Pull in conversations which would otherwise be impossible. And you're about to witness a peak Davos moment because live here at the NDTV studios in Davos, I am delighted and privileged to be joined by one of the world's leading technologists, chairperson and chief executive officer of IBM, somebody with very proud and strong Indian origins, Arvind Krishna. Thank you so much for joining us. And with us is somebody our viewers in India know very well but also somebody who went to the same IIT Kanpur. So both these are Kanpur boys from IIT. So a lot for the IIT Kanpur circle to be chatting about as well. I'm joined by Minister Ashwini Vaishnaw who's leading the Indian delegation at the World Economic Forum. Our minister for IT, information technology, information and broadcasting and railways. Thank you very much for joining us. Fantastic to have both of you. This is the one conversation I've most been looking forward to. I want to start by asking you Mr. Krishna about your view on what Ashwini Vaishnaw and the Indian government are doing at this moment in terms of trying to build up an AI ecosystem. We are in the AI age. We've seen all the tech developments being done by the tech giants in the United States, IBM amongst them. We're seeing low-cost high innovative value systems coming out of China. Where do you see India shape up in this global AI race from your lens?
Yeah, look, it's a great question. So, first it's a privilege to be here with you and with Ashwini ji. So, let's talk about this. And these are conversations I've been having with him I think for almost 3 years now. It actually comes down to the word independence or sovereignty. I think you would call it Viksit Bharat. When something is so important that is going to be used for the next many decades, it could be incredibly expensive or perhaps it's not in your control if you get it from outside then you must have capability within the country and that's what I've always encouraged Ashwini ji to invest in for India whether it's in the public sector or whether it's incenting the private sector. It doesn't have to be everything because I believe that small models are catching up to the capability of large models for perhaps 95% of all that you do. So how do you encourage that ecosystem be it the underlying semiconductors be it the model building and actually most importantly which the world doesn't worry about yet the usage applications which is where I think the breakthrough will come for India locally.
No. So the big concern in India amongst many who are watching is that you know we produce so many technologists, IT techies, chief executives like yourself running the world's biggest companies, we don't produce a DeepSeek. Why is it okay so we understand what's happening in Silicon Valley but why is it that when that breakthrough cutting edge innovation happened it came from China and not from India and what do we need to do now to try and ensure that the next DeepSeek moment comes out of India?
Look, putting aside a huge amount of pride which one should have there is a 20-year lag. It's purely I believe it is fundamentally a question of only two things. Economics meaning how much capital can you put at risk because doing things like DeepSeek means nine out of 10 fail. In the Indian ecosystem 90% failure is still not smiled upon. Some failure is smiled upon but not 90%. So there is the economics coupled with the risk-taking appetite of people to go ahead and do that. Now going from a trillion dollar GDP to three to five that unlocks a lot more. What it unlocks is if you kind of take how much is left at 5 trillion perhaps almost a trillion can be applied towards all this. When you're at 1 trillion you have zero. So this is just that question. So that's why more of it will happen now and as the next 2-3 years unfold I believe even more of that will happen. So those are really important considerations.
Mr. Krishna, in the last few years you've made an effort to try and build out an ecosystem a around semiconductors b around ensuring that we have enough compute power for Indian companies to be working on. How confident are you at this moment that even if DeepSeek didn't happen in India going up against the Chinese models and the world having to pick between them that India will be a third pillar in this race?
So so many people today in Davos they say that it's going to be a T3 world. So technology three big technology countries in the world and I'll pick up from where Arvind said 95% of the work today is done by small models and the models that the 12 models that we are building they are in the range of 50 billion to 120 billion parameters. A 50 billion parameter model can run the AI needs of every enterprise in the world. So we have to look at the entire AI ecosystem. The topmost layer the usage the applications that layer is where the maximum ROI is going to be return on investment. Or AI application India lead. We have been a very good IT services industry in our country. What matters is how do we quickly understand the enterprise way of working, understand the technology and apply that technology in the enterprise and that's where our IT companies are doing so well today. They have all pivoted they are their AI hiring has increased 33%. Second is the model. I said we are building 12 models and very soon we'll be launching three or four of them very soon. Third the chip model. the semiconductor layer. Our semiconductor program has progressed very well and today in our country we are already we have 10 plants in which construction is going on. Three plants pilot production has started. Four plants will be completing their pilot production and commercial production in 2026. Then comes the infra layer. $70 billion worth investment is going in infra layer. Then comes the energy layer. In the energy layer, we have opened up the nuclear energy to the private sector through the Shakti act. And that's where this entire stack of AI is going to have a major play in a country.
So let's deconstruct this one by one, right? For example, the application layer. Now, one critique can be that this is of course the use case, but this is where the Indian companies always restrict themselves. They're not doing fundamental core research and innovative breakthrough which is happening typically out of a Silicon Valley or in some cases out of China. Do you think that criticism is fair and if it is fair why does it not happen?
It's actually not fair. Not fair because unfortunately the majority of the work in India has historically been not actually in creating new applications but in maintaining and evolving them. So actually if you construct the new applications there is a lot of innovation because then you have to deeply understand the enterprise and you have to deeply understand the AI itself. So that's why I actually think that that criticism is not fair. The value every tech evolution not just AI it starts at the infrastructure and semiconductor layer typically lasts 2 to 5 years then moves into what we would call the middle enabling layer. In this case it is models that is always unique. In the case of PCs it was operating systems. iPhone and Android was also operating systems. And then you get to the applications. In all cases in less than 10 years 90% of the value shifts to the innovative applications not the maintained ones. And so that is what Ashwini ji is talking about. And if you take advantage of that it'll unlock a huge boom of innovation and monetary return.
Let's now come to the large language models because there are two ways of looking at it. One theory is that LLMs get commoditized and there's no point in countries like India spending a lot of compute power, time, energy and effort in building it. You can build on other models that have been built. The other is very core to the idea of data sovereignty and AI sovereignty that you can't depend on the Americans or the Chinese. You have to build this out yourself. What have you been re how do you come out on this?
I've been very public on this and that is not just my words that is the actions of our company and what we recommend to our clients. I believe that the very large LLMs because even 50 billion and 100 billion is a LLM 2 years ago but the trillion parameter LLMs I actually believe will be a commodity. A commodity can have value but a commodity means there is no switching cost between them which means price comes down to whatever is the minimum bearable and the reason is the amount of investment those are going to cost you $50 billion a year to get built and if that's the ongoing investment you can look at the economics and say maybe there'll be a half dozen people who will supply them but if switching cost is low which is what we observed price comes down the models then that can be infused with national data with Indian local health data with defense data with how work is done in the Indian legal system those can be much smaller and incredibly effective for that purpose if those can be built in India 95% of the work will be done there and it's fine for the other 5% use a big one.
You've been interacting Mr. Krishna very closely with the Indian companies building out these LLMs. A couple of them get released literally in a few weeks from now. Do you see any USP anything? One is the point that Arvind makes that because there's so much data in the Indian context, their languages, they can be useful there. Do you see anything which India can export and make of use to the rest of the world?
I would like to add to what Arvind said. There are two points here. First point is about the size of the model which is required for 95% of the work. Our models which are built by the 12 startups that we have funded and supported and one of the IITs which have supported are good for practically every application that we need our country. I wouldn't take any names but I can tell you in full confidence that during one major defense scenario in a country in the recent times the sovereign model used in a developed in a country used in our in that particular operation the results it produced was phenomenal phenomenal I'm saying so if that is the condition if that is the scenario that we have experienced so far The idea of having as Arvind said the trillion parameter models will become commoditized. What is important is have those bouquet of smaller models which are very focused, very solution specific, sector specific and start using them so that the productivity gains start coming in the economy and that's what our focus is. The second part is about the innovations which are going to happen today. If you look at all the 10 large models in the world, they are all looking at they are all in a very small bracket of performance, there is very little differentiation between one model from the another. So that means the next level of innovation will come from something very different. It might be an engineering innovation, it might be a math innovation, it might be totally different way of looking at or designing a model. And that's where some of our very smart startups are focused on. There is one startup which has today instead of doing a sound to text to sound they have come straight from sound tokenization to so sound to sound. So that kind of innovation is going to come from any part of the world and I'm sure India will be among the leaders in that.
You know we're seeing two broad approaches play out and I want you to share your insights on how viewers on any TV should wrap their heads around it. One is American tech companies betting on AGI, artificial general intelligence, hoping that that's literally around the corner. So do you want to give us your sense on how far this bend is and what lies when we get on the other side? And the second is what the Chinese are trying to do which is to embed AI into all different kinds of processes. How do you look at these two approaches and what would you recommend for a country like India?
So look first let's just look at the spectrum of this is belief almost like religion dharma as opposed to science. So there is a set of people who believe that AGI can be done. You just need to scale more. You need to train more. You know, if we could just do 10 times more compute, we would get there. There's another kind and I lie in that camp who believe that these current technologies are not going to get to AGI. You are going to need to infuse knowledge into models and we don't quite know how to do that yet. So why do I say that? To give you some intuition. Today if you were to say does a ball go from up to down or down to up the models will all tell you it goes from up to down but they don't know that. What they're doing is they're actually doing an incredible amount of compute to go do that. If that is the gap then actually the human is way smarter because the human knows that they're not rediscovering it each time you ask that question. So that is how you fuse those two. Nobody knows how to do that. When we learn how to do that, maybe we'll get to towards AGI. And a lot of people in academia have my belief that you need to fuse knowledge with these statistical techniques that are inside the models. And so, so then my belief becomes we've got to use these big ones as commodities. You got to have the specialized local ones. Actually, I'll quantify it. A 50 billion parameter model is roughly a 100 times more cost effective to run and 100 times more cost effective to train than a very large model. So the number of variations also of it you can do for all different purposes is massive. The big one you're going to have like one or two at that cost. So that is why I believe that this is the correct path. A lot of small models, domain specialized, industry specialized, language specialized, and that can unlock innovation because now you could get a startup who can get access to some of the infrastructure that provides and with 10-20 million worth of expense can actually innovate something new. If you're on the other side of AGI, if you don't have the hundred billion, you're not trying to compete with that. But I actually believe that's not going to happen.
So one theory is bigger the better. The more compute power, the more money, the more muscle, the higher the probability of innovation. You're taking Ashwini a different approach. The question though could be given how much China backs their industry and how much of a head start they've had in this process. They've literally been doing this for two decades. Do you feel we're late in trying to get onto this bus and can we catch up adequately to actually make this a three car race?
I think AI revolution is going to be a multi-decade phenomenon. It's just a beginning of the entire innovation which is happening and as so many people say this the brain just uses three watts to do what it does whereas a AI data center uses megawatt and megawatt of hundreds of megawatts of power. When that level of innovation is still left in the journey, I'm sure India will be among those who innovate very well. I fully agree with Arvind's point that and there are so many other very mature and very wise technology advisers who have all said that focus on smaller models which can be very effectively at a very low cost low cost of compute low cost of power and low cost of training can be deployed and the efficiency gains which come with it the productivity gains the innovation gains which come with it that is where the focus has to be and have a suite of sovereign models can be used even if any country tomorrow denies the availability of those.
Can I jump in with the sports analogy?
Exactly. You know what analogy to come up with? Baseball doesn't work here. No. Let's go to cricket. Let's imagine that there are two countries playing and it's going to be a five test matches because in some of your questions you're actually assuming there is much more progress than it is. In a five test match series. We are on the first day of the first innings of the first match in AI.
That is a good analogy. That is where we are. The it has started the players are known but we don't know who's hot, who's cold, who's going to do what. That is kind of the analogy of where we are in AI which is why different approaches can actually both win. By the way, China's not chasing AGI to our knowledge. Maybe there's a small science experiment somewhere. They are much enthused about applying it everywhere and I do think that that is the correct approach for India.
So both of you went to Kanpur and I want to ask you this question. Do you think you could have done to do over three and a half decades at IBM if you'd stayed back in India and if not what's your message to a young boy studying in IIT Kanpur at this moment he wants to be the next Arvind Krishna but he wants to do it back home.
Yeah, look I think about this a lot and I would tell you so I graduated from Kanpur in 85. He graduated after me. Let's just leave it at that. I got in 1985 and I did actually try for a short time working there but the conditions were that people were very risk-averse they didn't really want to spend capital and they just wanted to copy not really invent. I think by 2000 not by 2025 that had changed and so I think that the conditions have changed where when I look at the graduates for example from IIT Bombay half of them are doing startups in Bombay or somewhere in India. If I look at the startup ecosystem that we work with in India it's lively. So today you have opportunities in both places. And I would say by the way the analogy on me you go look at somebody like Chandra didn't he get all the way up to the head of Tata and Sons by India. So there is that I think that the number of tech companies that are coming up now.
