India is considering a potential collaboration with chip-making giant NVIDIA to acquire its graphics processing units (GPUs) and Neural Processing Units (NPUs) at subsidised rates. These would be provided to local startups, researchers, academic institutions, and other users, aiming to enhance the AI infrastructure in India.

India is projected to incur approximately ₹10,000 crore for this initiative, which is currently in its preliminary phases. A decision is anticipated after the conclusion of the general elections of 2024, as reported by The Economic Times.

NVIDIA commands a considerable share of the GPUs market, positioning itself as the preferred option for India's AI compute infrastructure requirements. Globally, nations and enterprises investing in AI computing infrastructure are prioritising the acquisition of NVIDIA GPUs, particularly the H100 chips.

India is investigating two approaches to procure and supply AI compute-capable NPUs and GPUs. These include a "rent-and-sublet" model and a marketplace model.

In the rent-and-sublet model, the government will provide subsidised GPUs to startups and researchers. In the marketplace model, companies will negotiate directly with Nvidia for rental or subletting deals, with incentives tied to increased productivity, similar to India's PLI scheme.

The scarcity and high cost of GPUs pose significant challenges for companies, particularly startups. NVIDIA's H100 GPUs typically command a price of around $50,000 each, while its newer Blackwell cards are priced at approximately $40,000.

Even for basic AI data centres capable of running existing AI models, a minimum of 100 to 300 GPUs is required. For meaningful computing capabilities suitable for creating, developing, and training large language models, the demand rises to 5000-10,000 GPUs.

For instance, OpenAI reportedly utilised 720,000 H100 GPUs valued at $36 billion to develop and train Sora, its AI video generator. Similarly, Meta is currently employing around 350,000 H100 units valued at $17.5 billion, with plans to acquire additional GPUs.

India's supercomputer 'AIRAWAT' at the Centre for Development of Advanced Computing (C-DAC) in Pune ranks 75th globally, with 640 GPUs. However, compared to the world's fastest supercomputers, which have over 30,000 GPUs, India aims to close the gap to stay competitive in AI research and development.

Alternatively, a marketplace model, similar to a production-linked incentive (PLI) scheme, is under consideration.

AI startups and prominent CEOs in India are urging government investment in computing capacity to bolster the nation's competitiveness in the global AI landscape.

India's AI mission prioritises both fostering innovation and safeguarding the sovereignty of Indian data. The limited availability of GPUs domestically has compelled many companies to rely on accessing cloud resources located abroad.

Image source: FirstPost