Something significant shifted in the Indian technology landscape around mid-2024. The conversation stopped being about whether India could build its own artificial intelligence and started being about which Indian AI would win. That transition marks the true beginning of the Indic LLM era.
The Language Problem Nobody Wanted to Solve
For years, the dominant AI models spoke English — not just in the sense that their interfaces were in English, but in the deeper sense that their training data, reasoning patterns, and cultural assumptions were rooted in an English-speaking world. For a country where fewer than 12 percent of people are comfortable in English, this was not a minor inconvenience. It was a structural exclusion from the benefits of the AI revolution.
India has 22 officially recognised languages and over 1,600 dialects. Hindi alone has more native speakers than the entire population of the United States. Tamil, Telugu, Bengali, Marathi, Kannada — each carries centuries of literature, commerce, and daily life. Building AI that genuinely understands these languages is not a translation problem. It is a civilisational one.
The researchers who understood this earliest were not in Silicon Valley. They were in IIT Madras, in Bangalore research labs, and in small offices in Hyderabad. By 2026, their work had produced something the world had not seen before: a cluster of foundation models built from the ground up for the linguistic reality of South Asia.
Sarvam AI and the OpenHathi Moment
When Sarvam AI released OpenHathi in late 2023, the response from the Indian developer community was immediate and electric. Here was a Hindi-first language model that did not feel like a translated version of something else. It felt native. The idioms landed correctly. The cultural references made sense. The model understood that a question about "jugaad" was not asking about a word but about an entire philosophy of resourceful improvisation.
OpenHathi was built on a foundation of Llama 2, but Sarvam's team — led by Pratyush Kumar and Vivek Raghavan, both veterans of the AI4Bharat research initiative — had done something more than fine-tuning. They rethought the tokenisation strategy for Hindi, which dramatically improved the model's efficiency on Indic text. Where a standard tokeniser might break a Hindi sentence into dozens of inefficient fragments, Sarvam's approach preserved linguistic structure, reducing compute costs and improving coherence.
The Series A funding of $41 million that followed in early 2024 was not just a vote of confidence in Sarvam's technology. It was a signal that global investors — Lightspeed and Peak XV among them — believed the Indic AI market was real, large, and defensible.
Krutrim: The Ambition of a Full Stack
Bhavish Aggarwal's announcement of Krutrim in December 2023 was characteristically bold. He was not just building a language model. He was building India's own AI stack — from custom silicon to consumer applications. The name itself, derived from the Sanskrit word for artificial, signalled an intent to own the entire narrative.
Krutrim's approach differed from Sarvam's in important ways. Where Sarvam focused on research-grade models and enterprise APIs, Krutrim aimed at the consumer market from day one. The Krutrim assistant launched with support for all 22 official Indian languages, a feat that required not just multilingual training data but careful attention to code-switching — the way Indian speakers naturally blend languages in a single conversation.
The $50 million funding round that valued Krutrim at $1 billion made it India's first AI unicorn. But the more significant milestone was what it represented: proof that an Indian AI company could attract the kind of capital that had previously flowed only to American and Chinese AI labs.
The Data Advantage India Did Not Know It Had
Building a good language model requires three things: compute, talent, and data. India's digital transformation over the past decade has generated an extraordinary corpus of Indic language text. WhatsApp messages in Tamil. YouTube comments in Telugu. Government documents in Marathi. News articles in Bengali. E-commerce reviews in Kannada. This data exists nowhere else in the world in this volume and variety.
The challenge has been collection, cleaning, and curation. Projects like AI4Bharat's IndicCorp have done the painstaking work of assembling these datasets, and their open-source release has given Indian AI startups a foundation that would have taken years and hundreds of millions of dollars to build independently.
Enterprise Adoption: Where the Money Actually Is
The consumer applications get the headlines, but the enterprise market is where Indian Indic AI companies are finding their earliest and most sustainable revenue. Banks and insurance companies need to communicate with customers in their native languages. Healthcare providers need AI that can understand a patient describing symptoms in Gujarati. Government agencies need systems that can process applications in any of the 22 official languages.
Sarvam's enterprise API business has grown faster than its consumer products, a pattern that mirrors the early trajectory of OpenAI. The B2B path to scale is less glamorous but more predictable, and it is funding the research that will eventually power the consumer products.
What Comes Next
The next phase of Indic AI development will be defined by multimodality. Text-only models, however good, cannot serve a population where voice is the primary interface for hundreds of millions of people. The combination of speech recognition, language understanding, and speech synthesis — all working natively in Indic languages — is the real prize.
India's AI revolution is not a story about catching up with the West. It is a story about solving problems the West never had to solve, building capabilities the West never had to build, and in doing so, creating technology that will eventually be exported back to the world. The billion voices are finding their AI. And the AI is learning to listen.