The government announced plans to support homegrown large language models (LLMs) developed by Indian startups and companies, days after a Chinese company’s demonstration of cost-efficient AI capabilities led to a trillion-dollar market value drop for American tech giants.
Union IT minister Ashwini Vaishnaw said on Thursday that six Indian entities had already shown promise and could release their foundation models in the next ten months. “What the world has seen today is that algorithmic efficiency matters a lot. Algorithmic efficiency can deliver a model at a much lower cost and in a much lesser time than the world has seen. Many of our researchers have been working on similar concepts. We are confident that using those concepts, we will be able to have a world-class foundational model in the next few months,” Vaishnaw said.
Foundation models, which form the basis of AI systems such as ChatGPT, are algorithms that train on vast amounts of data to develop broad capabilities in language understanding and generation. Unlike traditional AI programs designed for specific tasks, these models learn general patterns and knowledge that can be adapted for various applications — from writing code to analysing scientific papers. They serve as a base layer of AI that can be refined or “fine-tuned” for specific uses.
The development comes days after DeepSeek, a subsidiary of Chinese algorithmic trading company High-Flyer, released its open-sourced R1 model built at a reported cost of $6 million, significantly lower than the $100 million spent by American companies like OpenAI on comparable models such as GPT4. Vaishnaw said that DeepSeek’s open-sourced models will be hosted on Indian servers to address data security and privacy concerns. He, however, did not specify if the government was already in conversation with the company about this.
The ministry of electronics and information technology (MeitY) will accept proposals on a rolling basis, with applications received by the 15th of each month to be evaluated within 30 days. This process will continue for six months or until sufficient proposals are selected, according to IndiaAI CEO Abhishek Singh. Shortlisted applicants will receive either direct funding or investment agreements after detailed presentations.
The government’s initiative emphasises developing indigenous LLMs using Indian datasets to integrate local context, languages, and culture while removing biases. Singh added that applicants must demonstrate their model’s potential to address societal challenges.
To support AI development, the ministry announced the ten companies that submitted the lowest (L1) bids to provide common compute capacity across different graphics processing units (GPU) categories and thus could be empanelled under the IndiaAI Mission. The selected bidders have pledged 18,693 GPUs, exceeding the target of 10,000 GPUs, with 12,896 being top-end Nvidia H100 GPUs. Yotta Data Services leads with 9,216 GPUs. The other nine companies are Jio Platforms, Tata Communications, Ctrls Datacenters Limited, E2E Networks, NxtGen Datacenter and Cloud Technologies, CMS Computers, Locus Enterprise Solutions, Orient Technologies and Vensysco Technologies.
Each of these companies, after being deemed technically fit, was allowed to make financial bids. The ministry is expected to send a “letter of offer for empanelment” to the ten companies in a couple of days, a person aware of the matter said. Each company will then be given about a week to meet the lowest bid in each category to finalise their empanelment terms, this person told HT on the condition of anonymity.
Providing compute capacity is a major component of the government’s ₹10,371.92 crore outlay for the IndiaAI Mission; 44% or ₹4,563.36 crore is earmarked for providing compute capacity of more than 10,000 GPUs over a period of five years.
The compute facility will be available “within a couple of days” after approvals and contract signings, according to Vaishnaw.
Apart from compute capacity, providing India-specific datasets that can be used to train the models is also important, the Vaishnaw said. The ministry will soon make an announcement about the Indian Datasets Platform, another pillar of the mission, he said.
The lowest bids have already been made at significant discounts, averaging at 42% below their listed price. On an average, it will cost companies ₹115.85 per GPU hour and ₹150 per GPU hour for high-end compute capacity—substantially below the global average of $2.5-3 per GPU hour. Academia, researchers, students, and startups will receive an additional 40% government subsidy, bringing the cost to under ₹100 per GPU hour for them.
The government has readied a portal where entities can apply for access to this common compute capacity. This portal is currently undergoing security checks and audits, a second person aware of the matter said, and is expected to be ready in a week.
Additionally, Meity will fund 18 applications in agriculture, learning disability, and climate change under the Applications Development Initiative pillar of the IndiaAI Mission.
The government is also establishing an AI Safety Institute (AISI) using a collaborative hub-and-spoke model to enable participation from universities across the country in developing AI safety frameworks and tools, the minister announced.
HT reported in October that the government was considering setting up an AISI to set standards, frameworks and guidelines for AI development without acting as a regulatory body or stifling innovation.
Vaishnaw and Singh told HT that the AISI will work in a dispersed manner. For collaboration and knowledge sharing with other AISIs, such as those in UK, USA or Japan, IndiaAI will lead the discussions, Singh said. Effectively, IndiaAI within MeitY will act as the secretariat for coordination of the AISI’s work, the second person cited above said.
The institute is a part of the mission’s ‘safe and trusted AI pillar’. This pillar was allocated ₹20.46 crore, which officials say can be increased.
Eight previously selected projects focusing on responsible AI development will now fall under AISI’s purview. These include projects about machine unlearning, synthetic data generation for bias mitigation, AI bias mitigation in healthcare systems, and AI algorithm audition tools.
In December, Meity had invited proposals for five AI safety projects, including real-time deepfake detection, AI-generated content watermarking, ethical AI frameworks, and red teaming AI models, which will now be incorporated into AISI’s scope.