Google is looking to provide its tensor processing units (TPUs) for the Rs 10,000 crore IndiaAI mission in the next leg of applications, sources familiar with the development said. Oracle, too, is also evaluating its participation by providing computing infrastructure support.
Both tech majors are in talks with the ministry of electronics and IT (MeitY) for the empanelment. Sources said Google missed out in the first round to provide computing infrastructure owing to a delay in legal and internal approvals. Notably, Google’s TPUs are AI accelerators built for data centres and cloud AI workloads and it competes with the likes of Nvidia and AMD for its graphic processing units (GPUs) used for large-scale AI model training and inference.
Google TPUs are designed by Google but manufactured by third-party semiconductor foundries, primarily TSMC (Taiwan Semiconductor Manufacturing Company). Queries sent to Google didn’t elicit any response till press time. Under the IndiaAI mission, a common computing facility will be utilised by startups, academia and researchers. The government will bear 40% of the cost.
A sum of Rs 4,563 crore has been earmarked for AI compute infrastructure.
In the first round, proposals came in from 10 bidders to provide 18,693 GPUs for the IndiaAI mission. The lowest price bidders are E2E, Jio Platforms, NxtGen Datacenter, CtrlS, and Locuz Enterprise Solutions. Other bidders such as Yotta, Tata Communications, CMS Computers India, Orient Technologies, and Vensysco will have to match the pricing by lowest bidders to get empanelled under the IndiaAI mission. Notably, Yotta has proposed the maximum capacity of 9,216 GPUs. Work on the compute access portal, from where startups can access available GPUs, is on.
The government is looking to invite proposals for another 20,000 compute units in the next round, sources said, pointing to the huge anticipated demand for computing infrastructure.
Of 18,693 GPUs offered by the bidders across 30 models in the first leg, 12,896 are Nvidia H100 GPUs, 1,480 are H200, 512 units are AMD Mi325X, and 230 units are Mi300X. Other models are by Intel and AWS’ Trainium.
Under the IndiaAI mission, the average rate per a compute unit is Rs 115.85 per hour, whereas the average rate per higher precision AI compute unit is Rs 150 per hour. The government said the average discount from market prices for AI computing services is 42%. Higher precision AI compute units, critical for foundational model training, are being provided at a discount of 47% from market prices.
According to government officials, the empanelment process will be continuous, and other entities can also be part of the process. It will soon issue a notice to procure more computing units.
