled the DEPA for AI Training framework in Bengaluru on October 30, 2026.
- The report says it could add $500 Bn to India’s economy by 2030.
- Finance and healthcare run the first two pilots; startups join as data users.
India’s IT ministry unveiled the DEPA for AI Training framework in Bengaluru on October 30, 2026, a plan to let startups train AI on protected data.
MeitY Secretary S Krishnan released the DEPA for AI Training Handbook at a roundtable. The report says the DEPA framework could add $500 Bn in AI value to India’s economy by 2030.
DEPA stands for Data Empowerment and Protection Architecture. It is the system behind India’s Account Aggregator, which already moves bank data with a user’s consent.
The framework is not new. The iSPIRT Foundation, a technology think tank, built DEPA for Training in 2023, and it has run as a pilot since. The handbook is the government’s first move to scale it nationally.
What the DEPA framework changes for AI startups
Indian AI startups get a legal route to data they could not reach before. A startup joins as a Training Data Consumer, a firm that asks for datasets to train a model. The data owners, called Training Data Providers, are banks, hospitals and government bodies. Startups never copy or own the raw data. They train inside a secure space and take out only the finished model. Middlemen called Data Collectives pool and clean datasets from many owners, then set the sharing rules. The aim is to end what the report calls a “Data Deadlock”, where useful data sits locked inside separate institutions.
The handbook makes the pitch to founders in one line.
“In this model, compliance is not a barrier to innovation; it is the feature that unlocks it.”
From the DEPA for AI Training Handbook, MeitY and the iSPIRT Foundation, October 30, 2026.
How the data stays private
The DEPA framework rests on four technical pillars. Confidential clean rooms are sealed computing spaces where data is processed but never exposed. Differential privacy adds statistical noise so no single person can be traced in the results. Electronic contracts set, in machine-readable code, who may use what and for how long. Zero-trust security checks and logs every request. Data owners anonymise and encrypt each dataset before they register it. Before a trained model can be deployed or scaled, it must pass independent checks for bias, fairness, safety and explainability. After launch, the model stays under watch, and its outputs must trace back to the data and contracts used to build it.
Where the money is: finance and health first
Finance and healthcare go first, with regulatory sandboxes running beside the pilots. The report pegs India’s financial services market at $550 Bn and its healthcare market above $600 Bn. It says 400 million people and small businesses could reach formal credit if lenders train better models. Small businesses, or MSMEs, already make up nearly a third of India’s economy, yet face a credit gap of Rs 20-25 lakh crore.
The data exists already. The access did not. The report points to a Penn Medicine and Intel project across 71 hospitals that trained a shared model to over 99% of the accuracy of one built on pooled data.
| Sector | Market size | What the data could open |
|---|---|---|
| Financial services | $550 Bn | Credit for 400 Mn people and MSMEs |
| Healthcare | $600 Bn+ | Shared models across 71-plus hospitals |
| EdTech | $29 Bn | Tools for 250 Mn students |
| MSME credit gap | Rs 20-25 lakh crore | Loans the current data cannot support |
The catch: consent and old systems
The DEPA framework has real gaps. It targets bulk, de-identified data, not fresh consent for each use, which sits uneasily with the consent-first spirit of the DPDP Act, India’s data protection law of 2023. The report admits many institutions run legacy systems that cannot support secure data environments.
Trust is the other hurdle. Once an owner shares data, it loses control, the very problem the clean-room design tries to solve. Rollout now depends on whether banks and hospitals actually join. The push is also part of a wider move. TRAI has sought public comments by November 27 on draft rules for responsible AI in telecom. More AI regulation is coming.
What this means for you: If you build AI in fintech or health-tech, regulated bank and hospital data may soon be reachable without buying or copying it. Watch the finance and healthcare pilots, and get ready to join as a Training Data Consumer.
StartupFeed Insight
The $500 Bn number will travel far, but it is a 2030 promise, not a 2026 result. DEPA for Training has existed since 2023 and still runs on pilots. The hard part was never the cryptography. It is getting a cautious bank or a state hospital to share real data, and trust a clean room it cannot see inside. Watch the finance pilots. If two large lenders and one public hospital network are training live models by mid-2027, the framework is working. If not, this stays a handbook. Founders should prepare now, but bank on nothing yet.
By Avinash Mishra, Business Correspondent
Have a tip? Write to us at editorial@startupfeed.in.
Liked this story? Follow StartupFeed on Google so our reporting reaches you first.
