Building Trusted AI at Scale Cities Startups & Digital Sovereignty – Keynote Vivek Raghavan Sarvam AI
20 Feb 2026 13:00h - 14:00h
Building Trusted AI at Scale Cities Startups & Digital Sovereignty – Keynote Vivek Raghavan Sarvam AI
Summary
Speaker 1 introduced Vivek Raghavan, co-founder of Sarvam, a company developing AI that understands India’s languages and context [1-3]. Raghavan opened by asserting that India is capable of training state-of-the-art models and delivering AI to a billion users [4-7]. He argued that while short-term technical leads such as model size or chip speed are fleeting, long-term national sovereignty in AI is essential and must be built domestically [10-13][19-20]. Raghavan warned that without indigenous AI, India risks becoming a digital colony dependent on foreign technology, emphasizing that AI is a core capability a country cannot forgo [21-27]. He highlighted India’s unique advantage of linguistic diversity-22 official languages and regional variations every 50 km-which must be captured for AI to reflect the voice of the people [28-34]. The large, cost-conscious Indian market provides both demand and the need for low-cost, scalable AI solutions, as illustrated by the success of UPI and the potential for AI to improve public services affordably [35-42]. Sarvam’s strategy is a full-stack sovereign AI platform comprising home-grown models, applications, and infrastructure designed for Indian scale [42-46]. Its models are built from scratch without external data dependencies, yet aim to be world-class state-of-the-art systems [46-52]. The SARAS speech model, trained on diverse Indian data, is claimed to outperform global competitors in Indian-language speech-to-text and text-to-speech quality [53-66][69-71]. Sarvam also offers a 3-billion-parameter vision model that surpasses larger international models on document digitisation and visual-grounding tasks, especially in Indian languages [84-90][91-92]. The company has produced several LLMs, including a small 32K-context model trained on 16 trillion tokens and a 105-billion-parameter model, both benchmarked as superior to comparable global open-source models [93-106]. These achievements were realised by a team of only fifteen young engineers, demonstrating the depth of talent available in India [112-119]. Sarvam is already deploying the models in real-time voice applications serving millions of minutes daily, supporting NGOs, content digitisation, and edge devices such as glasses, while building the compute infrastructure needed for nationwide rollout [120-135]. The discussion concluded that building sovereign AI is both a strategic necessity and a feasible path for India to lead in AI that serves its diverse population and economy [13-27][34-38][42-46].
Keypoints
Major discussion points
– AI sovereignty is essential for India’s future.
Raghavan stresses that reliance on foreign models would make India a “digital colony” and that long-term national security depends on building AI in-house, just as India created Aadhaar and the India Stack as open-source public infrastructure [10-13][14-19][24-27].
– India’s unique strengths make sovereign AI feasible and necessary.
He highlights the country’s linguistic diversity (22 official languages, regional variation every ~50 km) and massive, cost-conscious market that can drive demand for AI at scale [28-34][35-40].
– Sarvam is building a full-stack sovereign AI platform: models, applications, and infrastructure.
The platform is organized around three layers-home-grown models, AI-powered applications, and scalable infrastructure-to deliver “world-class, state-of-the-art” solutions entirely from India [42-46][47-52].
– Concrete AI breakthroughs demonstrate world-class capability.
• Speech-to-text (SARAS) and text-to-speech models trained on diverse Indian data outperform global competitors [53-65][69-71].
• Vision models (3 billion-parameter) excel at document digitisation and visual grounding, beating larger international models [84-89][90-91].
• Large language models, from a 32K-context 16-trillion-token model to a 105 billion-parameter LLM, achieve superior benchmarks against GPT-OSS, Gemini Flash, etc., while being trained entirely in India [92-106][108-110].
– Deployment focus: real-world applications and scalable infrastructure.
Sarvam already powers over a million minutes of multilingual voice conversations daily, supports NGOs, content dubbing, and is optimizing models for edge devices and custom hardware to deliver AI at “India scale and India cost” [120-136].
Overall purpose / goal
The discussion aims to convince the audience that India not only can but must develop its own sovereign AI ecosystem. By outlining strategic imperatives, showcasing Sarvam’s technical achievements, and illustrating tangible applications, Raghavan calls for continued investment and collaboration to ensure India’s independence and leadership in AI.
Tone of the discussion
The tone is consistently confident and rallying, beginning with a broad, patriotic appeal to sovereignty, moving into a data-driven exposition of India’s advantages, then shifting to a technical, demonstrative mode when describing models and benchmarks. It concludes on an optimistic, forward-looking note, emphasizing youth talent and the potential for even larger breakthroughs. Throughout, the speaker maintains an enthusiastic, persuasive stance without significant negativity or doubt.
Speakers
– Speaker 1
– Role/Title: Event moderator/host (introducing the keynote) [S1][S3]
– Area of Expertise:
– Vivek Raghavan
– Role/Title: Co-founder of Sarvam (AI company) [S4]
– Area of Expertise: Artificial Intelligence, sovereign AI, speech and language models, large language models, Indian language technology
Additional speakers:
Speaker 1 introduced Vivek Raghavan, co-founder of Sarvam, which is building artificial-intelligence systems that can speak India’s many languages and understand its local context [1-3].
Raghavan opened with a concise rallying cry: “India can train state-of-the-art AI models and deliver them to a billion users,” positioning AI sovereignty as a national mandate [4-7].
He argued that fleeting technical bragging rights-such as model size or chip speed-are transitory, whereas home-grown AI is essential to prevent India from becoming a “digital colony.” He cited his work on Aadhaar and the open-source India Stack as proof that publicly built, indigenous technology can scale to serve a nation [10-13][14-19].
India’s unique advantages make sovereign AI both feasible and necessary. First, the country’s linguistic diversity-22 official languages and dialects that shift roughly every 50 km-requires AI that can capture this variation [28-34]. Second, the massive, cost-conscious market creates demand for affordable, scalable solutions; the success of UPI illustrates how technology can achieve mass adoption while remaining inexpensive [35-40].
Sarvam’s response is a full-stack sovereign AI platform organised around three layers: indigenous models, AI-powered applications, and infrastructure built for Indian scale and cost [42-46]. All models are built from scratch with no reliance on external data, yet aim for world-class performance [47-52].
Speech – The SARAS speech-to-text model, trained on extensive Indian data, is best-in-class for Indian languages in blind tests. Its text-to-speech and dubbing capabilities also rank highest in the country, with dubbing that preserves speaker modality, offers precise duration control, and supports mixed-language output [53-66][69-71][133-135].
Vision – A 3-billion-parameter state-space model excels at document digitisation, language-layout understanding, visual grounding, and reading-order prediction, outperforming larger international models on both Indian-language and English tasks [84-90][91-92][136-138].
Large language models – Sarvam has (i) a compact model with a 32 k-token context, trained on 16 trillion tokens for real-time multilingual conversation, and (ii) a 105-billion-parameter LLM, the largest trained entirely in India. The LLM is on par with most open-source and closed-source models of its class and is superior to GPT-OSS 120 B and Gemini Flash in benchmark comparisons [92-106][108-110][139-141].
The development of these models was enabled by a grant from the India AI Mission [142-144].
All of the above were achieved by a team of just fifteen young engineers, underscoring the depth of talent available in India [112-119].
Sarvam’s Servum platform powers more than one million minutes of real-time voice conversation each day across eleven Indian languages; NGOs have generated a crore minutes of calls in a single month, and the platform also supports content digitisation, translation, dubbing, and enterprise and government use-cases [120-130].
To reach every citizen, Sarvam is optimising models for edge deployment on smartphones and augmented-reality glasses [145-147], and is building large-scale compute infrastructure that can deliver AI at “India scale and India cost” [148-150].
In closing, Raghavan reiterated that AI sovereignty is not a luxury but a national mandate. It safeguards India from dependence on foreign technology, leverages the country’s linguistic richness and market size, and capitalises on its youthful talent pool. He called for continued investment, policy backing, and collaborative effort to expand the sovereign AI ecosystem, positioning India to lead globally while serving its diverse population [151-155].
I move on to our next keynote speaker, who is Mr. Vivek Raghavan, the co -founder of Sarvam, a company building AI that speaks India’s languages and understands India’s context. In a world dominated by models trained on English language data, their work is a powerful demonstration that sovereign AI capability is not just a luxury, it is a necessity. So, ladies and gentlemen, please welcome Mr. Vivek Raghavan, co -founder of Sarvam.
I come here to say that India can. And I think that’s the message I want to say. India can. And India can train state -of -the -art models, bring AI to a billion Indians, and do it all. And that’s really the message of why we started Sarvam. I want to talk about the long arc. You know, when you look at, you know, today the world is moving so fast. Everybody talks about where is the largest model or who has the fastest chip. These are all transitory technical advantages. In the long run, it’s sovereignty that matters. And unless we build these things ourselves, it’s something that, you know, will be left behind in the race. In the past 15 years of my life, I worked on building Aadhaar.
Which is India’s… digital identity program. Prior to that, many of these technologies, many of these technologies were proprietary technologies. And we built this kind of self -created technology that is open source and a public infrastructure that is available to all of us. And that led to the creation of the India Stack. So when you look at it over the course of long periods of time, sovereignty will always trump technical leads that are short term. We have a mandate to build. It is not an option whether we want to build in these technologies. AI is a technology that has impact on every single aspect of human life. And it’s a core technology that a country like India must have.
And it’s a core technology that a country like India must understand. from the foundational level. Otherwise, we will become a digital colony which is dependent on other countries for this core, core technology. That’s something that is, it’s not an option. It is something that we must do. And we have unique advantages. And our unique advantage is actually our diversity. We have so many languages. We have 22 official languages. And in fact, you know, the way people speak in our country changes every 50 kilometers. And that diversity must be captured if we have to understand the voice of the people. And therefore, if we build AI from India, it must acknowledge that diversity and do this. The other thing, of course, is we are a huge economy.
There is demand. and if AI is there to help the citizen to do everything, this is we can be one of the largest consumers of AI in the world. And that demand is there and then we have to build. We know that we are a cost -conscious country, right? Everything needs to be at the lowest cost. So we need to build efficient AI that actually can be delivered at scale for the people so that the last person in the country can actually have a better experience, right? Today, if you look at UPI, one of the great success stories of the past decade, and if you look, it is for the first time we feel that in India, things can be better than everywhere else in the world.
But AI done the right way can make sure that every service to citizens actually is the best and the cheapest and actually done in the best possible way for the country and that’s the promise of AI and that’s why we said we need to do this in India. and I think it’s not about I mean we have companies which have globally when you look at AI companies they are massive companies but in the end we have to show a new model where we can actually build AI which helps everyone and then we can win for the people and our model can be adopted in the world and that is my belief of where we need to go on this thing So Sarvam has been building India’s full stack sovereign AI platform and we work with developers fundamentally India is a country of developers we have more developers and we work with enterprises and we work with governments and we have a full stack platform which I’ll talk about In fact the full stack platform consists of three different things one is models we need models that are built in India And that is the key thing and that’s what we’ll focus on, sovereignty and models.
And then we are focusing on applications. Applications is AI for everyday tasks, for making things better for people. And finally, we’ll talk about infrastructure and infrastructure at India scale. The first thing I want to talk about are sovereign models. Rule number one is they are built from scratch. They are not dependent on any other model that is there in the world. They are built from scratch. There is no data dependency on anyone else. But at the same time, the focus is these are world -class, state -of -the -art models that we have. So I’m going to talk a little bit about some of our models. The first model is actually the SARAS model, which is actually the speech model which helps recognize speech in Indian languages.
This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. This is native. In fact, if you see, this is actually best in class in Indian languages compared to any other global model in terms of that’s something to say that in the, these are extremely small models, but these are models which have been trained with significant amounts of Indian diverse data, which will actually lead to better performance on Indie voices.
So let me just take a play, an iconic moment in India, diarized using our model. Oh, sorry. Okay, I don’t think I can make this happen. Okay. Is audio, it’s not playing. let’s maybe we’ll come back to this and here we have a majestic liftoff of lbm3 m4 rocket carrying india’s prestigious chandrayaan 3 spacecraft this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket this is the rocket of lbm3 m4 rocket We want to create models which are naturally expressive Indian voices, and they have low latency streaming and actually production grade quality.
So, in fact, our models, our speech to text models are considered in blind tests are actually the most preferred voices in Indian languages. And this is something compared to all the global competitors, such as Levin Labs and Cartesia, etc. And we have actually the most preferred text to speech model in the country. We’ll also talk, we also have a dubbing capability, which actually preserves the speaker modality and has precise control over duration and supports mixed language things. So I will show, in fact, in this model as well. We see compared to any other model in the world, we are the most preferred as far as dubbing is concerned. I will show a small snippet of what happens here.
We have the remaining 13 bits in which the 12th bit is called the small a bit. Then the remaining 6 bits are compute bits. These 3 are the destination bits and these are the jump bits. Then we have the remaining 13 bits in which the 12th bit is called the small a bit. Then the remaining 6 bits are compute bits. These 3 are the destination bits and these are the jump bits. Then we have the remaining 13 bits in which the 12th bit is called the small a bit. Then the remaining 6 bits are compute bits. These 3 are the destination bits and these are the jump bits. So we’ve also built these vision models. And these vision models are actually very good for document digitization.
They’re very good at language layout understanding, visual grounding, and in fact, finding intelligence by visual components. And finally, reading order predictions. In fact, the vision model that we built is only a 3 billion parameter state space model which beats all other models in the world, not just in Indian languages, but in English as well. So therefore, it shows, and many of these models are many orders of magnitude bigger than our models, and still we are able to get world -class performance from them. Of course, in Indian languages, we are far and away ahead of the models that are there from the global comparison. Now we come talk about some of the LLMs. We have actually, India has started the training of LLMs from scratch.
And this was done through a grant from the India AI Mission that actually without which it would not have been possible for us to train these kinds of models through a GPU grant. In fact, it is a context, it’s a model which is an extremely small model which can run on a single GPU. And it has a 32K context length and is trained on 16 trillion tokens. And it’s extremely efficient thinking which actually gives better answers with lower. And the focus of this model is actually real time conversational applications that people will be able to, it will be able to generate conversations in all the Indian languages in a real time system. And some of the benchmarks show compared to models, global models of similar size.
such as QN30B or GPT -OSS. It is far superior in the ability in terms of various parameters, such as fluency, language and script, and usefulness and conciseness. So the important thing is that this is able to, at the size, it is again a global best. And then we’ll come finally to our largest model, which is the 105 billion parameter model. This is the largest LLM that has trained from scratch in the country. And it is basically better on par with most open source and closed source models of its class. And it can handle various kinds of complex reasoning tasks, as well as web searching and things like that. So it’s actually a fairly complex, fairly advanced model, which again works in all…
all Indian languages. From a benchmark perspective, these models, again, compared to things like GBT OSS 120 billion, as well as Gemini Flash, is actually superior in terms of the kinds of outputs it can generate. So therefore, this is really an example. While this is quite a small model, it is really, just to give you an idea, last year we had DeepSeek R1, which actually was launched, and that was 670 billion parameters. The numbers that we are getting are actually superior to what DeepSeek R1 had last year. Of course, the state of the art has also improved. But the goal is we can show that India can build these things. And I want this. The most important thing that I want people to understand is…
just because, and I think that the, you know, I would love that not just us, but many other people come and show that we can actually build world -class models from India, because that is the fact. And these models was built with a team of just 15 young people. And really it’s the game of the youth of India that have actually made this model what it is. I’m just the spokesperson. These young kids have actually made something like this happen. And if these kids can do it, we have so much talent in the country. And I think that’s, I’m very positive about given the right kind of support in the way that we have been given, that much bigger things can happen.
Moving beyond our models, we actually want these models to become useful. And so therefore we build applications. And I’ll talk very briefly. about the kinds of applications that we build. So in fact, we have an ability to actually converse, and this is our real -time voice conversation that happens. We do more than a million minutes of voice conversation in 11 Indian languages every day using Servum. So these models are actually being used to build things. So these models which have been trained fully in our control are actually now being used for conversations across enterprises and government use cases. In fact, in the last month, we actually took about 20 NGOs, actually did a crore minutes of calling within a month to really understand what is the real -time voice conversation and what people actually are saying at the ground.
Okay. So we actually have ways by which we can make this available for work tasks and enterprise tasks. And we also have the ability to do this for content, which is digitizing books, translating books, dubbing videos. These are all studio products that we have. And I think finally, I want to end with infrastructure. We actually are doing many interesting things. We are making our models smaller to work on the edge, to work on phones. We are making, we’re actually, many of you may have heard that we have also launched some of these glasses so that we are able to have these models run on different form factors and capture the intelligence, capture the voice of India at every point in there.
And finally, for these things to work, you need to have compute at a large scale and the ability to actually very efficiently deliver all these models to India at India scale at India cost. Finally, we are, of course,
“Home‑grown AI is essential to prevent India from becoming a “digital colony.””
The knowledge base includes a statement that India must develop its own core technology or risk becoming a digital colony, echoing Raghavan’s point [S14].
“India has 22 official languages.”
A source explicitly notes that India has 22 official languages [S56].
“UPI demonstrates that technology can scale rapidly and affordably for a massive Indian market.”
The knowledge base cites UPI processing over 20 billion transactions monthly and showing scalable, inclusive technology [S57].
“India’s massive, cost‑conscious market makes sovereign AI both feasible and necessary.”
Another source highlights India’s status as the world’s strongest growth market where AI’s deflationary nature aligns with development needs, providing economic context for the claim [S53].
“Building sovereign AI is a national mandate to preserve cultural and technological independence.”
A discussion of India’s need for its own foundation models frames the issue as cultural preservation and national capability, adding nuance to the sovereignty argument [S8].
“Sarvam’s platform relies entirely on indigenous models without external data.”
The knowledge base mentions the broader Indian push for heterogeneous compute and sovereign AI capabilities, underscoring the strategic emphasis on indigenous model development [S19].
“India’s linguistic diversity creates a need for AI that can handle many languages and dialects.”
A source notes India’s multilingual environment (22 official languages) and the importance of multilingual collaboration for AI, providing additional background [S22].
The discussion shows clear alignment between the introductory speaker and the keynote on the need for indigenous, sovereign AI for India, framing it as a strategic imperative to avoid digital dependence. Beyond this core point, there is little overlap on other themes such as linguistic diversity, market size, or specific technical achievements.
Moderate consensus limited to the sovereignty argument; this shared stance reinforces policy momentum for building a national AI ecosystem but indicates divergent focus on implementation details and broader AI applications.
The transcript shows strong alignment between the introductory remarks and Vivek Raghavan’s keynote. Both emphasize the strategic need for sovereign, indigenous AI to avoid dependence on external technology and to serve India’s linguistic diversity and large market. No substantive contradictions or opposing viewpoints are evident.
Minimal – the speakers are largely in consensus, indicating a unified stance on the importance of building indigenous AI capacity. This coherence suggests that policy discussions around AI sovereignty in this context are likely to progress without major internal contention.
The discussion was driven by a series of strategically placed insights from Vivek Raghavan that moved the audience from a broad, ideological stance on AI sovereignty to concrete demonstrations of technical capability, social impact, and future deployment. Each pivotal comment reframed the conversation—first by positioning sovereignty as a national imperative, then by leveraging India’s linguistic diversity, emphasizing cost‑effective design, insisting on building models from scratch, showcasing a world‑class large model built by a tiny team, evidencing real‑world usage, and finally envisioning edge‑centric, ubiquitous AI. These moments collectively shifted the tone from abstract advocacy to demonstrable achievement and forward‑looking ambition, deepening the audience’s understanding of how India can achieve independent, inclusive AI at scale.
Disclaimer: This is not an official session record. DiploAI generates these resources from audiovisual recordings, and they are presented as-is, including potential errors. Due to logistical challenges, such as discrepancies in audio/video or transcripts, names may be misspelled. We strive for accuracy to the best of our ability.
Related event
