India in AI Race: Behind US and China, but Focused on Local Needs
For the past three years, the global artificial intelligence race has centred on one question: which nation has developed the most advanced large language model? The United States has led with models such as OpenAI's GPT, Google's Gemini, Anthropic's Claude, and Meta's Llama. China has responded with DeepSeek, Qwen, and ERNIE. France has emerged with Mistral.
India, despite being one of the world's largest software talent hubs and having the highest number of weekly active ChatGPT users, has yet to produce a globally recognised frontier AI model. This raises the question: has India already fallen behind? And more importantly, does it need to win the same race?
Experts say the answer is nuanced. Yes, India trails the US and China in building the largest models, but that is only half the story.
Saket Dandotia, co-founder and CEO of Onetab.ai, notes that frontier AI models require enormous computing power. Training a competitive model today demands cluster-months on over 10,000 H100-equivalent GPUs, costing hundreds of millions of dollars. In comparison, India's Rs 10,372-crore IndiaAI Mission has committed approximately 34,000–38,000 GPUs across all participating startups. That is far below what a single frontier AI laboratory may consume during one major training run.
"If the yardstick is whether India has built something that tops the GPT or Gemini leaderboard, the gap is real," Dandotia says. However, he adds that this is not the complete picture.
Over the past 18 months, India has quietly built an AI ecosystem. Sarvam has trained large language models entirely on Indian computing infrastructure. BharatGen, Tech Mahindra's Project Indus, Gnani.ai, and Avataar's Varya have also entered the market with products focused on Indian languages, voice technologies, and enterprise applications. According to Dandotia, several of these systems already outperform much larger international models on Indian language benchmarks. Their goal was never to compete with ChatGPT but to address India's multilingual and cost-sensitive AI challenges.
The nature of the AI race itself is changing. The first phase rewarded whoever built the largest foundation model. The next phase may reward those who make AI useful in everyday life. "The race has been silently shifting," says Sanajaya G, CTO at VigyanLabs. He argues that AI competition is increasingly about deployment, data, and infrastructure rather than model size alone. "The winner will not necessarily be the company with the biggest model. It will be the one that deploys real-world AI at low cost using strong infrastructure and quality training data."
Sanajaya says India is well positioned for this shift. The IndiaAI Mission has allocated over Rs 10,300 crore, nearly 38,000 GPUs are being onboarded, and the country has access to one of the world's largest and most diverse data pools, generated by over 1.4 billion people. India also ranks third globally in Stanford University's AI Vibrancy Index, behind only the US and China, driven largely by its AI talent base.
The debate is no longer just about prestige. Experts argue that AI is becoming critical infrastructure—like electricity, telecom networks, or cloud computing. "If the intelligence running your banks, hospitals, and government sits on someone else's servers, they decide the price, the rules, and even whether you get access," says Dandotia. He points to recent export restrictions on advanced AI models and chips as evidence that access can become a geopolitical tool. The commercial stakes are equally significant.
In summary, while India may not lead in frontier AI models, it is building a niche in local-language and cost-effective AI solutions. The evolving race may favour India's strengths in data, talent, and deployment, though significant investment in computing infrastructure remains essential.
