“ChatGPT can write a perfect essay on Shakespeare, but if you ask it to read your handwritten 1998 land record document written in mixed Malayalam and English, it would start making up legal clauses.”
Though not certain if the conversation ever took place, it encapsulates the problem that has quietly shaped India's ambitions in AI technology.
India is among the world's largest digital nations. Every day there are billions of transactions made via India's UPI, Aadhaar, DigiLocker, ONDC, and dozens of other platforms in regional languages. And yet, the foundation models which power this modern age of artificial intelligence were largely developed and trained abroad, thus forming what many technologists call digital asymmetry — India has been generating huge digital activity while value captured by frontier AI models is elsewhere.
However, the issue was not so much in where those models were built, but in how they perceive language.
Foundation models tokenize input texts before processing them. Tokenizers are optimized for the datasets with the highest frequency of English, thus making many Indian scripts significantly less efficient. The more tokens, the higher the inference cost, latency, and context window size. Thus, in India, AI was not failing to translate — it was failing to understand the linguistic reality of India.
This realization led to a new approach.
Instead of building yet another chatbot, India set out to build a sovereign AI ecosystem.
The centerpiece of this initiative is IndiaAI Mission with the total budget of ₹10,372 crore. Much like how UPI democratized digital payments, the mission seeks to democratize the AI compute through the provision of startups, universities, researchers, and public institutions access to shared GPU infrastructure. Right now, the program has brought together more than 38,000 GPUs, reducing the cost of advanced AI development, providing subsidized access to the cloud compute with high-end artificial intelligence accelerators with the average cost of ₹65 per GPU per hour.
But sovereign AI is not about the hardware.
It is about building every layer of the stack from the compute and datasets to the foundation models, evaluation benchmarks, AI safety, and applications.
The ecosystem has already been forming.
Sarvam AI has been leading the efforts by building multilingual models tailored to the needs of Indian languages and workflows. Its latest models feature Sparse Mixture-of-Experts (MoE) architecture where only the specialized parts of the model are activated in response to the query. The request written in Malayalam or Manglish will therefore be routed to language experts optimized for it.
Sarvam is not alone in this. SoketAI develops multilingual foundation models for healthcare, defense, and education. Gnani.ai focuses on conversational speech AI, Gan AI develops advanced multilingual voice synthesis, while BharatGen builds open foundation models for governance, agriculture, healthcare, and education. Rather than backing the single national champion, India has opted for the ecosystem of specialized AI developers.
The applications extend well beyond chatbots.
Government offices, courts, banks, and hospitals still operate with scanned PDFs, handwritten documents, and multilingual records. In order to fix this, Sarvam has recently launched Sarvam Vision, an OCR model tailored for Indian document understanding. Sarvam Vision achieved state-of-the-art accuracy of 84.3% on the olmOCR-Bench (English only subset), outperforming such frontier models as Gemini 3 Pro and recent OCR models like DeepSeek OCR 2.
Another important frontier is the voice — hundreds of millions of Indians feel more comfortable speaking rather than typing. Technologies like Bulbul alongside advancements made by Gnani.ai and Gan AI make it possible to develop multilingual speech recognition and natural text-to-speech in Indian languages, making AI accessible to even more people.
These technologies are already being deployed to the real world. Multilingual AI is being incorporated in Aadhaar services, agricultural platforms, voice-first commerce through ONDC and Bhashini, and public-sector speech translation. These are not attempting to replace ChatGPT — instead, they are solving problems that global frontier models are not specifically designed to solve.
The momentum is growing. The IndiaAI Mission has shortlisted 20 indigenous sovereign AI models proposals, while developing AIKosh, Safe & Trusted AI initiatives, funding startups, and nationwide AI labs.
Private sector investments are flowing in this area, too, with India's sovereign AI ecosystem becoming soon production ready.
India may not be able to afford competing with OpenAI, Google DeepMind, Anthropic, and Meta at least for now. But that isn't the goal — India's goal isn't to build the world's largest language model. But rather the most useful AI ecosystem for one of the most linguistically and culturally diverse nations in the world.
Rather than focusing on developing a few mega language models, India pursues the ecosystem approach, where startups, academia, and researchers all over the country are building from the foundational models, data, and safety infrastructure to applications and semiconductors.
Whether this ecosystem-focused approach would ultimately rival the world's leading AI laboratories is yet to be seen.
But in the coming decade, the leaders of artificial intelligence may not be the countries with the largest models. They may be the countries with the strongest ecosystems around them.