Yotta Data Services is rapidly expanding its AI infrastructure at a time when India is stepping up investments in sovereign AI capabilities. Sunil Gupta, co-founder, CEO and MD of Yotta Data Services, spoke to Poulomi Chatterjee about the company’s GPU expansion plans, concerns over power and water for data centres, the need for sovereign AI models, and gaps in the IndiaAI Mission. Excerpts:
Q. What are Yotta’s plans for scaling compute?
A. We currently have around 10,000 GPUs operational at our Navi Mumbai data centre. By the end of July, we will deploy 8,000 Nvidia B200 GPUs, followed by 20,000 Nvidia B300 GPUs by September and another 10,000 in Noida by November to meet rising demand. By March-May next year, we plan to add either 36,000 Nvidia GB300 GPUs or a slightly smaller number of Vera Rubin GPUs. That should take our total installed base to about 85,000 GPUs by the end of FY27.
Q. Cloud is a difficult business because of vendor lock-ins. How is Shakti Cloud performing?
A. Vendor lock-ins are real, but concerns around data sovereignty, IP protection and geopolitical risks are encouraging startups to shift to sovereign cloud platforms despite high data migration costs. Shakti Cloud has gained traction after being selected to manage the National Informatics Centre’s Delhi data centre. Government ministries and departments now use our cloud by default, and we expect this alone to generate over Rs 2,000 crore in annual business. We now have more than 900 customers, largely enterprises, and are gradually expanding into the private sector. We believe offering a strong portfolio of around 75 services is more valuable than trying to provide 200.
Q. Does India have enough power and water to support large AI data centres?
A. Much of the concern comes from applying Western assumptions to India. India has about 512 GW of installed power capacity, with more than half coming from green sources. The challenge is not generation but distribution in some locations. In contrast, parts of the US now have waiting periods of up to five years for power connections for data centres. India’s data centre industry currently consumes only about 1.5 GW, or 0.3% of national capacity, and the grid is expected to expand to around 900 GW by 2033. Water is also less of a concern than many assume. Most Indian data centres use air-based cooling rather than water cooling. Closed-loop cooling systems require water to be replenished only about once a year, keeping consumption relatively low.
Q. Does India need its own foundational AI models?
A. We cannot afford to keep debating this. Recent restrictions on AI models and chips globally show that countries are increasingly treating AI as a strategic asset. India should build frontier models where feasible, while also developing mid-sized, small and domain-specific models for sectors such as healthcare, agriculture and education. Relying entirely on foreign AI models carries long-term risks.
Q. How can the IndiaAI Mission improve GPU utilisation?
A. The IndiaAI Mission has strengthened the ecosystem, but deployment has lagged behind announcements. The number of GPUs actually operational is far lower than the numbers announced, which is why Yotta today operates more than half of India’s advanced AI compute capacity. Another issue is contract tenure. Infrastructure providers receive contracts for periods as short as one to 12 months, making financing difficult. Four- to five-year contracts would provide lenders and investors with the visibility needed to fund AI infrastructure at scale.