India’s AI Mission Isn’t About Chatbots. It’s About 800 Mill…

When three entrepreneurs building India’s own foundation models under the IndiaAI Mission came together for a panel discussion, the venue at the Bengaluru Tech Summit 2025 had to be obviously packed. The audience was eager to listen to what they had to say about the country’s AI ecosystem. 

Moderated by Kalika Bali from Microsoft Research, the session included Vivek Raghavan from Sarvam AI, Sashikumar Ganesan of Zenteiq, and Ananth Nagaraj of gnani.ai. They discussed why India cannot depend on global systems and must build its own base models, and what this means for the next 800 million Indians who don’t live inside Bengaluru’s ring roads.

The three startups, among the 12 empanelled under the IndiaAI Mission, have all taken up different verticals for building foundational models, ranging from a linguistic focused model to voice and material discovery models.

Raghavan of Sarvam AI arrived at the point directly. “If we don’t put an effort in foundational models, we will become a digital colony,” he said. He explained the risks of depending on open models whose origins and training data are unclear. 

He warned that even an open model “can actually be poisoned with a very small amount of data,” and pointed out that many top-performing open models today come from China, like DeepSeek. The implication was obvious. For a country the size of India, relying blindly on such systems made no sense.

He also mentioned figures to back his argument. “Certain early open models had maybe sub 1% Indian data,” he said. Sarvam’s upcoming model, built under the IndiaAI Mission, will be a 120 billion parameter system with 17 trillion tokens, around 15-20% from Indian sources. 

Raghavan said that distillation and domain-specific SLMs will become normal as applications scale for agentic AI.

The Missing Link

Sashikumar Ganesan of Zenteiq widened the discussion with his calm approach. “If you do not know how to build the foundational models, then definitely in the next wave we will be behind,” he said. According to him, India missed the supercomputing wave and paid the price. Missing the foundation model wave would be another generational loss. 

ZenteiQ.ai, formerly Zentech AI, is building BrahmAI, a scientific foundational model for engineering intelligence, scientific computing, and industrial innovation. Its approach differs from typical LLMs: the model will understand and validate physics-based scientific questions.

“Our focus is Industry 5.0—applications in aerospace, automotive, EVs, energy, and pharma. India relies on foreign software for industrial R&D. BrahmAI is about sovereign scientific AI,” Ganesan had earlier told AIM.

Initial phases aim for 35 billion parameters, eventually scaling to 80 billion depending on compute availability. ZenteiQ has been allotted 2,128 H200 GPUs for the first year, but wishes to scale it further after that. 

Ananth Nagaraj of gnani.ai pulled the conversation away from the labs and into a village 200 kilometers from Bangalore. “AI is a necessity for the next 800 million people,” he said. The internet in his village is still used for WhatsApp and YouTube, but the real potential is elsewhere. 

“Unless we control the whole tech stack,” he said, “we are prone to all sorts of attacks.” His concern was beyond inclusion, about security, and sovereignty in the most literal sense. And it was about building tech that solves problems for people who don’t have the luxury of English, typing, or even silence around them.

Gnani.ai is developing a 14-billion-parameter multilingual voice AI model with 1.3 crore GPU hours.

Read: India’s AI Push Might Be Pointless Without National Language Standardisation

India’s Unique Problems

Nagaraj explained the everyday problems with a clarity that cut through the noise. Indian conversations are noisy, code-switched, full of background chaos. People talk from speakerphones in buses, fields and markets. “We handle close to 10 crore voice calls. One lakh audio calls every second,” he said. 

They need to give a response in 150 milliseconds. It is a different universe. Western benchmarks simply don’t apply.

This is why Indic foundation models must be voice-first and robust enough for railway stations, farms and government offices. Bali reminded the audience of a previous ASR deployment that collapsed instantly because the model had never seen noise like Indian railway stations. Ananth nodded. This wasn’t a theory. This was daily reality.

Ganesan broke down how scientific foundation models work and why they can’t rely on standard transformers. “You cannot mix match. The next operator is not a probabilistic operator,” he said. Physical laws matter. Equations matter. Encoders must understand scientific structure, not only tokens. 

India needs such models for materials, energy, climate and manufacturing. He said India’s biggest challenges — energy, EVs, climate threats, materials — won’t be solved by generic chatbots. They need scientific reasoning, not autocomplete.

Raghavan also added that use cases matter, but only platforms change the country. India Stack succeeded not because it solved one problem, but because it created rails for the future. AI, if built right, can play the same role. “AI is an accelerant,” he said. 

He warned about the AI divide being even worse than the digital divide. 

The only way to avoid that future is to make sure every citizen gets access to it. Not a few thousand engineers. Not a few million users. Everyone.

The post India’s AI Mission Isn’t About Chatbots. It’s About 800 Million People. appeared first on Analytics India Magazine.

Generated by RSStT. The copyright belongs to the original author.

Source