India's AI Race: Government Launches IndiaAI Mission to Develop Sovereign Foundation Models

India's AI race heats up as entrepreneurs, researchers, and startups vie for IndiaAI Mission funding

Stanford graduate Ayush Gupta who quit his job at Apple to start his own venture, Gurgaon entrepreneur Sahil Gulati, who has experience in B2B sales, and Patna-based engineer Aaditya Kirti are among the many likely contenders to build a sovereign artificial intelligence (AI) foundation model for India. There is interest not just among established startups and experts, but even 20-somethings and those with no tech backgrounds. The ministry of electronics and information technology (MeitY) allows groups of individuals not registered as a company, or even an individual, to send in proposals under the IndiaAI Mission. A meeting called by MeitY earlier this week was attended by close to 300, mostly startups, researchers, academics and students, all keen to work on the project, encouraged by the government's offer of Rs 1,500 crore (about $173 million) support and subsidised computing power. Many have drawn inspiration from China's DeepSeek, with its claimed shoe-string budget of $5.6 million. The ministry will accept proposals till the 15th of every month for the next six months, or till it collects enough proposals, whichever is earlier. The first deadline is Saturday. "This is a good opportunity to develop native AI capabilities," said Gupta, whose startup Genloop has presence in New Delhi and San Francisco. Genloop is working on an India-specific compliance large language model (LLM). It will be pre-trained with Indian cultural context, history and policies, sourced from domestic websites, Parliament proceedings, open-source datasets, legal judgements, newspapers, regional cinema and social media feeds, he said. Level playing field The LLM will support the 22 scheduled Indian languages, said Gupta of Genloop. While his budget is yet to be finalised, it may cost him about $1.5 million, said the 30-year-old. According to startup founders, compute is the costliest component in developing an AI model. With that being taken care of with the government's offer of subsidised access to over 10,000 graphics processing units (GPUs), it is a level playing field for them with some highly funded companies such as Sarvam AI, which has raised a total of $53.6 million to date. EJY Health, founded by Kirti, is testing its model. "We are conducting the pilot in five hospitals in Patna. Three are government hospitals, while two are private," the 29-year-old said. The startup, incubated at IIT-Patna and supported by the Startup Bihar initiative, aims to submit a proposal to the IndiaAI Mission by March 15. "Our solution is designed to enhance the efficiency and productivity of nurses, doctors and hospital administrators. Currently, 60% of their time is spent on documentation. Our AI model will reduce this to 20%, allowing healthcare professionals to spend 40% more time on patient care, thereby reducing burnout and improving overall healthcare delivery," Kirti explained. Gurgaon's Gulati, 25, offers tech solutions through his startup EliteNotes, but has no formal tech education. His one-and-a-half-year-old company too is building an AI-driven enterprise platform to enhance multitasking across workflow applications. "For the IndiaAI Mission, we'd like to build a multilingual multimodal small language model. It would support audio and video in multiple Indian languages. We will be submitting our proposal by March 15," he said. India may take nine to 12 months to build its own sovereign AI foundation model, Abhishek Singh, the IndiaAI Mission chief executive and an additional secretary at MeitY, told the meeting on February 11. IndiaAI Mission adviser Aakrit Vaish told the attendees that the funding would be in the form of grants, compute credits, or staggered equity funding. The team's expertise and track record when it comes to research, development and deployment; approach in building the model, unique angle and data strategy are among essential eligibility criteria, the officials said.

AI Mission CEO pegs funding window for models at Rs 1500 crore

Developers of foundation artificial intelligence models can access Rs 1,500 crore of funds under the IndiaAI Mission, said Abhishek Singh, its chief executive and an additional secretary at the Ministry of Electronics and Information Technology (MeitY). Singh said this on Tuesday while addressing a meeting of the IndiaAI Mission, attended by close to 300 startups, researchers, academics and students. The corpus is meant for a period of five years. When asked about the funds available at the request for proposal stage, Singh said: "It depends on your approach of building that. We have all seen that foundation models can be built with as little as $6 million to different ranges." The IndiaAI Mission in a January 30 notification called proposals for building India's foundation models. Proposals will be reviewed by the 15th of every month, until enough proposals are selected or six months, whichever is earlier. Also Read: India to develop own generative AI model: Ashwini Vaishnaw Addressing the AI Action Summit in Paris on Tuesday, Prime Minister Narendra Modi reiterated India's ambitions on AI. "India is building its own large language model considering our diversity. We also have a unique public-private partnership model for pooling resources like compute power. It is made available to our startups and researchers at an affordable cost," he said. Singh attended the IndiaAI Mission meeting virtually from the AI Action Summit in Paris. "I can tell you that the buzz here is that everybody here expects India to do a lot," he said. "We have seen in panels the swagger of the Chinese because of the DeepSeek model that they have made. Everyone we meet say where is India and ask what India can do. Ultimately solutions will be built by startups, academics and researchers like you." MeitY had informed the Lok Sabha in December that 19%, or Rs 1,971.37 crore, of the Rs 10,371.92 crore approved by the Cabinet for the IndiaAI Mission in March last year was allocated for the IndiaAI Innovation Centre (IAIC). The IAIC is mandated to develop and deploy indigenous large multimodal models (LMMs) and domain-specific foundational models in critical sectors. Also Read: Union Cabinet approves Rs 10,372-crore India AI mission "Depending on what you need the funding for, depending on how you are approaching the problem, what your timelines are, what your team size is, depending on that, an assessment of your capabilities will be done and the chance to present it will be given," Singh said. "Justification of expense details have to be part of the proposal. It can be $5 million, $7 million or $10 million. The determining criteria of funding will be your approach and way towards making the model," he said. "We will be supporting a sufficient number of teams to ensure that at least a few of them are able to make it." The call for proposals allows for LLMs, multimodal models, smaller models and large quantitative models, he said. "The foundation model comes in pillar number two of the IndiaAI Mission or what we call the IndiaAI Innovation Centre which is for the development of indigenous large models and small models," IndiaAI Mission adviser Aakrit Vaish said. Funding will be in the form of grants, compute credits, or equity funding over a period of time, he said. Expertise, and track record of the team when it comes to research, development and deployment, the approach in building the model, unique angle, data strategy and milestones for building the foundation models are essential eligibility criteria, Vaish explained. "Think about long-term sustainability and viability of whatever you are trying to build. Think about how this can be used by companies, developers and startups over a period of time, over many years to come, and not just one time," he said. The proposals can be made by a consortium of companies or a consortium of individuals and researchers. What is foundation AI model? Any AI model where large amounts of training data are applied to build a model that can be used for generative AI purposes -- horizontal or general purpose use, multimodal (video, audio) or small models, ground up base model to solve a specific vertical use case, like healthcare, etc. -- can be considered a foundation model, Vaish said. Singh said fine-tuning an existing model like a Meta's Llama 3 will not be considered a foundation model. India may take nine to twelve months to build its own sovereign AI foundation model, he added. Also Read: AI mission takes off: Chip & model plans on board

Building Foundational Models Under IndiaAI Mission: MeitY clarifies

Disclaimer: This content generated by AI & may have errors or hallucinations. Edit before use. Read our Terms of use "The Government is currently building platforms to house datasets with access to some sources; however, dependent on specific datasets required, they would search for the same in government sources", IT Ministry's Additional Secretary Abhishek Singh said during a Q&A session on call for proposals for Foundational AI Models under the IndiaAI Mission. Other individuals and entities can access these datasets as well, once the foundational model has been built. However, if entities wish to procure proprietary datasets, they must outline the associated costs in the proposal. These responses addressed a query about sourcing training data for foundational models and the ownership rights, which remain with the developing entity. While the entity can monetise the model, the government can utilise the same for public use across various sectors. The government aims to develop licensing agreements with the entities developing the models for this purpose. This session followed Union IT Minister Ashwini Vaishnaw's call for proposals to create foundational AI models trained on Indian datasets. Besides the previous points, Singh, alongside IndiaAI Mission advisor Aakrit Vaish, clarified other doubts from applicants for the Foundational Models proposals. Here's what you need to know: Offering clarification on what constitutes a foundational model, Vaish explained that it includes any model whether speech or video, alongside the above-mentioned subsets. "Anything where the concepts of large amounts of training data are applied to be a model that can be used for generative purposes" is considered, he added. Further, the mission will consider ground-up-based models to solve specific vertical use cases like in healthcare. However, the government repeatedly emphasized that fine-tuning existing models will not be classified as developing a foundational model. Such an attempt (for example, building on an existing Llama 3) instead categorises as an application development initiative. Elaborating on developing foundational models on open-source architecture, Singh said that would depend on "how ingenious [one is]". Further, the proposal should address key questions, including the method for distillation, among others. To clarify, distillation is a process that involves transferring knowledge from a pre-trained AI model into a smaller, more deployable version. The term has been gaining ground recently, with OpenAI accusing rival DeepSeek of stealing its data, with many arguing that distillation could have facilitated this alleged theft. Moving forward, Singh reiterated that the mission has empanelled service providers furnishing compute, part of whose funding will be subsidised as included in the request for empanelment (RFE). Fielding another question on the IndiaAI datasets platform, Vaish reiterated that the dataset developed is akin to Hugging Face. India's platform, which is currently in the beta stage, would host several datasets from the governance and private sectors. The Government first announced this initiative back in October 2024 and the first two dataset collections focus on astrophysics and biological systems, among others. Hugging Face is an open-source machine learning (ML) platform that helps users train, build, and deploy AI models. Without delving into much detail, the Government also reiterated that entities must comply with all existing legislations, as previously stipulated by the rules. This was in response to a question about whether the applicant holds liability for reviewing datasets to ensure that they comply with the Digital Personal Data Protection (DPDP) Act, 2023, and other laws. While the proposal has certain leeway given the evolving nature of the demands, applicants must be as prescriptive as they can in detailing the problem and their approach. Further, while they must explain the actual tactical nature of their model, the IndiaAI Mission would hold applicants accountable for milestones instead of minute pieces. Aspects like costs for mining and cleaning datasets, salaries of involved personnel, and compute for training should be included in the proposal. In tune with the previous sentiments on the development of ingenious models, the government contended that the models must reflect the Indian cultural context and be "India-specific", alongside addressing multiple Indian languages, if possible.