Indian universities face a GPU shortage as AI research demand grows. Institutions now explore sharing infrastructure to meet these computing needs. Prof. Ram Kumar Kakani said, "There is no point in quarrelling over technology." Many colleges also build dedicated facilities or partner with tech companies to get better access.

New Delhi: A shortage of affordable computing power is pushing Indian universities to explore sharing GPU (graphics processing unit) infrastructure, as growing demand from artificial intelligence (AI) research makes it harder for individual institutions to meet their computing needs.

GPUs are specialised chips that can process large amounts of data simultaneously. They are widely used in AI, machine learning, graphics processing and scientific computing. As research involving large language models (LLMs), particularly agentic AI applications, gathers pace, universities are finding it increasingly difficult to secure affordable and reliable access to the computing capacity they need.

Institutions are trying several routes: sharing infrastructure, building dedicated facilities and partnering with technology companies. The efforts underline how access to advanced chips and computing capacity has become a key constraint for universities looking to expand their AI work.

RV University (RVU), for instance, is exploring the possibility of sharing computing infrastructure with neighbouring institutions, including S-VYASA University and R.V. College of Engineering in Bengaluru. IIT Kanpur, meanwhile, is planning a ₹200-crore high-performance computing facility equipped with advanced GPUs to support AI research and scientific simulations.

"During the late hours, when our campus is closed, it is a shared resource. There is no point in quarrelling over technology or who carries the net exercise. The good part is that we are in a zone where we have at least three to four universities around us, and that is where the sharing happens," said Prof. Ram Kumar Kakani, vice-chancellor of RV University.

Cochin University of Science and Technology (Cusat) has opened its TEJASWI high-performance computing facility, equipped with 24 Nvidia A100 GPUs and two petabytes of storage, to researchers from other academic institutions. They can access its computing resources on a chargeable basis.

Beyond sharing infrastructure and building their own computing facilities, universities are also exploring partnerships with technology companies to access AI models and computing resources. Manipal Institute of Technology Bengaluru, for instance, has partnerships with technology companies that provide academic licences and access to AI models and computing infrastructure, either free of cost or at subsidised rates. IIT Kanpur is also exploring partnerships with global suppliers to strengthen its computing capabilities.

Shared infrastructure

The push for shared infrastructure comes amid questions over whether existing government-backed computing initiatives adequately address universities' requirements.

Under the IndiaAI Mission, the government has established a common compute facility through which researchers, academic institutions, students and startups can access GPUs at subsidised rates. Universities, however, say they need more predictable access to computing capacity, particularly for research projects that require sustained computing resources.

"Academia's share in GPU consumption and AI compute subsidies has been very small. One way to address this would be to bundle computing resources with research grants. For instance, every project sanctioned by the Anusandhan National Research Foundation (ANRF) could come with a specified allocation of GPU hours," Kakani said.

ISCost of computing

At a broader level, experts argue that a government-enabled framework could help make computing infrastructure more widely accessible. This should extend beyond large institutions such as the Indian Institute of Science (IISc) and the Indian Institutes of Technology (IITs) to colleges in tier-II and tier-III cities, they said.

"It will increase entrepreneurship. If students from tier-II and tier-III colleges and tier-II and tier-III cities also access some of this, then we should have a framework to enable that," said Omprakash Subbarao, chief executive officer of the Foundation for Science, Innovation and Development (FSID) at IISc.

The cost of GPUs is another major consideration for universities looking to build their own computing infrastructure. An Nvidia H100 80GB GPU, used for demanding AI workloads such as training large language models, can cost between ₹25.5 lakh and ₹50 lakh per unit in India. An Nvidia A100 80GB GPU, commonly used for AI model training and scientific computing, can cost between ₹20 lakh and ₹29.6 lakh, depending on the configuration and seller.

Experts warn that inadequate access to computing infrastructure could constrain the scope and quality of research, particularly in fields involving large datasets and extensive model training. It could also force researchers to scale back work on areas such as fine-tuning smaller AI models and developing applications for Indian languages, agriculture and healthcare.

Finding alternatives

Universities are also turning to technology companies for access to AI models and computing resources. Several companies offer free or subsidised academic access, allowing students and researchers to work on projects that require substantial computing power.

"Since we have a lot of students working on various projects, and they also have compute requirements, these companies give us access to some of those models and the compute infrastructure. Sometimes it is free, and sometimes it is at a subsidised academic cost," said Subbarao.

However, access to commercial AI services may not meet every research requirement, particularly when institutions want to train their own models rather than use existing ones. This has prompted some universities to consider investments in their own computing infrastructure.

IIT Kanpur, for instance, has begun discussions with domain experts, including alumni based in the Bay Area and Silicon Valley, to explore partnerships with global suppliers. It has also launched a $1-billion fundraising campaign, Greater Than Us, with a March 2029 deadline, and has secured commitments of approximately ₹4,155 crore so far.

The campaign includes a proposed ₹200-crore high-performance computing facility equipped with advanced GPUs to support AI research and scientific simulations.

Meanwhile, MIT Bengaluru, part of Manipal Academy of Higher Education (MAHE), already has eight GPUs shared among students and researchers for AI-related work. The institution plans to improve access to its existing computing resources and expand capacity to support more demanding research.

"The roadmap is to provide centralized scheduling, standard AI environments, a usage dashboard, researcher onboarding and a model and data repository as near-term priorities. Longer term, it proposes a larger cluster and an MAHE AI Compute Centre," said Dr. Dayanand, dean of the school of computer engineering at MIT Bengaluru, MAHE.