India should build sovereign AI capacity in three layers: compute it controls, data it governs, and models it can inspect and run itself. It should do this without refusing the best global models, because sovereignty means the ability to choose and to switch, not the habit of owning every component. The mistake would be to treat building and buying as opposites.
Why compute, models and data became strategic
Training and serving large models takes accelerators, power, data centres and fast networking. The supply of advanced accelerators is concentrated in a few firms and is subject to export controls set in other countries. A country that depends wholly on foreign compute and foreign model interfaces is exposed to price changes, access limits and decisions about what a model will or will not do. Data raises a parallel issue. Models learn from data and are used on data, so where it sits, who holds it and under which law determine what a ministry, a hospital or a bank can safely do.
The IndiaAI Mission, approved in 2024, treats compute as a public input. It has sought to assemble a large pool of GPU capacity, sourced from private providers who are empanelled and made available at subsidised rates to start-ups, researchers and public bodies. It also supports foundation model development and a platform for datasets. Renting rather than owning is a sensible design, since it lets capacity grow without the state running data centres. The risk is idle capacity. Subsidised compute is only useful to users who can afford the engineers, the power and the customers that go with it.
Indian-language data and the Data Protection Act
Most large models are trained mainly on English text. Indian languages have far less digitised text, speech data is scarcer still, and dialects multiply the problem. Bhashini, the national language translation mission, and academic groups such as AI4Bharat at IIT Madras have worked on open datasets and models for Indian languages. This matters for any service meant for people who do not transact in English, which is a very large part of the population.
Our position is that language data is the most defensible sovereign asset India can build. Nobody else has the incentive to collect it carefully, and its value rises with every service that has to work in Hindi, Tamil or Bengali rather than in translation from English.
The Digital Personal Data Protection Act, 2023, supplies the legal frame. It sets out consent-based processing of digital personal data, duties for data fiduciaries, rights for the people the data describes, and a Data Protection Board. On cross-border transfer it permits flows except to countries the government restricts, instead of requiring blanket localisation. That means law alone will not force compute onshore. For government and sensitive uses, domestic hosting will be a matter of procurement policy and risk appetite. How the Act’s consent and purpose rules apply to training data is still being worked out in practice, and developers will want clear answers.
Sovereignty in AI is the freedom to switch providers, not the obligation to build everything.
Build, buy, or both
The tension is real. Frontier models from a few global labs are better at general tasks than anything India is likely to build soon, and insisting on domestic models for everything would raise costs for users and slow adoption. Yet relying wholly on foreign interfaces for health records, tax systems or security-related work is not prudent either.
We favour tiers. Sensitive workloads should run on domestic compute, using open-weight models that can be hosted and audited locally. General work can use the best model available. Every contract, in either tier, should preserve portability, so that a change of provider is a procurement decision and not a rebuild. Public funding for foundation models should go to fewer, better efforts, tied to Indian-language and domain use, rather than a wide spread chasing frontier benchmarks.
Here India’s experience with digital public infrastructure gives it an edge. UPI, Aadhaar, DigiLocker and the Account Aggregator framework all follow one pattern: public rails with open specifications, and private firms building services on top. The same pattern fits AI. Public compute and datasets act as rails. Standards for consent and evaluation set the rules. Private firms compete on applications. The Account Aggregator consent model is a ready template for how personal data could move to a model with the person’s permission.
A second advantage is a habit of building for low cost per transaction. In a market with modest revenue per user, the price of running a model for each query decides what can be offered at all. We think India’s edge is likelier to lie in cheap, reliable, multilingual deployment than in the largest model.
We are watching two things. One is how fully the subsidised compute is used, and by whom. The other is whether Indian-language datasets are released on licences that let companies build products on them. The open problem is physical: dense GPU clusters need large, steady power supply and cooling, and the sites that have both are few.