AM Intelligence wants up to $3 billion in debt to pay for its Nvidia Rubin GPUs. Here is what the money is for

One of the largest committed orders for Nvidia's next generation AI chips outside North America now needs financing. AM Intelligence, the Hyderabad AI infrastructure company set up by the founders of renewable energy group Greenko, is seeking to raise as much as $3 billion to pay for the Nvidia Rubin GPUs it has already ordered, Bloomberg reported on October 30, citing people familiar with the matter. The package could combine bonds and bank loans, and the split is not yet final.

The company did not respond to CNBC's request for comment on the report. But the orders the money is meant to pay for are on the record, in AM Intelligence's own press releases and in reporting by Moneycontrol and BusinessLine since August.

What has been ordered

The first order came on August 25: about 9,000 Nvidia Rubin GPUs for a first AI factory in Hyderabad, to be deployed as Vera Rubin NVL72 rack scale systems in 30 MW of capacity, Moneycontrol reported. Mahesh Kolli of AM Intelligence told BusinessLine that the racks should start arriving in January and that the facility should be fully running by March or April 2027. The company says the Hyderabad site is being engineered for roughly 450 exaFLOPS of inference compute in NVFP4, Nvidia's new low precision number format.

On October 5 it added two more firm orders for about 20,000 Rubin GPUs, roughly 70 MW to be deployed in Malaysia and India with delivery scheduled for the second quarter of 2027. Together the three orders come to about 29,000 GPUs and close to 100 MW, which the company says puts it among the earliest and largest committed buyers of Rubin outside North America. It put the cost of that first 100 MW at $6 billion already committed.

The plan does not stop there. AM Intelligence says it wants to bring another 300 MW to market over the next 15 months, at a capital cost of more than $20 billion, as part of a 400 MW pipeline across India, the US, Europe and Malaysia. The longer term target is 1 GW of compute as a service inside a wider programme of 5 GW of powered AI data centres. Kolli told BusinessLine that the company plans to keep about 1 GW of that 5 GW for its own compute service and offer the rest to hyperscalers, AI cloud companies and other operators.

Why debt, and why now

The $3 billion would cover about half of the $6 billion the company says it has committed for the first 100 MW. Bloomberg's sources also said AM Intelligence expects about $300 million in advance payments from potential offtakers, the customers who would contract for the capacity. Under the compute as a service model described to BusinessLine, model developers and enterprises take dedicated GPU clusters on contracts of one to five years for training and inference, which is the kind of revenue lenders usually want to see before they fund hardware.

Debt is not the only route being explored. Moneycontrol reported on October 19 that the company is also evaluating an equity raise of more than $1 billion, with Morgan Stanley advising, and that it is open to selling a minority stake. One person quoted by Moneycontrol called it early days, with the amount not fixed and dependent on investor appetite.

Bloomberg framed the plan as the latest sign that the AI boom is driving a wave of borrowing. That matters for timing. Lenders are being asked to fund chips that will not produce revenue until the racks are installed and contracted in 2027, and AI hardware can lose value quickly as each new chip generation arrives. Offtake contracts and the equity raise are the two things most likely to decide the price of that debt.

The energy company advantage

AM Intelligence's pitch rests on power. The company describes its compute as backed by Greenko's own renewable generation and storage, and says most of the Rubin systems' cooling will be liquid, about 95 percent according to Kolli, which reduces the need for large volumes of water. Greenko, which is backed by GIC and ADIA according to Moneycontrol, was founded in 2006 by Anil Kumar Chalamalasetty and Mahesh Kolli.

It is not alone. Neil Shah of Counterpoint Research told CNBC that many Indian energy companies are turning into neocloud providers that lease compute to hyperscalers, naming Adani Group, Reliance and Tata Power. CNBC also noted that Mumbai based Yotta Data Services said earlier this year it plans to invest $2 billion in an AI hub built on Nvidia's Blackwell Ultra GPUs. Shah argued that Rubin, which is tuned for agentic AI, gives an early buyer a first mover edge as enterprises shift toward "running low-latency production agents" rather than only training foundation models.

What to watch

Three things will show whether the plan holds. The first is whether the debt and the equity round close, and on what terms. The second is the Hyderabad schedule: racks from January and a working 30 MW facility by March or April 2027. The third is customers, since the $300 million in advance payments Bloomberg's sources described would be the first public sign of who is actually buying the capacity. The company already pitches Hyderabad as one of Asia's first frontier AI compute clusters on Rubin, so the delivery dates will be the clearest test of that claim.