Guide · Sovereign AI
Sovereign AI in India: What It Means for Teams Deploying AI Agents
By Vikas Goel, Founder of ThinkerWave · Last updated
For an Indian organization, sovereign AI means keeping control over where sensitive data goes, who runs the model, who can change or withdraw it, and who can audit the results. It is not one product or a yes-or-no status. An India-hosted foreign model and a self-hosted open-weight model give very different levels of control, so match the setup to your risk.
In brief
- Data location is only one part of sovereignty. Control of the model and the audit trail matter too.
- DPDP has no blanket data localization rule, but specific data and sectors can face stricter limits.
- AI agents raise the stakes, because a single task can send data through many model and tool calls.
- Ask vendors exactly where each call goes, rather than accepting the word “sovereign”.
Four levels of control
| Setup | What you control | What you don’t |
|---|---|---|
| Foreign API, hosted abroad | Your keys and usage | Data location, model changes, availability |
| Foreign model, hosted in India | Data region, subject to the contract | The model, its updates and its availability |
| Open-weight model on Indian cloud | Model version and data region | The physical infrastructure |
| Open-weight model on your own hardware | Model, data, network and logs | Little, but you take on the operating work and capacity |
More control isn’t automatically better. Self-hosting costs more to run, and a smaller open model may be less accurate on your task. Test accuracy on the model you’ll actually deploy, because a result from a frontier model does not carry over.
What DPDP does and doesn’t require
Based on law-firm commentary updated in October 2026, the main points on data location are:
- The DPDP Act does not impose blanket data localization.
- Under Section 16, the central government can restrict transfers to specified countries.
- The government can require Significant Data Fiduciaries to keep specified personal data in India.
- Stricter sector rules still apply, such as RBI’s payment data storage requirements.
- The rules were notified on 13 November 2025, and the transfer framework is expected from 13 May 2027.
So “DPDP forces us to self-host” is usually too simple, and so is “DPDP lets us send anything abroad”. Your sector, your contracts and the data involved decide the answer. This is general information, not legal advice.
Compute and Indian models
Access to compute is improving. The government reported more than 38,000 GPUs on the IndiaAI compute portal in March 2026, offered at subsidized rates for Indian startups, researchers and academic institutions. Check current eligibility and rates on the portal before planning around them.
India’s model options are growing as well, with IndiaAI supporting efforts such as Sarvam and BharatGen. Before choosing one, check its license, its fit for your languages and domain, and its performance on your own test set. An “Indian model” label doesn’t tell you whether it suits your workload.
Why AI agents need extra care
A chatbot usually makes one visible call. An agent may make dozens, using several tools and passing your data along at each step. A setup can look sovereign at the main model and still leak through a web search, a logging service or a fallback provider. Map every outbound call, then test it by watching network traffic during a real run.
A checklist for buyers
- Where does each model, tool and logging call go during a run?
- Can we pin the model version, and what happens if the vendor changes or withdraws it?
- Who holds the keys, and can we revoke them ourselves?
- Does it work with outbound internet access blocked?
- Can our own team or regulator read a full record of each run?
- How accurate is it on the model we’ll deploy, measured using a proper evaluation?
Where ThinkerWave stands today
The ThinkerWave desktop app runs on your machine, and your problem data doesn’t go to us. Model calls go to the provider you configure, using keys you hold, and every run keeps a full trail. That is a strong starting point, but with a hosted foreign model it isn’t full sovereignty.
The engine can connect to OpenAI-compatible endpoints, the interface used by common self-hosted model servers. We’re still validating full runs on open-weight models, so we don’t yet claim a proven air-gapped deployment. If you need one, talk to us and we’ll tell you what has been tested.
Frequently asked questions
What is sovereign AI?
Sovereign AI is a country’s or organization’s ability to use AI under its own control, across data, compute, models and oversight. How much control is needed depends on the sensitivity of the work.
Does India’s DPDP Act require all data to stay in India?
No. Legal commentary describes no blanket localization rule in the DPDP Act. The government can restrict transfers to specified countries, may require Significant Data Fiduciaries to keep specified data in India, and sector rules can be stricter. The transfer framework is expected from May 2027. Take legal advice for your own case.
Is running a foreign AI model in an Indian data center sovereign?
It addresses data location, which can be enough for some work. It does not by itself give you control over the model, its updates or its availability. For stricter requirements, consider open-weight models on infrastructure you control.
Is ThinkerWave a sovereign AI product?
It is built with control in mind, but it does not make a setup sovereign by itself. The desktop app runs on your machine and your problem data does not go to ThinkerWave, but model calls go to the provider you configure. Full sovereignty also requires a model you control, and we are still validating full runs on self-hosted open-weight models.
Sources
Try it on your own problem
ThinkerWave works a hard problem from several angles, checks its claims against the evidence, and keeps a full record of how it got there. It runs on your machine, with your own model keys.