India’s answer to this challenge is the IndiaAI Mission, under which the government has created a shared national compute infrastructure and is offering subsidised access to GPUs for startups, researchers, educational institutions and other eligible users.
The programme – backed by an outlay of more than ₹10,000 crore – has expanded beyond its initial target of 10,000 GPUs, with over 18,000 GPUs empanelled in the early phases and more than 38,000 GPUs now available through the broader compute ecosystem at subsidised rates.
Compute access is available at around ₹65 per GPU hour, which is considerably below the international market rates.
The initiative focuses to reduce one of the largest barriers to AI development. Yet conversations with startup executives, researchers, cloud experts and infrastructure providers suggest that while subsidised compute is an important step, it is unlikely to close India’s AI capability gap on its own.
Who gets access, and should everyone be treated equally?
However, questions remain over how scarce high-end compute should be distributed.
Omprakash Subbarao, chief executive of FSID CORE, the R&D arm of the Indian Institute of Science (IISc), argued that a purely first-come-first-served approach could favour institutions that already possess proposal-writing and administrative capabilities.
“If I were to build a tiered system, I would start with researchers conducting work in multilingual NLP and safety to receive baseline allocations that require minimal friction to access,” he told Business Standard.
Subbarao suggested that startups creating products with public value should receive initial allocations and qualify for additional compute based on demonstrated outcomes.
Hari Balaji, partner, technology consulting at EY India, also advocated concentration of resources rather than broad distribution. “India should concentrate our subsidised GPUs with a few deserving players because scale compounds disproportionately in the world of frontier models,” he said, pointing to frontier AI startups, academic consortia and research-led laboratories.
Training models or powering applications?
Another point to factor in is how Indian companies are using compute resources. For many startups, the objective is not to build a foundational large language model (LLM) from scratch. Instead, they are adapting existing open-source models for commercial applications.
Dr Kanishk Agrawal, chief technology officer at Judge Group India, an IT solutions company, said his organisation primarily uses GPUs for fine-tuning and inference rather than training foundation models. Building an LLM from the ground up requires enormous compute resources and high-quality data, making it impractical for many startups, he said.
By contrast, fine-tuning existing models allows companies to create domain-specific AI systems for areas such as governance, financial services, customer support and multilingual applications.
According to Abhijeet Kate, co-founder and principal consultant at digiCloud Solutions, training still accounts for around 80 per cent of total GPU hours consumed, but demand is increasingly shifting towards inference.
“More than 60 per cent of startups requesting access want GPUs for inference, and that share is growing fast,” he said.
Kate described a three-tier market. Roughly 60 per cent of demand comes from inference-focused startups running chatbots and APIs. Another quarter involves fine-tuning workloads, while only a small segment comprises companies training large models over extended periods.
Chaitra Vedullapalli, an AI expert and president of Women in Cloud, said demand is strongest in Indic-language models, enterprise copilots and sector-specific applications in healthcare, finance, agriculture, education, governance and cybersecurity.
Can subsidised compute compete with global cloud providers?
Subbarao said the ₹65 per GPU-hour pricing made available through competitive bidding has provided researchers with a viable alternative to cloud services priced in dollars. However, he cautioned that affordability alone does not guarantee competitiveness.
Kate added that the voucher system typically covers 40-60 per cent of compute costs, with users paying the remainder. According to him, the programme works particularly well for academic institutions, government projects bound by data-sovereignty requirements and developers building Indic-language models.
Experts, however, drew a distinction between cheaper hardware access and a complete AI ecosystem.
“Subsidised computer processing power is an essential element; however, it is not enough on its own to compete with the capabilities provided by large, global hyperscalers,” Agrawal said.
Global cloud providers combine compute credits with developer tools, managed services, deployment infrastructure and technical support. India will need to build similar ecosystem capabilities if domestic AI companies are to compete globally, he added.
The implementation challenge
Experts pointed to operational issues that could determine whether the programme succeeds.
Kate said short contract cycles and pricing uncertainty remain major concerns for startups. Companies selling AI services often commit to fixed customer pricing over long periods, making sudden increases in compute costs difficult to absorb.
“For fine-tuning runs, the seven-day lease cap creates real operational risk,” he said, adding that founders are seeking predictable access rather than simply cheaper compute.
Subbarao highlighted broader structural concerns as he pointed to budget volatility, the need for multi-year compute commitments for foundation-model developers and an uneven hardware mix in which a major portion of available GPUs are more suitable for inference than frontier model training.
He also cited past downtime incidents at AIRAWAT, India’s supercomputing infrastructure for AI, arguing that reliable service-level agreements are as important as hardware procurement.
“Building compute is the easy part; running it reliably with SLAs researchers can actually plan around is what the Mission has not yet solved completely,” Subbarao said.
The infrastructure beneath the GPUs
Even if compute becomes widely available, India’s AI ambitions will depend on whether supporting infrastructure can keep pace.
Rajesh Chhabra, general manager, APAC large markets at Acronis, said AI-focused data centres require significantly higher power capacity and more sophisticated cooling systems than traditional facilities.
He said large AI clusters will require liquid or hybrid cooling technologies, high-density racks, resilient power systems, low-latency networking and scalable storage architectures. Investments in power grids, renewable energy and operational resilience will be as critical as GPU availability.
According to Chhabra, cities such as Mumbai, Chennai, Hyderabad and Bengaluru already possess many of the characteristics needed for AI infrastructure hubs, including connectivity, talent pools and policy support.
However, he warned that power availability and grid reliability could become constraints comparable to GPU shortages.
Beyond compute: The data and talent question
Perhaps the strongest message emerging from the interviews is that compute alone will not create globally competitive AI systems.
Agrawal identified access to high-quality, domain-specific datasets and experienced AI talent as larger long-term challenges. Building reliable AI systems requires clean, diverse and well-governed datasets, as well as engineers capable of training and deploying large-scale models.
Subbarao raised a similar concern. While nearly half of the IndiaAI Mission’s budget is directed towards compute infrastructure, only a small portion is earmarked for datasets, he noted.
“50,000 H100s will not save you if you have a thin training corpus for Indic language,” he said.
He also argued that India faces a shortage of frontier AI researchers working on new architectures, data curation and model interpretability, a gap that cannot be solved through broad-based AI literacy programmes alone.
Can subsidised GPUs close the gap?
The IndiaAI Mission has lowered the cost of accessing advanced hardware and expanded compute availability for startups and researchers. It has also created a national platform that could support public-interest AI, academic research and indigenous model development.
Yet, experts point to a longer list of requirements: predictable access, reliable infrastructure, datasets, skilled researchers, power capacity, cooling systems, cybersecurity and governance frameworks.
As Balaji of EY put it, “Procuring more chips is a small part of the answer.”
For India, the race to build sovereign AI capability may ultimately depend less on the number of GPUs it acquires and more on whether it can build the broader ecosystem needed to put those GPUs to productive use.
