The proposal, announced Sunday (Feb 1) by India’s finance minister Nirmala Sitharaman and flagged in a report by TechCrunch, would give companies a tax holiday through 2047 for running their artificial intelligence (AI) services through data centers based in India.
Companies in question “will, however, need to provide services to Indian customers through an Indian reseller entity,” Sitharaman said in an address to India’s parliament. “I also propose to provide a safe harbor of 15% on cost in case the company providing data centre services from India is a related entity.”
As the report noted, the announcement is happening as American cloud giants like Google, Microsoft and Amazon are scrambling to build up data center capacity to accommodate a storm of AI-related demand.
India, TechCrunch wrote, has become an attractive target for new investment, thanks to a healthy pool of engineering talent and rising demand for cloud services.
Google announced in October that it would spend $15 billion for an AI hub and expanded data center infrastructure in India, with Microsoft saying two months later it would invest $17.5 billion by 2029. Amazon is also boosting its spending in India. The company hopes to invest an additional $35 billion in India by 2030.
The report also pointed out potential challenges to scaling data center capacity in India — including water shortages, spotty electricity availability and high electrical costs, all of which could slow progress and drive up costs for cloud companies.
“For the past several years, the AI boom has been inseparable from a race in cloud capacity,” PYMNTS wrote last month. “Training large models and running inference at scale drove unprecedented capital expenditures across hyperscalers, reinforcing the idea that bigger models required bigger data centers.”
However, that report said, there is a growing body of research that challenges that assumption, contending the infrastructure requirements of AI have been shaped more by early architectural decisions and less by “unavoidable technical constraints.”
A recent study from Switzerland-based tech university EPFL argues that although frontier model training is still computationally intensive, many operational AI systems can be deployed without relying on centralized hyperscale operations.
“Instead, these systems can distribute workloads across existing machines, regional servers or edge environments, reducing dependency on large, centralized clusters,” PYMNTS wrote.