The tender aims to set up public AI cloud infrastructure with a capacity of at least 10,000 GPUs through a public-private partnership under the India AI Mission.
ET has seen a copy of the corrigendum issued on Wednesday.
It marks the second extension of the deadline, which was originally set for September 6 and later pushed to October 16. New clauses were added during the previous extension, with some companies raising concerns, chiefly regarding the requirement to achieve a power usage effectiveness (PUE) of less than 1.35.
Also Read: Firms scramble to build consortiums to be eligible for Rs 10,000-crore GPU plan
PUE is a measure of the energy efficiency of data centres, calculated by dividing the total power used by the data centre by the power used by its computing equipment. A lower PUE indicates higher energy efficiency.
Discover the stories of your interest
On October 9, ET reported that several data centre operators wanted the PUE criterion removed, citing the need for increased capital expenditure, which would be passed on to customers. Data centre executives said that achieving a PUE of 1.3 is technically challenging in India due to high ambient temperatures, especially with direct-to-chip cooling technology.When asked whether the tender’s clauses had been modified, the government official said that, as of Wednesday, the tender remains unchanged, though the concerns raised by some companies are under consideration. The corrigendum did not reflect any further modifications to the tender.
“The last date for bid submission has been extended to November 12. As of today, the clauses in the tender remain unchanged. A few companies have raised concerns, and we are reviewing them,” the official said.
Sunil Gupta, CEO of Yotta Data Services, said at the Nasscom Future Forge 2024 event in Bengaluru on Wednesday, “There are six or seven large conglomerates that are willing to invest heavily in AI infrastructure in India.”
Yotta Data Services is one of the bidders for the IndiaAI Mission’s GPU tender. Last November, it placed an order for 16,384 H100 GPUs and has deployed the first phase with 4,096 GPUs.
Gupta said that in addition to GPU investment, the country would require substantial investment in underlying data centres. “AI workloads are extremely high-density. A typical data centre rack consuming 6 KW of power would require 100 KW per rack,” he said.
“Globally, data centres consume nearly 3% of total global energy production. If you introduce thousands of GPUs, this figure is at risk of reaching double-digit percentages. This is a global concern,” Gupta said, pointing out that Singapore had imposed a moratorium when data centres began consuming 7% of the country’s total power.
However, Gupta argued that in India, where green energy is being promoted, data centres consume 1,500-1,800 MW—about 0.2%-0.3% of the country’s total energy production—so it isn’t a major concern. He emphasised that technologies such as direct-to-chip cooling and immersion cooling could help reduce power consumption.
