Should India build a sovereign, foundational AI model?

Does India need to build a sovereign foundational AI model for its local needs?

Pranesh Prakash: I believe that India should be working on foundation models, but not for the purpose of sovereignty. It is important to have people who are able to build foundation models and also to have people who can build on top of foundation models to deploy and build applications. We need to have people in India who are able to apply themselves to every part of building AI.

Sovereignty is something that we should be concerned about from the perspective of things such as sanctions, which we have seen the U.S. apply over specific kinds of chips, especially those that are used for training AI models, and export controls over things like software or models. Those may be coming down the road.

But I do not think you need to be worried about that. As things stand, what we have are not just proprietary models which are closed, under licence, and cannot legally be used without the permission of the person who created that model. We also have a lot of open weights models and a few open source models, such as DeepSeek R1. And because we have these, for which you do not need permission and can deploy locally, we do not need to approach these from the perspective of national security and sovereignty and concerned about sanctions, and so on.

If there are sanctions tomorrow which are applied to the usage of any models produced in the U.S., Indians can just “fork” the code of many of the existing open source models, refashion those, and build on top of them.

As long as these models exist, are state-of-the-art, and are competing well against proprietary models — and I believe we will soon come to a point where they will out-compete proprietary models — we do not need to be concerned about this from the perspective of sovereignty.

Tanuj Bhojwani: I would say that for sovereignty we should do this, but if they really wanted to block you, they would start with chips, as Mr. Pranesh pointed out. Even if you had your own model, you need Graphics Processing Units (GPUs) to run it. DeepSeek uses Huawei’s [Ascend HiSilicon] 910C chips — India cannot even put that up. We have no contract with manufacturers such as Taiwan Semiconductor Manufacturing Company Limited (TSMC) to produce chips of even a previous generation.

More than sovereignty, there is an element of pride here; we are ambitious people, and want our own model. We have done this in other areas; we have reached beyond, in GDP per capita terms, what our aukat (station) would be. Infosys co-founder Nandan [Nilekani]’s point and my point is this: do this if you see returns. Do this if anybody sees the returns on it, whether the pride is the return for the investment, make sure the investment is proportionate to that end.

Does the emergence of DeepSeek at low cost make this more attainable? Is it worth the expense?

Tanuj Bhojwani: Even with DeepSeek, this remains a game of a few hundreds of millions of dollars. DeepSeek V3’s training run cost $5.6 million. But in cutting-edge research, there are a lot of failed attempts at that goal. So we are still talking about hundreds of millions of dollars in investments. Even if you produce a foundational model in India, there is still a lot to do — paying salaries and so on — before you can make money back on that investment.

The market is still majority U.S.-dominated. About 60% of the market cap of the entire globe is in the U.S. Even today, business-to-business software-as-a-service founders in India are largely selling to the U.S. If we believe we will make an Indian model with local language content, you are capping yourself on the knee because the overall Indian enterprise market that will purchase AI is much smaller.

Spending $200 a month to replace a human worker may be possible in the U.S., but in India, that is what the human worker is being paid in the first place.

To succeed here anyway, the government needs to create an institute that has autonomy and spending power. 90% of the spending will go to failed experiments. Our public procurement system is not geared for that kind of error tolerance.

Pranesh Prakash: There is a lot we can do even in terms of software, in which India has been seen as a success. We need more focus on what is built on top of models — how we commercialise the AI models that are already in use. The enthusiasm among young developers is very much there, even if they are not always well-versed. Students nowadays in their first year are learning how to use six different development tools right from the very beginning.

We should also promote innovation based on constraints, as Alibaba and DeepSeek have shown.

Tanuj Bhojwani: I think the real question to ask is not whether we should undertake the Herculean effort of building one foundational model, but to ask what are the investments we should be making such that the research environment, the innovation, private market investors, etc., all come together and orchestrate in a way to produce — somewhere out of a lab or out of a private player — a foundational large language model (LLM). The real questions we are not paying attention to are: Why does it not happen? What is wrong with our research? What is wrong with our R&D investments, the spending both by the private market and government?

The IndiaAI Mission under the IT Ministry has announced that it is making GPU clusters available to startups and academia at subsidised rates. What is your take on this announcement and on government AI policy in general as an impetus to investment?

Pranesh Prakash: The capital expenditure is a good sign. I think the government should look at other ways of promoting the responsible usage of these technologies, including amongst them doing careful studies of what kinds of regulations ought to be put and, very importantly, what kinds of regulations ought not to be put.

Tanuj Bhojwani: This does bring down the cost of renting a GPU significantly downwards, but even then somewhere in all of this are the margins that we are paying to data centre operators, providers, etc., to do these things. It is my understanding that both DeepSeek and, of course, all the large labs definitely have dedicated contracts wherein they are buying, installing the GPUs and using them completely by themselves for training runs.

Even with its low training run bills, DeepSeek has over a critical threshold of captive GPUs — they are known to have at least one cluster of Nvidia H-100s, and a calculation doing the rounds puts their investments at $1.3 billion. The Big Tech firms are investing $80 billion a year on infrastructure.

We have resource constraints here; the budget for this is less than the cost of training Meta’s Llama 4 model, and that budget then has to be spread out equitably and fairly among the whole country. The government is not taking that concentrated bet. We are taking very sparse resources that we have and we are further thinning it out.

The subsidy is still a very good thing if people are using it for the right purposes. AI for Bharat is trying to train IndicTrans2, and a text-to-speech system for Indian languages. And those are critical things we need, and you can get away with, say, 500 or 1000 GPUs for that task, and you can get very good results compared to the state-of-the-art. So I think that is really the message that we should acknowledge that we have limited resources, and then we should think about what is the wisest way to spend them. Competing with a [ChatGPT] o3 or [DeepSeek] R1 may not be the best idea with the absolute quantum of resources we are talking about.

Listen to the conversation in The Hindu Parley podcast