India’s Sovereign AI Bet: What Sarvam AI Is Really Building

Source: Lunchbreak with Lightspeed / Lightspeed India and Southeast Asia — Pratyush Kumar and Vivek Raghavan of Sarvam AI, hosted by Hemant Mohapatra — Video ID: NHufWLprpv0

India is one of the world’s largest consumers of AI, but consumption is not capability. Sarvam AI’s wager is that India cannot become a serious AI power by exporting its data, importing intelligence, and hoping the most important platform shift in a generation accrues value somewhere else.

Who Are Pratyush Kumar and Vivek Raghavan?

Vivek Raghavan brings an unusual builder’s résumé to Sarvam AI: IIT, a PhD in the United States, roughly 20 years building software for semiconductor chip design, and then 15 years as a full-time volunteer helping build India’s digital public infrastructure. He worked on Aadhaar, was involved with GST, advised UPI, advised the courts, and spent the last several years focused on Indian-language AI.

Pratyush Kumar came from IIT Madras and Microsoft Research, with a research background in AI and systems. At Sarvam, he has led the company’s technical arc from Indian-language fine-tuning to pre-training models from scratch, post-training, reinforcement learning, GPU-scale engineering, and full-stack AI products. Together, they represent a rare founder pairing: public infrastructure depth plus frontier AI research execution.

The Founding Instinct: “We Should Build This Ourselves”

Sarvam AI did not begin with a neat market map or a fashionable sovereign AI slogan. Pratyush describes the company’s initial clarity more simply: “We should build this ourselves.” That phrase matters because it captures the psychological break required for a country, a company, or a founder to move from usage to ownership.

Vivek had not planned to do another startup. After returning to India in 2007, he expected to spend the rest of his career around open source, government systems, and public infrastructure. But generative AI changed the size of the question. He saw a technology so foundational that philanthropic projects and government advisory work were not enough. “This is so fundamental,” he argued, that India needed to do more than consume it.

Pratyush, then splitting time between IIT Madras and Microsoft Research, saw the same step change. Researchers were experiencing what he calls an “identity crisis” in the post-AI world. The old boundaries between research, product, infrastructure, and national capability were collapsing. The only honest response was to go “all in,” even before the phrase “sovereign AI” had become common.

The company’s first operating choice reflected that urgency. Sarvam began not as a two-person startup, but with roughly 15 people in an office at IIT Madras Research Park in Chennai. The team ran something like a boot camp: get everyone in one place, align on the ambition, and start building. Pratyush recalls an early meeting in his living room with about 30 people discussing models, applications, GPUs, inference, and what it would take to build across the stack.

That starting shape is important. Sarvam was not asking, “Which app can we ship quickly?” It was asking, “What capabilities must exist locally if India is to participate in the value creation layer of AI?” The answer was not one product. It was models, infrastructure, applications, and the talent density to keep moving as the field changed.

Why “Sarvam” Means More Than a Brand

The name Sarvam comes from Sanskrit and means “all.” For Pratyush and Vivek, the word carries two meanings. First, AI should reach everyone. Second, more people should be builders of the most important technology layer of the post-AI world.

Pratyush warns that AI could create a “K-shaped” society where a small number of people and organizations become dramatically more capable while many others get left behind. His preferred frame is not universal basic income, but universal basic intelligence: broad access to intelligence that empowers people directly.

The second meaning is more strategic. At the time Sarvam was forming, only a handful of companies — especially OpenAI and Google — appeared to be building the foundational layer of AI. Pratyush’s objection was blunt: the most important value-creating component of the post-AI world cannot be built by only two companies.

Sarvam’s thesis is that the model layer is the linchpin. It looks upward to the application layer and downward to infrastructure. Companies that truly understand and control the model layer can move in both directions. Companies that do not risk becoming dependent on someone else’s roadmap, pricing, availability, and policy constraints.

That has particular significance for India. Pratyush notes that India has built services, apps, and enormous usage, but has not historically built many value-accruing platforms. Sarvam is betting that the model layer is such a platform. That is why the company refuses to stay only at the model layer. It wants the full stack: bare-metal GPU management, models, inference, voice AI, agents, content tools, and eventually broader sovereign infrastructure.

Full Stack Is Not Sprawl If the Stack Compounds

From the outside, Sarvam can look like several companies at once. It trains small models for speech, vision, documents, embeddings, translation, and Indian-language tasks. It trains large language models for reasoning, math, coding, and agentic workflows. It operates GPU infrastructure. It builds voice AI for low-latency Indian-language conversations. It works on agentic products with connectors and skills. It develops content creation tools for dubbing, book translation, audiobooks, podcasts, and vision.

Hemant Mohapatra pushes on the obvious concern: is this too much? Why not focus like ElevenLabs in audio, or a specialized model company in one domain?

The answer is that Sarvam sees compounding across layers. Work on frontier reinforcement learning can improve vision models. Infrastructure lessons from large training runs can improve inference economics. Voice AI at scale teaches the company what production-grade systems require across models, applications, and infrastructure. The point is not random breadth. The point is to own enough of the stack that learning transfers.

Pratyush also argues that the definition of focus is changing. In AI, entire product categories appear and disappear quickly. The cost of experimentation has fallen. Teams can compose capabilities “at the speed of thought.” But he draws a sharp line between hacking and winning. “Victory in the market is very different from getting something that is working.”

That distinction matters for founders intoxicated by demos. Sarvam is comfortable experimenting at the top layer only because it is also investing in the hard systems beneath. Voice AI, for example, took real time: models, application logic, infrastructure, latency, cost, accuracy, and scale. You can prototype quickly, but production still demands what Pratyush calls the “packing after the hacking.”

This explains Sarvam’s 14-day launch run before the India AI Impact Summit. The team decided it could not wait for the event itself, where noise would be high. There were 14 days left, so they launched something every day. Pratyush admits the idea came together quickly, but the substance came from years of accumulated work. The launches converted internal capability into external proof, while compounding brand, talent, and capital.

The Sovereign AI Argument: Data Out, Intelligence In

Vivek’s strongest argument for sovereign AI is not sentimental. It is structural. AI, he says, is comparable to nuclear technology in one respect: the world will have AI powers and AI have-nots. The United States and China are already in the first category. No other country is fully there yet.

For India, the decision is whether to make the bet required to control its own AI destiny. Vivek frames the downside sharply: if India falls too far behind, it risks becoming a “digital colony.” His most memorable line captures the dependency problem: “Sometimes with pride, we say we have the largest number of ChatGPT users, but what are we doing? Really, we are exporting our data and importing intelligence.”

That is the core of Sarvam’s national argument. Usage is not enough. In fact, usage without ownership can deepen dependence. If the country’s people, companies, and institutions route their knowledge, workflows, queries, and decisions through foreign model providers, India may grow AI consumption without building AI leverage.

Pratyush adds that AI is not a summit to reach; it is a capability to have. Like weapon systems, countries do not “finish” capability development. They keep playing the game. OpenAI started in 2015 and put in years of hard work to prove scaling. Anthropic advanced agentic capabilities. Chinese companies showed frontier training and open-source momentum. India, given its size, civilization, and economy, must have substantial players at the table.

Importantly, Sarvam does not argue that every Indian company should train models. IT services firms can use existing models and make money. Application companies can build on APIs. But some bets must exist at the foundation layer. Pratyush says there should ideally be “5, 10 Sarvams” if capital and talent allow.

From 2 Billion to 105 Billion to 1 Trillion Parameters

Sarvam’s technical journey has been deliberate. The company started by working with Llama, fine-tuning models to add Indian-language skills. That work taught the team a key lesson: “You can’t bolt on language to an existing model.” Language capability must be deeply integrated, not superficially patched.

The work led to a collaboration with Meta. Pratyush joined Meta’s advisory board and says Sarvam contributed to Llama 3.1 as the only Indian company involved in that effort. From there, Sarvam moved to pre-training. When the first thousand Hopper-series GPUs were connected in India, Sarvam put its own money to work and pre-trained a 2 billion parameter model. That taught the company about pre-training data, Indian-language data creation, and running GPU workloads.

Infosys commercially bought a fine-tuned version of that model for IT operations, giving Sarvam an early market signal. The next phase was post-training. Sarvam took Mistral’s roughly 30 billion parameter small model and learned supervised fine-tuning, reinforcement learning, Indian-language skill addition, reasoning, hybrid thinking and non-thinking behavior, and scaled RL. The team published models and detailed technical blogs in the open, earning attention from both Meta and Mistral.

The next leap was training from scratch. Over six to eight months, Sarvam trained 30 billion parameter and 105 billion parameter models. It also built an in-house 3 billion parameter model strong in audio and vision. The hardest part was not merely compute; it was building the data pipeline. Pratyush says frontier-scale training often needs 15–20 trillion tokens, and Sarvam created that dataset internally rather than relying on a data vendor.

The 105 billion parameter milestone is central to Sarvam’s credibility. Pratyush says that, to the best of his knowledge, outside North America and China — leaving aside Mistral — nobody had pre-trained from scratch a 100 billion-plus parameter model. Sarvam’s 105 billion parameter model performed on math, reasoning, and STEM at the level of a 600 billion parameter DeepSeek R1 model, illustrating the compression of intelligence into smaller models.

Now the target is a 1 trillion parameter model. Vivek calls larger models “table stakes” because they are not only useful directly; they distill into better smaller models. Pratyush’s goal is a “no-regret alternative” across small models, multimodal models, and large reasoning models within six to nine months, while reducing the current frontier gap from about nine months toward six.

The capital required is serious but not OpenAI-scale, at least for Sarvam’s current ambition. Pratyush describes the requirement as “hundreds of millions of dollars” to build a no-regret alternative stack. For national infrastructure, he says India needs “tens of thousands of Blackwells” to make AI available at scale. The goal is to drive token consumption up by four or five orders of magnitude so every student can have a tutor and every person can have access to a doctor-like assistant.

Adoption: The Real Chasm Is Human Alignment

Hemant invokes Geoffrey Moore’s Crossing the Chasm to ask how Sarvam moves from early adopters to the early majority. The question is not only technical. AI must be packaged into forms society can adopt: tools, workflows, pricing, trust, and use cases that produce obvious productivity gains.

Pratyush’s answer is that adoption will differ across individuals and organizations. Individuals with agency will move quickly because the proof is adjacent. They try something, see it work, and adopt. A small or medium business using voice AI can feel value immediately. In his view, being on the leading edge will compress from years to months to weeks.

Organizations are harder. The “tax,” as Pratyush puts it, is human alignment. Leadership must believe the change matters. Employees must believe it is not simply a job-loss machine. Society must believe innovation is worth pursuing. These are not technical objections, but they shape deployment more than benchmarks do.

Vivek adds that Sarvam’s success should ultimately be measured by token share. If a significant percentage of all tokens used in India over the next two years come from Sarvam, the company will have achieved something meaningful. That would imply not only model usage, but AI adoption at a scale that affects daily life.

The market may misunderstand Sarvam because it is playing multiple games at once. Some see a voice AI company. Some see a foundation model lab. Some see a sovereign infrastructure company. Pratyush accepts the “elephant metaphor”: each observer touches one part and names the whole differently. Sarvam is comfortable with that because the ambition is integrated. It wants voice AI as infrastructure, self-hosted model experiences with sovereignty and lower cost, creator tools rooted in language, frontier models, and data center partnerships.

The company calls its next phase Sarvam 2.0. The technical capabilities are not limited to India, but the current market position is India-specific. Global expansion will depend on business focus and the right partners. The capability may be general; the go-to-market must be chosen.

Key Lessons

Why This Matters for Diffie

Diffie is not building sovereign national infrastructure, but the Sarvam story maps cleanly to the problem of building an AI browser testing tool for frontend engineers. The lesson is not “train a trillion-parameter model.” The lesson is to own the capability layer that customers cannot afford to treat as a toy.

For frontend teams, browser testing is becoming one of those capability layers. Modern web apps change constantly: UI states, responsive layouts, auth flows, checkout paths, permissions, dashboards, visual regressions, and edge-case interactions. Engineers do not want another brittle test recorder. They want confidence that the product works before users find the breakage.

Sarvam’s “no-regret alternative” is a useful positioning frame for Diffie. The buyer should feel that choosing Diffie does not require sacrificing quality, control, or workflow fit versus manual QA, Playwright-heavy internal systems, or generic AI agents. For a technical founder selling to engineers, that means the product must prove three things quickly: it understands the app like a browser-native tester, it produces actionable failures instead of noisy screenshots, and it fits into CI, PR review, and developer debugging loops without creating process drag.

The “packing after the hacking” lesson is especially important. AI browser testing can demo beautifully: click through a page, describe a bug, maybe generate a test. But market victory will come from the unglamorous layers: deterministic reproduction, stable selectors, auth/session handling, flaky-test control, visual diff precision, network logs, console traces, CI reliability, and clean handoff to engineers. Diffie’s trust layer is its infrastructure layer.

Sarvam’s adoption insight also applies directly to GTM. Individuals with agency adopt first. For Diffie, that likely means frontend leads, founding engineers, QA-minded product engineers, and CTOs at fast-moving teams who already feel regression pain. The early ICP should not be “all engineering teams.” It should be teams where one person has enough agency to plug Diffie into a workflow and immediately see value: fewer broken flows, faster PR confidence, less manual release checking, fewer embarrassing production bugs.

The token-share ambition has a Diffie equivalent: test-coverage share of critical user flows. If Diffie becomes responsible for a meaningful percentage of the browser checks a team trusts before shipping, it becomes infrastructure rather than a tool. That is the strategic move: start with a wedge that feels obvious, then compound into the testing layer that watches the frontend continuously.

Finally, Sarvam’s full-stack argument suggests a product strategy. Diffie should experiment at the top layer — AI-generated checks, natural-language bug reports, autonomous exploration, visual summaries — but only where those features compound on a strong execution substrate. The winning product will not be the flashiest AI agent. It will be the one that makes frontend engineers say: “I can ship faster because this catches what I would have missed.”

For Anand’s current challenge around ICP, GTM motion, and outbound, the practical takeaway is clear: sell agency and confidence, not generic AI automation. Lead with a painful, concrete workflow — “catch broken critical browser flows before merge” — and target the people who can act without a six-month evaluation cycle. Like Sarvam, Diffie should earn the right to broaden by becoming indispensable in one high-value layer first.