The reason is that LLMs are a commodity.
“ChatGPT is available to everyone practically equally,” he said. “If you can pay for its APIs (application programming interface), you get the exact same thing. Anybody can build a wrapper on top of these things. So, is it possible that somehow Indian startups can use these technologies to create breakthroughs? Unlikely.”
However, disruption is part of Indian startups’ DNA and there are enough problems in India that GenAI could solve, exactly the way traditional AI technologies have helped scale up engineering systems at Zerodha, Swiggy, Zepto or Ola, he said. “We the developers, the engineers in our tech team including me, are using LLMs for coding heavily,” he said. “It has made our lives much easier, saving us immense amounts of time. But will the presence of AI automatically mean that it’s possible to create lean teams? No.”
A 30-member tech team formed over seven years built India’s largest stock broking platform at Zerodha, even before LLMs existed, he said, adding. “That’s an organisational management philosophy.”
Discover the stories of your interest
A champion of open-source software, Nadh is of the view that the answer to affordable AI lies in this, rather than “renting” it.
“Proprietary AI systems belong to a certain company,” he said. “They can cut off access at any point. Open-source models are the only way to have widespread decentralised innovation, equitable access and de-risking. Otherwise, you’re just paying rent every month to an AI company giving you AI APIs.”
Nadh, who cofounded the Free and Open Source Software (FOSS) Foundation as a young hobbyist developer in 2002, aims to support its growth globally. Zerodha has launched a $1 million a year FLOSS (Free/Libre and Open Source Software) fund for developers and communities who create and maintain projects.
“We built Zerodha out of nothing from scratch using open-source systems–the best possible database system in the best programming language available in the highest quality,” he said. “That’s exactly what every tech company does. You pick any tech unicorn in India–I’d say probably 90%-plus of their entire tech stack will be made of open-source components. But hardly anyone acknowledges it.” As for GenAI, Nadh said LLM performance has hit a wall. The graph of model improvement has flattened in 2024. “We had crazy breakthroughs in 2023,” he said. “Every week there was something big that made something from the prior week obsolete. But that has slowed down. It’s not easy to predict but reflecting on 2024, it seems to have plateaued.”
The models are getting better, but not exponentially better with every release like it was happening in 2022-23.
“So any improvements are going to be incremental but usage is going to increase,” he said.
