Indian Universities Face Critical Shortage of AI Computing Resources

Indian Universities Desperately Need More AI Compute

The AI compute capabilities of Indian universities are not in great shape for providing assistance for AI research. IIT alumni should definitely take charge to donate to their universities for funding AI research and building better AI compute infrastructure. In one such recent case, Krishna Chivkula, an alumnus of IIT Madras donated INR 228 crore to the university, which is the largest donation ever. Another great example is Nandan Nilekani's investment in building AI4Bharat at IIT Madras, which is enabling the best AI research currently in the country. Meanwhile, the IITs and IISc computing infrastructure is abysmal. If more money goes toward acquiring compute, then the research from the institutions can definitely see a rise. A lot of researchers often head away from the country as there is no way foundational research can be done using such little compute availability. Hosting models on GPUs accessible via cloud hosted APIs is not cheap. Ask researchers from Indian universities and they will tell you. For example, 8 NVIDIA A100s on Microsoft Azure cost close to $20k per month, which is around INR 17 lakh per month. Given the amount of funding these universities receive, half of it would just be spent on buying compute for research. Prime Minister Narendra Modi recently led the inaugural governing board meeting of Anusandhan National Research Foundation (ANRF), which is the future of Indian research, bridging the gap between academia and industry, pushing research to production. "This is the best time for research and innovation in India," said Modi. While the government is taking steps in the right direction, the compute shortage still needs to be addressed. Few months back, the government announced that the country is planning to procure 10,000 GPUs within the next 18 months, with an investment of INR 10,300 crore. The process seems to be still underway, which is also in line with the predictions of $5.1 billion spent on AI infrastructure by 2027. And when it comes to government institutes, the case is even worse. "IIT profs have to beg the government for compute whereas private institutions are buying H100 nodes by the dozen because the latter have money," said Raj Dabre, a prominent researcher at NICT in Kyoto, adjunct faculty at IIT Madras and a visiting professor at IIT Bombay. "Dozens of H100s still doesn't cut it. We need 100s," said Rahul Madhavan, PhD candidate in theoretical ML at IISc. IIT-H had partnered with NVIDIA a year back to procure three NVIDIA DGX-1TM systems and two NVIDIA DGX-2TM, which according to researchers are now also used by other IITs. But in most cases in many universities, there are not even research grade NVIDIA GPUs. Some are making it work with barely consumer grade GPUs and CPUs from NVIDIA and Intel. According to guesstimates, India's largest AI computer infrastructure is currently being built by C-DAC, which has around 656 GPUs, which is currently being used at IIT Kharagpur. Even if India secures the computing infrastructure, allocation and provisioning of it would be a task in itself, which would definitely hinder the quality of research coming out of the country. Allocating most of them to universities would not be beneficial for the government directly and they would benefit from giving it to startups and enterprises. It is undeniable that Indian researchers have the talent and potential to do groundbreaking research, but there needs to be more funding and support provided to the institutions to bring in the sovereign AI revolution. But the government has to look at this potential with faith in research to reap benefits in the long run. At the same time, the task for building the Nalanda 2.0, also called Ekagrid, which would have been the best AI institute in the country, is also now shut because they could not get enough attention from the government and investors. Few months back, Fei-Fei Li concerningly revealed that Stanford's Natural Language Computing lab only has 64 GPUs and the academia is "falling off a cliff" relative to industry. Even UC Berkeley had the same issue with GPUs that can be counted on two hands. Compare it with giants like Meta, OpenAI, and Google, who have tens of thousands of GPUs for research. Most recently, Oracle's OCI said that it is offering up to 131,072 NVIDIA B200 GPUs, which are far better than NVIDIA H100s. Meanwhile, Yotta is also building its AI supercomputer in partnership with the Telangana government in Hyderabad with 25,000 H100s, but currently only 4,000 have been received. The purpose would probably serve enterprises for their AI compute, and not universities, though the latter is desperately needed. Ola Krutrim's plans for building AI chips for India are underway but they are also slated to release in 2026, which would be a massive setback of 2 years or more for Indian research.

IIT Alumni Should Start Donating GPUs to Power AI Research

The AI compute capabilities of Indian universities are not in great shape for providing assistance for AI research. IIT alumni should definitely take charge to donate to their universities for funding AI research and building better AI compute infrastructure. In one such recent case, Krishna Chivkula, an alumnus of IIT Madras donated INR 228 crore to the university, which is the largest donation ever. Another great example is Nandan Nilekani's investment in building AI4Bharat at IIT Madras, which is enabling the best AI research currently in the country. Meanwhile, the IITs and IISc computing infrastructure is abysmal. If more money goes toward acquiring compute, then the research from the institutions can definitely see a rise. A lot of researchers often head away from the country as there is no way foundational research can be done using such little compute availability. Hosting models on GPUs accessible via cloud hosted APIs is not cheap. Ask researchers from Indian universities and they will tell you. For example, 8 NVIDIA A100s on Microsoft Azure cost close to $20k per month, which is around INR 17 lakh per month. Given the amount of funding these universities receive, half of it would just be spent on buying compute for research. Prime Minister Narendra Modi recently led the inaugural governing board meeting of Anusandhan National Research Foundation (ANRF), which is the future of Indian research, bridging the gap between academia and industry, pushing research to production. "This is the best time for research and innovation in India," said Modi. While the government is taking steps in the right direction, the compute shortage still needs to be addressed. Few months back, the government announced that the country is planning to procure 10,000 GPUs within the next 18 months, with an investment of INR 10,300 crore. The process seems to be still underway, which is also in line with the predictions of $5.1 billion spent on AI infrastructure by 2027. And when it comes to government institutes, the case is even worse. "IIT profs have to beg the government for compute whereas private institutions are buying H100 nodes by the dozen because the latter have money," said Raj Dabre, a prominent researcher at NICT in Kyoto, adjunct faculty at IIT Madras and a visiting professor at IIT Bombay. "Dozens of H100s still doesn't cut it. We need 100s," said Rahul Madhavan, PhD candidate in theoretical ML at IISc. IIT-H had partnered with NVIDIA a year back to procure three NVIDIA DGX-1TM systems and two NVIDIA DGX-2TM, which according to researchers are now also used by other IITs. But in most cases in many universities, there are not even research grade NVIDIA GPUs. Some are making it work with barely consumer grade GPUs and CPUs from NVIDIA and Intel. According to guesstimates, India's largest AI computer infrastructure is currently being built by C-DAC, which has around 656 GPUs, which is currently being used at IIT Kharagpur. Even if India secures the computing infrastructure, allocation and provisioning of it would be a task in itself, which would definitely hinder the quality of research coming out of the country. Allocating most of them to universities would not be beneficial for the government directly and they would benefit from giving it to startups and enterprises. It is undeniable that Indian researchers have the talent and potential to do groundbreaking research, but there needs to be more funding and support provided to the institutions to bring in the sovereign AI revolution. But the government has to look at this potential with faith in research to reap benefits in the long run. At the same time, the task for building the Nalanda 2.0, also called Ekagrid, which would have been the best AI institute in the country, is also now shut because they could not get enough attention from the government and investors. Few months back, Fei-Fei Li concerningly revealed that Stanford's Natural Language Computing lab only has 64 GPUs and the academia is "falling off a cliff" relative to industry. Even UC Berkeley had the same issue with GPUs that can be counted on two hands. Compare it with giants like Meta, OpenAI, and Google, who have tens of thousands of GPUs for research. Most recently, Oracle's OCI said that it is offering up to 131,072 NVIDIA B200 GPUs, which are far better than NVIDIA H100s. Meanwhile, Yotta is also building its AI supercomputer in partnership with the Telangana government in Hyderabad with 25,000 H100s, but currently only 4,000 have been received. The purpose would probably serve enterprises for their AI compute, and not universities, though the latter is desperately needed. Ola Krutrim's plans for building AI chips for India are underway but they are also slated to release in 2026, which would be a massive setback of 2 years or more for Indian research.