The fourth round of the graphics processing units (GPU) tender for the IndiaAI Mission is underway, and at least five new compute providers have expressed interest during the pre-bid process, the mission’s chief executive officer, Abhishek Singh, told ET.
The government-led IndiaAI Mission aims to deploy 38,183 GPUs (graphics processing units) of compute capacity and develop Indian-language foundation models.
As per the IndiaAI compute portal, 21,986 GPUs (57.58%) had been allocated AI workloads by Wednesday. Also, a total subsidy of Rs 760.93 crore has been allocated so far.
According to Singh, the portal’s allocation rate will increase, with many more GPUs being added via upcoming rounds.
Around 8,000 B200 GPUs are expected in January, sources said, adding that the existing providers, such as E2E and Neysa, are augmenting their fleets.
E2E has placed an order for 4,000 Blackwell GPUs, sources said.
Sources added that Sarvam’s six-month GPU allocation is ending and will be reallocated, and compute availability will continue to rise.
IndiaAI Mission’s tenders remain platform-agnostic, Singh said, allowing providers such as Cerebras and Sambanova to offer GPUs, wafer-scale solutions, or other compute platforms. The mission emphasises user choice and broad eligibility for all established technologies, creating customisation in AI training and inference, he said.
Singh disclosed that apart from the 12 already announced companies building foundation models, five more are in the final selection stage, pending compute resource security.
"We have done the third round for foundation models. Five more are in the final stage of the selection. Right now, we have announced 12. Apart from 12, we have five more, but we are just trying to secure compute before we announce. Otherwise, the expectations go up," he said.
The mission’s goal is to focus on both high-end and regular GPUs for training and inference, with inferencing needs set to rise across sectors.
"We have continued with the existing model of the tender in which we are focusing on both high-end and regular GPUs, for both training and inference," said Singh. “In fact, our inferencing needs will go up as we go forward. Not only for Bhashini, even after Sarvam builds up its model, the actual use cases of that model, whether we are giving, say, agri-advisory services or healthcare services, or various use cases, that will definitely go up.”
Responding to skepticism about the international significance of Indian foundation models, especially those trained on local languages, Singh invoked the example of instant payment system the Unified Payments Interface (UPI)’s global impact despite initial doubts. He stressed that as long as AI models solve India’s problems, they will be valuable domestically and potentially globally, given India’s billion-plus population.
Sources said that for norms, such as the 10% threshold for AI-generated content labeling, feedback from industry is driving ongoing discussions. They said labeling is a standard regulatory practice not expected to hinder benign uses of AI, and that risk-based classification is being considered for future regulation adjustments.
"The cyber law division has got a lot of good feedback from industry and everyone. Ultimately, labeling is required. Whether it's 10% or 5% is a matter of detail," sources said. “All kinds of labeling happens. When you go to watch a movie, you get that disclaimer that animals were not killed in this movie. So, similarly, this is something that's not too big a problem. Benign uses of AI is something the committee will be looking into whether we need a risk-based classification or not."
Also Read: Eight local firms may get IndiaAI foundational AI models incentives
The government-led IndiaAI Mission aims to deploy 38,183 GPUs (graphics processing units) of compute capacity and develop Indian-language foundation models.
As per the IndiaAI compute portal, 21,986 GPUs (57.58%) had been allocated AI workloads by Wednesday. Also, a total subsidy of Rs 760.93 crore has been allocated so far.
According to Singh, the portal’s allocation rate will increase, with many more GPUs being added via upcoming rounds.
Around 8,000 B200 GPUs are expected in January, sources said, adding that the existing providers, such as E2E and Neysa, are augmenting their fleets.
E2E has placed an order for 4,000 Blackwell GPUs, sources said.
Sources added that Sarvam’s six-month GPU allocation is ending and will be reallocated, and compute availability will continue to rise.
IndiaAI Mission’s tenders remain platform-agnostic, Singh said, allowing providers such as Cerebras and Sambanova to offer GPUs, wafer-scale solutions, or other compute platforms. The mission emphasises user choice and broad eligibility for all established technologies, creating customisation in AI training and inference, he said.
Singh disclosed that apart from the 12 already announced companies building foundation models, five more are in the final selection stage, pending compute resource security.
"We have done the third round for foundation models. Five more are in the final stage of the selection. Right now, we have announced 12. Apart from 12, we have five more, but we are just trying to secure compute before we announce. Otherwise, the expectations go up," he said.
The mission’s goal is to focus on both high-end and regular GPUs for training and inference, with inferencing needs set to rise across sectors.
"We have continued with the existing model of the tender in which we are focusing on both high-end and regular GPUs, for both training and inference," said Singh. “In fact, our inferencing needs will go up as we go forward. Not only for Bhashini, even after Sarvam builds up its model, the actual use cases of that model, whether we are giving, say, agri-advisory services or healthcare services, or various use cases, that will definitely go up.”
Responding to skepticism about the international significance of Indian foundation models, especially those trained on local languages, Singh invoked the example of instant payment system the Unified Payments Interface (UPI)’s global impact despite initial doubts. He stressed that as long as AI models solve India’s problems, they will be valuable domestically and potentially globally, given India’s billion-plus population.
Sources said that for norms, such as the 10% threshold for AI-generated content labeling, feedback from industry is driving ongoing discussions. They said labeling is a standard regulatory practice not expected to hinder benign uses of AI, and that risk-based classification is being considered for future regulation adjustments.
"The cyber law division has got a lot of good feedback from industry and everyone. Ultimately, labeling is required. Whether it's 10% or 5% is a matter of detail," sources said. “All kinds of labeling happens. When you go to watch a movie, you get that disclaimer that animals were not killed in this movie. So, similarly, this is something that's not too big a problem. Benign uses of AI is something the committee will be looking into whether we need a risk-based classification or not."
Also Read: Eight local firms may get IndiaAI foundational AI models incentives
