Yotta prepares a $7.5 billion order for 50,000 Nvidia Rubin GPUs in Delhi

Yotta Data Services is preparing to order 50,000 Nvidia Vera Rubin graphics processors and 45,000 GB300 chips from the Blackwell family, cofounder and chief executive Sunil Gupta told Business Standard. He put the Rubin tranche, including networking and related gear, at about $7.5 billion. Building the 120-megawatt D4 data centre in Delhi that would house the machines would take another $750 million. If the purchase closes on those terms it will be the largest GPU order yet placed from India.

Until this week the domestic marker was AMI Intelligence, part of the Greenko group, which has a binding order for 9,000 Rubin GPUs for an intelligent factory in Hyderabad. Yotta's Rubin count is more than five times that figure. Gupta described the paperwork as in process rather than signed. Nvidia, on its own earnings call the same week, said it already holds purchase orders for Rubin from every major hyperscaler, AI cloud and system OEM and expects the platform to deliver about 20 percent of data-centre revenue in the current quarter.

Why the building matters as much as the chips

A 120-megawatt hall is a power problem before it is a silicon problem. Delhi's grid, land and water permits will decide whether those 50,000 GPUs ever sit in racks. Gupta separated the $7.5 billion chip-and-network bill from the $750 million construction bill for a reason. One is a purchase order to Nvidia and its integrators. The other is concrete, substations and cooling in the National Capital Region.

India's semiconductor and AI infrastructure push has been running on announcements for two years. Piyush Goyal spent this week in Japan talking up Semiconductor Mission 2.0 and a $15 billion central outlay. The Yotta number is a private-sector test of whether that talk converts into installed megawatts. Rubin racks list in a wide band, often cited between $5 million and $7.8 million per NVL72 cabinet. At those prices, 50,000 GPUs imply hundreds of cabinets and a hall that has to be live when the crates arrive, not two years later.

Supply is the binding constraint

Nvidia's Jensen Huang said demand for the company's products is running ahead of the 70 percent revenue growth it just guided for fiscal 2028, and that memory shortages cap how fast it can ship. A $7.5 billion Indian order sits in that queue with Microsoft, the first commercial Rubin deployer, and with sovereign projects such as Japan's Noetra consortium, which plans 27,500 Rubin GPUs in a 140-megawatt factory.

Yotta already operates large halls and has been one of the more aggressive Indian buyers of earlier Nvidia generations. Scaling from existing clusters to 50,000 Rubin units is a step change in capital, power and customers. The company has not named the tenants who would fill D4. Without contracted demand, a $8.25 billion combined bill is a wager that Indian and regional AI users will rent the capacity as fast as it is bolted down.

Gupta's phrasing, in the process of ordering, leaves room for the count or the mix to move. The direction is clear. Indian operators are no longer placing thousand-chip orders and calling them national infrastructure. They are talking in the same tens of thousands that U.S. and Japanese buyers use. Delivery dates, not press quotes, will decide whether D4 is a hall or a headline.