# Our Compute Cost Was Lesser Than DeepSeek, Built 3 Models For $20 Million: India’s Sarvam AI > Source: > Published: 2026-07-27 12:49:02+00:00 India is one of the countries that has also thrown its hat into the AI model creation right, and while it might not currently be close to the frontier, it says it’s managing to build models for cheaper than what most people expect. [Sarvam AI](https://officechai.com/ai/indias-sarvam-raises-234-million-at-1-5-billion-valuation/) co-founder and CEO Pratyush Kumar has said the Bengaluru startup’s total compute spend across three model builds came in well below what DeepSeek reportedly spent on its older releases. Speaking in an interview, Kumar put a number on it, something AI founders have rarely done in public. “The compute cost was definitely lesser than what DeepSeek took by significant margin,” Kumar [said](https://www.youtube.com/watch?v=S81CzYD1Ldo). “In fact, the whole compute was about 20 million, or a little bit less. And we trained three models.” Kumar was quick to clarify that the figure isn’t the cost of a single training run, but the cumulative spend across all the work that goes into getting a model out the door. “It’s not about one run that it costs. It’s a lot of exploratory work. It’s a lot of data creation work that gets layered. But the bottom line is, it is not billions of dollars. It’s a single-digit million dollar model,” he said. When the interviewer pressed him on whether this meant a startup running on single-digit million dollar funding could realistically build a competitive model, Kumar didn’t hedge. “Absolutely,” he said. “And the bottom line is this, right? You need to have the talent.” The claim lands at an interesting moment for Sarvam. The company recently [raised $234 million at a $1.5 billion valuation](https://officechai.com/ai/indias-sarvam-raises-234-million-at-1-5-billion-valuation/) in what is being called the largest Series B round in Indian startup history, with the fresh capital earmarked for training its next frontier model and buying additional compute. That round came months after Sarvam released [Sarvam 105B](https://officechai.com/ai/india-ranks-6th-among-countries-with-ai-models-with-sarvam-105b/), a Mixture-of-Experts model with 105 billion total parameters that was pre-trained entirely from scratch in India, pushing the country to[ sixth place globally](https://officechai.com/ai/india-ranks-6th-among-countries-with-ai-models-with-sarvam-105b/) on the Artificial Analysis Intelligence Index. Much of that training relied on compute the company didn’t have to pay for outright. Sarvam was the first company selected under India’s sovereign AI mission, and the government agreed to [provide access to 4,096 Nvidia H100 GPUs](https://officechai.com/ai/indian-govt-to-take-equity-in-sarvam-ai-in-exchange-for-providing-compute-says-co-founder/) for six months in exchange for taking an equity stake in the startup. Kumar’s framing pushes back on a narrative that building a genuinely capable model requires the kind of balance sheet only a handful of companies in the world can afford. Sarvam’s numbers suggest the entry price is lower than that, provided a team has access to cheap or subsidised compute and knows how to use it efficiently. Whether $20 million across three models holds up as a template other Indian startups can follow remains to be seen. Sarvam had the advantage of government-backed GPU access that most private companies don’t get, and its models are still some distance behind the true frontier labs in the US and China. But for a country trying to prove it belongs in the AI conversation at all, a number that low is the kind of headline India’s AI ecosystem has been waiting to make.