Modi government offering 100% subsidy for foundational AI model development
23 Jun 2025

The Indian government is offering a 100% subsidy on the compute infrastructure costs for companies developing foundational AI models, according to Moneycontrol.

The move is aimed at boosting the country's artificial intelligence (AI) sector.

It is important to note that this subsidy is only for foundational model development. The existing 40% subsidy will support other GPU-intensive activities like inference and applications.

Clarification on earlier subsidy announcement
Subsidy details

Back in January 2025, the government had announced a coupon-based system to subsidize GPU access by up to 40%.

So far, over 34,000 GPUs have been empaneled by the Ministry of Electronics and Information Technology (MeitY) to facilitate compute access for start-ups and researchers.

The clarification is important as it was previously assumed that this 40% subsidy also applied to foundational model development.

Four start-ups approved for developing foundational models
Start-up approvals

So far, four start-ups have been approved by the government in order to develop foundational models.

These include Sarvam, Gnan. AI, GAN.ai and Socket.

Sarvam has already been given access to 4,096 GPUs, while the other three are expected to get a similar number of GPUs for their respective models.

Over ₹111.85 crore in subsidies distributed under GPU allocation scheme
Subsidy distribution

The IndiaAI Mission has already distributed over ₹111.85 crore in subsidies under its GPU allocation scheme.

The biggest single allocation of GPUs and subsidy was made to Sarvam AI's co-founder, Pratyush Kumar, who got 4,096 NVIDIA H100 SXM GPUs through Yotta Data Services.

This highlights the government's commitment to promoting AI development in the country.

Shift needed toward inferencing and real-world applications of AI
Expert opinions

Experts have suggested that the government's focus should not just be on training large AI models but also on inference and real-world applications of AI.

Inferencing is the use of pre-trained models to perform tasks like answering questions, translating languages, or recognizing images.

This shift would require GPUs to run at full capacity as cloud service providers (CSPs) are designed for 100% usage to ensure a return on investment within three years, after which the hardware starts losing value.