In this authored article, Fujitsu’s Amitkumar Shrivastava explores what sovereign AI really means for India, arguing that sovereignty extends beyond where models and data are hosted. He examines the importance of domestic R&D, engineering capability, infrastructure, evaluation, talent and the ability to adapt or replace AI systems without creating new dependencies. Shrivastava also looks at India’s linguistic diversity, the economics of AI and the growing need for transparency and control as AI systems become increasingly autonomous. Here’s what he says:
What Sovereign AI Means for India
When I think about sovereign AI, I do not start with where a model is hosted. A model can run in an Indian data centre and still leave an institution with very little control over how it evolves. The licence may limit modification, the software stack may make migration difficult, and the expertise required to improve the system may sit elsewhere. Location matters, but it is not the same as capability. Data jurisdiction, regulatory control and local infrastructure all matter, but they are dimensions of sovereignty rather than substitutes for technological capability.
That distinction is becoming more important as India moves from being primarily a consumer of AI to becoming a creator of it. Indian startups and research teams are building models, language technologies, AI platforms and deployment capabilities of their own. This is an important shift. But a model launch, by itself, is not sovereignty. The deeper question is whether that capability can be improved, sustained and replaced without creating a new form of dependence.
For me, sovereign AI comes down to three engineering capabilities: the ability to operate critical systems reliably, the ability to improve the underlying technology, and the ability to change models or suppliers without unacceptable disruption. These are more useful tests than simply asking whether a system is domestic or imported. They focus attention on what an institution can actually do when technology, economics or circumstances change.
The economics of AI make this especially relevant. Stanford’s 2025 AI Index reported that the price of querying models matching GPT-3.5-level performance on one knowledge benchmark fell by more than 280-fold between November 2022 and October 2024. That does not mean every AI project became cheaper at the same rate. It does show how quickly a capability that once looked scarce and expensive can become widely accessible.
This creates an apparent paradox. If world-class AI is becoming cheaper to access, why build domestic capability at all? I see the answer in the difference between access and agency. Access gives us the ability to use today’s technology. Agency gives us the ability to adapt it to local needs, understand its limitations, keep critical systems running and shape what comes next. At the same time, sovereignty should not become an argument for rebuilding every layer of the technology stack. For many routine applications, the best engineering decision may still be to use the strongest global tool available.
The more durable opportunity is to build strength where it compounds: frontier research, efficient model development, high-quality datasets, shared compute, evaluation infrastructure and the engineering talent required to turn models into dependable systems. Innovation does not always mean training the largest model from scratch. It can mean making an existing model dramatically more efficient, designing a better architecture for a specific class of problems, or creating a deployment approach that works at India’s scale and cost constraints.
Language is a good example of why this matters. A 2023 NeurIPS study found differences of up to 15-fold in token counts between translated versions of the same text, depending on the language and tokenizer. Tokens are the units a model processes, and they can influence both latency and cost. The finding should not be read as a universal penalty on Indian languages today. Its significance is broader: seemingly small design choices deep inside an AI system can translate into very real economic differences for users.
India’s linguistic and cultural diversity therefore need not be treated only as a complexity to manage; it can become a research advantage. A system that performs well across regional languages, accents, code-switching, noisy audio and highly varied real-world contexts has solved a harder engineering problem. That is why product claims such as ‘supports multiple Indian languages’ tell us very little on their own. What matters is measured performance on the tasks and conditions people actually encounter.
This is where evaluation becomes central to sovereignty. Benchmarks need to move closer to real use: the quality achieved on an unfamiliar local task, the cost of completing that task to an acceptable standard, the amount of human correction required, and the ease with which the underlying model can be changed. In critical systems, even a periodic migration exercise can be revealing. If changing a model causes months of disruption, the system may be technically operational but strategically brittle.
The same principle extends beyond institutions to the people affected by AI systems. As automated decisions become more consequential, technical capability has to include traceability, explanation, contestability and a clear path to a responsible human. Sovereignty should expand choice, not merely change who controls the dependency. A domestic system that cannot be questioned, audited or replaced can be just as constraining as an external one.
There is also a physical reality that software discussions can easily overlook. AI depends on electricity, cooling, chips, networks and specialised software. No serious technology ecosystem is completely self-contained, and international partnerships will remain essential. The engineering objective is not isolation. It is to understand critical dependencies, develop credible alternatives where they matter, and retain enough knowledge and skill to keep improving the systems on which important services depend.
This challenge will become more significant as AI moves from answering questions to taking actions. Systems will increasingly be able to submit applications, coordinate workflows, purchase resources or trigger other software. At that point, sovereign AI is not only about who built the model. It is also about who defines its authority, how its actions are recorded, what limits are enforced and how that authority can be revoked. Control over AI will increasingly be a question of system design, not model ownership alone.
The opportunity for India is therefore larger than producing a national model or matching a global benchmark. It is to build an ecosystem that can create, adapt, evaluate and operate AI on its own terms while remaining connected to the best ideas and technologies in the world. Over time, that capability could also become an export strength: affordable AI systems that other countries and organisations can inspect, adapt and operate with greater confidence. That is the more meaningful idea of technological leadership – not simply owning more technology, but creating technology that gives its users more choices.