India will offer the cheapest compute in the world at less than $1 per hour for high-end chips that power generative artificial intelligence (GenAI) as the government's ₹10,000 crore IndiaAI Mission comes into play from Thursday, said Ashwini Vaishnaw, minister for electronics and IT.
The government will also incentivise the development of local language models built by academia and industry with investment capital and other support, Vaishnaw said. The move is aimed at building up Indian language foundational model muscle.
Proposals for model development were invited on Thursday. At least six startups and developers that can do it within the next 10 months have been identified, he said. ET was first to report on January 23 that India will back indigenous foundational models.
“The real value will come from two things—algorithmic efficiency and the quality of training datasets,” Vaishnaw said, adding that Chinese AI company DeepSeek has proven to the world that a cost-efficient model can be developed. “DeepSeek was trained on 2,000 GPUs (graphics processing units),” Vaishnaw said. “We have now 15,000 high-end GPUs. (OpenAI’s) ChatGPT version 1 was trained on about 25,000 GPUs. So this gives us a huge compute facility, something which will really give a boost to our ecosystem.” GPUs are high-capacity chips needed to run complex AI development tasks. Vaishnaw said India is not late to the AI party and will play a big role in the innovation taking place in the field globally. Since India has now incentivised compute, the models will follow, he said.
The rates for common compute in India are “phenomenally competitive”, he said. The government will provide a 40% subsidy to those accessing GPUs. It will bring down the cost to below $1 per hour for GPU access, one-third of the global average of $2.5-3.
Under the IndiaAI Mission, commitments have been received for 18,693 GPUs, out of which 15,000 are high-end GPUs and 10,000 of those are available (with the bidders) as of today, said Vaishnaw. “We put huge thrust on getting the common compute facility developed,” the minister said. “This gives us a huge advantage vis-a-vis many other countries. Because now we will have this compute facility available for so many innovative and new architectures that people would like to create and test.”
The frontrunners in GPU capacity empanelment include Jio Platforms, E2E Networks, NxtGen Datacenter and Cloud Technologies, Locuz Enterprise Solutions and CtrlS Datacenters out of a total 10 companies offering 18,693 GPUs. Locuz is a partner of hyperscaler Amazon Web Services (AWS).
The average bid price during the commercial round was ₹115.85 ($1.34) per GPU hour. Of this, 40% cost will be subsidised by the government over the next three years with a total budget allocation of ₹4,563.36 crore. Further, ₹689 crore has been earmarked for AI application development. “The second big mission of the IndiaAI mission was to develop an AI model,” Vaishnaw said. “We have now created the framework which will be launched today. We are calling for proposals to develop our own foundational models, where the Indian context, the Indian languages, the culture of our country (are considered), where the biases can be removed… where the datasets, for our country, for our citizens… that process will start today.”
The minister expressed confidence in India's ability to develop such models.
“We believe that there are at least six major developers who can develop AI models in the next six to eight months on the outer limit, and four to six months on a more optimistic estimate,” he said.
Vaishnaw said DeepSeek’s AI model will be hosted on Indian servers.
“Good thing is that DeepSeek is an open-source model and we are very soon going to host DeepSeek on Indian servers, the way we have hosted (Meta’s) Llama, so that data privacy parameters can be addressed,” he said. “Already, our team has worked out the details of how much capacity, how many servers are required.” Vaishnaw addressed past scepticism about India's capacity to develop foundational models. “The innovations that are happening in the world are humongous,” he said.
Budget with ET
Fraught juggling act draws near for Nirmala Sitharaman. What will Feb 1 bring?
Budget 2025 needs to pull off a big growth trick for the untapped half of its firepower
2 key events coming up for India as world readjusts to Trump 2.0
Proposals for model development were invited on Thursday. At least six startups and developers that can do it within the next 10 months have been identified, he said. ET was first to report on January 23 that India will back indigenous foundational models.
“The real value will come from two things—algorithmic efficiency and the quality of training datasets,” Vaishnaw said, adding that Chinese AI company DeepSeek has proven to the world that a cost-efficient model can be developed. “DeepSeek was trained on 2,000 GPUs (graphics processing units),” Vaishnaw said. “We have now 15,000 high-end GPUs. (OpenAI’s) ChatGPT version 1 was trained on about 25,000 GPUs. So this gives us a huge compute facility, something which will really give a boost to our ecosystem.” GPUs are high-capacity chips needed to run complex AI development tasks. Vaishnaw said India is not late to the AI party and will play a big role in the innovation taking place in the field globally. Since India has now incentivised compute, the models will follow, he said.
The rates for common compute in India are “phenomenally competitive”, he said. The government will provide a 40% subsidy to those accessing GPUs. It will bring down the cost to below $1 per hour for GPU access, one-third of the global average of $2.5-3.
Under the IndiaAI Mission, commitments have been received for 18,693 GPUs, out of which 15,000 are high-end GPUs and 10,000 of those are available (with the bidders) as of today, said Vaishnaw. “We put huge thrust on getting the common compute facility developed,” the minister said. “This gives us a huge advantage vis-a-vis many other countries. Because now we will have this compute facility available for so many innovative and new architectures that people would like to create and test.”
The frontrunners in GPU capacity empanelment include Jio Platforms, E2E Networks, NxtGen Datacenter and Cloud Technologies, Locuz Enterprise Solutions and CtrlS Datacenters out of a total 10 companies offering 18,693 GPUs. Locuz is a partner of hyperscaler Amazon Web Services (AWS).
The average bid price during the commercial round was ₹115.85 ($1.34) per GPU hour. Of this, 40% cost will be subsidised by the government over the next three years with a total budget allocation of ₹4,563.36 crore. Further, ₹689 crore has been earmarked for AI application development. “The second big mission of the IndiaAI mission was to develop an AI model,” Vaishnaw said. “We have now created the framework which will be launched today. We are calling for proposals to develop our own foundational models, where the Indian context, the Indian languages, the culture of our country (are considered), where the biases can be removed… where the datasets, for our country, for our citizens… that process will start today.”
The minister expressed confidence in India's ability to develop such models.
“We believe that there are at least six major developers who can develop AI models in the next six to eight months on the outer limit, and four to six months on a more optimistic estimate,” he said.
Vaishnaw said DeepSeek’s AI model will be hosted on Indian servers.
“Good thing is that DeepSeek is an open-source model and we are very soon going to host DeepSeek on Indian servers, the way we have hosted (Meta’s) Llama, so that data privacy parameters can be addressed,” he said. “Already, our team has worked out the details of how much capacity, how many servers are required.” Vaishnaw addressed past scepticism about India's capacity to develop foundational models. “The innovations that are happening in the world are humongous,” he said.
