Govt Offers 100% Subsidy under IndiaAI Mission to Start-ups Building Foundational AI Model

The Indian government is providing a 100% subsidy on compute-infrastructure costs for companies developing foundational AI models, Moneycontrol reported. This initiative aims to strengthen India’s AI sector.

It is critical to clarify that this 100 % subsidy applies solely to core model development. The existing 40 % subsidy supports other GPU-intensive operations, such as inferencing and applications. This distinction is significant, as it was previously assumed that the 40 % subsidy for GPU access also covered foundational-model training.

“This is not a 40 % scheme like others—foundational-model efforts are being given 100 % subsidy on compute,” said a person directly familiar with the development. “Other AI workloads, like inferencing or vertical applications, will fall under the 40 % bracket.”

In January 2025, the government announced a 40 % subsidy for GPU access through a coupon-based mechanism. To date, the IT ministry has allocated over 34,000 GPUs to support businesses and researchers.

The IndiaAI Mission has disbursed approximately ₹ 111.85 crore in subsidies under this GPU-allocation scheme to enhance AI computing access for start-ups, researchers and government entities.

Four companies, Sarvam, Gnan.AI, GAN.ai and Socket have been approved to develop foundational models. Sarvam, led by co-founder Pratyush Kumar, secured the largest single allocation, receiving 4,096 NVIDIA H100 SXM GPUs from Yotta Data Services, with a subsidy of ₹ 98.68 crore against a project cost of ₹ 246.71 crore. The other three start-ups are expected to receive a comparable GPU allocation.

GPU-Allocation Scheme

This underscores the government’s commitment to advancing AI development in India. However, some experts argue that the focus should shift towards inference and real-world AI applications rather than solely training large AI models. Inference involves using pre-trained models for tasks such as question answering, language translation and image recognition.

Such a shift would require GPUs to operate at full capacity, as cloud-service providers are designed for 100 % utilisation to ensure a return on investment within three years, after which hardware begins to depreciate.