Used to run applications that require a lot of processing power, compute is a technology stack that combines a hardware layer of graphics processing units (GPUs), an infrastructure layer of data centres and server optimisation algorithms, and a software layer of development frameworks. Compute capacity can be measured by the total number of GPUs or in terms of floating-point operations per second (FLOPS).
In March, the government allocated Rs10,372 crores for the IndiaAI mission. Nearly half of this amount, Rs4,563 crores, will be used to build compute capacity across the country. India’s strategy has two main goals: Increase the country’s compute capacity to meet growing demand and make AI more accessible and affordable by subsidising it. To achieve this, the government plans to build a national AI compute capacity of over 10,000 GPUs, which is going to happen through partnerships between the government and private companies.
In August, the Ministry of Electronics and Information Technology (Meity) published a request for empanelment (RFE). It invited applications from companies to provide AI compute and cloud services to make the GPUs available to startups, researchers, students, and academicians.
On December 3, Meity informed that 19 bidders had submitted proposals, showcasing participation from cloud to data centre service providers. The bidders include Jio Platforms, Tata Communications, Yotta Data Services, Sify, E2E Networks, a few companies from AWS, and Oracle. The technical evaluation committee will evaluate these bids based on the eligibility and technical criteria specified in the RFE document.
“I’m confident that we’ll exceed our initial demand of 10,000 GPUs. I’m expecting that even if only 10 out of the 19 bids are empaneled, I’ll still get access to around 20,000 GPUs. The minimum offer is 1000 GPUs, and some companies are offering as many as 5000-7000 GPUs,” Abhishek Singh, additional secretary of MeitY, tells Forbes India.
The 19 bids will be evaluated by mid-January 2025. The RFE process allows for a running empanelment, where new companies can submit bids every three months under the same terms and conditions. At least three-four companies, including Google and Intel, were unable to submit bids this time around due to various reasons.
The high cost of compute capacity is a concern, but the government is stepping in to subsidise it. For instance, 1000 GPUs are equivalent to Rs300 crore. “Other countries, such as the UAE, Saudi Arabia, and France, are also investing heavily in AI. If India doesn’t keep up, it risks becoming a mere user of AI rather than a developer,” adds Singh.
The demand for GPUs in India is building up now. Until six or seven months ago, there was hardly any GPU capacity. People were not attempting to build their own models. They were using existing models and applying them to specific tasks, explains Sunil Gupta, co-founder of Yotta Data Services.
Now that GPUs are available in India, startups, government organisations, and research-orientated educational institutions have begun building AI models focused on Indic languages.
India can benefit from existing models, fine-tune them for Indic languages, customise them to meet local requirements, and build applications and use cases on top of them. While this may not require thousands of GPUs concurrently, it will still necessitate tens and hundreds of GPUs. With thousands of potential use cases, India may eventually require millions of GPUs, explains Gupta.
“This projection may seem absurd at present, but given India’s vast potential for use cases, the country will likely require significant compute power in the future. This compute power will primarily be needed for inferencing rather than training models,” Gupta adds.
There is still a question about spending capacity. While large government organisations and educational research institutions may receive funding from the Department of Science and Technology or other budgetary allocations, startups may struggle to mobilise the necessary resources.
“The subsidies from the India AI mission will be instrumental in supporting startups that have good ideas and products but can’t afford to pay for GPUs,” says Gupta.
As the lead chair of the Global Partnership on Artificial Intelligence, India also hosted the Global INDIAai Summit this year, bringing together delegates from 15 member countries to discuss AI and compute-related issues. The momentum continued with the presence of Jensen Huang, CEO of Nvidia, who emphasised India’s potential to become a compute powerhouse.
Experts suggest the current target capacity may not be sufficient to meet India’s strategic AI goals. For comparison, companies like Meta will have much larger compute capacities, equivalent to 600,000 H100 chips by 2025. Building foundational models in AI requires massive amounts of compute. For example, training OpenAI’s GPT-3 model required 3,640 petaflop/s-days.
India’s plan for increasing its computing power is still in its early stages. Nevertheless, as Neil Shah, co-founder of Counterpoint, puts it, AI empanelment is an important step in the right direction from the government’s journey towards AI mission and democratises access to innovative AI compute for startups and researchers.
Image: Shutterstock