India must build the foundational models underpinning the artificial intelligence (AI) age or risk becoming merely a consumer of the disruptive technology, according to the country’s only not-for-profit AI entity, which is also the largest benefactor of the government’s $1.2-billion AI mission.
BharatGen, a consortium of AI labs across nine institutes, was set up by Indian Institute of Technology (IIT) Bombay in December 2024 as a non-profit, and received a ₹1,058.82 crore grant under the Centre’s AI mission on 18 September 2025. The entity is developing one of India’s first sovereign large language models (LLMs)—foundational AI systems that mimic human writing and speech by analyzing vast amounts of data.
“We can only truly understand AI models when we build them from scratch,” said Ganesh Ramakrishnan, chair professor for computer science engineering at the IIT-Bombay and founding director of BharatGen.
He explained that just as knowing a recipe allows you to tailor a dish to your health needs, knowing exactly what goes into an AI model, especially its data, is essential for building effective tools around it.
“For AI adoption to scale, users have to move from the last mile to the first mile. An ecosystem built solely around using AI tools will remain at the consumption layer. We need people who understand, build and improve the technology itself.” he added.
Growing chorus
Ramakrishnan is among a growing number of industry leaders, including Anand Mahindra, chairman of Mahindra and Mahindra, that have said India needs to build its own LLMs rather than depend on those built by global giants such as Google, OpenAI and Anthropic.
Mahindra wrote in a letter to shareholders, published Wednesday, “Frontier AI is not just a commercial technology, it is a strategic capability, increasingly being shaped by questions of trust, regulation, national interest and sovereignty. India cannot be only a consumer of intelligence built elsewhere. It must also be a creator, shaper and trusted deployer of intelligence for its own society and for the world.”
The push for sovereign AI models in India gained momentum after 12 June, when the US restricted access to Anthropic’s latest models outside the country. The move, enacted for "national security" reasons, has sparked widespread concern over potential future blockades.
While India’s ₹10,372-crore AI Mission was founded with the objective of building foundational models using public funds, progress under the mission has received mixed responses from the industry. In just over two years since being notified, the mission has backed 12 startups with ₹2,194.29 crore in committed graphics processing units (GPUs) and cash grants. Key benefactors include India’s first AI unicorn Sarvam, listed data analytics firm Fractal, and Makers Lab, a subsidiary of Tech Mahindra.
BharatGen has received access to 1,400 GPUs through the mission, Ramakrishnan told Mint. The company has used this to release Param-2, a 17-billion-parameter foundational model trained from scratch. “The next step is scaling to 70-100 billion parameters and eventually to one trillion parameters. Alongside this, we’ve also built Sooktam2 for text-to-speech, Shrutam2 for speech-to-text, and Patram for understanding documents. The next generation of all these models is already in development.”
Following this development, global open-source AI consortium AI Alliance invited BharatGen to showcase India’s advancements on the world stage on 21 June. In an official statement, the alliance noted that BharatGen will anchor India’s initial participation in Project Tapestry, a global open consortium dedicated to building frontier AI. Project Tapestry is led by Yann LeCun, one of the world’s foremost AI researchers and the former AI chief of Facebook parent Meta Platforms.
Ramakrishnan declined to comment on the net cost of building the models, noting that the development was largely funded by grants. He said BharatGen currently employs a multidisciplinary team of researchers, engineers, and linguists to build text, speech, and vision AI models.
However, talent remains scarce. Ramakrishnan said that rather than just being AI users, students need more exposure to the core engineering required to build foundational models. While he acknowledged that top AI talent is currently concentrated within the premier IITs, he emphasized that the BharatGen consortium aims to democratize this expertise by bringing more institutions into the fold.
Plugging the AI gap
Many industry stakeholders said BharatGen could solve a core industry challenge for India’s AI push. “It’s not just about building foundational models; BharatGen is focusing on developing AI in local languages, which is where there is a gap. There aren’t many AI models that can power voice conversations in local languages,” said Jibu Elias, country lead for responsible computing at non-profit internet services body Mozilla Foundation.
Kashyap Kompella, an AI analyst and founder of tech consultancy RPA2AI Research, noted that BharatGen could focus on research in the long term, distinct from Sarvam AI's role in India's sovereign AI mission. “BharatGen should not be trying to do the same thing (as Sarvam). Sarvam, being a venture-backed private company, is naturally focused on taking AI products to customers and commercializing technology. BharatGen’s strength lies in building the foundational layers of the AI ecosystem—creating datasets, benchmarks, research and open-source technologies,” he said.
Elias, however, said that alongside BharatGen, India’s AI ecosystem needs a holistic overhaul. ”We need infrastructure, the right talent, data ecosystem, and energy capacity. That's the challenge. Our talent pipeline is a mess: we keep saying Indians are third in terms of AI papers published, but many of those achievements come from our researchers leaving India because they don't have the funding or R&D facilities here,” he said. He added that India lags well behind the US and China on sovereign AI.
“We need many things, not just language models. We need patient capital. In India, we often fund replica startups—Uber being copied as Ola, or quick-commerce ventures. The idea that building data centres and language models automatically gives you sovereignty is misplaced.”