India’s AI sovereignty debate may have been looking at the wrong thing. The question is not simply whether India can build its own AI models. It is whether India can remain in control of the technology those models depend on.
Abhishek Bhatt, Secretary General of the Bharath Digital Infrastructure Association (BDIA), has a striking analogy for where India’s real strategic leverage could lie. He describes data as India’s “Strait of Hormuz”.
“India’s strait of Hormuz is data. We don’t build compute. We have to be transparent. We can build softwares because Indians and India in general has been good at softwares. So we can build LLMs. There’s no doubt about it. But for building LLMs, we need compute,” Bhatt tells AIM.
The analogy gets to the heart of India’s AI sovereignty problem. Just as the Strait of Hormuz is a critical chokepoint for global energy flows, data could become one of the most important strategic assets in India’s AI economy. But controlling data alone will not be enough if the compute, cloud, hardware, connectivity and deployment layers remain dependent on systems that India cannot control.
Bhatt argues that sovereignty should therefore not be confused with owning every piece of technology.
“Sovereignty is not AI. Sovereignty is about control,” Bhatt says, pointing to five critical layers of the technology stack—foundational connectivity and subsea cables, hardware, cloud, applications and AI.
No country, he argues, can realistically ensure complete sovereignty across all these layers. The objective is to retain enough legal, economic, technological and policy leverage to control what matters most.
Sourav Jena, CEO, Coredge, takes a similar view. “Yes, but I would not frame it as a choice between owning models and owning infrastructure. Sovereignty is becoming a question of control across the AI value chain,” Jena explains.
“I would define AI sovereignty as the ability to control the critical dependencies that determine how AI is built, trained, deployed and used—rather than simply where the original model was created,” Jena says.
Compute Remains India’s Biggest Weakness
India’s model-building efforts are already changing the narrative that the country has no domestic AI capability. Under the IndiaAI Mission, 20 indigenous AI models are in development, including 12 LLMs and eight smaller models, with five already released.
Sarvam has open-sourced models with 30 billion and 105 billion parameters, while BharatGen and Gnani are developing indigenous AI systems.
But building models is only one part of the problem.
“AI needs three things. One is the compute, other is the data, and third, the algorithm, the power to run that. These are the three input raw material for any AI,” Bhatt states.
For India, compute remains a particularly difficult dependency because the economics extend far beyond GPUs.
“The first challenge is compute economics. Sovereign AI requires affordable and reliable compute at scale. GPUs are only one part of the equation — power, cooling, networking, storage, utilisation and inference efficiency ultimately determine the cost of AI,” Jena says.
This means that simply deploying thousands of GPUs does not automatically create sovereign AI. If compute remains too expensive, unreliable or inefficient for sustained use, domestic AI capability could still depend on external infrastructure and economics.
Namit Sureka, Chief Analytics & AI Officer, Straive, argues that the debate itself needs to move beyond models versus data.
“The framing of data-versus-models is a false binary. Sovereignty is built across four layers: models, data, compute, and the deployment layer where AI becomes work. India is building all four.”
India does not need to build everything
India is increasingly building its own models, but there is also a question of how much of the frontier it should try to reproduce.
“India should build foundational and domain-specific models, but we should be selective about where we invest. We do not need to recreate every frontier model at enormous cost simply to establish ownership,” Jena clarifies.
That could mean prioritising Indian languages, local context and strategically important sectors rather than attempting to reproduce every frontier model built by the largest technology companies.
The economics make this distinction important. Frontier AI development costs billions of dollars, while the technology itself is moving quickly.
“In a category where frontier development costs run into billions and the leading model changes every quarter, whatever India spends today can be recovered by a well-timed wait,” Sureka states. The more immediate question, therefore, is what India needs to control even when it uses technology developed elsewhere.
Data Residency Does Not Equal Sovereignty
Data is central to that question, but keeping data inside India does not necessarily mean India controls the AI system using it.
“It’s definitely not about data residency. Data residency has got nothing to do with sovereignty or with respect to control,” Bhatt says.
A system can physically operate in India while relying on foreign-owned hardware, software, cloud infrastructure, intellectual property or technology providers.
Jena points out that sensitive data does not necessarily have to be embedded into model weights. Retrieval, private inference and fine-tuning can allow models to work with controlled datasets.
But those mechanisms create another dependency if the infrastructure handling them remains outside domestic control.
“If these techniques only preserve sovereignty if the retrieval and inference layers themselves sit under domestic control. Otherwise, you haven’t removed the exposure; you’ve simply moved it — from the model’s weights to every individual query,” Jena says.
For government and enterprise systems, this distinction becomes particularly important. Keeping data in India means little if every query, inference request or retrieval operation can still depend on an external provider.
Bhatt identifies four levers that determine whether India can actually exercise control.
“Do we have the requisite enabling legal structure to exercise control? That is the first part. Do we have the economic levers to exercise control? Means if you want to operate out of here, you need to pay taxes. Do we have technological levers? And the fourth is do we have the policy levers?”
The Bigger Risk is Being Switched Off
The concern around foreign technology is ultimately less about where a company is headquartered and more about whether access can be withdrawn.
“If the access to a technology can be cut off, so that is a big reminder that why only that technology? Even the internet technology, even the Chrome that you are using, these access can be pulled off.”
Tarun Nandwani, CEO, Newgen, makes the same point about critical infrastructure.
“All these will need to be tried, and we will find the right fix, because we cannot depend on foreign infrastructure for sure. All governments have to do that to take care of, because see tomorrow if somebody switches off a plug for you. You cannot leave your systems non-functioning,” he states.
But sovereignty also has an economic limit.
“I think the direction is absolutely required and the investments have to go there. But finally, we have to find the right balance with the cost and the output,” he adds.
That is where India’s AI strategy becomes more complicated. The country needs domestic capability, but it cannot realistically isolate itself from global AI research, hardware, cloud platforms, capital and technology.
“The objective is not self-sufficiency. It is strategic autonomy: the ability to choose the best technology without surrendering control of the critical dependencies around it,” Jena states.
India’s strength in software and applications also does not automatically translate into expertise across the deeper layers of AI. Building sovereign capability means developing the ability to build, adapt, optimise, secure and operate systems when required.
Control first, indigenous capability next
For Bhatt, India does not have to wait until every component is indigenous before it can exercise sovereignty.
“I’ll give you a simple analogy. Mercedes is not our engine. But when it runs in India, who exercises control over it? Government of India,” he adds.
He applies the same logic to AI.
“So whoever built the model is definitely wishful thinking that we, it’s a good thing that we should have our own domestic model, that is absolutely right. But till we don’t get there, there are various means and mechanisms to ensure that control is there. So we first must achieve control. Then the next priority is it should be indigenous,” he states.
That puts the real AI sovereignty question somewhere beyond the model itself. India can use the best global models and still pursue sovereignty, provided it has enough control over the data, infrastructure, compute, deployment and critical dependencies that sit around them.
The model may be the most visible part of AI. But for India, the more consequential battle could be over everything that allows that model to operate.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.