To accommodate start-ups, MeitY relaxes AI compute procurement norms

The move is part of the Rs 10,370 crore IndiaAI Mission to establish a computing capacity of more than 10,000 GPUs and also help develop foundational models with a capacity of more than 100 billion parameters trained on datasets covering major Indian languages for priority sectors like healthcare, agriculture, and governance. The idea is that if such an infrastructure exists in the country, start-ups could plug into it for developing AI systems.

Computing capacity, or compute, is among the most important elements of building a large AI system, apart from algorithmic innovation and data sets. It is also one of the most difficult elements to procure for smaller businesses looking to train and build such AI systems, given the high costs.

In a corrigendum to a tender published in August, the IT Ministry has also lowered the capacity of computing power that successful bidders would have to provide. Initially it said that 1,000 GPUs should have a performance threshold of 15 TFLOPS for FP32, 300 TFLOPS for FP16, and 40 GB AI compute memory. In the corrigendum, the government has reduced FP16 to 150 TFLOPS (from 300 TFLOPS) and reduced AI compute memory requirement from 40 GB to 24 GB.

TFLOPS is a unit that measures the computing power of a system. For example, if a system has 10 TFLOPS of FP16 performance, it means it can carry out 10 trillion FP16 calculations every second.

While there have been some relaxations, the ministry has also introduced a new technical criterion of a company’s experience in offering AI services to customers. Now, companies will have to submit the number of companies they have previously offered AI compute services in in the last three financial years (2020- 21, 2021-22 and 2022-23). Minimum billing should be Rs 10 lakh, with companies receiving additional marks for servicing a certain number of customers.

Another new requirement for sourcing components for cloud services locally has been introduced. The components used for providing AI cloud services by bidders should be procured from either a Class I local supplier or a Class II local supplier, in accordance with the ‘Make in India’ initiative guidelines, the corrigendum said.

A class I supplier offers goods and services with domestic value addition of at least 50 per cent, whereas a class II supplier’s local content is between 20-50 per cent.

“The bidder shall ensure the ‘Make in India’ initiative guidelines of Department for Promotion of Industry and Trade, Ministry of Commerce and Industry and Ministry of Electronics and Information Technology,” the corrigendum said.

Successful bidders will have to ensure availability of AI compute capacity for consumption – a demand of up to 100 AI compute hours shall be met immediately and up to 500 AI compute hours shall be met within two days and demand of more than 500 hours of AI compute shall be met within a week.

One key condition of the tender is that all AI services are to be delivered from data centres in India. “Data uploaded to their cloud platform by end users should not be sent outside the sovereign territory of India in any form (anonymous/pseudonymous/encrypted, etc.),” the tender said.

Of India’s Rs 10,370 crore plan, the implementation of computing infrastructure will be done through a public-private partnership model with 50 per cent viability gap funding. If the compute prices come down, the private entity will have to add more compute capacity within the same budgeted amount to meet increased demand. Of the total outlay, Rs 4,564 crore has been earmarked for building computing infrastructure.