AI will reshape how we live, work, and operate in profound ways: NVIDIA’s Jigar Halani

India stands at an unusual intersection in the global AI landscape. The country does not share the same starting point as many Western economies. While the West often struggles with labour shortages, high wages, and limited skilled manpower, India’s reality is quite different.

In this context, Jigar Halani, Director of Solution Architect and Engineering, NVIDIA, shared a grounded view of what truly shapes India’s AI trajectory—not in shortage of talent or manpower, the constraints lie elsewhere, in areas that demand a uniquely Indian approach to building and deploying AI systems.

“We as a country do not have a problem or a dearth of resources. We have absolutely good skills and manpower… Our challenges are slightly different. Our challenges are language first and then the cost,” he explained at TechSparks 2025.

The realities of India’s multilingual society have long limited digital access. Initiatives such as Bhashini and AI4Bharat seek to bridge this gap by creating open-source foundations in Indian languages because inclusion begins with communication.

Cost is the second pillar of India’s challenge, according to him, as low labour costs make AI adoption difficult to justify unless it delivers overwhelming savings. A call centre seat in India can still cost far less than in many countries, so AI must operate at very low per-transaction costs and high reliability.

The last pillar is scale, Jigar said, adding that India is already the largest user base for several global platforms, which means failures or inefficiencies have amplified consequences.

“We are the largest when it comes to X, OpenAI, Perplexity, number two probably for Gemini, largest for WhatsApp, largest for everybody. That is our volume,” Jigar noted. Any AI system designed for India must therefore plan for national scale from the outset.

Scalable and efficient AI infra

These challenges naturally lead to the need for infrastructure that can withstand India’s demands. AI in India must be powerful yet affordable, high throughput yet resilient, and capable of running across cloud, on-premise and edge environments.

“The architecture needs to be sorted out. It should be so pervasive that it should work everywhere. It should be built with that resilience in play. The ROI matters, and what the technology curve that’s futuristically going to come in, all these things need to be catered to it essentially.”

Efficiency is a major driver. When a single GPU rack can process over a million tokens per second, inference becomes affordable even at a population scale. However, it is not enough. The infrastructure must be robust enough to support hundreds of millions of concurrent users, because downtime has an outsized impact in a country as large as India.

This combination of cost efficiency, scale, and resilience is not optional; it forms the foundation on which all practical AI deployments must be built.

Data curation and specialisation

While infrastructure attracts much attention, Jigar emphasised that effective AI is not driven by hardware alone. Data curation forms the true backbone of any meaningful system. Models trained solely on publicly scraped data remain generalists and cannot meet the level of precision that real organisations require.

“Data curation is the largest of the largest lot, both in terms of heavy lifting and compute hungriness…Everybody is a generalist because everybody has web data. A specialist comes when you are adding your own data,” he remarked.

Curation means cleaning, filtering, and aligning data so that models learn domain-specific understanding. It is a demanding and resource-intensive process, but it creates a genuine advantage. In India, this work is even more vital because curated data must reflect local languages, cultural nuances, and on-ground practices.

Open-source datasets and curated libraries reduce barriers, but each organisation must still refine and expand its own specialised datasets if it wants to differentiate its systems from generic models.

AI agents in industry

Jigar highlighted that AI agents are not hypothetical ideas. They are deployed today in industries that rely on speed, accuracy and cost-effectiveness. Some examples illustrated how efficient modern agentic systems have become.

Domino’s Pizza uses AI agents to compute the fastest routes for drivers across the United States. 

Jigar noted, “See what is the fastest route I can make it reach. That just runs on one GPU for the whole of the US for the entire pizza delivery throughout the day.” This demonstrates not only efficiency but also the maturity of practical AI deployments.

A more local example appeared in India’s judicial system, where courts are piloting AI that transcribes courtroom proceedings, summarises arguments, and suggests applicable legal sections, which supports faster decision-making and reduces manual burden.

“Capturing everything: who is speaking, what language they are speaking, transcribed, summarised, connected with laws, and personalised for each judge,” Jigar described.

Enterprises are also adopting agents for customer service and engineering support. These systems can read documentation, extract relevant details and provide guidance, often replacing hours of manual research.

Agentic and physical AI

The next phase of AI development reaches beyond digital workflows into the physical world. Agentic systems are beginning to collaborate with real-world robotics, autonomous vehicles, and digital twins to bridge simulation and physical environments. This expands the impact of AI into warehouses, hospitals, factories, and homes.

As agents gain the capacity to learn from feedback, adapt to evolving workflows and interact with physical systems, they will become part of everyday operations.

“AI started with a very simple intent of solving some very complex problems on the data sets that we have. It is going to evolve and revolve the whole world in a very different way, how we are living, breathing, operating and working together,” he explained.