The Sovereign Push Fueling India’s AI Aspirations with Payal Malik
What conditions can allow AI startups to flourish in India?
In this episode, journalist Puja Mehra speaks to economist and competition law expert Payal Malik about India’s push to become an AI powerhouse. While India has advantages in data and digital talent, Malik argues that inadequate computing infrastructure, limited R&D spending, and over-reliance on big tech threaten to stall its progress.
They discuss how initiatives like AIRAWAT aim to build indigenous capacity, but without strong competition policy and sustained investment in innovation, India risks repeating past mistakes—ceding control of critical digital infrastructure to global giants. With countries like China far ahead in AI funding, Malik warns that unless India builds open, contestable markets and ramps up support for startups, its AI ambitions may remain just that—ambitions.
Tune in for analysis of the political economy shaping India’s AI future.
NOTE: This transcript is done by a machine. Human eyes have gone through the script but there might still be errors in some of the text, so please refer to the audio in case you need to clarify any part. If you want to get in touch regarding any feedback, you can drop us a message on [email protected].
—
TRANSCRIPT
Puja Mehra: Payal, hi, thank you for coming on the show again. Thank you, Puja. I want to talk to you today about AI and get a sense of where does India stand in terms of AI technologies.
We hear a lot about what China is doing, what US is doing, and how these two sort of tech giants are competing with each other. But we also hear a lot about India, but not exactly what it is that we are doing. So if you could just take us through that to begin with.
Payal Malik: Yeah, sure. Puja, because you know, the geopolitics is really, really unsettled right now. And technology seems to be in the crosshairs of geopolitics.
And there is an issue regarding not being dependent on a certain bunch of companies. Say, for instance, the companies, the big techs, which are mostly US companies. Because recently there have been even executive orders basically limiting the ability of a sovereign nation state, say a country like India, to even implement the rule of law when it comes to these big tech companies.
And not only that, even prior to that, I think India was very clear that they don't, and Indian government was very clear that they don't want to repeat the mistakes of the past. And in just being dependent on a handful of companies for their technology needs. And so therefore, the India AI mission was started in order to bring in some kind of a self-reliance when it comes to technology.
So the government through its India AI mission started investing in computing infrastructure, which includes high performance GPUs and data centres, because that was very, very critical inputs for training and deploying AI models. And in the absence of compute capabilities, advancements in AI remain constrained. So India's ability and India's comparative advantage is in the application layer of the AI stack.
So to the credit of the Indian startups who have really done a great job when it comes to entrepreneurship, even in the previous startup ecosystems, which were app-based in the web 3.0, which has just gone by before AI technology came in. A similar story is being repeated here, where the Indian startups are extremely competent, tuned in to the needs of not only the consumers, but also the needs of the businesses to develop applications which harness AI technology. But for doing that, and just to let the listeners also know that at Icareer, we did a complete study on AI markets.
And a crucial component was that the stakeholders insisted and stakeholders consistently told us that as far as innovation in the application layer is concerned, India is very well geared. India can perhaps overtake China at some time, hopefully, because creating specific models and specific applications for India's needs. But the constraining factor was the computing infrastructure.
In that sense, India's AI mission tried to dovetail with India's private sector capacity or the startup capacity in the AI application layer by providing infrastructure support through the AI India mission. Of course, we can discuss that it may be not enough. A lot more needs to be done.
And commentators are talking about that India and Indian government specifically needs to be actually paranoid. And it is paranoid, but not sufficiently enough when it comes to having some kind of a self-reliance. And that word or industrial policy has again crept in the AI technology ecosystem as well.
So that was one, compute. And so the second... So there was a lot of debate with many naysayers, basically, including the doyens of technology, I don't want to name, who were basically saying that India does not need to create its own large language model.
And large language model, by now, everyone knows, whosoever is tuned in to the newspapers, and you need not be an expert, that large language models or foundational models are the core infrastructure, so to speak, when it comes to artificial intelligence and its use cases. And large language models were mostly, and are till date, mostly the Western models in the sense, trained on Western data, trained on data which is not Indian. And the government was cognisant of this drawback of those large language models.
So what happens is, when it comes to creation of applications or creations of now what we are calling agentic AI models, you need these foundational models, which are these large models trained on trillions of tokens and billions of parameters to create a very, very use specific, either a small language model or an AI agent.
Puja Mehra: Okay, so I don't know what is agentic model and I don't know how small language model is different from a large language model. So I'm sure our listeners would benefit if you could explain those.
Payal Malik: Yeah, sure. So small language models are mini large language models because they will be solving a certain specific problem. Say, for instance, if a business have an HR capable artificial intelligence model, so it need not just use the large, it has to fine tune a large language model to develop HR monitoring or HR task capabilities within its business.
And that small language model will be catering specifically to the requirements of that particular business to deal with its HR management problems. But yes, it would be riding on a large language model or it will use a large language model, say a chat GPT or Meta's Lama or a Gemini, etc. And these large language models which I'm talking about are all either Meta or OpenAI or Google's models.
But Indian industry and Indian application layer providers were creating and are creating very, very use specific small language models which do not require so many parameters. As I said, there are some parametric inputs into creating these models. They do not require so many parameters.
So they train on small datasets, real time datasets, but on the back of these large language models. That's one. AI agentic models are now, AI agents are now the latest revolution in the AI space.
That is where you don't even have to give prompts to what you require. You just say, just give one prompt and that agent would behave as if you're interacting with a human being. To just give you an example, supposing you're planning a summer vacation, say you just skip the dates, say between 15th, so it's like a concierge service, say between 15th May to 31st of May, 2025 in Paris.
That agentic model will prompt you your airline tickets, will prompt you the hotels, will prompt you all the places you want to go to, will prompt you with the entry ticket. So it will give you everything in one package without you have to continuously prompt them. So the last bastion of AI revolution is now currently AI agents, which are performing quasi-human activities.
And for that also, again, these are very context specific. That means when I'm talking of small language models and they are dependent upon the local data. So that's why the Indian government, but as I said, they are still dependent on large language models.
And so therefore Indian government in its wisdom thought that an Indian large language model would be very, very essential to harness and to allow for innovation in the downstream layers, that is be the small language models, be it AI agents or AI application. So the Indian AI mission talks about creating local language, local dialects datasets because these language models are very, very dependent on datasets. But in India, languages are so many, the dialects are so many, and just to therefore depend upon a metaslama.
And so having said that, they too have now got into the space, given that India is a large market, would not be sufficient. And Honourable Minister Vaishnav had declared that many months back that India should have its own large language model, which will be Indic data specific, means using Indic languages and training on Indian data to give you better results. And so therefore recently, Sarvam AI, I think that report was in the news, has been chosen and been selected by the Indian government to develop India's large language model.
Puja Mehra: Yeah, Minister Ashwini Vaishnav had said that and he's been quite excited about AI developments, expectedly so. But I'm also reminded of a statement that another minister, his colleague, Mr. Piyush Kaur had made, which sort of ties in with what you said. He said something which triggered a huge public debate on how innovative are Indian startups.
I don't think he said it specifically for the AI ecospace, but asking you specifically about the AI ecospace, you said that many startups are doing a lot of innovation, coming up with a lot of innovative applications and that is where India's strength for now seems to be. And that is probably where India will make a mark. Are you seeing that kind of high quality innovation that Mr. Piyush Goyal has ambition for India?
Let's not go into the debate where many people said that a lot of innovation also depends on the policy ecosystem and also the market structures, which again depends on policy, etc. We'll come to that later. But for now, are we seeing a lot of very innovative work being done by the application makers?
Payal Malik: I would say yes, and they have moved up the value chain, at least when I'm now just talking specifically of the AI application layer. Yes, the first round of the startups of now, we can easily say startups of the year, were in internet companies just solving intermediation problems between consumers and businesses. And I guess that's what triggered the minister's statement without getting into that.
But I think AI is one area where I think the startup ecosystem has actually tried to meet those expectations. But what would be required is complimentary investments by government. It cannot happen only by depending and leaving it only to the private sector or only leaving it to the startups, because a lot of money goes into this and a lot of patient capital is required.
So the previous cycle in which the startups mushroomed were heavily dependent on venture capital and foreign capital inflows. The domestic capital was not supporting them. And now that needs to change.
So there needs to be a lot of domestic patient capital given the fact that innovations are happening. For instance, I was in discussion with Art Park, ISC Bangalore-based startup, which is looking at creating Indic language SLMs, that is the small language models, using or creating also open data sets, that is data sets, which are, because data is, just think data as the new oil, means I hate to use that analogy, but data is the oil for AI. As data was important in the web 3.0 genre of technology development, here too data is required. But it is very important that the data is contextual. The data is indigenous. The data is to solve Indian problems.
So there again, governments attempt through Bhashini, that is open data set, which allows these developers, these application layer startups to dive in and be able to create solutions for local problems is going to be extremely important. So it cannot be just, it's left to the private sector. Now, increasingly, and because of many reasons, one reason is, of course, that you need to, and so the word sovereign, the word industrial policy is now back in the tech policy jargon, and it is now no longer a cuss word because in the competition policy domain, when you say industrial policy, basically you think that the government is trying to distort markets.
But industrial policy's role has changed considerably, especially when it comes to AI. First of all, of course, every government is responding to what is an AI hype and it's kind, so some commentators have said that it has become like more of an arms game. But it is important, therefore, that the state support or the industrial policy support to startups helps them move up the value chain.
And this recent development, which as I said, where Sarvam AI was selected by the government of India to build the country's first truly indigenous large language model under the India AI mission is a step towards that. So that's one reason. That is, you want to reduce dependencies.
Secondly, every country, for instance, having some sovereign, so technology and sovereignty now are seeing a marriage made in heaven, so to speak. For instance, I was recently in discussion with one of the EU commissioners, and there is a specific position of the EU commissioner and her title read Commissioner for Tech Sovereignty. EU Commissioner for Tech Sovereignty.
The reason is because technology is now becoming very, very strategic, and it's a strategic imperative when the global order is in a state of flux. And so, therefore, selection of Sarvam AI is significant because, as I said, it's a purely indigenous company founded in 2023, and it brings deep expertise at the intersection of cutting edge AI research and public scale infrastructure design. So it was also a startup to begin with, just a two-year-old company.
But yes, it required some handholding, and that handholding is to be in the form of a finance or a subsidy, and that's the area where government is focused on. Of course, we can debate that the money is less, etc., etc., but yeah, so we are moving in the right direction.
Puja Mehra: So when we talk of India's AI potential for growth, growth in AI, that is, you've spoken about the importance of data. We also hear a lot about talent, how that's absolutely critical and an advantage India has in this field. But you briefly mentioned the role of computing infrastructure.
How critical is that, and how well are we doing on that front?
Payal Malik: So again, here, for instance, even Sarvam has a partnership with Microsoft Azure when it comes to computing and cloud storage. So dominant firms like AWS, Google Cloud, and Microsoft Azure are the frontrunners when it comes to compute infrastructure. And what happens is, because of the dependency on these handful of companies, it does create high switching costs and therefore reinforces market control.
There are reasons why, again, computing infrastructure becomes important because data centres and GPUs are very essential for training and deploying AI models. So once India has announced that it wants to have its own LLM, that is a large language model, it had to get some compute infrastructure also. And so money has been put in, projects have been announced, where many startups will have access to the compute infrastructure, which would be subsidised by the government of India.
So that basically means that the government does recognise the strategic importance of AI. And for instance, the development of AIRWATS, that is A-I-R-A-W-A-T. It's a supercomputer to build indigenous compute capabilities.
Puja Mehra: Named after the mythological elephant, I think, of Lord Indra, if I'm not wrong, right?
Payal Malik: I guess so. So India will always have these kinds of acronyms, which are India specific, and rightfully so. For instance, for a cloud infrastructure, they had made, the name was BPI, that is a digital public infrastructure in cloud known as Megaraj.
So I guess mythology will play its role here. So Indian startups, particularly in tier two cities and sectors like healthcare, fintech are struggling with high GPU costs and limited data centre access. And access for on-premise deployment costs for sensitive workloads.
So access, so that, as I said, will be the binding constraint. So that constraint needs to be relaxed and government is playing its role to relax that constraint. Because access to affordable and scalable compute infrastructure would reduce the barriers for startups.
Again, back to Sosa Samao, it goes back to the statement made by the minister, but it is going forward. It has to be a very healthy, complementary public-private partnership between the government and the startup ecosystem to achieve a situation where India innovates much more. A lot needs to be done when it comes to innovation.
We are doing well when it comes to AI, but as I said, still there would be a time going in the future where we would say, why aren't we doing much more than now? And just to give you one metric, I used that metric way back in 2009 when I was doing India's innovation ecosystem. That is government's expenditure on R&D was something like 1.2%. And recently, there have been articles after articles saying that India needs to do, that is, I'm talking about opinion pieces in newspapers, that India needs to do much more. And so therefore, the aspiration of the Indian politicians that our startup ecosystem rises up to the occasion cannot happen till it is a mixed public-private. So private sector also needs to invest.
So I'm not just putting the onus at the door of the public sector. That 1.5% of GDP is now down to only 0.7% of GDP. So these aspirations will just remain aspirations until and unless we ramp up this.
0.7% is with little traction in patenting or deep tech commercialisation. So not only there has to be an increased investment or increased R&D expenditure, but it has to also shift gears that R&D expenditure is in patenting. R&D expenditures is in commercialisation because I'm doing another project in standard essential patents where startups are innovating.
Startups do have good technologies, but they can't take it to market or they can't commercialise because the whole process of patenting is so adverse and onerous on them.
Puja Mehra: In fact, you know, the reason I invited you for this conversation on AI is because of market structures. I find that in the Indian public discourse on anything related to the economy, we don't pay enough attention to market structures. And you're an expert in competition, competition law and how competitive markets are.
And from just what you just said, and otherwise, it's been clear to me that for AI technologies to grow in India and for startups to be truly innovative as our ambition is, a key role is going to be played by how the market structures are. You're saying that there has to be public-private partnership. You're saying that financing is a constraint.
You're saying that infrastructure is a constraint, which may be related to the financing side of it. And then you're now also saying that commercialisation of research is something that needs attention. Is one way of doing that to get the market structure right?
So I want you to explain a bit about the role of market structures and competition in this AI space and how important it is for growth, AI growth.
Payal Malik: Yeah, so I think we don't need to repeat the mistakes of the past. World over, be it EU, and EU has really taken the heat from the recent geopolitical developments. There is a call for creating what is at least new EU stack.
And in India, therefore, taking a cue from, and India was a front runner in creating digital public infrastructures. Taking its cue for creation of these digital public infrastructure, it has to go to the next level where digital public infrastructures are created even for AI. And the DPIs for AIs, in the sense, again, to make them more open, contestable, accessible infrastructures so that the core infrastructures do not get dominated by a few handful of companies.
So government is pushing the envelope in that case. One or two recent innovations, so to speak, in the AI digital public infrastructure is perhaps going through what is known as open cloud computing. That is, rather than one or two companies providing, say be it AWS or Microsoft, because cloud computing would be very, very important.
Already said that for AI, rather than a handful of companies providing that and increasing your dependencies on that. Indian cloud computing firms, so we have good stories. We have Yota.
We interviewed a few small Bharat Cloud, small cloud computing companies which are indigenous. And this is no longer a nationalist agenda. It is also a competition agenda.
So for instance, if you had asked me this 20 years ago as an economist, I would have said, we are an open economy. Trade should happen wherever goods comes from or wherever services come from at a comparative advantage, which are cheapest, we should get that. But technology and the recent technology changes have not allowed that trade paradigm now to survive.
That trade paradigm is losing ground. Therefore, the same practises are happening again and again. To give you an example, the last time when we met, we talked about Google's search monopoly and how all the operating, that is the Android operating system, was using Google search as default.
And most of the Indians, because it was a free product, there were revenue sharing agreements with OEMs. Most of the Indians didn't realise their data was flowing to Google because of that free product. Currently, again, the same thing is going to happen and history should not repeat itself. So therefore, you need a multi-pronged strategy when it comes to devising the market structure in AI.
A, you, of course, do need strong competition laws to look out for the same kind of problems, say, be it self-preferencing, be it tying and bundling. So Google has struck a deal with, say, Samsung phones, which would then integrate Google's Gemini in those phones, which again means that it may, as an operating system being a distribution platform, again, it may be and Android phones are the most used phones. And also, AI is now moving onto your mobile device.
AI will no longer require a lot of cloud computing. That means where computing can be on your device, that is known as edge computing. So the kind of AI jobs which a common man would need, a kind of a chatbot or as I mentioned about an agentic AI to solve your...
Those would be just like your apps, which we are using nowadays. Those things would require that these practises, these anti-competitive practises do not happen again. For instance, Google has got into an agreement with Samsung to integrate Gemini in all the Samsung phones.
It would mean that other AI agents, other small language models and other applications in AI may not get a fair treatment. So all the previous problems related to barriers to entry, that has limited access to key inputs because again, that perpetuates the data advantage. So for instance, when search was in your handset, you used to take it for granted that Google is the only search engine in the world.
Though there was, Bing was trying to come in and Bing was being used as Dcuk duckr. Go, etc. Because that was creating a status quo bias and entry barriers. Similarly, barriers to entry could be created in the AI ecosystem.
These big tech companies which have enjoyed advantages in the AI market, stemming from their previous advantages, share the same characteristics that they are able to. So for instance, if through search you can leverage into so many domains. Similarly, through your large language model, but by not competition on merits, it is through a billion dollar deal with Samsung.
Whereas a billion dollar deal with Apple has been sanctioned against both in India and in the US, the last Google case which we discussed. But the same thing is being repeated. So the government and the regulator needs to be quite up to the mark when it comes to looking at these behaviours of the already entrenched firms.
So that again, we don't end up with a very concentrated AI market. So currently we are at that stage where there is a race to capture the ecosystem. So ahead in this race are still these big tech companies owing to their dominance in various domains which are important for AI.
For instance, Google in the operating system and search market, Microsoft and Amazon in the cloud market, that is the cloud computing market. So they already have their dominance and it is time for them to again get into the adjacent AI application markets through anti-competitive conduct, which as I gave you one example has starting to show up such as tying in bundling, self-referencing, and of course, complemented by mergers and acquisitions. And now the recent kid on the block is partnerships.
So these companies are getting into partnerships without any equity purchase in the firm of interest, of their material interest. For instance, Microsoft is in a partnership with OpenAI. So OpenAI is an independent company or a not-for-profit company and of course, it is trying to become a for-profit company, but Microsoft has a partnership.
What are the contours of this commercial relationship, et cetera, is not very clear. The other things for leveraging, which are being used, for instance, again, as in the cloud computing space, which is already concentrated. And this came out from our discussions, both in the cloud study, as well as in the AI studies.
Most of our Indian startups are being given what are known as cloud credits. And these cloud credits are nothing else but free purchase of cloud infrastructure is given for a certain usage of that cloud infrastructure. And of course, it's chargeable after that.
But that means that you get dependent on these cloud providers, be it AWS, Google Cloud, or Microsoft Azure. And they bundle this with using the cloud marketplace. So for instance, the Apple or the Google Play Store, and the Play Store examples and Play Store cases are now well known, that the Google Play Store housed all the applications and charges a hefty commission of 30% from these applications for accessibility, visibility, and downloads of any digital material from their applications.
Similarly, there is a possibility of the rise of platform ecosystems or marketplaces centre either around the model layer of the stack, the model layer is, as I said, and that's why an Indian, indigenous model layer was important, again, both for contestability as well as for sovereignty, which is likely to favour the vertically integrated cloud providers and the dominant large language model providers. So the idea, therefore, should be to have a multi-pronged strategy. And that is why I guess it must be, and if I would like to end on that, it must be a little controversial, but I guess the minister was trying to be provocative and provocative because there is an understanding that until and unless our startup ecosystem ramps up, its innovative capability goes just beyond providing service solutions, we are perhaps doomed to be locked in, again, in this big tech ecosystems. And China, therefore, in the sense of always the comparison is with China, we will see that and for instance US, because India is not only behind US, which spends about 8.8 billion dollars for its AI mission. China spends 215 billion dollars just to give some numbers. South Korea, Taiwan, Ireland, they're all ahead of India when it comes to R&D. So in the end, if I say one thing which will take India to a totally different level would be sustained and long-term investments in knowledge and innovation and India's failure to internalise this principle is going to perhaps hurt it in the future.
Puja Mehra: Thanks. Thanks, Payal.
Payal Malik: Thank you so much. Thanks.
