For months, the debate around Indian AI has pitted stakeholders into two predictable camps. One argues India lacks the compute to compete with the US and China on building frontier models and should focus on applications instead. The other insists open-source models have already made the race irrelevant.
Meanwhile, BharatGen has quietly been doing the work.
The IndiaAI Mission-backed consortium has secured access to 1,400 GPUs, joined a global alliance to build distributed frontier AI, assembled one of India's largest academic collaborations, and laid out an ambitious roadmap to develop a 1 trillion parameter model.
“Our intent is to have these models before March 2027,” Ganesh Ramakrishnan, Founding Director of BharatGen and professor at IIT Bombay, tells AIM.
Prof Ramakrishnan doesn’t see India’s AI journey as a binary choice between chasing frontier models and settling for smaller ones. Instead, he describes it as a continuum.
“Frontier minus one is a very important milestone. Frontier minus two is an important milestone," he says, referring to AI models that are one or two generations behind the industry maximum and echoing what Vivek Raghavan of Sarvam AI earlier told AIM. "The frontier will always be evolving."
The real objective, he argues, is to keep moving towards the frontier while constantly improving capabilities through cross-pollination of ideas, engineering practices, and research. That philosophy also explains why BharatGen is simultaneously working on current models while laying the groundwork for its trillion-parameter ambitions.
BharatGen’s Sovereign AI Strategy
The call for building a sovereign AI model has gained momentum ever since the US government restricted access to Anthropic's most powerful models—Claude Fable 5 and Mythos 5—for foreign nationals. Now, the Donald Trump administration has also reportedly asked OpenAI to delay the release of GPT-5.6.
For BharatGen, the roadmap to achieving AI sovereignty starts from models in the 100 billion parameter range, then to 500 billion parameters, before eventually reaching 1 trillion, provided compute continues to be provisioned.
That makes BharatGen one of India’s most serious frontier AI efforts. But Ramakrishnan believes the project is trying to solve something much larger. For him, sovereignty is fundamentally about participation.
“Sovereignty means you don't think of the last mile as the last mile, but bring the last mile as the first mile by design and not a retrospective effort,” he notes.
Only a few days after BharatGen announced it had joined Project Tapestry—a global open-source initiative by the AI Alliance with Yann LeCun to enable distributed, globally federated training of frontier AI models while retaining ownership of their data—many questioned whether collaborating globally somehow diluted India’s sovereign AI ambitions.
Ramakrishnan argues the opposite.
"The whole idea of federated learning is bringing the strength of complementarity," Prof Ramakrishnan notes. "The sum of parts greater than the whole."
Rather than every country independently repeating the same engineering effort, BharatGen hopes each participant contributes unique expertise while preserving control over its own assets. "What BharatGen and what India gains with Project Tapestry is the ability to move towards frontier AI models without having to do everything in one place."
That philosophy also mirrors India's broader digital public infrastructure story.
Just as UPI created interoperable payments rather than a single payments company, Ramakrishnan believes frontier AI can evolve through distributed collaboration instead of centralised ownership.
The Compute Question
India's compute deficit remains one of the biggest criticisms of sovereign AI.
Prof Ramakrishnan doesn't deny it. Instead, he questions whether endlessly chasing larger GPU clusters should be the objective. Thanks to the IndiaAI Mission, BharatGen now has access to 1,400 GPUs from the national pool of more than 38,000 GPUs.
"We must accept that there are going to be limitations. It's not even clear whether we should keep shooting an elusive target of infinite compute,” he shares.
Instead, he argues for decentralisation, bringing computation closer to where data is created and used.
"If you centralise too much, you're taking the source and consumption further away from each other,” he notes.
That philosophy also explains BharatGen's enthusiasm for Project Tapestry, where distributed model development reduces duplication while allowing countries to retain sovereignty.
What Comes Before Training
Several companies, such as Sarvam, Avataar, Tech Mahindra, Gnani, and others, are building foundational models. While most focus on the model itself, BharatGen, arguably, is obsessed with everything that comes before the model.
Prof Ramakrishnan spent far more time describing multilingual data pipelines, provenance, safety engineering, bias correction, compliance systems, taxonomy mapping, deduplication, benchmark contamination detection, and personally identifiable information removal than discussing model architectures.
The consortium has already collected around 22 trillion multilingual tokens as part of its data pipeline. "What BharatGen brings is a very unique combination of research and engineering,” he asserts.
He mentions sophisticated workflows for multilingual quality control, automated compliance, ethical scrutiny, and culturally grounded data collection. The objective is not simply to train a larger model. It is to ensure the model is trustworthy before training even begins.
"Data cannot be an afterthought. Data should not be treated as just a means. It could actually be a first-mile problem."
For Prof Ramakrishnan, sovereignty is less about keeping servers inside borders than understanding every stage of how models are built. "Sovereignty gives you alternatives,” he comments.
Building India's AI Through Academia
Unlike most frontier AI projects driven by a single company, BharatGen resembles a national research network. The consortium currently includes IIT Bombay, IIT Madras, IIT Mandi, IIT Kanpur, IIT Indore, IIIT Hyderabad, and several other institutions, backed by nearly ₹989 crore under the IndiaAI Mission.
But Prof Ramakrishnan says that is only the beginning. "We are a consortium of nine academic institutions. We are now trying to expand it... touching three dozen institutes."
One particularly striking example comes from IIM Indore. Instead of merely collecting datasets, researchers travel to smaller cities and libraries, educating custodians about the value of their archives before requesting participation.
That collaborative model has become something of a template for Project Tapestry itself.
Making AI Affordable
Perhaps BharatGen's biggest differentiator is that it is thinking beyond training.
Inference costs remain one of the largest barriers to widespread AI adoption, particularly in countries like India. According to a FinOps Foundation report, the average enterprise AI budget has grown from $1.2 million a year in 2024 to $7 million in 2026.
Prof Ramakrishnan believes future breakthroughs will come less from brute-force scaling and more from better representations, efficient architectures, and frugal engineering. "The real north star is making these models... much more compliant, removing a lot of the bloating," Prof Ramakrishnan said.
"Our recent studies have shown the cost per million tokens on our models is around ₹5." Comparably, DeepSeek’s V4 Flash model has lowered its input cost per million tokens to $0.14, or ₹13.
While BharatGen’s figure has not yet been formally published, the ambition is clear.
India's frontier models cannot merely compete on capability. They must also become affordable enough for nationwide deployment.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.