Yotta Adds $6B to Blackwell Buildout as India Chases Sovereign AI Compute
Yotta Data Services is adding $6B to its AI infrastructure push, expanding from a large Indian GPU buildout into a national compute platform with frontier-scale ambitions.
The company is increasing its planned Nvidia Blackwell deployment from 20,000 to 30,000 GPUs. It also plans to bring 8,000 Nvidia B200 GPUs online within a month and is evaluating another 36,000 to 37,000 GB300 or Vera Rubin-class GPUs next year.
The first 30,000 Blackwell GPUs require about $3B of investment. The additional next-generation tranche could cost another $4B. The separate 8,000 B200 deployment is nearly $600M.
The Schedule
The first 8,000 B200 GPUs are expected to become operational within a month. The initial 20,000 Blackwell GPUs are scheduled to go live by September, with the remaining 10,000 units following by November. The larger GB300 or Vera Rubin deployment is planned for May 2027.
That puts Yotta on a cadence closer to a hyperscaler than a regional data center operator.
Sovereign AI Means Racks
India’s AI policy debate has often focused on local models, language coverage, and data sovereignty. Yotta’s bet points lower in the stack. The bottleneck is physical compute: GPUs, power, cooling, and procurement relationships that cannot be recreated with a policy announcement.
Yotta already operates a 2GW data center campus in Mumbai and a 250MW facility in Delhi. It also supports government workloads through National Informatics Centre data centers. That combination gives it a plausible role as infrastructure supplier for both domestic AI firms and sovereign workloads.
The Model-Lab Gap
Yotta is not building a foundation model. That matters. India can have model startups, government datasets, and AI mission funding, but without local access to accelerator clusters, the value chain stays dependent on US and Chinese infrastructure decisions.
The strategic question is whether Indian demand can fill the capacity fast enough. If it can, Yotta becomes a compute utility for a national AI market. If it cannot, the company will need global customers to absorb a GPU order that now sits in the same conversation as second-tier hyperscaler deployments.