Case Study on Sarvam AI: Story Behind India’s Leading Sovereign AI Company

Introduction

The challenge that India faces with language is unlike anything in the world. With 22 official languages, hundreds of dialects, and a linguistic reality where the way people speak changes every 50 kilometers.

Global AI models were not built with this reality. ChatGPT, Gemini, and their counterparts were trained predominantly on English-language data, with other languages filling the gaps. India was at genuine risk of becoming a consumer of AI rather than a builder of it.

That is the gap Sarvam AI was founded to close.

When the Union Cabinet approved the IndiaAI Mission in March 2024 with an investment of ₹10,372 crore over five years, it created the foundation that companies like Sarvam needed.

This case study on Sarvam AI discusses how a startup less than three years old became a leading contender for India’s AI future.

What is Sarvam AI?

Sarvam AI is India’s full-stack sovereign AI platform. It is a Bengaluru-based startup building foundational AI models, enterprise applications, and developer APIs entirely within India. It is the infrastructure layer that allows Indian businesses, developers, and government bodies to deploy AI that actually understands local language.

For Sarvam, “AI for all from India” is not just a tagline. It is an engineering mandate.

Who Founded Sarvam AI?

Dr. Vivek Raghavan and Dr. Pratyush Kumar founded Sarvam AI in August 2023. They were working at the intersection of technology for Indian languages and large public infrastructure.

Dr. Vivek was Chief Product Manager and Biometric Architect for Aadhaar at the UIDAI, where he worked for more than a decade to enroll more than a billion people. He understands how to build technology that scales to India, not just in the lab.

Dr. Pratyush was the lead researcher at AI4Bharat at IIT Madras, one of the biggest open-source initiatives to develop language models for Indian languages. His work directly countered the oversight of global AI companies: that Indian languages are not a side project; they are the first language of hundreds of millions of people.

Where Does Sarvam AI Stand in the Indian AI Industry?

India’s AI ecosystem is bigger, faster, and more diverse than most people realize. From healthcare diagnostics and fintech automation to enterprise SaaS and sovereign language models, Indian AI companies are starting to set the agenda.

But most of them are built on building at one layer. They are integrating existing global models into workflows, rather than building the foundational infrastructure on which those models run.

That distinction is exactly where Sarvam AI stands apart.

Sarvam’s Competitive Position

In April 2025, the Indian government selected Sarvam AI to build India’s first homegrown sovereign LLM under the IndiaAI Mission.

What separates Sarvam from the rest of the field is not just what it builds, but how it builds. Sarvam is the only Indian company building a complete vertical stack, from raw foundation models and speech systems to enterprise APIs and consumer applications.

Just as UPI created the rails on which hundreds of fintech apps were built, Sarvam is positioning itself as the AI infrastructure layer on which India’s next generation of AI products will be built.

Sarvam AI Platform Architecture

Most AI companies build one thing well. Sarvam AI built three layers and then connected them into a single platform. Its architecture spans foundation models, applied AI products, and developer APIs, with every layer designed specifically for Indian users. Here is what that looks like in practice.

Layer 1: Foundation Models

The base of the Sarvam stack is two large language models, both built from scratch.

• Sarvam-30B: a 30-billion parameter model based on a Mixture-of-Experts (MoE) design, activates approximately 1 billion parameters per token with a 32,000-token context window.
• Sarvam-105B: a 105-billion parameter model activating approximately 9 billion parameters per token, with a 128,000-token context window, positioned for complex reasoning and enterprise applications.

Both models are open-sourced under the Apache 2.0 license and hosted on Hugging Face, making them freely accessible to developers and researchers. They were trained using computing resources provided under the IndiaAI Mission, with infrastructure support from data center operator Yotta and technical support from Nvidia.

Layer 2: Speech, Vision & Multimodal Systems

Above the foundation model sits a suite of specialized AI systems for speech, document intelligence, and content, each optimized for the realities of Indian language and script.

• Saaras V3: an automatic speech recognition model supporting streaming and batch recognition across 22 official Indian languages plus English, trained on 1 million+ hours of curated multilingual audio data.
• Bulbul TTS: a text-to-speech model covering 11 Indian languages with 39 distinct speaker voices, built for natural-sounding voice output across regional accents and tones.
• Sarvam Vision: a vision model that can extract text, tables, charts, and structured data from documents and images, supporting OCR across Indian language scripts.
• Sarvam Dub: an AI dubbing model that takes existing audio or video content and dubs it into other Indian languages using zero-shot voice cloning to retain the speaker’s original voice characteristics across languages.

Layer 3: Enterprise Products

This is where Sarvam’s platform translates into direct business value. Three products sit at the enterprise layer, each addressing a distinct operational need.

• Sarvam Samvaad: a conversational AI platform to deploy omnichannel agents across voice, WhatsApp, and web, connecting to enterprise tools, using business data, taking actions, and providing analytics.
• Sarvam Studio: a native platform for translating and dubbing content across Indian languages, combining video dubbing and document translation capabilities into a studio-level workspace where users can upload source content.
• Sarvam Akshar: a document intelligence workbench built on the Sarvam Vision model, capable of extracting data from complex, unstructured Indian documents.

Sarvam AI Use Cases

Sarvam’s platform is not a proof-of-concept stack; it is already deployed across enterprise and government operations at scale:

• Cart recovery and sales follow-ups: Samvaad agents proactively reach customers via voice and WhatsApp in their preferred language to recover abandoned transactions.
• Appointment booking: AI voice agents handling scheduling across healthcare, banking, and government services without human intervention.
• Payment follow-ups and loan servicing: Tata Capital uses Samvaad across its consumer loan products for multilingual customer engagement, including detecting sentiment and escalating to human agents when needed.
• Government document digitization: Akshar is digitizing land records, historical manuscripts, and official documents for state governments.
• Sovereign AI in governance: Sarvam has announced partnerships with the governments of Odisha and Tamil Nadu to build large-scale AI-optimized compute facilities.
• National Language services: SEBI deployed Sarvam tools for an AI-driven public outreach campaign; the National Commission for Women uses Sarvam Studio for translating training materials

Deployment Options & Security

For enterprise buyers evaluating Sarvam for production use, the platform offers three deployment models:

• Sarvam Cloud: fully managed, auto-scaling cloud infrastructure; fastest path to deployment
• Private Cloud / VPC: dedicated infrastructure within a secure private environment for data-sensitive organizations
• On-Premises (Air-Gapped): physical isolation of compute for regulated industries such as defense, banking, and government, ensuring that sensitive data never leaves the client’s environment

On the compliance front, Sarvam holds ISO certification and SOC 2 Type II accreditation. It is the enterprise-grade security standards that B2B buyers in BFSI, healthcare, and government typically require before signing deployment contracts.

How Sarvam AI Stacks Up Against ChatGPT and Gemini

This is not a winner-takes-all comparison. ChatGPT excels at general-purpose reasoning, coding, and content generation. Gemini leads on multimodal tasks and Google ecosystem integration. Sarvam AI is built for something neither of them prioritized: making AI actually work for Indian languages, Indian documents, and Indian users at scale.

Where Sarvam Wins

Document OCR and Intelligence: Sarvam Vision scored 84.3% accuracy on olmOCR-Bench, outperforming Gemini 3 Pro at 80.2% and ChatGPT at 69.8%.

Speech Recognition: Saaras V3 achieves approximately 19.31% Word Error Rate on the IndicVoices benchmark, lower than figures reported for Gemini 3 Pro and comparable systems.

Token Efficiency: Global models require 4–8x more tokens per word for Indian scripts than for English, quietly inflating API costs for Indic workloads.

Offline Capability: Sarvam-Edge processes speech eight times faster than real-time on-device, without internet connectivity.

Where Global Models Still Lead

ChatGPT and Gemini retain real advantages in general reasoning benchmarks (MMLU, GPQA), long-context document processing, global language coverage, coding workflows, and out-of-the-box enterprise integrations. For workloads that are broad, global, or heavily technical, they remain strong choices.

Sarvam AI’s Defining Moment

Every company has a moment that changes how the world sees it. For Sarvam AI, that moment came in February 2026 through a calculated two-week marketing campaign that repositioned it from a promising Indian startup to a national AI infrastructure player.

Sarvam ran what it called a “14 days, 14 launches” campaign around the India AI Impact Summit at Bharat Mandapam, New Delhi. It echoed OpenAI’s rapid-release strategy, executed at a fraction of the budget and with considerably more to prove.

What Was Launched

The 14-day campaign covered every layer of Sarvam’s platform:

• Sarvam-30B and Sarvam-105B
• Indus
• Sarvam Vision
• Sarvam Akshar
• Saaras V3
• Bulbul V3
• Sarvam Dub
• Sarvam Samvaad
• Sarvam Studio
• Sarvam Edge
• Sarvam Arya
• Sarvam Kaze
• Qualcomm, HMD, and Bosch partnerships
• SEBI deployment

Sarvam AI Business Model

Vivek Raghavan has described Sarvam as “largely B2B,” working with enterprises and governments. That positioning reflects a deliberate revenue architecture built around four streams, each serving a different buyer while reinforcing the others.

How Sarvam AI Makes Money

1. API Consumption: The self-serve API platform is Sarvam’s lowest-friction entry point. This consumption-based model has driven 150% year-on-year API adoption growth into 2026. Open-source model releases on Hugging Face feed this funnel directly.

2. Enterprise Platform Contracts: Above the API layer sit large-scale deployments of Samvaad, Studio, and Akshar for BFSI, healthcare, and government clients. Tata Capital’s Samvaad deployment across consumer loan products is a live example.

3. Government Contracts: Chanakya, announced March 2026, is Sarvam’s vertical for environments where failure “is not an option”, such as defense, regulated finance, and government departments.

4. Cloud Marketplace Distribution: Listings on Microsoft Azure and Google Cloud, plus system integrator partnerships with TCS and Infosys, extend Sarvam’s reach into global enterprise procurement channels.

Sarvam AI Pricing

All plans include ₹1,000 free credits to start. Paid plans scale from ₹0 to ₹50,000:

• Speech-to-Text: ₹30 per hour
• Text-to-Speech: ₹15–30 per 10,000 characters
• Translation: ₹20 per 10,000 characters
• Enterprise tiers: Samvaad, Akshar, and on-premise deployments are custom-priced via direct sales
• Startup Program: up to 12 months of API credits, priority engineering support, and production infrastructure access for early-stage builders

Revenue & Net Worth

Sarvam’s annual revenue stood at ₹29.1 crore as of March 2025, reflecting a company that spent its first 18 months building infrastructure rather than selling it. The $350 million fundraise in April 2026, backed by Nvidia, Bessemer, Amazon, and Accel, implies roughly 14x valuation appreciation in under 30 months. With this, Sarvam is widely reported to be targeting unicorn status as the round closes.

Challenges Sarvam AI Still Faces

Sarvam has made strong progress in less than three years, but it still faces significant challenges ahead and recognizing them makes the story more credible.

1. Capital Asymmetry vs. Global AI Labs

The global LLM market is dominated by tech giants like Google, Microsoft, and OpenAI, which invest hundreds of billions annually. Raghavan himself acknowledged that matching Gemini or Claude at scale requires capital of a categorically different magnitude.

2. Compute Dependency and Infrastructure Gaps

India’s push for sovereign AI is complicated by persistent infrastructure gaps and compute limitations. A global shortage of GPUs and long wait times for advanced chips are significant hurdles, forcing reliance on international cloud providers and limiting startup scalability.

3. Talent Retention

Building population-scale AI in India requires sustained talent retention within the country and that is easier said than done. US AI labs routinely offer compensation packages that Indian startups structurally cannot match.

4. Scaling Beyond the Indic Edge

Sarvam’s India-first strategy makes its models highly relevant for domestic governance, enterprise, and vernacular applications but also means sacrificing some universality and global generalization.

Final Word: How Does the Future of Sarvam AI Look?

As the IndiaAI Mission scales compute infrastructure, Sarvam AI is positioned to translate that expansion into deployed public services across health, agriculture, and governance. Sarvam also plans to bring its models to Nokia and HMD feature phones, cars, and Kaze smart glasses.

Indonesia, Nigeria, Brazil and every linguistically complex emerging market face the same gap India had. Sarvam’s architecture is a replicable playbook, not a one-country story.

Case study on Sarvam AI confirms that this time India is not watching the AI race from the sidelines. It is building its own models, for its own people, on its own terms. The foundation is built. The market is ready. The biggest returns are still ahead.