India's Frugal AI Startups Target 800M Users
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Sarvam AI, co-founded by Vivek Raghavan and Pratyush Kumar after their work on AI4Bharat at IIT Madras, targets bringing generative AI to 800 million Indians with smartphones through voice-based, low-cost models in Indian languages.
- SarvamM, the company's flagship 24-billion parameter LLM, is trained across 10 Indian languages and powers vernacular learning assistants that handle codemixed queries for math and programming instruction.
- OpenHathi bolts Indian language skills onto existing open-source models like Meta's LLaMA and France's Mistral—starting with Hindi—rather than building expensive models from scratch, and releases outputs on Hugging Face.
- Hindi sentences require 3-4 times more AI tokens than the same English text, making every Indian-language AI interaction roughly five times costlier; Sarvam addressed this by building better tokenizers and high-quality datasets.
- Krutrim, launched in April 2023 by Ola Cabs co-founder Bhavish Aggarwal, was trained on over 2 trillion tokens and supports 22 Indian languages while running efficiently without supercomputers.
- AI4Bharat, launched at IIT Madras in 2020, anchors the frugal approach—lightweight systems running on low-end smartphones and low bandwidth, designed for India's 1,600+ dialects and 22 official languages.
- In healthcare, Sarvam AI has deployed voice-enabled multilingual conversational agents via WhatsApp that let rural patients access medical advice, schedule appointments, and consult doctors through low-bandwidth interfaces.
Why it matters: With 22 official languages and 1,600+ dialects, India can't rely on English-centric AI—and the tokenization economics prove it: Hindi costs five times more to process than English, making frugal design an economic necessity, not a philosophy. If SarvamM and Krutrim succeed at scaling vernacular AI to 800 million smartphone users, they offer a replicable blueprint for the Global South that sidesteps Silicon Valley's compute-heavy model entirely.
