From first check to first frontier conference
Hemant Mohapatra · August 2026
“We must have the stubbornness to accept our gladness in the ruthless furnace of this world.” — Jack Gilbert

When we wired Lightspeed’s first check into Sarvam in mid-2023, the company had two founders, an idea, and a belief. The entire thesis came down to one word: token production. India was about to become the largest consumer of intelligence on the planet and we produced almost none of it. Every token we consumed originated elsewhere, priced elsewhere. Lightspeed became Sarvam’s first institutional investor. We led the seed round and the series A: all of this before the models were built, before the IndiaAI Mission, before sovereign AI was a topic. I wrote the Sarvam Manifesto earlier this year to explain why we believed Sarvam will become one of the most important AI companies in the world. Consider this Part-II.

The world we saw in 2023

Three beliefs, really. (1) Intelligence would follow the internet’s curve in India: from scarce and imported to cheap and everywhere, and whoever produced it locally would compound for decades. (2) Voice would beat text; hundreds of millions of Indians will never type, but everyone talks (and we love talking!) (3) None of this works as a wrapper on someone else’s model; it had to be full-stack, built for our languages at Indian price points, at population scale.

And then the founders. Vivek had built and exited a few startups before and then spent 12 years volunteering inside Aadhaar. Building population-scale infrastructure for India was muscle memory for him. Pratyush left a professorship at IIT Madras behind to do it all over again from zero. And oh, both are PhD scholars. In 2023, pouring tens of millions into two scholars was called brave. Some called it something less polite :)

“Intelligence made in India.” - Lightspeed’s seed thesis on stage 3yrs later

The stumble, and the grind

I won’t pretend the line from there to Epoch was straight. In May 2025, weeks after the IndiaAI Mission selected Sarvam to build the country’s sovereign model, the team shipped Sarvam-M: a fine-tune on an open base. It was an experiment for us, but when you put something out on the internet, it creates its own memory. And the internet was not kind and for a fortnight the whole country debated whether India could build frontier AI at all. Board calls in those weeks were not fun. What struck me was how unmoved Vivek and Pratyush were. It’s a stepping stone, they kept saying, and then they went dark. Two quarters of hard slog later, they had pre-trained 30B and 105B parameter models from scratch on Indian compute. In February came 14 launches in 14 days. Probably one of the best launch sequences I’ve ever seen. The world noticed Sarvam. Naysayers tuned in. The Prime Minister wore Sarvam’s Kaze glasses on stage. This team powered through. They persevered.

Epoch — the crescendo

Which brings me to two weeks ago. I spent the Sunday before the keynote judging the Epoch buildathon: ~200 builders, 8 hours, everything on the Sarvam stack, top-10 teams earning the main stage. Then two days and twelve launches. As an investor, we are incentivized to sing praises of the companies we invested in, but some of the launches truly blew me away - it felt like we touched global quality many times in these two days. A few of my favorite ones below

On speech. Every ASR system on earth breaks when two Indians talk over each other… which is roughly whenever two Indians talk :) The standard fix is a three-pass pipeline that chops the audio and loses the overlaps. Saaras V4 Multi-Speaker does diarisation and transcription in a single pass, one model, and the team ran it live against ElevenLabs and Deepgram cleanly pulling apart a male speaker and a female anchor talking over each other in real time. I thought this tech should have immediate PMF in all parliamentary systems! 

Bulbul V4 was the other half of the speech models, a single reel that went from theatrical narration to a whisper to something sombre, steered with simple emotion and emphasis tags in the text. There were definitely goosebumpy moments here: when I heard sun lo duryodhan! It immediately took me back to my childhood in Chandigarh when streets would clear out on Sundays because Mahabharata was playing on DD-1. Then another voice declared chattrapati shivaji maharaja cha parakrama! in marathi and I thought of my years spent in mumbai as a young student in college, visiting those crowded ganesh puja pandals, and those many, many afternoons taking trains to and from “CST” (Chhatrapati Shivaji Terminus) to Kanjurmarg! Voice is such an important part of our lives - voices are memories. There is nothing more humanizing and grounding than hearing someone speak your mother tongue; there is a reason the word “mother” sits in there. AI for everyone should speak to everyone as they are. Sarvam’s mission finally felt real, personal. 

On inference. One of the most important, long-term bets is inference, from India for India. Sarvam Inference serves frontier open models, including their own 105B alongside GLM and Gemma, entirely from Indian infrastructure, at ~$0.80 per million tokens vs $4.50+ for the comparable US offering. Banks, hospitals, and ministries here increasingly cannot ship data abroad; DPDP made residency a feature. Vivek called it token sovereignty. A country that imports 100% of its tokens is as exposed as one that imports 100% of its crude.

And on the trillion. And finally, announcing an SF research office, with Devendra Chaplot joining as advisor to help build globally competitive models from India. Pratyush said Sarvam would start with a 1T+ parameter model, trained from scratch in India, aimed at coding, cybersecurity, simulation and science, targeted to go live in six months.

Within five years, we believe no serious AI workload in India will run without a Sarvam layer somewhere in its loop.

Lightspeed’s journey with Sarvam

At the Builder Day

Working with this team has been one of the great joys of my investing life. I’ve watched Pratyush grow from a professor to founder-CEO without losing the professor’s precision, and seen Vivek carry his years building Aadhaar into building a team and a company that thinks in populations by default. I am but a small member of a much larger Lightspeed team that has tirelessly worked on anything and everything Sarvam needs. Abha & Sid who have been helping hire a world class team, Anuj & Vidur who have helped Sarvam raise $350M so far, Shashank & Gaurav who have inserted themselves almost as full-time members of this team on special projects, our legal and finance team with Amit, Pavani and Vinayak tirelessly handling all the complex work Sarvam throws our way :)  

My favorite poet Jack Gilbert wrote in his poem A Brief for the Defense (which I'd highly recommend) that We must have the stubbornness to accept our gladness in the ruthless furnace of this world. Startups are one of the most intense experiences one could go through in their professional lives - no less than a furnace in its own way. But a furnace is also a forge and it’s been an absolute joy to see Sarvam grow stronger through the years and do what they do best in a way that defines class, quality, and craftsmanship. 

Onwards.