The billion-dollar public-private partnership model to set up graphics processing unit (GPU) capacity in the country is likely to be structured in a way that the government doesn't have to pay data centre companies for any capital expenditure.
This is unlike the $10 billion incentive scheme for semiconductors, where the government is paying chip companies to set up plants in the country.
The industry has advised the government that it should not invest directly in the capex for GPUs. However, it should subscribe to the GPU capacity put up by domestic data centre companies and allocate it to startups, according to Sunil Gupta, CEO and cofounder of Yotta, a data centre services company.
Apart from other hindrances in building home-grown foundational models, India lacks domestic compute capacity. To solve that issue, one of the aims of the IndiaAI mission is to make over 10,000 GPUs available in India. IndiaAI is the government's AI ecosystem building initiative.
GPUs, which are primarily designed render images and videos quickly and efficiently, are the hardware on top of which AI models can be trained. They are primarily housed in a data centre.
Gupta said that data centres have a quicker turnaround time in terms of cash flows from operations and hence don't need any viability gap financing. These businesses would instead prefer to have a steady and large customer in the government.
Meanwhile, Yotta has already received delivery of the first tranche of an order for 4,000 H100 GPUs from Nvidia. Gupta, who was in San Jose for Nvidia's annual developer conference, told journalists that these GPUs will be set up in a Mumbai data centre and go live by June.
Pramod Verma, former chief architect of India's digital identity programme Aadhaar, has similar views on the government’s role in procuring GPUs. He suggests the creation of a financial instrument or bonds to democratise access to GPUs.
Verma, who is also the chief technology officer of the Nandan Nilekani-backed EkStep Foundation, underscores the inherent challenges in traditional procurement, which is laced with time-consuming bureaucratic and logistical issues that come with government-led purchases.
He believes that the government should refrain from purchasing GPUs physically, as it would result in a delay in acquiring the latest technology. In essence, this approach would mean that the GPUs so procured would likely be outdated by the time they are deployed, as the provisioning process takes time.
“They (government) don't need to buy; why should they buy,” Verma told Moneycontrol, adding, “10,000 (GPUs) is a good starting point but beyond 10,000, they can still create financial instruments to sort of create affordable pricing models or vouchers for startups to use and they can negotiate as a pooled demand.”
Verma has many feathers in his cap, including being the co-creator of the open-source Beckn protocol, which is the base for India’s Open Network for Digital Commerce (ONDC), and Namma Yatri, a ride-hailing service app.
He recommends that the government go to hyperscalers, such as Amazon Web Services or Microsoft Azure, and demand a fixed number of GPUs for a period of, say, five years, thereby pre-committing the money. By pre-committing funds through these GPU bonds, venture capitalists and startups can wield significant influence in bargaining for competitive rates, which would be impossible individually, he noted.
Verma cited neighbouring China as an example, pointing out that it negotiated with large tech players to give vouchers to startups instead of buying.
“We have enough muscle power, we have large data, and a lot of demand. So if we can create a pooled mechanism to negotiate, then we can also create some good access to compute instead of fully buying everything by government,” Verma said.
He spelled out the potential logistical issues that could crop up in case the government buys GPUs. “After buying, provisioning takes time. Which startup will apply? Who will they prioritise? How will they price it to the startup once you buy the actual hardware,” he asked.
Also read: GPUs under AI Mission will be available in 18-24 months: MeitY Secy
The government should learn from the challenges encountered during the deployment of PARAM (Parallel Machine) supercomputers to renowned academic institutions, he said. “It was supposedly given to all the IITs. Why was it so delayed? It took 2-3 years and they were still not provisioning it because of the bureaucracy involved,” Verma said.
Priority and pricing will emerge as the biggest challenges for the government when procuring GPUs. “Will there be a 2G-like scam if I under-price it and somebody will pick up… saying I could have priced it more,” he said, in jest.
Furthermore, Verma advocated the importance of creating a level playing field for Indian startups against Western companies, wherein AI startups can thrive without being disadvantaged by Big Tech companies, which have most of the data. “If the government can create a level playing ground with minimum bureaucracy and minimum overhead, that's probably the best way to do so,” he said.
Also read: Tech wants government to prioritise access to data under IndiaAI Mission
Discover the latest Business News, Sensex, and Nifty updates. Obtain Personal Finance insights, tax queries, and expert opinions on Moneycontrol or download the Moneycontrol App to stay updated!
