Sovereign AI In 2025: India’s Search For Homegrown LLMs

Till the end of 2024, India’s stance on building homegrown foundational or large language models (LLMs) was ambiguous. One year later, all that has changed. Sovereign AI and homegrown large language models became a dominant theme for AI in India in 2025. 

Despite the government announcing the INR 10,300 Cr IndiaAI Mission early last year, timelines for investment were vague, execution plans were still evolving, and many of India’s top tech leaders openly argued against India’s need for building LLMs.

But 2025 changed the equation. Beyond the buzz around agentic AI, the defining theme in India’s AI landscape this year became the government’s hunt for LLM builders. The urgency intensified after China’s DeepSeek unveiled its indigenous DeepSeek-R1 at dramatically low training costs, forcing India to move faster and with more clarity.

Soon, the IndiaAI Mission began identifying and backing companies capable of building sovereign models. 

The first set of selections included Sarvam AI, Soket AI, Gnani.ai, and Gan.AI. The next tranche expanded to both enterprise and smaller players, such as BharatGen, Fractal Analytics, Tech Mahindra, Avataar AI, Zeinteiq Aitech Innovations, Genloop Intelligence, NeuroDX (Intellihealth), and Shodh AI.

Though most of these startups and enterprises are yet to launch their fully developed indigenous LLMs after months of groundwork, the ecosystem insists progress has been significant.

Ganesh Gopalan, cofounder and CEO of Gnani.ai, told Inc42, “We need to approach the matter from this lens: not just India, but every country in the world are forming their sovereign AI strategies. To some extent, this has to do with geopolitics, while it’s also how several government departments are seeing advantages in it. Investments in data centres are increasing too. So, all the preparations are on track for what is to come.”

Some examples include BharatGen, which publicly rolled out its first 3-Bn parameter model in July this year. Or Sarvam-M, a 24-Bn parameter open-weight hybrid language model was launched in May, albeit built on top of Mistral Small. 

Fractal also launched its 14-Bn parameter open source LLM built with Deepseek. Other similar innovations fostering India’s sovereign AI ambitions have also accelerated this year.

But before looking ahead as to what awaits the country, it is worth looking back at how India has progressed so far, the delays and preparations, and how 2025 pushed India to start competing with Silicon Valley and other global giants on this front.

The Emergence Of India’s LLM Brigade

When the GenAI wave swept through the world, India was late in catching up with the pace. However, as the country started realising the mammoth impact of the wave, the first thing that the country focussed on was building applications leveraging the largely available foundational models like Gemini, OpenAI’s GPT, Llama, and others. The private capital also flowed in there.

Despite the huge cost of building the LLMs and hardly any government support at that point, many students and alumni of the IITs strongly believed in the urgency for India’s sovereign models. 

With geopolitical tensions on the rise and the GenAI-induced reality check, the Indian government has also started prioritising the matter. 

So far this year, all 12 companies building foundational models – from LLMs to domain-specific SLMs – have been supported by the IndiaAI Mission with access to graphic processing units (GPUs), which is the lion’s share of the cost in building LLMs, and via some grants and other means.

Soket AI, which aims to release a 120 Bn parameter open-source text model within the next 12 months, has already been allocated around 1,500 GPUs and promised grants in the form of both equity and debt, its founder and CEO Abhishek Upperwal told Inc42.

IIT Bombay-incubated BharatGen has also received GPU allocation and grants to support its other costs from IndiaAI Mission. In fact, before MeitY, this not-for-profit organisation had raised INR 235 Cr (around $30 Mn) from the Department of Science towards the end of last year, which has funded its first 3-Bn parameter model building.

Currently, it is in the process of building a 14 to 20 Bn parameter model, which, as per its executive vice president Rishi Bal, is being built completely from scratch, with a quarter and a half of the data coming from Indian sources, compared to the foreign models that are built with around 2-3% of Indian data.

“We have a team that’s on the ground approaching publishers and different data sources in India for data to build these models. So, we’re trying to grow a data ecosystem in the country altogether,” Bal told Inc42.

Meanwhile, Gnani’s Gopalan, without revealing many details, also emphasised that it has received all the compute support from the IndiaAI Mission to build its 16-Bn parameter speech-to-speech model from scratch, which is one of its kind, globally.

While most of these companies are in different stages of their building phase, and some are building from scratch, others, including Sarvam and Fractal, are building on top of other LLMs, the enthusiasm is unmistakable. 

However, with a new ecosystem emerging, challenges are meant to follow, as they have in this case, too.

Addressing The Gaps & Risks

The year has been as much about bridging the gaps in India’s sovereign AI mission as acknowledging the ones that persist. The government has tried to ease one of the biggest bottlenecks – compute – by facilitating GPU access for the selected LLM builders and reducing the cost burden of model development. Yet, as several startups told Inc42 earlier, access to cutting-edge GPUs of the kind used by global AI leaders remained limited.

Though this gap is gradually narrowing and model builders now have access to one of NVIDIA’s top GPUs, H100 clusters, the overall supply might become uneven from time to time. A few companies with early approvals or deeper partnerships can secure large GPU blocks, but emerging startups could find themselves waiting in line.

And the selection numbers make this visible. Out of more than 500 applicants, only 12 companies have been chosen so far by the IndiaAI mission, underscoring the emerging resource constraint that the LLM builders in India might face.

Interestingly, fellow startups are also playing a role here in curtailing any imbalance. Soket AI said that of the 1,000+ GPUs allocated to the company, around 25% are in use. 

“We didn’t want to take 100% of the allocation on day one, as it might lead to waste. We are taking a subset of the total capacity, setting up our testing and training stack, and then we will gradually scale up,” said the startup’s CEO Upperwal.

A unified effort by almost every stakeholder is also visible to address two of the other interrelated areas of challenge — capital and talent.

Due to their limited fund sizes, private investors showed less appetite for funding LLM builders, preferring to back AI applications instead. In India, this capital gap had to be filled by the government — and to a large extent, its interventions (with GPUs and smaller grants) have done exactly that.

“We do recognise that these investments by IndiaAI Mission are a great start, but it is still much smaller and not at the same scale of leading AI economies like the USA or China… Although India started with a slow base, we have seen decent progress in the amount of effort in building Indic models and some positive movement with the release of some of these models,” says Suraj Amonkar, Fractal’s chief AI research and platforms officer.

In fact, interacting with the top LLM builders, Inc42 found that the Indian players recognise their limitations in capital, and hence, are building teams and every block of the puzzle more mindfully. Talks around retaining and improving India’s tech talent are also gathering steam. 

The Talent Hunt & Retention

NITI Aayog, for instance, has noted in a recent report that the tech workforce in India faces headwinds such as productivity improvement, competition from onshore AI-native service firms, and macroeconomic geopolitical instability.

In fact, amid the increasing job losses in the country and the past-paced AI adoption, it has proposed a unified “India AI Talent Mission” that will engage with IndiaAI Mission. 

This proposed talent mission has three targeted recommendations for government and academia — become a global AI talent magnet, build an AI skilling engine for the current workforce, and embed AI in the education system from school, undergraduate programs.

But how far is this achievable, and how could this foster the sovereign AI model-building endeavour?

The AI native startups in India have little doubt that India has solid tech talent. These companies are now trying their bit to secure them, despite the continuing departure of top talent to other countries.

BharatGen’s Bal said that the reason the entity was started as a consortium with IITs across Bombay, Madras, Kanpur, Kharagpur, and other top Indian universities was to ensure that the R&D talent from universities could join to build India’s foundational model.

“It’s not just about building models. For us, it’s about raising an ecosystem of people that can lead the path for India to be a creator. By partnering with all of these institutes, we have found people who have turned down PhDs and Master’s programs abroad and chosen to stay here with BharatGen.”

Gnani’s Gopalan says retaining the country’s research talent is one of the biggest challenges as tech giants outcompete startups in terms of pay. “We expect new institutions to come up in India, many of them in coordination with the global colleges, to offer world-class Master’s and PhD education in the research field. As we speak, some of the US colleges are also looking to open campuses in India.

But Soket AI’s Upperwal sees a silver lining. “The current shortage gives us an opportunity to learn things from scratch, just like the Western nations did.”

India’s AI Economy Takes Shape

The momentum can’t slow down and the existing gaps must be bridged because building sovereign models is no longer an option for the country, but a must, the experts believe unanimously.

And in that endeavour, they say, building LLMs is not enough. Smaller, domain-specific language models can be a big boost for the Indian AI economy.

Fractal, for instance, has been selected for an initiative to build reasoning models that will have the ability for advanced intelligence in STEM areas and in healthcare-based reasoning. Amonkar says that the company aims to build and deploy its models with these abilities in 2026.

NeuroDX is also building a clinical LLM while Zeinteiq is working on a large-scale AI model for scientific simulations. 

In fact, the second batch of selection by the IndiaAI Mission is more domain-specific than the first batch that are doing both LLM and domain-specific SLMs.

“This approach is fantastic for India’s future sovereign AI capability. Large models are absolutely necessary for generalisation, to make artificial general intelligence (AGI) a reality in the future, but small models or domain-specific models have a big role to play from the usability point of view,” added Upperwal.

The ecosystem players are also stressing on continuing efforts of public-private partnerships to foster the growth ahead.

BharatGen’s Bal believes that the private-public partnership has been a good starting point in terms of what an India model of building LLMs can look like.

“Most models that are built in the US are on the backs of profit centres, like Google, Amazon, that are making profits on one hand, and using it to subsidise the creation of these models. That’s the American way — a high capital system, big risk takers. The Chinese model appears to be more of a closed ecosystem with a very competitive market that is funded by the government. In India, we need to build our model of building models, and we have got examples like UPI — a highly successful public-private partnership.”

If India can keep strengthening this model, highly subsidised US-made AI LLMs entering the country cannot kill the local ecosystem. Each AI stakeholder has to keep investing in R&D to remain competitive in deeptech and create the right conditions for innovation.

[Edited By Nikhil Subramaniam]