Sarvam AI: How one startup is building India's unique path in artificial intelligence

INTELLIGENCE WAS NEVER a virtue that humans aspired to — till we got artificial about it. And in this ever-intensifying race for supremacy in artificial intelligence (AI), the world has largely been a spectator to a Silicon Valley spectacle. From OpenAI’s astronomical funding rounds to Anthropic’s quiet but calculated ascent, the landscape was largely shaped by western tech behemoths. The status quo was disrupted by China, with DeepSeek emerging as a formidable player. The company revealed that it trained its V3 model in just two months, spending under $6 million and using the relatively cheaper Nvidia H800 GPUs. In other words, it operates at 5% of the cost of traditional models.

But what about India? While the country has long been a thriving hub for software services, with a deep talent pool that has also powered global tech giants, there was a crying need for an AI-first company that builds from the ground up — not just tinkering at the edges of innovation but owning the full stack. That is exactly what Sarvam AI wants to achieve — it “seeks” to make a “deep” and lasting change in the AI narrative.

Founded in July 2023 by Vivek Raghavan and Pratyush Kumar, it aims to make Generative AI accessible at scale in India. Both co-founders were previously involved with AI4Bharat, a research initiative focussed on open-source Indian language AI. Raghavan brings over a decade of experience from the UIDAI, which oversees Aadhaar, while Kumar, a PhD graduate from ETH Zürich and an IIT Bombay alumnus, co-founded AI4Bharat to advance Indian language AI applications.

Kumar’s vision is clear: India needs its own foundational AI models, ones that are not just adapted from western counterparts but built from scratch, trained on Indian data, and deployed securely within the nation’s borders. “Come 2040, India must have companies capable of training and deploying these models independently,” he tells Fortune India.

The urgency of this mission was underscored by what Kumar refers to as the “DeepSeek moment”. When China opensourced its AI model for a fraction of what was previously believed necessary, it sent ripples through the global AI community. The notion that only companies with billions of dollars at their disposal could build cutting-edge AI models was suddenly called into question. “DeepSeek proved that you don’t need billions to build a competent model. This fundamentally alters what is considered defensible in AI,” Kumar explains.

For Sarvam, this shift was both an affirmation and an opportunity. The only difference for Kumar is that sovereignty in AI is not just about developing foundational models — it is about owning the entire stack, from data generation to deployment.

Unlike many AI startups that are content with fine-tuning existing models, Sarvam has set out to build its own. Initially, it collaborated with Meta to refine Llama models for Indian languages. But Sarvam soon realised its true calling — building India’s AI stack from scratch.

Axonwise Pvt. Ltd, the legal entity of Sarvam, is setting out to establish a truly sovereign AI ecosystem for India, one that is built in India, trained on India’s vast linguistic and cultural diversity, and deployed on India’s terms. In 2023, Axonwise unveiled its flagship large language model (LLM) model, Sarvam-1, which was trained on 2 trillion tokens, followed by a much more advanced Sarvam 2B, a 2-billion parameter LLM trained on a proprietary dataset of 4 trillion tokens, focussed on 10 Indian languages, including English. Now, the company is planning to take a far bigger leap — building a 70-billion parameter model, with an estimated price tag of $40–50 million.

Business model

If foundational models are Sarvam’s backbone, its business model is its heart. Unlike OpenAI, which is still navigating its transition from a nonprofit research lab to a revenue-generating juggernaut, Sarvam has been built as a for-profit enterprise from day one. But its monetisation strategy is not simply about licensing models — it is about embedding AI deeply into the fabric of India’s economy.

Sarvam is building AI-powered applications that cater specifically to India’s linguistic and operational realities. The first major product line focusses on conversational AI. In a country where digital adoption has surged but literacy barriers remain, voice-based AI has the potential to revolutionise customer service. For instance, Sarvam has developed Shuka v1, which focusses on audio understanding, enhancing AI’s capabilities in processing spoken Hindi. Companies in quick commerce, financial services, and government-citizen interactions are prime candidates for adoption. “For companies such as Zepto or Zomato, customer service is expensive. AI is much cheaper. It can speak to people in their preferred language, resolving queries in seconds,” Kumar explains.

Large consumer-facing companies in India — banks, airlines, telecom, and similar industries — employ tens of thousands in customer-facing roles. “Automating these roles with AI could be transformative, especially in a multilingual market like India. So, while it’s early days, I see a strong potential for Indian language LLMs to create meaningful value in sectors with heavy customer interaction,” says Harshjit Sethi, MD at Peak XV Partners, which was part of a consortium that invested $41 million in 2023.

The second major revenue stream focusses on AI-driven productivity tools for the government. Sarvam is working with NITI Aayog to develop AI-powered systems capable of analysing vast amounts of public data, generating reports, and assisting policymakers in real-time. “The government is sitting on vast troves of data. We’re building an AI that can make sense of it — at scale,” Kumar says. If successful, this could become India’s next digital public good, akin to UPI.

Sarvam is not trying to compete with SaaS companies that are layering AI onto existing enterprise products. It is doing something far more ambitious: making AI a foundational layer of India’s digital economy. Unlike OpenAI or Anthropic, which have built global models with limited localisation, Sarvam AI is focussed on India’s unique needs.

Kumar describes Sarvam’s approach as radically different from the AI-as-a-Service model popular in western markets. He and his team are investing in infrastructure, ensuring that models are trained using locally available compute resources, and designing systems that can be deployed within Indian institutions without relying on external cloud services. “I cannot name them but we’re deploying models within critical government premises. It’s about creating an AI backbone that India controls,” he says.

For Sarvam, the next few years are critical. The company is already engaging with over 25 enterprises, including conglomerates, digital-native startups, and government bodies. Sethi adds: “They’ve already secured pilot projects with several top Indian conglomerates across different sectors, which highlights the potential for AI in localised, voice-based solutions.”

While the interest is strong, Kumar acknowledges that business adoption in India takes time. “The traction is promising, though doing business in India takes time. We are seeing strong inbound interest, and this year will be an inflection point,” he notes.

The company will need to continue raising funds — it has secured $41 million so far, a fraction of what its global competitors have raised — but the long-term vision is unwavering. “Building AI models in India is a capital-intensive race, and regular fundraising is inevitable,” says Kumar. If his vision holds, Sarvam AI won’t just be another AI company. “The next decade is uncertain, but AI is the new semiconductor race — we must build the capability to stay independent,” says Kumar.

In other words, Sarvam aspires to be the foundation of India’s AI independence.