When Chinese startup DeepSeek unveiled its open-source foundation model DeepSeek-R1 earlier this year, it sent shockwaves across the AI world—and a wake-up call through India’s tech community. The model, developed with significantly fewer resources than its Western counterparts, highlighted what many Indian AI builders had long suspected: the barrier to AI leadership may not be resources alone, but resolve.
In Bengaluru, 20-year-old Adithya Kolavi, founder of CognitiveLab, saw the announcement as a moment of validation. “If DeepSeek could do it, why not us?” he asked, echoing the growing sentiment among India’s ambitious AI entrepreneurs. For others, like Abhishek Upperwal of Soket AI Labs, who had struggled to develop India’s early multilingual model Pragna-1B on a shoestring budget, the moment was bittersweet. “If we had been funded two years ago, there’s a good chance we’d be the ones building what DeepSeek just released,” he reflected.
These diverging perspectives capture the dual mood within India’s AI ecosystem—hope fueled by ingenuity, and frustration constrained by systemic gaps. As MIT Technology Review reports, India, despite its global reputation as a tech hub, has lagged in developing homegrown AI models due to chronic underinvestment in R&D, limited compute infrastructure, and a fragmented linguistic landscape.
A Government in Motion
The DeepSeek-R1 release jolted the Indian government into action. Within days, the Ministry of Electronics and Information Technology (MeitY) issued a call for foundational model proposals. By April 2025, India had committed to developing six large-scale AI models and 18 sector-specific applications, spanning agriculture, education, and climate resilience.
Key to this push was MeitY’s partnership with private-sector providers, unlocking access to nearly 19,000 GPUs—critical for AI training. In a significant move, Indian startup Sarvam AI was chosen to lead the charge with a 70-billion-parameter model focused on Indian languages and voice fluency, powered by over 4,000 Nvidia H100 GPUs.
Gautam Shroff of IIIT-Delhi likened the rapid mobilization to India’s Mars mission success: “India could do a Mangalyaan in AI.”
The Language Labyrinth
India’s rich linguistic diversity—22 official languages and hundreds of dialects—poses a unique challenge in training large language models (LLMs). Unlike English-dominant countries, India lacks the digitized, high-quality datasets necessary for training robust multilingual models. Standard global tokenizers often fail on Indic scripts, leading to bloated inputs and degraded performance.
But local innovators are fighting back with smart engineering. Upperwal’s Pragna-1B, trained for just $250,000, used a technique called “balanced tokenization,” enabling it to perform comparably to much larger models in Hindi and Gujarati. Meanwhile, Sarvam AI released OpenHathi-Hi-v0.1, a major open-source Hindi language model built on Meta’s Llama 2.
Another standout is Krutrim-2, a 12-billion-parameter model designed for India’s linguistic complexity. The team built a custom Indic tokenizer and optimized training for multimodal and voice-first applications—essential in regions where text-based AI tools fall short.
Building an AI Ecosystem
India’s new $1.25 billion IndiaAI Mission is designed to nurture this emerging ecosystem. It supports compute capacity, national datasets, deep-tech R&D, and AI startups. The program includes partnerships with companies like Soket Labs, Gan AI, and Gnani AI, and aims to fund at least a dozen startups focused on foundational models.
However, debates have emerged over whether these publicly funded models should be open-source. Sarvam’s flagship model is closed, despite its public backing—unlike DeepSeek-R1, which was released with open weights. AI policy expert Amlan Mohanty argues, “True sovereignty should be rooted in openness and transparency.”
MeitY’s Abhishek Singh maintains a neutral stance: “India has always supported open standards and open source, but it’s up to the teams. The goal is strong Indian models, whatever the route.”
Infrastructure and Innovation
While India still faces a severe GPU shortage and depends heavily on imported chips, officials are optimistic. Singh believes new infrastructure investments and public-private partnerships will ease the compute bottleneck within months.
India also enjoys cost advantages: building a hyperscale data center here costs about half what it would in the US or Europe, thanks to cheaper land and labor.
In the meantime, startups are focusing on smaller, efficient models and fine-tuning techniques to do more with less. “The absence of infrastructure doesn’t mean the absence of innovation,” notes Emory University’s Hancheng Cao.
A Different Game
Industry veterans like Infosys chairman Nandan Nilekani warn against chasing OpenAI-sized ambitions. “Let the big boys in the Valley do it,” he said at a Meta summit. “We will use it to create synthetic data, build small language models quickly, and train them using appropriate data.”
Others, like Paras Chopra—founder of the AI residency program Lossfunk—are investing in open-source science and homegrown talent to ensure India develops its own path.
As MIT Technology Review underscores, India’s AI ambitions are not about catching up—they’re about charting a new course that aligns with the country’s linguistic realities, economic constraints, and social needs. The ultimate goal: strategic self-reliance in AI that benefits not just the elite, but the masses.
Whether through large models like Sarvam’s, grassroots innovation like Pragna-1B, or nimble solutions like Krutrim-2, India’s AI builders are beginning to believe that the next AI frontier might just speak their language.
