A year after being launched by a consortium led by IIT Bombay as the country’s first government-backed sovereign AI initiative, BharatGen secured over Rs 900 crore in funding by the Ministry of Electronics and Information Technology (MeitY). As the concept of ‘sovereignty’ in AI catches further steam, CEO of the non-profit, Rishi Bal in an interview with Poulomi Chatterjee, spoke at length about how the AI terrain is changing, why India shouldn’t distill AI models like China and what is next for BharatGen. Excerpts:

India’s constraints around building foundational AI models have been funding and compute. How has the progress been?

When we started our journey a couple of years ago, we were one of the first startups to get access to GPUs. Gradually, we started speaking about expanding to hundreds of GPUs. Now, discussions have shifted to thousands of GPUs. I am similarly optimistic about our trajectory in funding. But I think we just need to have some patience. We can do amazing things but they will also take time. There’s a lot of pressure to get things done yesterday. And the risk is that it will lead to short-term optimisation rather than long-term structural investment. We have to think about deep investment in research, papers, publication, data, people who actually have experience. There are a lot of building blocks. So, if the iceberg is the AI model, thats visible only because of everything underneath.

Why is India building AI models from ground up when advanced open AI models like Kimi-K3 can be built from distillation faster?

I think it depends on what your horizon is in terms of success whether its a six-month or a one-year window you optimise differently. If you look at a five-year window to make India can be genuinely competitive in AI, then you have to build in a structural manner. When you don’t build the technology from ground up then you don’t actually understand how it works. There is also a certain aspect around capability building, data, people, and long-term viability that’s involved. It’s like wanting to build an aircraft and then buying the engines from somewhere else, which of course can be done. But it doesn’t make you competitive in the long run. Secondly, there’s an an aspect of supply security where today’s open models can easily be tomorrow’s closed models. There’s already discussions around whether Chinese AI models should publish their research. If you want to have control over how your AI works and behaves and if you want to have a national capability it is essential that we build our own models.

What is the kind of demand that BharatGen models are seeing and from which sectors?

Demand has been very encouraging for us as there’s a lot of interest from both enterprises and government for our models. We are engaged with the central, state governments and verticals like education and BFSI to help solve real-world problems. Most use-cases are around accessing large data sets, processing and creating documents, quickening workflows and accelerating productivity. Our AI models have been downloaded two lakh times on Hugging Face. There’s also seeing a lot of interest from the open source community. This is just the beginning- we’ve got a lot of interesting work in the kitchen right now and we will continue to roll out. There are no plans currently around a coding-specific model. For now, we are sticking to creating niche domain-specific models.

Our Param-2 models are very popular but we have speech and voice AI models that have found usage in enterprises and governments. We will continue to release a steady stream of AI models in the next quarter.

What are the obstacles around building AI in a price-sensitive market like India? How will this shape the landscape?

Generally, cost cannot be an afterthought when building in India. Companies that are simply trying to take LLMs and adapt them to work in India will struggle. Pricing has to be built into their business models. The challenge is also obviously to build a sustainable business because you can get users to sign-up at a subsidised price but when that number is hiked, will they stay? We will continue to see a significant depreciation of cost. There will also be new hardware and architectures that will enable this.

Additionally, when you’re operating in India, you will have razor thin margins. Hopefully, AI firms in India will be able to scale successfully like other great manufacturing or other legacy companies we have. Lastly, not all of coding requires a frontier level model. I think what will emerge is that we will have smart routing so varying levels of complexity can be intelligently routed to appropriate AI models. Smart harnesses in the background will be more popular and the world will look more heterogeneous.