The Sarvam Moment: How the local AI start-up could reshape India’s tech push

Days after the India AI Impact Summit in New Delhi, and fresh from the launch of Indus—a conversational interface—Vivek Raghavan enters the meeting with BT in his Bengaluru office in high spirits. The co-founder of Sarvam AI appears relaxed and quietly pleased—the company, after all, unveiled India’s first foundational AI models during the summit. But our conversation quickly moves beyond product momentum to something far more consequential. “Sovereignty is important in the long run,” he says. “It’s not about what we build today, tomorrow, or even 10 years from now. You have to look at the long arc.” He is blunt. Without domestic AI capability, India risks becoming merely a consumer in a future shaped elsewhere, a new form of “technological colonisation.” That long arc started taking tangible shape at the summit, one of the largest global AI gatherings, which drew delegations from over 100 countries. There, Sarvam unveiled the first homegrown foundational models—Sarvam-30B and Sarvam-105B—built from scratch for India’s scale and diversity. Designed using a mixture-of-experts architecture for high efficiency, the models support more than 22 Indian languages and have been built with a clear premise—AI must work for India’s 1.4 billion people, across languages, voice interfaces, and real-world applications. A mixture-of-experts is a neural architecture that increases efficiency by activating only a subset of experts for an input rather than the entire model. The ambition is to move beyond elite, English-first AI and build for Bharat. But in a landscape dominated by global heavyweights such as OpenAI, Google, and Anthropic, what will it take for India’s sovereign AI push to move from technological capability to sustained strategic and economic relevance? According to the Stanford University 2025 Global AI Vibrancy Tool, India ranks third globally in AI competitiveness, after the US and China. Gnani.ai unveiled major advances in voice technology, including Vachana TTS for human-like text-to-speech in 12 Indic languages, and Inya TT, a foundational text-to-speech model delivering human-quality speech synthesis with zero-shot voice cloning capability (zero-shot voice cloning can operate with just a short audio sample of the target speaker&#039;s voice—sometimes as brief as a few seconds). The company serves over 200 enterprises, powering tens of millions of daily voice interactions. CEO Ganesh Gopalan says voice AI in Indian languages has shifted from experimental to mission critical. Its models are built to handle accents, dialects, and code-mixed speech, though true AI self-reliance will require stronger compute infrastructure and long-term capital. -Ganesh Gopalan,CEO, Gnani.ai True Sovereignty Real AI sovereignty is not just about building models. It requires control over the entire stack right from data and algorithms to compute infrastructure and hardware. Without making their own graphic processing units (GPUs), countries will remain hostage to export restrictions, undermining technological independence. At the same time, sovereignty is not only about hardware. Only a model trained on domestic, culturally relevant data and governed by local regulations—especially when deployed within controlled environments—can retain sovereignty over its behaviour, privacy, and output. For Raghavan, the challenge is best understood as a layered problem. “The most fundamental layer is energy. Running AI efficiently will increasingly depend on power availability, which is why smaller, more efficient models that consume less energy will matter. But overall, significant energy infrastructure will be needed.” The next layer is semiconductor chips. “We are not in that race today. We will have to import. Over the long term, may be five years, that could change. After that comes the question of how efficiently you run these imported systems at scale. That’s another gap in India today. Then come the models. We are playing in that layer. But ultimately, the models have to be orchestrated into platforms, and platforms into vertical applications. Sovereignty has to be built across all these layers.” India’s Test Case Founded in 2023 by Vivek Raghavan and Pratyush Kumar, also the CEO, Sarvam AI has emerged as one of the most closely watched players in India’s sovereign AI push. The Bengaluru-based start-up has raised $54 million, including from Khosla Ventures and Lightspeed Venture Partners. It was also the first start-up to receive support under the IndiaAI Mission, the government’s Rs 10,370 crore programme aimed at building indigenous AI capabilities. Both the founders have deep experience of building technology at national scale. Kumar, an adjunct faculty member at the Indian Institute of Technology, Madras, had earlier co-founded AI4Bharat, an open-source initiative focused on advancing AI for Indian languages. Raghavan played a key role in scaling Aadhaar at the Unique Identification Authority of India, where he served as Chief Product Manager and Biometric Architect. The founders told BT at the AI summit that their ambition goes beyond building a competitive model. The larger goal, they said, is to make AI accessible at population scale, drawing inspiration from digital public infrastructure playbooks of Aadhaar and UPI. This shapes how the company measures progress. Despite rapid advances, Sarvam has avoided traditional start-up metrics such as revenue or valuation. Instead, the focus has been on technical capability. “We are on a deep-tech journey. For us, the focus isn’t business model innovation but deep technical innovation. That means setting technical rather than financial goals,” says Kumar, recalling a conversation with venture capitalist Vinod Khosla. “He told me he spends most of his time looking at technology and not finance. That’s the deep-tech mindset.” Sarvam’s core metric is the extent to which the knowledge required to build a globally competitive large language model (LLM) can be developed in-house, he says. “Throughout our journey, the goal has been to maximise that fraction,” Kumar says. “We’ve reached an inflection point. We’ve trained a model that is competitive.” At the India AI Impact Summit 2026, BharatGen unveiled a 17-billion parameter sovereign, multilingual foundation model designed to support 22 Indian languages with multimodal capabilities. Developed under the broader PARAM-2 initiative, the model aims to reduce India&#039;s dependence on foreign AI systems and strengthen domestic technological control. BharatGen was the largest beneficiary in the third tranche of the IndiaAI Mission, receiving Rs 989 crore. It has brought together leading academic and research institutions, including IIT Bombay, IIT Madras, IIIT Hyderabad, IIT Kanpur, IIT Hyderabad, IIT Mandi, IIM Indore, IIIT Delhi, and IIT Kharagpur, creating a collaborative ecosystem of researchers, academia, government, and industry to build indigenous AI capability and secure India&#039;s AI future. -Ganesh Ramakrishnan,Principal Investigator, BharatGen The company built its models using 4,096 NVIDIA H100 SXM GPUs, provisioned through Yotta Data Services, under the IndiaAI Mission. Yet, the core team is small. “At Sarvam, 30–40 people are working on the models, and the core group building the LLMs is about 15. That’s what it takes—a small team of committed people willing to put in the time and dedication to make something like this happen.” Equally central to Sarvam’s approach is frugal innovation. “I think frugality is often misunderstood. Sometimes, it’s seen as jugaad, a shortcut. But for us, it means building in a resource-efficient way, without compromising the fundamentals.” Training frontier-scale models was once seen as requiring billions of dollars and so limited to a handful of global technology giants. Sarvam’s experience challenges that, he says. The shift has been enabled by a combination of factors. “Part of it is the hard engineering work we’ve done, and part of it is that the economics of AI are improving, the cost of building these systems is coming down,” he says. The idea that a few companies will build models for the world is changing. “AI has to be built more broadly, it has to become a shared capability, not the monopoly of a few,” he says. Can India catch up? “We are the world’s most ingenious users of technology, focusing on real-world utility over theoretical hype,” says Nipun Kalra, India Leader, BCGX, adding that India’s edge lies in its ‘value-first’ DNA. With a projected $17 billion market by 2027 and a massive talent pool, the real ‘secret sauce’ is the country’s diffusion rails. “Our digital public infrastructure allows us to distribute AI updates to 1.4 billion people overnight,” he says. The Business Case Having a sovereign AI model is a matter of pride. But what about use cases and business proposition? For Sarvam, the near-term opportunity lies less in consumer AI and more in enterprise and institutional deployments around large-scale human interactions. Raghavan says the most immediate revenue pools will emerge from B2B2C (business to business to consumer) and B2G use cases, particularly in sectors that rely on large and distributed workforces. Industries such as insurance, banking, telecom, and FMCG employ millions of field agents and sales representatives, many of whom operate in regional languages. AI tools that assist these agents—through real-time prompts, training, customer interaction support, or even partial automation—could improve productivity and conversion rates. Sarvam recently partnered with SBI Life Insurance to deploy multilingual generative AI for enhanced customer engagement and with Tata Capital for scaling its multilingual voice AI. Raghavan says the revenue opportunity is even larger in Tier II and Tier III markets. “The final interaction in India almost always happens in voice and in a local language,” he says, pointing to AI’s potential to both assist and analyse these exchanges. Government outreach can be another revenue source. From welfare delivery to grievance redressal and citizen feedback, AI-powered voice interfaces can help public agencies understand citizen needs at scale. Sarvam also sees value in conversation intelligence, helping enterprises monitor how effectively their frontline teams communicate, promote products, and engage customers. Raghavan says sovereign AI’s business case will be built not on generic chatbots, but on optimising India’s vast last-mile workforce and citizen interface ecosystem. One of the first AI companies to go public, Fractal Analytics recently launched Vaidya 2.0, the next generation of its healthcare-focused reasoning model. According to Srikanth Velamakanni, Co-founder, Group CEO and Executive Vice Chairman, the platform is designed to function as a comprehensive healthcare operating system. It goes beyond answering medical queries, enabling medically grounded reasoning across complex, multi-step clinical and operational scenarios aligned with real-world workflows. Velamakanni says demand is emerging across consumer, enterprise, and institutional segments, though adoption patterns vary. Consumer use cases have moved fastest due to low friction and immediate value. Enterprise adoption, however, is expected to drive the most durable spending. -Srikanth Velamakanni,Co-founder, Group CEO and Executive Vice Chairman, Fractal Analytics The Scale Gap Compared with global peers such as ChatGPT, whose underlying models run trillions of parameters, Sarvam’s models are significantly smaller. The reasons are structural—limited access to high-end GPUs and constrained capital in a domain where frontier AI development requires billions of dollars. Yet, experts say the achievement must be viewed in context. Jaspreet Bindra, founder of AI &amp; Beyond, says any discussion on sovereignty must acknowledge India’s resource realities. From a sovereign foundation model perspective, India has three key efforts—Sarvam, Gnani, and BharatGen—with Sarvam emerging as the most comprehensive among them, he says. In total, five companies working under the IndiaAI Mission announced their models at the India AI Impact Summit, including Gnani.ai BharatGen, Fractal, and Tech Mahindra. “All credit to these teams,” he says, pointing out that the Indian models have been built under severe capital, compute, and infrastructure constraints. “Developing models of this scale, with a relatively low cost and limited GPUs, is a significant achievement.” Bindra cautions against directly comparing Indian models with global leaders such as GPT-4, Gemini, Claude, or DeepSeek purely on size or completeness. “The point is not how large the model is. The real achievement is that India now has sovereign capability—of whatever scale—built in-house. That is strategically important.” What stands out is Sarvam’s approach of building beyond just a foundation model, he says. Alongside the core models, the company has launched a conversational interface, Indus, as well as edge AI capabilities designed to run on devices. For a country such as India, such smaller, application-focused models that work offline or with limited connectivity could be far more relevant than large, general-purpose systems. The company has also introduced applied capabilities, including visionOCR, text-to-speech, and hardware experiments such as smart glasses Sarvam Kaze, signalling a shift towards real-world deployment. OCR stands for optical character recognition. Bindra says the focus remains firmly on Indian use cases, Indic languages, and sector-specific applications. His only concern is the scale of resources available. India’s sovereign AI efforts are being pursued under tight financial and infrastructure constraints, he says. Even Sarvam’s training cluster, about 4,000 GPUs, is modest by global standards. “Sometimes we take too much pride in frugality,” he says. “These are (companies that have built LLMs in India) small teams doing remarkable work. With the right level of capital and compute, they could operate at a different scale and become far more relevant for India’s enterprise ecosystem.” So, how will Sarvam compete globally? Bindra says one should not focus on whether the homegrown models will be used outside India or not as “maybe with their capabilities, they can make something for some of the other countries of the Global South that require frugal engineering and pricing. We should focus on building a huge number of applications on top that are used not only in India but also elsewhere.” But what about enterprise adoption? According to Hemant Mohapatra, Partner at Lightspeed India Partners, also an investor in Sarvam, enterprise AI adoption in India is at an early stage. “Enterprise spending on AI is extremely small, perhaps tens of millions of dollars, and low hundreds of million at best,” he says. “But the potential market runs into tens of billions. The opportunity is massive, but the current base is tiny.” That gap means the market is wide open, he says. However, large-scale adoption will depend on meeting specific regulatory, security, and infrastructure conditions. Tech Mahindra is among the 12 companies selected under the IndiaAI Mission to make homegrown AI models. The company has come up with Hindi-first education-focused LLM. According to Malhotra, the model is trained exclusively on Hindi tokens and over 37 Hindi dialects, achieving 100% token density. The current iteration shifts the architecture to a Mixture of Experts model, focused deliberately on education for grades VI to XII, beginning with physics, to avoid overextension. -Nikhil Malhotra,Chief Innovation Officer &amp; Head of AI and Emerging Technologies, Tech Mahindra Sarvam could have an advantage in sectors such as banking and financial services as regulatory requirements often mandate local models, locally stored data, and domestic hosting, says Mohapatra. Government use cases come with even stricter norms and require fully localised infrastructure across the stack. In sensitive environments such as defence, systems may need to operate in air-gapped environments, completely isolated from the internet and without reliance on foreign cloud providers or external dependencies. For large enterprises and conglomerates, though, adoption is far from automatic. Vendors must compete through rigorous technical evaluations against global cloud and AI providers. “These are not default wins,” says Mohapatra. “Enterprise decisions come down to technical bake-offs against hyperscalers and other AI players. In some cases, you win on price, in others, on performance or solution design. Sometimes, relationships matter as well.” However, there are risks that could prevent sovereign AI from creating economic or strategic value. Kalra says the biggest risk is the AI-RoI gap—only 25% executives report finding real value in AI. Innovation depth is another challenge, as despite 2,000+ AI start-ups, India contributes less then 1% of global AI patents. Staying competitive, however, will require huge capital. Raghavan says the scale of ambition Sarvam has set for itself will demand sustained funding. India’s sovereign AI push will not be judged by model launches but by what gets built on top of them. The real test lies ahead—whether enterprises adopt at scale, whether capital deepens, and whether applications reach millions of users across languages and sectors. For Raghavan, the journey is measured in decades. If India can turn early capability into a thriving ecosystem of infrastructure, developers, and real-world deployments, it could move from being an AI consumer to a creator. If not, the risk he warns about remains—a future shaped elsewhere, with India merely adapting to it. @PalakAgarwal64

The Sarvam Moment: How the local AI start-up could reshape India’s tech push

How India's first sovereign model can give it an edge in the age of AI