The second phase of the IndiaAI Mission is all about finally delivering the GPUs that were promised earlier, while also adding new promises to the table. While the first four startups are slowly beginning to receive their hardware, the eight new ones are already queued for allocations.
The government is now moving into full-scale compute deployment, making this the country’s biggest AI GPU programme. At the centre of it is BharatGen, led by an IIT-Bombay consortium, which has received the single largest single GPU allocation in the country to build sovereign large language and multimodal models, surpassing even Sarvam AI, one of the initial selections.
BharatGen has been allocated 4,096 NVIDIA H100 GPUs for two months, 8,192 H100 GPUs for 10 months, 440 H100 GPUs for speech models over a year and 912 H100 GPUs for vision-language models across two six-month phases. The project is supported with ₹988.6 crore, along with up to 25% additional funding for non-compute costs. The goal is a trillion-parameter model, an attempt no Indian team has made before.
“This is not just about building models but understanding why they behave the way they do. Trust is important, ethics is important, and Indian languages need the right representation,” said professor Ganesh Ramakrishnan of IIT Bombay, who is leading BharatGen, while speaking at AIM’s Cypher 2025. “If we don’t take this seriously, we risk losing not just Indian languages but also Indian content.”
The Bigger GPU Rush
BharatGen is part of a much larger GPU push. In less than a year, IndiaAI has empanelled over 34,000 GPUs, four times its original target of 10,000, with another 6,000 in the pipeline. This brings the total close to 40,000, making it one of the largest AI compute programmes outside the US and China.
Alongside BharatGen, several other companies are drawing from this pool. Fractal Analytics has secured 4,096 H100 GPUs for over nine months with a 40% concession to build India’s first large reasoning model, scaling up to 70 billion parameters. The focus is on structured reasoning and decision-making in healthcare, drug discovery, national security and education.
Alongside it, ZenteiQ.ai (formerly Zentech AI) is building BrahmaAI, a science-driven foundation model for engineering and scientific computing. It has 2,128 H200 GPUs over a year, supported by ₹74.7 crore, to deliver models from eight billion to 80 billion parameters. The project will cover engineering simulations and a multilingual science-education chatbot for non-invasive diagnostics.
In the healthcare space, NeuroDX, under IntelliHealth, is building a 20 billion parameter multimodal EEG model with 368 H200 GPUs over 18 months with ₹12.5 crore in support. The aim is early detection of dementia, personalised treatments for depression and anxiety and future brain-computer interfaces. All outputs from the project will be open-weight and open-source.
Meanwhile, Genloop is pursuing a smaller-scale, India-focused approach with three models—Yukti, Varta, and Kavach—each with around two billion parameters. The company aims to support all 22 scheduled Indian languages. Backed by just 16 H100 GPUs and ₹1.32 crore in funding for 12 months, Genloop is focusing on conversational AI for rural healthcare, inclusive banking and content moderation.
Tech Mahindra’s Makers Lab is working on an eight-billion-parameter model for Hindi dialects and an agentic AI platform, Orion. Its allocation is 32 H100 GPUs over nine months with ₹1.06 crore in support. Orion will extend Project Indus, which began with Hindi large language models (LLMs), into agritech, edtech, rural finance and healthcare.
Avataar.ai and Shodh AI complete the GPU landscape of IndiaAI’s second phase.
Avataar.ai is taking a multimodal route, building large multimodal models that range from 1.5 billion to 70 billion parameters across image, video and text. Its ‘Avataars’, a suite of domain-specific and distilled AI models, are intended to power key sectors like agriculture, healthcare, education and governance. The project, supported by an allocation of 768 H100 GPUs over six months, will also develop an agentic platform to enable sector-specific applications.
Shodh AI, meanwhile, is focused on a seven-billion-parameter foundational model for material discovery, compressed from an 80-billion-parameter scientific LLM. The model will power an autonomous system for hypothesis generation and experiment design, creating a materials-science AI assistant across electronics, semiconductors, healthcare and defence.
To achieve this, Shodh AI is allocated up to 128 H100 GPUs for a period of eight months.
When it comes to the GPUs already allocated from Phase 1, Sarvam AI, which is building a 120-billion-parameter model for Indic languages. Sarvam AI secured the highest order of 4,096 H100 GPUs over a period of six months and nearly ₹99 crore in subsidies. It is expected to release India’s first LLM next year.
Initially, the firm had received 1,536 GPUs from the initial tranche. However, Sunil Gupta from Yotta previously confirmed to AIM that the company has now received all of the allocated GPUs.
Similarly, Soket AI Labs is charting a path to a 120-billion-parameter Indic model for healthcare, defence and education. Its journey begins with a seven-billion-parameter model over six months, before it is scaled up to 120 billion parameters. Gupta also confirmed that 1,536 GPUs are reserved from Yotta’s end for the firm, which will be deployed soon.
Gnani.ai has a ₹177 crore contract for 1.3 crore GPU hours, equivalent to 1,536 GPUs across H100 and H200 units from E2E Networks for one year. The firm is building a 14-billion-parameter voice AI model for multilingual real-time speech processing. The timeline for its GPU allotment is yet to be revealed.
As for Gan.ai, while there have been no announcements about GPUs or the roadmap yet, the company aims to create a 70-billion-parameter model to achieve ‘superhuman text-to-speech’ capabilities.
Why BharatGen Stands Out
BharatGen alone accounts for approximately 13,640 H100 GPUs. Taken together, the disclosed allocations so far amount to roughly 22,776 H100 GPUs and 2,496 H200 GPUs with a mix of 3,072 H100 and H200 GPUs, or just over 28,000 GPUs—around 71% of the available pool. That leaves 11,000-12,000 of India’s GPUs, around 29%, either unallocated or undisclosed.
Seeded initially by the central government’s science and technology department (DST), BharatGen is now a nine-member consortium with funding and GPUs higher than those of Phase 1 beneficiaries. It has already released Param-1, a 2.9-billion-parameter bilingual model with 25% Hindi data, as well as several domain-specific models in agriculture, law, finance and Ayurveda.
It has also developed Shrutam, an automatic speech recognition system, and Patram, a seven-billion-parameter vision-language model for document understanding. These smaller-scale projects are the foundation for its trillion-parameter roadmap.
At Cypher 2025, Ramakrishnan explained why sovereign models matter. “The government has taken this very seriously. We would like all this to balloon into a massive creation of Indian language content,” he said. “The models cannot be built by merely tokenising Indian languages into English-heavy models. I don’t buy into such a vision. Indian languages are as important as English. This is about building models on our own terms.”
This philosophy is baked into BharatGen’s open approach. All models will be open-source, open-weight and open recipe, and the IP will be retained by IIT-Bombay and IndiaAI. Use cases range across governance, finance, healthcare, agriculture, education and law. The underlying bet, however, is on sovereignty. The goal is to make India a producer, not just a consumer, of AI technologies.
“This is actually the seat of India’s AI ecosystem, having our feet on the ground through applications, while building models which are not just aping Western models,” Ramakrishnan added.
The post How BharatGen Took the Biggest Slice of IndiaAI’s GPU Cake appeared first on Analytics India Magazine.
