In an effort to strengthen the country’s AI infrastructure, India is considering a possible collaboration with the world’s largest chip manufacturer, NVIDIA, to purchase its GPUs and NPUs and subsidize their costs for use by researchers, academic institutions, local startups, and other users.
This is an early initiative that will come at a cost of approximately ₹10,000 crore for India. The Economic Times reported that the ultimate decision is anticipated to be made after the 2024 general elections are over.
With a sizable share of the GPU industry under its belt, NVIDIA has become the obvious choice for India’s AI compute infrastructure requirements. Around the world, nations and companies making investments in AI computer infrastructure are lined up to buy NVIDIA GPUs, primarily the H100 models.
India is investigating two strategies to procure and provide GPUs and NPUs with AI computation capabilities.
Specifically, a marketplace model and a “rent-and-sublet” model.
Under the rent-and-sublet concept, researchers, entrepreneurs, and other organizations will receive GPUs from the government at a discounted price.
Under the marketplace model, the government would incentivize businesses to enter into direct negotiations with Nvidia for leasing or subleasing agreements. Based on the additional productivity gained while utilizing the GPU, incentives would be given out; this is comparable to the methodology used in India’s PLI program.
Acquisition of GPUs has become a major hurdle for enterprises, especially startups, due to their scarcity and high cost. NVIDIA’s more current Blackwell cards retail for roughly $40,000, while its H100 GPUs often sell for roughly $50,000 per.
We are looking at at least 100 to 300 GPUs for the most basic AI data center, which is limited to executing current AI models. We are looking at 5000–10,000 GPUs for any significant processing power that may be applied to the design, development, and training of big language models.
For instance, it appears that OpenAI spent $36 billion and 720,000 H100 GPUs to develop and train Sora, an AI video producer. In a similar vein, Meta is operating with roughly 350,000 H100 units, valued at $17.5 billion. Meta intends to purchase more GPUs of this type.
To put things in perspective, the 640-GPU supercomputer known as “AIRAWAT” housed at the Centre for Development of Advanced Computing (C-DAC) in Pune, India, is the 75th fastest supercomputer in the world. To be competitive in AI research and development, India is working to close a sizable gap with the fastest supercomputers in the world, which have over 30,000 GPUs.
The Union Cabinet gave its approval in March for the Rs 10,372-crore India’s AI Mission, which aims to use public-private partnerships to install 10,000 GPUs. The government intends to rent-and-sublet GPUs to researchers, startups, and other eligible parties at a concession price.
As an alternative, a marketplace model—which is comparable to a production-linked incentive (PLI) scheme—is being explored, in which performance may be objectively assessed and rewards are allocated accordingly.
Globally, businesses and governmental organizations buy GPUs. But it is improbable that any nation will be given preference over another based only on the federal structure when such limited resources are allocated. In order to maximize its strategy, India is investigating a range of options and interacting with stakeholders.
With a market valuation of $2.16 trillion, Nvidia is now the third most valuable business in the world, behind Apple and Microsoft, thanks to the strong demand for GPUs. Reliance Group, Tata Group, and Yotta Infrastructure are among the Indian corporations who have already signed agreements with Nvidia to purchase GPUs for internal use.
Prominent CEOs and AI firms in India have been pushing for increased government investment in processing power to make their country more competitive in the global AI market.
India’s AI mission seeks to protect the sovereignty of Indian data in addition to promoting innovation. Owing to the scarcity of GPUs in the country, a lot of businesses are turning to foreign cloud resources.
