# DEPA AI Training: MeitY Sealed Rooms Open Bank Data to AI > Source: > Published: 2026-11-04 04:43:37+00:00 DEPA AI training is a framework the Ministry of Electronics and Information Technology (MeitY) launched on 31 October 2026, in which an AI developer sends a model into a sealed computing environment, trains it on a bank's or hospital's data, and receives only the trained model back. The [PIB release](https://www.pib.gov.in/PressReleasePage.aspx?PRID=2330756) describes the Bengaluru launch of the Handbook on Data Empowerment and Protection Architecture (DEPA) for AI Training, and the handbook itself claims the approach could [unlock $500 billion in AI-driven value for India's GDP by 2030](https://cio.eletsonline.com/artificial-intelligence/meity-depa-ai-training-framework-sealed-rooms-sensitive-data-ispirt/76548/). A bank pilot has reportedly handled [more than 2 million requests in two weeks](https://www.livemint.com/news/india-pilots-data-grid-to-train-ai-on-sensitive-info-without-seeing-it-11793362037802.html) without exposing raw customer lists. The handbook is a draft for discussion, not a law. **Last verified:** 3 November 2026 DEPA stands for Data Empowerment and Protection Architecture. NITI Aayog introduced it in 2020 as a consent framework that lets citizens share their own data, and that model [already underpins the Account Aggregator system in finance](https://www.livemint.com/news/india-pilots-data-grid-to-train-ai-on-sensitive-info-without-seeing-it-11793362037802.html). The new handbook extends the idea from individuals to institutions: banks, hospitals and public agencies that hold data, and startups, researchers and enterprises that want to train models on it. MeitY Secretary S. Krishnan unveiled the handbook at a roundtable in Bengaluru. iSPIRT had published [technical updates and sample training scenarios earlier in 2026](https://pn.ispirt.in/depa-training-tech-updates/), so this release is the government-facing step rather than a new invention. Per the launch presentation, the data stays with its owner. A sealed room, formally a Secure Data Environment, runs only what a digital contract permits, and nobody can inspect the data inside, including the company operating the system. Principal Scientific Adviser Ajay Kumar Sood summed it up as models travelling to the data instead of the reverse, as [reported by Mint](https://www.livemint.com/news/india-pilots-data-grid-to-train-ai-on-sensitive-info-without-seeing-it-11793362037802.html). The handbook rests on four pillars. The status column below is our assessment from launch coverage, not an official rating. | Pillar | What it does | Status today (our analysis) | |---|---|---| | Electronic contracts | Machine-readable permissions that bind a request to a specific purpose | Drafted in the handbook | | Sealed rooms | Controlled environments where training runs under contract limits | Piloted in finance and health | | Privacy safeguards | Techniques such as differential privacy limit what leaks from outputs | Specified, depends on each deployment | | Data provenance | Traces origin, permissions and use of data through training | Specified, not yet tested at scale | This worked flow is our reading of the handbook's design as described by [Elets CIO](https://cio.eletsonline.com/artificial-intelligence/meity-depa-ai-training-framework-sealed-rooms-sensitive-data-ispirt/76548/): In some cases the consumer receives only a decision, such as a loan-application outcome, rather than a model. The handbook calls the problem a data deadlock: useful data is split across domains and locked behind institutional and regulatory walls. Hospitals and banks earn little from sharing, and a leak can attract [penalties of up to Rs 250 crore under the DPDP Act, 2023](https://cio.eletsonline.com/artificial-intelligence/meity-depa-ai-training-framework-sealed-rooms-sensitive-data-ispirt/76548/). Stripping names and ID numbers does not solve it either, because anonymised records can often be re-identified and lose analytical value. The same compute-and-data question shapes India's wider AI buildout. Our pieces on [IndiaAI compute for agentic AI](https://dev.to/articles/agentic-ai-tutorial-indiaai-compute) and the [Vizag AI data centre](https://dev.to/articles/vizag-ai-data-centre-andhra-pradesh-512mw-am-ai-factory) cover the hardware side. DEPA addresses the data side. Two were presented, according to Mint. A bank working with a fintech partner since March 2026 identified which fintech customers qualified for its credit cards before asking for consent, so it approached only people with a genuine offer. The pilot processed more than 2 million requests in two weeks. Separately, a consortium of research hospitals and a health-tech startup trained an MRI tumour-detection model on scans the startup's engineers never viewed. Both are self-reported demonstrations. Neither has been independently audited in public. **Builder's box (our analysis)** Consent-driven data flows are also reshaping payments, as covered in [UPI's new rules for AI agents](https://dev.to/articles/upi-new-rules-2026-ai-agents-pay-on-upi). For the policy backdrop, see [India's semiconductor mission and AI governance](https://dev.to/articles/india-semiconductor-mission-2026-ai-agents-governance). This decision rule is our analysis, not part of the handbook. | Situation | Better option | Main tradeoff | |---|---|---| | Needs real, sensitive records and a regulated partner will contract | DEPA-style sealed room | Draft framework, no registered Data Collabs yet | | Needs broad statistics, not row-level detail | Bought anonymised data | Re-identification risk and reduced usefulness | | Needs volume for early prototyping | Synthetic data | May miss real-world edge cases | In short: prototype on synthetic data, then move to a sealed room when accuracy depends on real records. The stated roadmap starts with pilots in finance and healthcare and regulatory sandboxes, followed by cross-border agreements, capacity building and incentive models. Sectoral regulators such as the RBI are expected to define data sensitivity, and the Data Protection Board of India to oversee compliance, while the handbook ties into the IndiaAI Mission's AI Kosh dataset repository, so catalogued datasets can carry DEPA e-contracts, as [coverage of the document reports](https://beatsinbrief.com/2026/10/31/meity-depa-handbook-ai-training-consent-based-data/). The handbook's economic claims should be read as its own estimates, not independent forecasts. **Q: Is DEPA AI training legally binding?** **A:** No. The handbook is marked as a draft for discussion and proposes a framework. Binding rules are expected to come from sectoral regulators, much as with Account Aggregator. **Q: Can my startup use DEPA AI training now?** **A:** Not as a production route yet. Pilots exist in finance and health, but Data Collabs and standard contracts are still to be established, so the practical step is to follow the consultations and approach a data holder about a pilot. **Q: Who pays data providers under DEPA?** **A:** Nobody, formally, for now. Providers receive benefits in kind, such as priority use of resulting models, while royalties and access fees are listed as later work. **Q: How is this different from buying anonymised data?** **A:** With anonymised data you receive a copy, and it can often be re-identified. With a sealed room you never receive the data at all, only the trained model or a decision, and every use is tied to a contract.