Sarvam AI Success Story: How India's Homegrown AI Startup Is Building Sovereign Artificial Intelligence

Introduction

Most of the AI models people talk about were built in the United States or China and trained mainly on English or Chinese text. India has 22 scheduled languages, hundreds of dialects and a habit of mixing languages in a single sentence, so a model built around English can serve it poorly. Sarvam AI, a Bengaluru-based Indian AI startup working on multilingual AI and what is often called sovereign AI, was founded to address that gap.

Sarvam was established in August 2023 by Vivek Raghavan and Pratyush Kumar, both formerly with AI4Bharat at IIT Madras, and is headquartered in Bengaluru (Wikipedia). In June 2026 it announced a first close of $234 million in a $300 million Series B at a $1.5 billion post-money valuation, which made it an Indian AI unicorn (Sarvam).

This article tells the Sarvam AI success story, from the founders to the funding, using the company's own announcements, press coverage and public records. It covers who founded Sarvam, what it builds, how it has been funded, what the government's role has been, and what remains unproven. Where a figure comes from the company itself, the article says so, and where something is not publicly known, such as revenue or profitability, it says that too.

Quick facts

What Is Sarvam AI?

Sarvam AI builds AI models and products for Indian languages. Its catalog, as listed on its own models page, includes:

• Large language models: Sarvam 30B and Sarvam 105B, plus Sarvam-M, an open-weight hybrid reasoning model for Indic languages.
• Speech-to-text: Saaras V3, described as streaming speech recognition for 22 Indian languages with low-latency decoding and code-mixed support.
• Text-to-speech: Bulbul V3, described as natural, expressive text-to-speech across 11 languages.
• Vision: Sarvam Vision, a 3-billion-parameter state-space vision-language model for document digitization, OCR and visual understanding in Indian languages.
• Translation: Sarvam Translate, an open-weights model supporting 22 Indian languages (Sarvam models).

It also offers products built on these models, including a consumer chatbot called Indus, listed on Wikipedia among its products, and a first hardware device, the Kaze smart glasses.

Why Indian languages are hard for AI

A few features of the Indian language landscape explain why a dedicated multilingual AI effort made sense:

• Many languages. India's constitution lists 22 scheduled languages, and Sarvam's speech model and open models target all of them.
• Multiple scripts. Languages such as Hindi, Tamil, Telugu and Bengali use different writing systems, which affects how text is split into the tokens a model processes.
• Code-mixing. Many speakers switch between a local language and English within a sentence. Sarvam describes its speech model as supporting code-mixed input.
• Tokenizer efficiency. A tokenizer that splits Indian-language text into many small pieces makes models slower and more expensive to run. Sarvam says its custom tokenizer is especially efficient for low-resource Indic scripts (Sarvam).

When Sarvam raised its first round, Raghavan told TechCrunch that its models were more efficient, in terms of tokens consumed, for understanding and generating Indian languages than any existing LLM (TechCrunch). That is a company claim from 2023 and not an independent benchmark.

Sarvam AI Founders

Vivek Raghavan

Raghavan previously spent more than a decade at UIDAI, India's identity authority, according to TechCrunch's 2023 profile, and he worked at AI4Bharat, an initiative backed by Nandan Nilekani (TechCrunch). When Sarvam raised its first round, he was quoted as saying: "I have seen firsthand the enormous value in innovating at foundational layers and deploying at population scale."

Pratyush Kumar

Kumar also worked at AI4Bharat and has an AI research background. TechCrunch described him as having experience in AI research. Wikipedia lists both founders as formerly with AI4Bharat at IIT Madras (Wikipedia). Kumar was also the voice of the Kaze launch announcement.

What the founders had in common

Both founders came out of AI4Bharat, a research effort focused on Indian-language AI, before starting Sarvam. That gave them a background in the specific problem the company set out to solve. Beyond what the sources above say, this article does not attribute private motives or conversations to them.

Sarvam AI Timeline

Sarvam AI Funding Journey

The first round (December 2023)

Sarvam announced about $41 million raised across seed and Series A rounds. Lightspeed led the Series A, Peak XV Partners and Khosla Ventures took part, and the company was about five months old with roughly 18 employees (TechCrunch). Valuations for this round were not disclosed in the sources reviewed.

The Series B (June 2026)

Sarvam's announcement says the Series B is $300 million in total, with a first close of $234 million at a $1.5 billion post-money valuation. HCLTech invested $150 million as lead strategic investor, Bessemer Venture Partners participated, and existing investors Khosla Ventures and Peak XV Partners continued their support (Sarvam).

The company says the money will fund continued research on training its next frontier model for agentic, coding and cybersecurity use cases, along with access to compute at scale.

What the totals show

Adding the two announced rounds gives about $275 million raised (about $41 million plus $234 million). This is an arithmetic sum of the figures above, not a number the company has stated, and it does not include the balance of the Series B that has yet to close.

The Government Connection: IndiaAI Mission

Sarvam's relationship with the Indian government is one of the most distinctive parts of its story.

In April 2025 India selected Sarvam to build the country's first sovereign large language model under the IndiaAI Mission. Sarvam's own post says it will develop multi-modal, multi-scale foundation models from scratch in three variants: Sarvam-Large for advanced reasoning, Sarvam-Small for real-time applications and Sarvam-Edge for on-device tasks. The stated goals are models that are capable of reasoning, designed for voice, fluent in Indian languages, secure and ready for population-scale deployment (Sarvam).

Compute was part of that arrangement. Sarvam says the 30B and 105B models were trained entirely in India on compute provided under the IndiaAI Mission (Sarvam). Press reports have put the allocation at 4,096 Nvidia H100 GPUs, but Sarvam's own announcement does not state a number, so that figure should be treated as reported and not confirmed.

Why this matters

Training large models needs a large amount of computing power that few startups can fund alone. Access to government-provided compute lowered that barrier for Sarvam. It also tied the company's work to a national goal, which is sometimes called sovereign AI: the ability to build and run AI systems within a country using local data, infrastructure and expertise. The trade-off is closeness to government priorities, which can both help and constrain a company.

Sarvam's Large Language Models for India

Sarvam 30B and 105B

Sarvam released both models on March 6, 2026 under the Apache 2.0 license, which allows broad commercial use. According to the company's technical post:

Both use sparse expert layers with 128 experts and support all 22 scheduled Indian languages (Sarvam). The models are available on Hugging Face, AIKosh, the Sarvam API dashboard and the Indus app.

What "Mixture-of-Experts" means

In a Mixture-of-Experts model, only part of the network runs for each token it processes. A 30-billion-parameter model that activates 2.4 billion parameters per token does less computation per step than a dense model of the same total size, which can reduce the cost and delay of running it. The trade-off is that the full model still has to be stored in memory.

Sarvam-M and earlier models

Sarvam-M is listed as an open-weight hybrid reasoning model for Indic languages, tuned with supervised fine-tuning and reinforcement learning (Sarvam models). The Wikipedia entry also lists an earlier Sarvam-1 model.

Benchmark claims

In its Series B announcement, Sarvam says its 105B model matches larger reasoning models on benchmarks. That is the company's statement, and benchmark results depend on which tests are used, so independent evaluations are the better guide to how the model compares.

Speech, Vision and Translation

Language models are only part of Sarvam's catalog.

• Saaras V3 is described as streaming speech recognition for 22 Indian languages with low-latency decoding and code-mixed support.
• Bulbul V3 provides text-to-speech across 11 languages.
• Sarvam Vision is a 3-billion-parameter vision-language model aimed at document digitization and OCR in Indian languages.
• Sarvam Translate is an open-weights translation model for 22 Indian languages.

(Sarvam models.) The company's emphasis on voice is consistent with its stated aim in 2023 of building models with voice as a primary interface.

Usage figures the company reports

In the Series B announcement Sarvam gives several usage figures. They are company statements and have not been independently audited:

• Sarvam Vision has digitized more than 35 million pages from insurance and land records.
• Its conversational platform handles over 2 million interactions a day.
• Its inference platform processes 10 million API calls a day.
• Its speech models transcribe over half a million hours of audio each month.
• Deployed voice agents collected data from 17 million farmers for India's Ministry of Agriculture.
• An insurance campaign supported policy renewals for 45 million policyholders.

(Sarvam.)

These numbers suggest the company's tools are being used at scale in sectors such as insurance, agriculture and public records. They do not disclose revenue, and they do not name the commercial terms behind these deployments.

Sarvam Kaze: Moving Into Hardware

Sarvam's first hardware product is Kaze, a pair of smart glasses. The company unveiled them at the India AI Impact Summit 2026 in New Delhi, where Prime Minister Narendra Modi was the first person to try them. Sarvam says the glasses are engineered to listen, understand, respond and capture what the wearer sees, with cameras embedded in the frame and voice-based interaction. At the unveiling Kumar said, "Designed in India, built in India, fitted with AI from India. All in your hands this May" (Indian Startup News).

At the time of that coverage, detailed specifications and pricing had not been announced, and the sources reviewed do not confirm shipping details or sales. Coverage framed Kaze as entering a market currently led by Ray-Ban Meta and Oakley Meta glasses.

Business Model: What Is Known and What Is Not

Sarvam's public materials point to several ways it can earn money:

• An API and model platform. The company has an inference platform and an API dashboard, and says it processes 10 million API calls a day.
• Conversational and speech products. Sarvam describes a conversational platform that handles over 2 million interactions a day, plus speech models used in voice agents.
• Document digitization. Sarvam Vision is used for large-scale document processing.
• Hardware. Kaze adds a consumer device.

What the sources reviewed do not provide is revenue, pricing detail or profitability. Sarvam is a private company and has no obligation to publish financial statements. Any article that quotes specific revenue projections or margins for it is estimating, and this one does not.

Competition and Context

Sarvam operates in India's AI ecosystem, a field with very large global competitors. Global labs such as OpenAI, Google, Anthropic and Meta build general-purpose models that support many languages, and open-weight models are widely available. Sarvam's position is that a model built for Indian languages, with an efficient tokenizer and a voice-first approach, can serve the Indian market better and more cheaply.

Two cautions apply:

1. Claims of superiority should be checked. Company benchmarks are useful but not neutral, and general-purpose models may keep improving their coverage of Indian languages.
2. Open weights cut both ways. Releasing models under Apache 2.0 builds adoption and goodwill, and it also lets others build on them, including competitors, without paying Sarvam.

For another example of an open-model infrastructure company and how it was funded, see our Together AI success story. For another Indian deep-tech company that grew out of the country's public institutions, see the Skyroot success story.

Risks and Open Questions

• Revenue and unit economics are unknown. Large valuations rest on expectations, and the sources reviewed provide no revenue figures.
• Compute is expensive. The Series B is meant partly to buy compute for the next frontier model, and training large models is capital-intensive.
• Government dependence. Support under the IndiaAI Mission has helped, and priorities and policies can change.
• Competition from global labs and open models. Larger companies can improve their Indian-language performance.
• Hardware is a new business. Making and selling consumer glasses is very different from selling software, and Kaze's commercial results are not yet known.
• Company-reported metrics. Usage and benchmark claims come from Sarvam and should be treated as claims until independently verified.

What Entrepreneurs Can Learn

These are observations from the record above, not guarantees.

1. Specific problems can justify a new company. Sarvam focused on a gap, Indian-language AI, that general models served unevenly.
2. Founder background helped. Both founders came out of AI4Bharat, and one spent over a decade at UIDAI, which gave them relevant technical and institutional experience.
3. Early funding backed the team. Sarvam raised about $41 million when it was five months old and had around 18 employees, which suggests investors were backing the founders and the thesis.
4. Public compute can reduce a startup's biggest cost. Training under the IndiaAI Mission removed a major barrier, and it also tied Sarvam to a national program.
5. Open weights can build adoption. Apache 2.0 releases let developers adopt the models freely, though they do not themselves produce revenue.
6. Strategic investors can bring distribution. HCLTech's $150 million investment is described as a strategic investment from a large IT services company, which may help with enterprise relationships, though the sources reviewed do not detail the commercial arrangement.
7. Be careful with metrics. Usage numbers are a start, but investors and customers eventually look for revenue.

Frequently Asked Questions

What is Sarvam AI? Sarvam AI is a Bengaluru-based company, established in August 2023, that builds AI models and products for Indian languages, including language models, speech-to-text, text-to-speech, vision and translation models (Sarvam models).

Who founded Sarvam AI? Vivek Raghavan and Pratyush Kumar founded it, both formerly with AI4Bharat at IIT Madras (Wikipedia).

How much funding has Sarvam AI raised? It raised about $41 million in seed and Series A funding in December 2023 and $234 million in the first close of a $300 million Series B in June 2026, a combined figure of about $275 million by simple addition.

What is the Sarvam AI valuation? The Series B first close was at a $1.5 billion post-money valuation (Sarvam).

Who invested in Sarvam AI? Lightspeed led the 2023 round with Peak XV Partners and Khosla Ventures. In the Series B, HCLTech invested $150 million and Bessemer Venture Partners took part, with Khosla Ventures and Peak XV continuing.

What are Sarvam 30B and Sarvam 105B? They are open-weight Mixture-of-Experts language models released on March 6, 2026 under Apache 2.0, supporting all 22 scheduled Indian languages. Sarvam 30B has 2.4 billion active parameters (Sarvam).

Is Sarvam AI a unicorn? Yes. The Series B first close at a $1.5 billion post-money valuation in June 2026 put Sarvam above the $1 billion unicorn mark (Sarvam).

What is the IndiaAI Mission's role? India selected Sarvam in April 2025 to build its sovereign LLM, and Sarvam says the 30B and 105B models were trained in India on compute provided under the mission.

What is Sarvam Kaze? Kaze is Sarvam's first smart glasses, unveiled at the India AI Impact Summit 2026, with cameras and voice interaction. Specifications and pricing had not been announced at the time of the sources reviewed.

Is Sarvam AI open source? Some of its models are. Sarvam 30B, Sarvam 105B, Sarvam-M and Sarvam Translate are listed as open-weight, and the 30B and 105B models use the Apache 2.0 license.

Is Sarvam AI profitable? Sarvam has not published financial statements, and the sources reviewed give no profitability or revenue figures.

How does Sarvam AI make money? Its public materials point to API usage, conversational and speech products, document digitization and hardware, but the company has not disclosed revenue or pricing details.

Which languages does Sarvam support? Its speech-to-text and open models target the 22 scheduled Indian languages, while its text-to-speech model covers 11 languages (Sarvam models).

Conclusion

The Sarvam AI success story so far has three parts: a focused problem, Indian-language AI; a founding team with deep roots in that field; and unusual support, from private investors, a large Indian IT services firm and the government's compute program. Its open releases and the usage figures it reports suggest real activity, and its Series B shows that investors are willing to back that thesis at a $1.5 billion valuation.

What the public record does not yet show is how that activity turns into revenue, how its models compare in independent testing as global models improve, and whether its hardware move will succeed. Those open questions, more than the headline valuation, are what will decide whether this Indian AI startup's success story keeps growing.

Sources

• Sarvam, Sarvam raises $300M Series B (June 2026)
• Sarvam, Open-sourcing Sarvam 30B and 105B (March 2026)
• Sarvam, Sarvam to build India's sovereign large language model
• Sarvam, Models
• TechCrunch, India's Sarvam AI raises $41 million from Lightspeed, Khosla, Peak XV (December 2023)
• Wikipedia, Sarvam AI (founding, founders, IndiaAI selection, product list)
• Indian Startup News, Sarvam AI launches its first AI-powered smart glasses, Kaze
• HCLTech, Sarvam raises $234 million in first close of $300 million Series B (read through search results; page not directly readable)