Building the Future of Voice-First Sovereign AI: Sarvam & NVIDIA

本站添加: 2026-06-03

[00:00:00]We always believed that AI is such a critical technology that a country of the size of India, we should be building AI grounds up in India. So the two main focus areas for us are the ability to do full stack. We go from uh the data sets required to the models to be built and also the applications uh that can be built on these models. And the second thing is we want to do it with foundational research uh which can improve each of these layers while being completely sovereign about it. We took up the problem of doing it for Indian languages and doing it completely in open source. We understood the nuances of languages, the long tale of challenges and so on. And that's where Serum was born. We have an API platform today uh which serves more than 4 million API calls a day. By far the largest AI API effort out of India. And we're doing all of this on technologies.

[00:00:50]We're using the entire Nvidia stack for training these models and inferencing them at scale. So doing this entire thing of the platform, the model and the applications gives us lot more levers of optimization and quality. For all the work that we do at Saram, we start with data. Uh and data requires curation entire pipelines around ensuring that the quality of data is good and we've been extensively using the Nemo curator platform. In fact, we have trained now large language models from scratch and all that data for tens of trillions of tokens for millions of hours of audio, billions of images. All of that has flown through Nemo curator and that tool has really scaled and now we understand that deeply and the value it brings. The training itself we have done with the Nemo framework. The pre-training, the fine-tuning, the reinforcement learning.

[00:01:38]In fact, reinforcement learning is something that's giving consistent dividends at scale and we've been using the Nemo RL framework for that. And of course we have been um doing inferencing with our models at at some fair scale.

[00:01:49]So we've been using training and inferencing stacks extensively primarily on the hopper series of GPUs. In India having such a large developer base I think should be building AI not just consuming AI and that requires being expert in the new stack of software for accelerated compute. Nvidia stack is one great example of where people can do this. I genuinely believe developers should think about this as the core of development going ahead because whatever we build will hit these genative AI models. With NVIDIA, we are looking really forward to building models that represent the diversity of India and also serve them at scale so that it actually is a population scale effort rather than a few people just using them.