As the new government settles down, it is imperative that it takes this mission up as one of its key priorities. (File Photo)
It was Vladimir Putin who famously put AI on every country’s priority list when he declared back in 2017 that the nation that leads in AI “will be the ruler of the world”. Every world leader has echoed this in some way. For China, AI is a national priority, the US had an executive order by President Joe Biden on AI, the UK convened a global summit on AI at the iconic Bletchley Park, and both the G7 and G20 have the development and safety of AI among their top agenda items.
To its credit, India was an early mover in AI. The Niti Aayog published the National Strategy for Artificial Intelligence in June 2018. India’s strategy focused on “AI for All” and set out healthcare, education, agriculture, smart cities and smart mobility as the five priority sectors. It then recommended some clear action points for the government, some of which are a part of the IndiaAI Mission, which the government followed up with in March 2024. As the new government settles down, it is imperative that it takes this mission up as one of its key priorities. With the rapid proliferation of generative AI and newer and more powerful Large Language Models (LLMs) being introduced thick and fast, countries and corporations across the world are racing to keep pace with the rapid-fire advances and jockey for leadership positions. India needs to be a frontrunner in that race.
Towards that, let us look critically at the AI mission and its various pillars. The Cabinet has allocated Rs 10,372 crore (approximately $1.3bn) towards this initiative. It has a few key pillars: Building dataset platforms, innovation and application development centres, and future skills, along with startup financing and focusing on safe and trusted AI. However, the central pillar of this mission, with nearly half the sum (Rs 4,568 crores) earmarked for it, is to build cutting edge compute capacity for the country. The objective is to scale up compute capacity for local demand, and, importantly, bridge the “AI divide” by offering inexpensive compute for the prioritised sectors.
Since this seems to be the biggest focus and spend area of the mission, it deserves a more detailed examination. The official statement by the Ministry of Electronics and Information Technology declares that “a cornerstone of this effort is the India AI Compute Capacity, envisioned to erect a cutting-edge, scalable AI computing infrastructure by deploying over 10,000 Graphics Processing Units (GPUs) through strategic public-private collaborations”. While this is a creditable objective, it would be pertinent to examine the manner in which this will be done, and whether that meets India’s longer term objectives.
At first glance, the number itself seems very low. To put it in context, a private company, Meta, will have 60 times more GPUs than this, and even small AI startups globally have more GPUs. That is because LLMs and GenAI need massive amount of compute capacity for training and inference. As an example, it took OpenAI 3,640 petaflops of compute to train GPT3, and 10,000 GPUs translate to just around 25 petaflops. And GPT 3 was six years and three generations of models back. Secondly, it would be important to understand the governance of the same. While India has some stellar examples of national bodies building capacity and shining, like ISRO or BARC, mostly, creating a central repository of resources managed by the bureaucracy has not been the best model. There is mention of doing this through a PPP mechanism, but the details have not been spelt out yet.
My biggest question, however, is on the centrality of buying GPUs, at a huge expense, and clearly insufficient in number. The technology is moving at lightning speed, with Nvidia bringing out new models every 12 months. The game is inevitably going to move beyond Nvidia, with Intel, AMD, startups like Cerebrus, and every Big Tech company bringing out its own chips tailored towards AI. Perhaps a better and more strategic idea would be to hedge our bets and consider buying or renting compute from AI cloud providers.
All hyperscalers and every other cloud provider is incorporating newer chips as they come along and offering a choice to its customers. Considering what the budget would buy in terms of AI GPU capacity, rather than just GPUs; given that all GPUs would not be used all the time, the same budget could stretch and buy much more capacity depending on when it is needed. This approach, therefore, could be a more prudent alternative. It is understandable that AI is a strategic technology, and the country needs to have control of the same. Therefore, the government should certainly have some strong contractual alliances, and a clear mandate that this sovereign AI capacity should be hosted within the nation’s borders. At this time, India has a strong hand geopolitically especially vis-à-vis the US and could lay down the terms of how it manages its own critical AI infrastructure.
Compute is far more complex than just the hardware; it involves local data centres, cloud adoption, and building software required to build and manage AI systems. As a part of the PPP philosophy, India would be well served to take a holistic view of compute, rather than just focusing on bringing in hardware with resources at its disposal. Both the AI for All strategy and the IndiaAI Mission are laudable steps by the government; it now has to work with industry and geopolitical friends to make sure we execute its tenets correctly and become one of the leaders in this inevitable age of AI.
The writer is the founder of Tech Whisperer Ltd, UK and teaches at Ashoka University
