In the Race to Build Sovereign AI, is India Missing Out on Easy Wins?

Sovereignty is a spectrum. At one end are countries that control every layer of the AI stack—from silicon and compute infrastructure to foundation models, talent, and applications. 

Somewhere in the middle are countries like China, which have built world-class AI models but still face constraints from export controls and limited access to the most advanced chips.

At the other end are nations that remain largely dependent on foreign technology, whether for cloud infrastructure, AI models, or silicon. That’s where India, and its application-heavy market, lies. However, this poses clear strategic vulnerabilities, as was the case when the US government restricted access to Anthropic's Fable. 

This has reignited calls for India to achieve digital sovereignty. But a more nuanced debate concerns the level of sovereignty India should pursue.

Since the ₹10,372-crore IndiaAI Mission was approved a couple of years ago, the government has largely focused on backing startups that are building indigenous foundation models. Critics, however, question whether that should be the immediate priority.

If building frontier models from scratch is ultimately a function of capital, compute, and talent, can India simultaneously post-train open-weight models, adapting them with Indian data and deploying sovereign AI applications at scale?

Sovereignty Through Access, Not Ownership

In an interaction with AIM, Pranesh Prakash, a tech policy researcher and affiliated fellow at Yale Law School, argues that sovereignty is not about where a model is built, but about whether a country has unrestricted access to use, modify, and deploy it. “At a minimum, what is required for digital sovereignty when it comes to AI is access to open models. Whether a model is Indian-made or not does make a difference in that regard,” he says.

In contrast to proprietary reasoning models, open-source and open-weight models allow users to freely download, modify, fine-tune, and deploy them without requiring permission from the original developer.

This raises a broader question as to why India has not yet produced a widely adopted AI product built on top of open-weight models and adapted for Indian users and enterprises.

India has already experimented with this approach. AI4Bharat’s Hindi chat model Airavata, Sarvam’s Hindi model OpenHathi and audio language model Shuka, and more recently, multilingual model Sarvam-M were all built through varying degrees of adaptation and post-training of existing open models such as Llama and Mistral. Yet none have achieved mass adoption among consumers or enterprises on the scale of ChatGPT or Claude.

India's Missing Layer: Productisation

India now has more than 100 million weekly active ChatGPT users, making it the company's second-largest market globally. Users under 30 account for nearly 80% of activity on the platform, while work-related tasks make up roughly 35% of all messages sent by Indian users.

Much of India’s private sector continues to rely on foreign AI providers, even as government-backed platforms such as BHASHINI support 36 Indian text languages and process billions of translation, speech, and text requests.

Startups such as Meesho, PhysicsWallah, and Cars24 have adopted ElevenLabs for multilingual voice experiences despite the availability of Indic-language alternatives. Even Prime Minister Narendra Modi’s podcast conversation with Lex Fridman was translated into Hindi using ElevenLabs, while companies including ixigo, Bajaj Finserv, Unacademy, MakeMyTrip and Practo are building integrations around OpenAI models.

Prakash argues that the hardest part of the AI stack is not model creation but productisation. “People seem to think that once a highly intelligent AI model is available, everything else automatically falls into place. That's simply not true. The practical application of AI across different sectors is still extremely hard,” he says.

Even as the government commits to building indigenous AI capabilities, it is ironically underutilising a growing pool of highly capable open models that are readily available.

“We already have resources available at no cost,” says Prakash. “There are low-hanging fruits that just haven't been plucked, and people are thinking about building skyscrapers.”

Using cybersecurity as an example, he argues that India, instead of seeking access to Mythos, could explore what could already be achieved with open models. He points to recent work by Linux kernel maintainer Greg Kroah-Hartman, who used a completely offline setup on a consumer-grade Framework Desktop powered by an AMD Ryzen AI Max+ processor to uncover complex bugs on the Linux kernel, which models like Mythos are expected to exploit

The debate, however, often only focuses on foundation models rather than the silicon that powers them. 

Despite progress under the India Semiconductor Mission and investments from Tata Electronics, Micron, and others, India remains years away from producing leading-edge AI chips, let alone building foundational models on them. 

“There is no open-source silicon,” says Prakash. “Even if you have ISAs like RISC-V (which allows anyone to design and manufacture chips), knowing how to build them is hard, and getting the technology to do that is hard. ASML is a monopoly for a reason. TSMC and Samsung are in the positions that they are in for a reason.”

The Case for Building Foundation Models From Scratch

One company attempting to build sovereign AI models is Sarvam, a beneficiary of both IndiaAI funding and private venture capital.

Recently, it raised $234 million in its Series B funding round, and its multilingual AI chatbot Indus has crossed 1 lakh downloads on Google Play Store. While users yearn for more features and an adaptive interface, it has largely been met with positive reviews.

The company also released open-weight large language models like Sarvam-30B, Sarvam-105B, and BulBul, which it claims, based on its own testing, outperforms ElevenLabs on tasks like handling complex Indian speech, code-mixing, and accurately pronouncing regional names and numerics.

In a recent interaction with AIM post the funding round, CEO and Co-founder Vivek Raghavan outlined reasons for when building a foundational model from scratch should take precedence over post-training open-weight models. However, he agreed that there is no single answer to the debate, noting that training from scratch is necessary under certain conditions.

The first factor is linguistic. “If you actually have a tokeniser that is better suited to Indian languages, then it becomes more efficient for queries that are in Indian languages,” he said, explaining that a model’s tokeniser cannot be modified later during fine-tuning.

Tokenisers are part of the model’s core architecture, and they cannot be modified once the model has been trained, at least not without retraining the model from scratch or extensively retraining it

While Raghavan agreed that post-training open-weight models will be sufficient in ‘99% of use cases’, he arguef they are not feasible for national security use cases. “My definition of a sovereign model is a model where the sovereign can actually audit the data that was used to train the model,” he noted.

While open-weight models allow organisations to host, run, and modify the model locally on their own infrastructure without the risk of an external provider revoking access, they do not enable a complete audit or verification of the underlying dataset used during the initial pre-training phase.

Raghavan cited Anthropic research conducted with the UK AI Security Institute showing that as few as 250 poisoned documents can implant a hidden backdoor into a model. Because the attack depends on the absolute number of corrupted documents rather than their proportion within the dataset, he argues that full visibility into pre-training data remains important for national-security use cases.

However, the Sarvam co-founder acknowledged the value of using existing open-weight models after training to solve business problems, but noted the significant technical hurdles involved. However, he also argued that effectively adapting and reinforcing models at scales exceeding 100 billion parameters remains a difficult engineering challenge that only a few organisations can currently execute.

Beyond the technical complexities of post-training, Raghavan challenged the prevailing view that India should focus exclusively on small models, advocating instead for developing "frontier minus one" systems.

As global frontier models reach scales of 3–10 trillion parameters, he argued that India must aim to build at least a trillion-parameter model to achieve comparable performance for local, specialised use cases. Post-training a standard 7-billion-parameter model simply will not suffice to bridge that gap.