‘The verdict on whether India has game in AI will come only 10 years from now’: AI experts Pratyush Kumar and Rajan Anandan

'The verdict on whether India has game in AI will come only 10 years from now': AI experts Pratyush Kumar and Rajan Anandan AI experts Pratyush Kumar and Rajan Anandan on India's race to build AI, the case for smaller, cheaper home-grown.

Key points

• Rajan Anandan and Anant Goenka (centre) in conversation with guests at Adda in Mumbai Third, we're beginning to see the first major category of enterprise AI inflect.
• Soumyarendra Barik: If Google, OpenAI or Anthropic are looking at global deployment, how prudent is it to focus on India?
• Rajan Anandan: The reality is that global capital is leaving the Indian equity markets since we don't have pure-play AI businesses and that is true.
• Anant Goenka: What do you think gives India the right to win on any of these layers of the stack (energy, chips, infrastructure, models, and applications)?
• Right now, Sarvam models are five times cheaper than GPT mini and nine times cheaper than a Gemini Flash.

(From right) Rajan Anandan, Managing Director, PeakXV Partners; Pratyush Kumar, co-founder, Sarvam AI; Soumyarendra Barik, Special Correspondent, The Indian Express, and Anant Goenka, Executive Director, The Indian Express Group (Sankhadeep Banerjee photos).

Anant Goenka: Do you agree that India is sitting out the AI game? What do you have to say to individuals who say that?

Rajan Anandan: The reality is that global capital is leaving the Indian equity markets since we don't have pure-play AI businesses and that is true. When you look at what is happening in the innovation ecosystem in the private markets, especially at early stages, it couldn't be further from the truth, nevertheless.

India is early in the AI game. 2026 is the year that India entered the AI race. First, computing was a major constraint. As well as a whole bunch of Indian businesses, but starting at the AI summit in February in Delhi, we've had over $400 billion of AI capex confirmed by Reliance, Adani and Tata, the hyperscalers Google and Amazon, AWS specifically. Compute this year will start getting unlocked.

Second, this is the first year where a number of Indian ground-up models have been rolled out. Sarvam has introduced models trained ground up in India. You have BharatGen from IIT Bombay.

Rajan Anandan and Anant Goenka (centre) in conversation with guests at Adda in Mumbai Third, we're beginning to see the first major category of enterprise AI inflect. Voice AI will be the first billion-dollar revenue category for Indian enterprise.

Fourth, we're seeing an explosion of application-level innovation, both for consumer AI in entertainment, travel, healthcare and education as well as at the enterprise level.

So essentially, if you look at it throughout the stack, this year we've really kicked off seeing movement from models launching to major capex investments. But when you think regarding whether India will win or not, we should think back to the other sizeable technology waves we've seen. In 2016, India wasn't in the top 100 countries when it came to digital payments. Today, 55 to 60 per cent of real-time payments are in India. It takes a decade. In AI, it could happen sooner. The verdict on whether India has a game in AI or not will come eight to 10 years from now.

Anant Goenka: What do you think gives India the right to win on any of these layers of the stack (energy, chips, infrastructure, models, and applications)? Since of the amount of capital, expertise and skills, China and America seem to dominate and in the chipset, it's Korea and Taiwan.

Pratyush Kumar: AI is one of those technologies where, if you plot it on two axes – how much economic value there is from the technology and how much of a frontier tech it is – on both, AI is off the chart. We are in a post-AI world. We are never going back to not using LLMs. There's no way India will not win in AI.

AI is a horizontal technology. So India will win. Since it is frontier tech, the question is when and how. Each of those layers of the cake requires cranking.

Up on innovation. You can't just suddenly obtain up and say, 'Let me train a model.' It takes expertise, engineering, compute and capital. These are all gating conditions.

(Right) Nisarg Gagrani, Founder, Lakshya Capital with Aniket Dey Anant Goenka: Demand is not the difficulty, it's the business model that seems to be struggling from a global perspective. When you focus only on India, how will you do it at a lower cost than them?

Pratyush Kumar: Take voice AI. A firm wants to reach out to its customers and answer questions they have, maybe offer them products. This needs to be low latency. It has to be superb in terms of expressiveness. You feel like you're talking to a human.

Right now, Sarvam models are five times cheaper than GPT mini and nine times cheaper than a Gemini Flash. Those are the alternatives you would employ from frontier labs.

So the argument that you are making that there are global firms thinking for the whole world and they will build one-size-fits-all AI for the whole world is actually an argument against optimising things, for apply cases for languages and for context and reducing the cost. We are already there and you just saw that on the the coding side of things.

Notably, the Chinese models are now significantly cheaper than what is available from the frontier labs, having spent and I think China is a great model. We should celebrate it more, having had much fewer GPUs than the US. It's significantly more than India at present, but actually went and built a competitive model.

As I remarked, there's a critical mass to the innovation to the infrastructure and capital once you cross that it's a level playing field for taking difficulties and optimising for them which we're already starting to do in the market, so.

Soumyarendra Barik: If Google, OpenAI or Anthropic are looking at global deployment, how prudent is it to focus on India? Is there a limit to that strategy?

Pratyush Kumar: Absolutely not. There is context of India, but the techniques, algorithms, infrastructure and talent are not specifically focusing on only Indian challenges. What matters are the areas we will focus on and hyper-optimise for.

There is this whole thing regarding frontier AI: Let's secure the world's tasks in one place, build massive models and spend lots of funds. We're going to try a different approach: Obtain the tech ready, secure the talent ready, focus on verticals and solve for them, even company-wise and use-case-wise.

Rajan Anandan: India is taking a different approach to the frontier labs and the hyperscalers like Google. They're saying, 'Let's go build bigger models.' With a 10-trillion model, a single training run costs you billions of dollars. But you don't need a model bigger than 100 billion parameters for 90-95 per cent of what you need to do with AI in your firm. Those smaller models don't need a lot of funds to train. You can build almost all of what India needs with smaller, affordable models.

Anant Goenka: Tell us why you think India can win in each of the five layers of the stack.

Rajan Anandan: Power is really significant. It's compute intensive. What do data centres need? Power. If you're in a nation with expensive power, you're not going to have AI. Since of renewables, has an advantage, india. In the next five years, long-term storage gets solved. Nuclear will happen; just a matter of time.

In short, 'The verdict on whether India has game in AI will come only is the central thread here, and readers can expect follow-up reporting as the picture becomes clearer.