Daybreak - India's AI mission could be transformative. But it's dead set on becoming Deepseek

Episode Date: April 2, 2025

Back in January, when China’s Deepseek R1 model stunned the world with its performance and low training cost, India was thinking only one thing – how do we beat it? How do we become a global AI superpower? But when it comes to the AI race, India has been stuck at the starting line for quite a while now. Its approach has largely been to throw things at the wall in the hope that something eventually sticks. Now, Deepseek  has really amped up the pressure. India’s electronics and IT ministry, or Meity, has swung into action. It has been announcing housekeeping steps for the country’s year-old AI mission at a speed that can match the language model advances hitting the headlines. But in the process, the actual goal of the mission has become more incoherent than ever. Tune in. Daybreak is produced from the newsroom of The Ken, India’s first subscriber-only business news platform. Subscribe for more exclusive, deeply-reported, and analytical business stories.

Hi, this is Rohan Dharma Kumar. If you've heard any of the Ken's podcasts, you've probably heard me, my interruptions, my analogies, and my contrarian takes on most topics. And you might rightly be wondering why am I interrupting this episode too? It's for a special announcement. For the last few months, I and Sita Raman Ganeshan, my colleague and the Ken's deputy editor, have been working on an ambitious new podcast. It's called Intermission.

We want to tell the secret sauce stories of India's greatest companies. Stories of how they were born, how they fought to survive, how they build their organizations and culture, how they manage to innovate and thrive over decades, and most importantly, how they're poised today. To do that, Sita and I have been reading books, poring over reports, going through financial statements, digging up archives, and talking to dozens of people. And if that wasn't enough, we also decided to throw in video into the mix. Yes, you heard that right. Intermission has also had to find its footing in the world of multi-camera shoots in professional studios, laborious editing, and extensive post-production. Sita and I are still reeling from the intensity of our first studio recording.

Intermission launches on March 23rd. To get an alert as soon as we release our first episode, please follow Intermission on Spotify and Apple Podcasts or subscribe to the Ken's YouTube channel. You can find all of the links at the ken.com slash I am. With that, back to your episode. Back in January, when China's DeepSeek R1 model stunned the world with its performance and low training cost,

India was thinking only one thing. How do we beat it? How do we become a global AI superpower? Now the thing is, this is AI we're talking about. This is a technology that is evolving at breakneck speed. Every week, if not every other day, some massive breakthroughs are being reported. And India, meanwhile, has been stuck at the starting line for quite a while now. Its approach has largely been to throw things at the wall in the hope that something eventually sticks.

Multiple experts who have contributed to the government's India AI mission project even attested to the sentiment. But now, Deepseek has really amped up the pressure. India's Electronic and IT ministry, or METI, has swung into action. It's been announcing housekeeping steps for the country's year-old AI mission at a speed that can match the language model advances hitting the headlines across the world. But in the process, the actual goal of the mission has become more incoherent than ever before. A computer science professional we spoke to who participated in a high-profile 20-profile 20% global virtual consultation organized by the ministry said as much.

They said that the India AI team were in a hurry to spend money. They also seemed to know how to spend the money, but they didn't actually have a vision. They didn't know what will come out of all of this. And that was in November. The election year budget had sanctioned $5.51 crore rupees for the mission. And it had to spend that money before 31st March, the end of the financial year, lest it laps.

And it did, because in the end, the government ended up spending only 173 crore rupees. The government has earmarked 2,000 crore rupees for the program for FY26. Now, the idea, according to people associated with the ministry, is to have something between the NPCI or the

National Payments Council of India and Isro. In the mix, they also throw in manufacturing's production linked incentives or PLI as a model. But turns out, the actual AI mission is none of the above. Welcome to Daybreak, a business podcast

from the Ken. I'm your host Rahal Philippos and I don't chase the news cycle. Instead, every day of the week, my colleagues, Nikda Sharma and I will come to you with one business story that is worth understanding and worth your time. Today is Wednesday, the 2nd of April. The Indian Space Agency is basically the poster child of efficient governance. Isro knows how to get things done. But the difference is that for a good part of Isro's work, the science is clear and so are its goals. With AI, meanwhile, particularly with large language models or LLMs, which METI is so hung up on, things are a little more fuzzy.

Neither is the science clear, nor has India laid out any problem statement. So it's been calling for people to come forward with proposals, but that whole effort can easily be summed up with one word, BYOP. Bring your own problem. There's clearly a need to step back and look at the big picture, instead of just focusing on fixing a problem that may or may not exist. Jan Lecun, Meta's chief AI scientists, hinted at just that when he visited India in October.

He said Indian PhDs and scientists also need to focus on AI research, apart from just engineering and development. Now, contrasts this with NPCI, which is a quasi-government body with large banks as its shareholders. It also has a clear goal, to enable digital payments and create a secure, payments infrastructure. With PLI, it's even more precise. Companies applying for the incentives know exactly what they are committing to make. Then think about Métis plan of action in comparison.

It has one loosely defined end to take on the likes of OpenAI's Chad GPT, Google's Gemini and China's Deep Sea Car 1. That could mean either building one killer LLM or a bunch of foundational models before the year ends. For context, foundational models are trained on large data sets of text and code, whereas LLMs are a subset of more versatile foundational models and are trained on text data sets to generate grammatically correct text. Both contribute to enhancing chatbot capabilities. Now, the ministry has said that it is simply seeking proposals from anybody who wants to build models and in turn it will put its money behind them either as a grant or by seeking equity. If you look at the India AI website, you'll see some broad processes,

things like the government will not own any intellectual property or asset. If it's a startup then it should have commercialization and long-term sustainability plans in place and so on. But we spoke to at least a dozen startups and large companies in the space and no one had the slightest clue about what was going on. What is interesting though is that the Méti has been distancing itself from Bharajen, a foundational model initiative at IIT Bombay, which is widely being touted as India's answer to Deepseek. It even got a $2.35 crore-rupy grant from the government science and tech department. We learned that Ganesh Ramchrishan, who leads Bharajan, has been building at a top metas open-source model Lama for several months now. It's essentially a multi-modal large language model initiative specifically for Indian languages.

But beyond that, there isn't a lot of clarity on what it is. Then there are also models like Bharatj GPD, Sarvamai's Openhati, Ganga and others under development at different. institutions and companies. But the thing is, across the world, there are now anywhere between 50 to 100 open source pre-trained models which are only getting better. So the question here is, should India put its already scarce resources behind building new models, or should it focus instead on post-training the models towards more focused outcomes?

Unfortunately, that isn't really the goal here. One person from Méti told us that of the total outlay, some $300 million is bookmarked for building foundational models. The ministry is counting on funding three to four models, the math being that it could take $50 to $100 million to develop just one. The catch is that whoever builds this will have to give unlimited access to the government. The other thing is that the models will have to be hosted on AI Kosha, which is India's own hugging face.

Now, this is a hub that offers developers, researchers and data scientists, the ability to share open source models, data sets and apps. It's basically an AI community that allows people to build on existing pre-trained models. Apart from 70 odd text to speech or generative models in Indic languages from AI for Bharat

and AI Research Lab at IIT Madras. The marketplace today also lists Microsoft smaller five series of models and a few specialized non-LLM models too. To download the latter, one has to go to hugging phase. As for the datasets, out of nearly 340 listed on the site, very few are actually LLM usable.

And a handful that are just aren't comprehensive enough. The ministry knows this is not high quality or AI-first data, but it hopes to figure it out soon. Stay tuned. METI has received close to 70 applications for foundational models. The ministry has been bragging about how a ton of them are primarily focused on social sectors. And it does make sense.

India is a country with a large population and grave resource constraints, which means there is a huge opportunity to make a real difference to the common man with the help of technology. Of course, politicians and bureaucrats have leaned into this spiel big time. But the real question here is to what end will those applications be funded? Like one analytics professional pointed out to us, you have to do the population scale implementation. Otherwise, it's just a proof of concept.

It has to be something that tries to solve a real problem that affects a massive chunk of people. For instance, say we build a simple LLM and deploy it across all primary care centers. It asks patients a basic set of questions about their health and then directs them to the correct hospital they should go to. Or in agriculture, it could be an LLM that looks at the salinity of soil across the country and suggests which crop to sow when rainfall is 80 to 90% of the normal range. It's quite likely that such proofs of concept have already been done, but in smaller pockets. For instance, take Wadwani AI. This is a non-profit which many in the industry describe as an extension of the government.

It will now work with Ames, the All India Institute of Medical Sciences, to build out AI solutions in healthcare. Using AI to screen for diabetic, retinopathy and TB via chest scans is one idea on the table. But the funny thing is, this isn't a new idea. These two AI products have already been developed in India and commercialized by two startups, Cure AI and Remedio. Ames, meanwhile, seems to be reinventing the wheel. And it's also not alone. The India AI portal is cluttered with old projects and news. But that's not necessarily a bad thing.

You see, it's not that the old paradigm of AI has become unsexy. It's ironic that large infrastructure investments by big companies is actually giving a new boost to old AI, like computer vision. Raghu Galapali, the network director of global partnerships and development at LV Prasada I Institute, put it this way. He said, what was theoretically possible then is certainly possible now. The old models have matured and don't hallucinate anymore, which is how his institute managed to start building a myopia model, which allows it to tell children in just a visit or two

how their myopia will progress as they get older. And with that data, they're able to create certain mitigations, like what kind of glasses they should wear or whether they need to wear contact lenses overnight. And what's really incredible is that LVPEI is paying under $150,000 to develop this model. Relatively low cost, high impact. But to pull something like this off needs two things. Deep thinking and meticulous planning. And so far, that's not the approach India AI has adopted. The AI mission should be focusing on creating strategic AI capabilities.

It should start by picking fewer battles. Instead, the likes of Amitab Kant, India's G20 Sherpa have said things like, there's no shortage of money. In the process, this has turned into some sort of hobbyist's pursuit. But now more than ever, intelligence is getting commoditized. So it's on METI to fix its accountability and lay out some deliverables apart from browbeating about India's own LLM. Daybreak is produced from the newsroom of the Ken

India's first subscriber-focused business news platform. What you're listening to is just a small sample of our subscriber-only offerings. A full subscription unlocks daily long-form feature stories, newsletters and podcast extras. Head to the ken.com and click on the red subscribe button on the top of the website. Today's episode was hosted by Rahil Filippo's and edited by Rajiv Sien.