Sarvam AI's voice-first Indic bet: $41M seed, then India's newest AI unicorn at $1.5B

Sarvam bet India's AI would be voice-first and small, priced agents at ₹1/min, and became India's newest AI unicorn at a $1.5B valuation in June 2026.

Sarvam AI

它在赌什么That India's AI future is voice-first and small: open Indic-language models priced at ₹1 per minute will beat text-heavy frontier models for a billion daily interactions.在扩

做的是什么生意

A Bengaluru AI lab building Indic-language foundation models, voice agents and enterprise AI for banks, insurers and government bodies.

启动资金：$41M across seed and Series A by December 2023 (Lightspeed, Peak XV, Khosla Ventures), then a $234M Series B first close at a $1.5B valuation in June 2026, targeting $300M total.

起因

Founded in 2023 in Bengaluru by Vivek Raghavan and Pratyush Kumar, who had worked on Indian-language AI at AI4Bharat. They argued text-only chatbots fail India's 22 official languages and 19,000-plus dialects, so in August 2024 they launched a voice-enabled AI agent supporting over 10 Indian languages, built on the open-source Sarvam 2B model; customer Sri Mandir had already processed more than 270,000 transactions through the voice agent.

新闻TechCrunch: Sarvam's ₹1/min voice-agent bet

经过

After the December 2023 raise, Sarvam released the 2B-parameter Sarvam 1 in October 2024. In April 2025 the IndiaAI Mission selected it to build India's sovereign LLM with dedicated compute. In February 2026 it unveiled 30B and 105B mixture-of-experts models trained from scratch on about 16 trillion tokens, plus speech and document models, and committed to open-sourcing them.

新闻YourStory: India taps Sarvam for sovereign LLM 新闻TechCrunch: Sarvam's open-source 30B/105B models

结果

In June 2026 Sarvam became India's newest AI unicorn, raising $234M at a $1.5B valuation in the first close of its Series B, led by HCLTech ($150M) with Bessemer, Khosla Ventures and Peak XV. Its conversational platform handles more than 2 million interactions a day and about 10 million API calls daily; multilingual voice agents have collected data from 17 million farmers and supported policy renewals for 45 million policyholders.

新闻TechCrunch: $234M round, $1.5B valuation

背景

Sarvam AI was founded in 2023 in Bengaluru by Vivek Raghavan and Pratyush Kumar, veterans of AI4Bharat, the IIT Madras Indian-language research group. Their core claim: a text-only chatbot is the wrong interface for a country with 22 official languages and over 19,000 dialects, where typing in local scripts is hard. So they bet on voice-first, small, open models trained on scarce Indic data, and priced agents at ₹1 per minute — roughly a cent — to make AI cheap enough for millions of daily interactions.

By December 2023 the pair had raised $41M across seed and Series A from Lightspeed, Peak XV and Khosla Ventures. In August 2024 Sarvam launched a voice agent supporting more than 10 Indian languages, built on Sarvam 2B, a model trained on 4 trillion tokens of cleaned synthetic data because the open web lacks enough Indic-language text. It claimed the model would cost a tenth of comparable systems, and open-sourced it.

The bet kept compounding. In April 2025 India's IndiaAI Mission selected Sarvam to build the country's sovereign foundation model with dedicated compute. In February 2026 it unveiled 30B and 105B mixture-of-experts models trained from scratch on roughly 16 trillion tokens, open-sourcing them alongside speech and vision models. In June 2026 HCLTech led a $234M first close at a $1.5B valuation — making Sarvam India's newest AI unicorn — with plans for a $300M Series B.

Scale followed the interface bet: 2 million-plus conversational interactions and about 10 million API calls a day, 500,000 hours of audio transcribed monthly, 35 million-plus pages digitized, and voice campaigns touching 17 million farmers and 45 million policyholders.

这件事要成立，得有什么

Sarvam treated India's linguistic fragmentation as the product: voice-first agents in 10-plus languages instead of a text chatbot in a couple.
Small open models priced per minute made AI affordable at India's scale, deliberately rejecting the expensive frontier-model playbook.
Synthetic data turned a scarcity — too little Indic-language text on the open web — into a training strategy for Sarvam 2B.
A government compute grant plus a strategic investor (HCLTech) converted a startup bet into national infrastructure and a $1.5B valuation.

可借鉴之处

Sarvam won by betting against the frontier-model playbook: voice-first, small, open, and priced per minute, proving that Indic-language scale is a market rather than a niche.

后续进展

As of 1 September 2026, Sarvam is private and scaling: its Series B remains open toward $300M; models and products (including Sarvam for Work and the Samvaad agent platform) are deployed across banking, insurance and government services; and it is expanding into inference infrastructure and next-generation agentic, coding and security models. No public revenue figures have been disclosed.

资料来源

发现哪里写错了？告诉我们。

轮到你了

你刚读完一家。说说你在做什么，看看谁在赌同一件事。

免费账号 · 3 次免费提问 · 不用绑卡