Sarvam’s Mission to Build a UPI for AI

“What does it take to pull off a UPI for AI in this decade?”, that’s the question Vivek Raghavan and Pratyush Kumar, the co-founders of Sarvam AI, posed during their first interaction with us. While we have had the privilege to interact with many amazing teams building in Artificial Intelligence (AI) over the course of the last several quarters, the scale of ambition of Sarvam’s team left us both intrigued and inspired. We have had the joy of working with these founders since day one, co-leading their seed round. Today, we are excited to formally announce our partnership with them.

As we deliberated this question over the course of several subsequent interactions with Vivek and Pratyush, we appreciated the simplicity and contrarianism of their insight. While the most ambitious teams working on AI across the world were caught up in a race on model size and performance, Vivek and Pratyush recognized the deflationary power of AI to commoditize technology and services that are prohibitively costly today. They shared with us their dream of building a population-scale AI system.

Sarvam’s core thesis is very simple, yet powerful: we are in the early days of this exponential adoption curve of AI, and we need to build population-scale infrastructure to enable rapid adoption. This requires cutting-edge innovation on foundational models and a fundamental rethinking of the technology stack.

The founders often pose a simple question when asked about their vision, “How can we deliver $10 worth of AI value to 500 million Indians by 2030?”

The skeptics will point to the challenges in this quest, from realizing value from selling software in India to the challenges in competing with multibillion dollar companies. But the last many years have shown us the technological success India can have in building systems of enormous scale—be it the 97 billion authentications enabled by Aadhaar, the country’s nationwide resident identification system, the 11 billion monthly transactions on UPI, or the 50 million concurrent viewers watching the India-Pakistan cricket match.

This aligns closely with our thesis: our competitive advantage lies in helping startups build systems of scale that can deliver value and reliability.

The use cases that AI can enable are immense. From rapidly democratizing access to quality education in classrooms across the country, to enabling financial transactions through voice-enabled by the identity and security of Aadhaar, to helping millions of Indians access preventive and personalized healthcare advice. AI’s use cases are perhaps limited only by one’s imagination. 

Beyond this revolution for consumers, India also has a thriving knowledge economy which will be transformed by AI—from pharmaceuticals discovering new drug combinations to IT services modernizing the world’s codebase using it as a co-pilot; from BPOs automating processes with quality assurance provided by humans-in-the-loop, to manufacturing safety and standards maintained by introducing AI on factory floors. AI will be present across many aspects of our daily personal and professional lives in the next few years.

Vivek and Pratyush are the accomplished technologists embarking on this mission. Pratyush has been conducting world-class research in systems, from GPU design and programming to language models at Microsoft Research, IIT Madras (as faculty), and IBM Research before starting Sarvam. 

We came across his work on AI4Bharat and were lost in the depths of his blog on spiritualism and AI. In him, we noticed a rare combination of a world-class researcher with boundless ambition and an unusual clarity of thought about how to serve India’s AI revolution. We have not been surprised that in the last few months that we have been partners, several of his best students have left lucrative jobs across the world to join him on his mission at Sarvam. 

Vivek, on the other hand, had played a crucial role in architecting the world’s largest and most sophisticated biometrics database at Aadhar. We had learned how Aadhar built world-class face recognition and deduplication models through Suvrat Bhooshan, the founder of one of our portfolio companies, Gan.ai. Suvrat had worked with Vivek on this. 

It was only during our subsequent interactions with Vivek that we realized that he is a deep technologist and a serial entrepreneur himself. During his PhD at Carnegie Mellon University (which he finished in three years instead of the usual five), he designed SPICE, an analog electronic circuit simulator, which is now the default method used across the semiconductor industry. He started two companies in the electronics design space, and both were acquired by Synopsys. He had retired after leading large business units there and then came forward to volunteer to build Aadhar. 

Sarvam is assembling a world-class team to deliver on a bold and ambitious promise to innovate on two key fronts:

1. Models: Sarvam will launch pre-trained foundational models, including focus on Indian languages. Not only is this a strategic imperative in the AI race where leading global models’ training datasets have less than 1% Indic language content, but a tokenizer trained on Indian languages can make AI-serving dramatically efficient and cheap. Data is the most important ingredient in model performance, and India has unique long-tail content that can be leveraged to serve local needs much better than other horizontal models. On top of that, our demographic dividend also allows us to undertake the world’s biggest and most sophisticated reinforcement learning from human feedback (RLHF) and steering the model in ways that are controlled by the Indian community.
   
2. Platform and ecosystem: AI is eating software. The way software is created, deployed and consumed will change dramatically. India will play an important role in developing and maintaining post-genAI software stack used by the world. This not only includes deep innovation on infrastructure at a systems level (chips, compilers and the works), but also crafting a beautiful developer experience that gives them intuitive control in authoring genAI applications that can be shipped to production reliably. This full-stack platform will be able to provide multiple knobs for robustness and efficiency, along with baked-in integrations with common APIs and protocols for Indian audiences. 

We are excited to partner with Sarvam on their journey, as they enable India to leapfrog the world in AI adoption and help create multiple new  businesses built on this state-of-the-art platform.