Tech Mahindra’s AI Chief Thinks India Is Building the Wrong AI

India’s AI ambitions today are easy to measure. Count the GPUs, count the “upcoming” sovereign AI models, and count the billions being poured into data centres and compute infrastructure.

However, Nikhil Malhotra believes India is measuring the wrong things. The Chief Innovation Officer and Global Head of AI and Emerging Technologies at Tech Mahindra says India’s AI future will not be decided by how quickly it builds another LLM, but by whether it can produce original research, solve uniquely Indian problems, and create breakthroughs that the rest of the world follows.

“There has been no Indian contribution” to the transformer architecture that powers today’s LLMs, Malhotra tells AIM in an interview. “India’s ambition should be building more research-based startups rather than just startups.”

His criticism comes at a time when almost every discussion around the IndiaAI Mission revolves around sovereign models, GPU clusters, and frontier AI. While he believes India must build its own foundation models to remain technologically independent, he argues that sovereignty alone cannot fuel innovation.

“The end goal is to actually produce something which only India has the ability to produce.” That, according to Malhotra, requires a very different mindset.

Lack of Reward Culture for Research

Malhotra believes India has become too focused on reproducing technologies invented elsewhere instead of asking what entirely new problems it can solve. He readily admits that the ecosystem, including Tech Mahindra itself, has contributed to the problem.

Also read: In 2016, PM Modi Saw an AI Demo. 10 Years Later, Its Creator Says India Still Doesn’t Get AI

“I am as guilty as anyone else for not putting papers on arXiv,” he says. “Corporates also need to allow people to put in more and more of these research papers because we look at patentability faster. We don’t put our research first.”

It is a rare admission in an industry that often celebrates product launches, valuations, and funding rounds more than scientific breakthroughs.

For Malhotra, India does not lack talent. It lacks a culture that rewards research.

He argues that the country needs organisations built around long-term scientific work, similar to Google DeepMind, where researchers spend years pursuing difficult problems rather than chasing quarterly product launches.

“India’s ambition should be building more research-based startups rather than just startups.” But perhaps the most interesting part of his vision is not about research labs. It is about where AI should actually be deployed.

He repeatedly returns to what he calls the “diffusion of AI.”

According to him, India has barely scratched the surface when it comes to applying AI to real-world problems.

“We have not seen diffusion of AI,” he notes. Instead of building yet another chatbot, he believes AI should quietly solve everyday problems that affect millions.

He points to public transport as one example.

“Do I really need a ticket to get into the metro? I don't. My face is a ticket enough.” The technology already exists, he argues. What is missing is widespread deployment.

Agriculture offers an even bigger opportunity. "Farmers... would they be very happy if they get a large model or a small model? I don't think they care."

What matters is whether AI can predict rainfall, detect crop diseases, improve yields, or help farmers make better decisions. "There are hundreds of such problems available in India," he adds. "All we need to do is look, make a problem register and start solving them."

Also read: Ashok Jhunjhunwala Thinks India Must Solve Its Hardest Problems First

He believes India's diversity is its advantage. Indian farms, languages, healthcare systems, and public infrastructure generate unique edge cases. Building AI for those challenges could eventually create technologies that become globally relevant.

That philosophy also shapes how he thinks about the next generation of AI itself.

Move Beyond LLMs

While much of the industry remains obsessed with sovereign LLMs, which Malhotra believes are important for critical industries, the conversation needs to move beyond just text generation for other use cases. 

“Let's stop calling them large language models now,” he asserts. “The world now needs to shift into large action models.”

He explains this using a telecom network, where thousands of alarms and support tickets are generated every day. Instead of merely summarising information, an action model should recognise similar incidents it has seen before, recommend the next course of action, choose the right software tool, instruct the appropriate AI agent, and analyse the outcome before deciding the next step.

According to Malhotra, many organisations are trying to deploy AI agents before building the contextual intelligence those agents actually require.

He argues that enterprises are jumping directly from data foundations to orchestration while skipping the "knowledge layer" that includes domain knowledge, relationships between data, dependency graphs, and contextual understanding.

Without that layer, agents simply execute instructions without understanding why they are doing so.

"They are the hands and legs of a nervous system which is run by the language model itself." His vision of AI is therefore less about replacing people and more about building systems that genuinely understand their environment.

He believes that context will become increasingly important as AI evolves.

Take an HR department, for example.

A general-purpose model may know everything about world capitals and prime ministers, but it will not understand an internal company acronym unless it is trained on that specific domain.

That is why Tech Mahindra increasingly fine-tunes smaller models using enterprise-specific data before asking them to perform actions. "While big models are great, I think in our specific context, smaller models will work much better."

Malhotra's thinking extends well beyond enterprise software. He believes the AI industry itself may have become too dependent on brute-force computing.

"I completely agree that mathematics alone is not going to build an intelligent system."

His research focuses on combining neural networks with symbolic reasoning inspired by how the human brain works.

He illustrates this with a simple example from his own research.

While experimenting with the classic Snake game, his team introduced symbolic rewards alongside traditional reinforcement learning. Instead of rewarding the AI merely for eating more food, they rewarded it for reaching the food faster. Overnight, the snake unexpectedly began moving diagonally—a behaviour it had never been explicitly trained to perform.

For Malhotra, that was evidence that symbolic reasoning could produce new forms of emergent intelligence.

If similar ideas scale to larger AI systems, he believes the economics of AI itself could change. "Our brains do all of this using only 25–40 watts of power, in AI, we are spending megawatts."

Then came perhaps the strongest statement of the interview.

"If somebody was to tell me AI is sustainable, I tell them ‘no, it's not’. It's actually the most unsustainable form of tech."

AI may improve not from endlessly adding GPUs, but from understanding how the brain learns with very little data and very little energy.