For two years the Digital Personal Data Protection Act sat on the shelf as a law without teeth. Passed in August 2023, it had no rules, so there was nothing concrete to comply with. That changed on 13 November 2025, when MeitY notified the DPDP Rules. The clock is now running, and the first hard date, the Consent Manager framework, lands on 13 November 2026. If you build AI products that touch Indian personal data, this is no longer a legal team side project. It is a date on your roadmap.
The timeline, finally
The framework rolls out in phases instead of flipping on all at once. The dates that matter:
| Date | What takes effect |
|---|---|
| 13 Nov 2026 | Consent Manager framework goes live; consent tracking obligations begin |
| 13 May 2027 | Core obligations: data principal rights, security safeguards, breach reporting |
| 13 May 2027 | Significant Data Fiduciary duties: Indian-resident DPO, periodic audits, impact assessments |
Full compliance lands 18 months from the rules notification, on 13 May 2027, and there is no grace period after it. MeitY has also consulted on compressing the window from 18 months to 12, so treat every date above as a ceiling, not a target. If you are not already in motion, you are late.
What DPDP actually requires
The Act is built on principles that sound obvious until you apply them to a modern AI stack:
- Consent with purpose limitation. You collect personal data for a stated purpose, and you cannot quietly reuse it for another one. Training data, fine-tuning data, and inference prompts are three different purposes that need three different answers.
- Data minimization. You hold what you need and nothing more. An inference provider that silently retains every prompt it ever sees is not minimized under any reading.
- Data principal rights. Access, correction, and erasure are not aspirational. If a user asks you to delete their data, you have to actually do it, including what lives in vector stores, logs, and caches.
- Significant Data Fiduciary obligations. Large-scale processors carry the heavier load: a DPO resident in India, periodic data audits, and data protection impact assessments.
- Penalties that scale to fund size. Fines run up to ₹250 crore per breach, a number chosen to matter to a startup and a bank alike.
None of this is exotic. What is new is that it now has a compliance date attached.
Where inference runs is now a compliance question
Here is the part that matters to AI teams specifically. A large share of the AI tooling Indian companies rely on today routes prompts to inference APIs hosted on another continent. A “regional endpoint” often turns out to handle only the API dispatch, while the model weights actually run somewhere else. A data processing agreement can promise residency, but a contract cannot enforce what the architecture does not do.
DPDP does not ban cross-border transfer outright, but it puts the burden on the data fiduciary to demonstrate lawful, accountable handling. When a customer, a bank, or a hospital asks “where does the data actually go,” the answer can no longer be “the vendor says it is fine.”
This is the difference between compliance by policy and compliance by architecture. A policy is a document. Architecture is where the compute physically sits. If the GPUs that run your inference are in India and there is no path in the routing layer that sends data abroad, then residency is a property of the system, not a promise bolted on after the fact. That is the bar enterprises are starting to ask for, and it will only get more common as 2027 approaches.
A practical checklist for AI teams
If you process Indian personal data, this is the minimum to have in motion before the end of 2026:
- Map every place personal data lands. Prompts, logs, vector stores, fine-tuning sets, analytics. If you cannot name where each one lives, start there.
- Ask your inference provider the residency question directly. Not “are you compliant,” but “where do the GPUs that run my requests physically sit, and can data leave that region.”
- Treat retention as a setting, not a habit. Zero-retention inference removes an entire class of erasure problems.
- Prepare for erasure across the stack. A delete request has to reach embeddings, caches, and backups, not just the primary database.
- Appoint a DPO if you are approaching SDF scale. Even if you are not there yet, start thinking about who fills that role.
- Watch the acceleration proposal. If the window compresses to 12 months, your May 2027 work becomes November 2026 work.
IndieRouter is built for this exact question
This is the reason IndieRouter exists. It is an OpenAI-compatible inference API where models run on GPUs in Indian data centres, and the routing layer is built so that prompts never leave the country. There is no “abroad” in the architecture to route to. Billing is in rupees with GST invoices, which also removes the forex and cross-border payment friction Indian finance teams deal with on foreign APIs.
We built it because the compliance conversation was heading exactly where it has now gone: from a checkbox on a vendor questionnaire to a date on a calendar. If you want to make the “where does my data run” question easy to answer, the quickstart is the fastest way to see it work.