India’s AM Intelligence Orders 9,000 Nvidia Vera Rubin Systems for an $8B AI Factory
On this page
AM Intelligence, an AI infrastructure company started by the founders of India’s Greenko Group, has placed a binding order for roughly 9,000 Nvidia Vera Rubin systems — the first large-scale commitment to Nvidia’s next-generation platform anywhere in Asia. The GPUs will be installed as Vera Rubin NVL72 rack-scale systems inside a 30-megawatt AI factory in Hyderabad, with deliveries scheduled for the first quarter of 2027. The purchase is the opening move in an $8 billion plan to build one gigawatt of rentable computing capacity across India, the United States, Finland and Malaysia.
I have been watching Indian data-center announcements pile up for two years now, and most of them are memorandums of understanding that quietly expire. This one reads differently. AM Intelligence says the order is binding, it names a specific chip generation that Nvidia has not shipped in volume yet, and it comes with a customer already signed for the first tranche of capacity. That does not guarantee the racks will hum on schedule, but it is a firmer signal than the usual press-release ambition.
What was actually announced
The company, promoted by Greenko Group founders Anil Chalamalasetty and Mahesh Kolli, disclosed the order in a statement on August 25, 2026. Bloomberg first reported the figure, and it was quickly picked up by Business Standard, Yahoo Finance and a wide spread of Indian outlets. The core claims are consistent across all of them:
- About 9,000 Nvidia Rubin GPUs, deployed as Vera Rubin NVL72 rack-scale systems.
- A first site in Hyderabad rated at 30 MW of IT load, targeted to come online in Q1 2027.
- A binding order, not a letter of intent.
- Part of a broader 1 GW “Compute-as-a-Service” roadmap backed by more than $8 billion in planned capital spending.
- Initial capacity at the Hyderabad factory already bought by a US customer whose name is withheld under a non-disclosure agreement.
Nvidia’s own literature puts the Vera Rubin NVL72 at around 450 exaFLOPS of NVFP4 low-precision inference compute per full deployment of this size, paired with HBM4 memory and a claimed reduction in token-serving cost of up to 10x versus the current Grace Blackwell generation. Those are vendor numbers, measured in a format tuned to make them look their best, so treat the multiplier as a direction rather than a promise. The Hyderabad site is being designed with liquid cooling, high-throughput RoCE networking and large parallel storage, and AM Intelligence says it is built to train and serve trillion-parameter models and agentic AI workloads.
Quick reference
| Detail | What AM Intelligence says |
|---|---|
| Order size | ~9,000 Nvidia Rubin GPUs (Vera Rubin NVL72 systems) |
| Order type | Binding |
| First site | Hyderabad, India — 30 MW |
| Delivery window | Q1 2027 |
| Total program | 1 GW across India, US, Finland, Malaysia |
| Planned capex | More than $8 billion |
| Near-term deployment | ~200 MW |
| Backing | Founders of Greenko Group (renewable energy) |
| Stated use | Trillion-parameter models, agentic AI, sovereign Indian models |
| Anchor customer | Undisclosed US company (initial Hyderabad capacity) |
The renewable-energy angle is the real pitch
The thing that makes AM Intelligence more than another GPU reseller is where it comes from. Greenko is one of India’s larger renewable-power developers, with a portfolio that spans solar, wind, hydro and pumped-storage projects across more than 20 states, and a stated ambition to build out roughly 50 GW of energy infrastructure over time. Chalamalasetty framed the compute business as a natural extension of that work, saying the group has spent “over two decades” focused on “transforming electrons into value.”
That is a genuine advantage. The single hardest constraint on AI data centers right now is not chips, it is power — specifically, firm power that can be delivered to a site on a schedule that matches a two-year build. A company that already develops generation and storage assets can, in principle, co-locate compute with supply and skip years of grid-connection queuing. Whether AM Intelligence can actually match a 30 MW factory (growing toward hundreds of megawatts) with round-the-clock clean power, rather than topping up from a coal-heavy grid when the wind drops, is the question worth holding them to.
Why the timing matters
The order landed days before Nvidia’s quarterly earnings, and it fits a pattern the company has leaned into all year: large, headline-friendly commitments from customers outside the familiar US hyperscalers. India is a market Nvidia’s leadership has courted repeatedly, and “one of Asia’s first Vera Rubin clusters” is exactly the kind of line that travels well in an earnings call.
There is also an export dimension that nobody has fully resolved in public. Rubin-class accelerators sit near the top of what US export rules scrutinize, and the framework governing advanced-chip shipments to India has shifted more than once in the past 18 months. AM Intelligence is planning capacity in the US and Finland as well as India, which spreads that risk, but a Q1 2027 delivery date is far enough out that the licensing picture could look different by the time the crates are meant to arrive. I would not assume the schedule is locked.
For readers who have followed the broader infrastructure story on this site — from Nvidia and OpenAI’s Ohio data-center plan to Cerebras pushing its wafer-scale inference systems — this is the same trend reaching a new region. The demand for training and inference capacity is running well ahead of the world’s ability to build and power it, and the companies that control energy are increasingly the ones writing the biggest cheques.
What to watch next
Three things will tell you whether this is real. First, a ground-breaking or a construction update on the Hyderabad site before the end of 2026 — binding orders still need buildings. Second, the identity of that anchor US customer, which would signal how serious the demand is. Third, any public detail on the power contracts behind the site, because the renewable pitch only counts if the electrons are actually clean and actually firm. Until then, this is a strong statement of intent from a company with an unusual and relevant background, not yet a running cluster.
Frequently asked questions
What is the Nvidia Vera Rubin NVL72?
It is Nvidia’s next-generation rack-scale AI system, the successor to the current Grace Blackwell NVL72. Each rack combines Rubin GPUs with Nvidia’s Vera CPUs, HBM4 memory and high-speed interconnect, and is sold as a single liquid-cooled unit rather than as loose cards. Nvidia has shown the platform but has not shipped it in large volumes yet.
How much is AM Intelligence spending?
The company has cited more than $8 billion in planned capital expenditure for its full 1 GW program. It has not broken out the cost of the 9,000-GPU Hyderabad order on its own, and Nvidia system pricing at this scale is negotiated rather than list-based.
Who is behind AM Intelligence?
It was founded by Anil Chalamalasetty and Mahesh Kolli, the founders of Greenko Group, an India-based renewable-energy developer. The pitch is that AM Intelligence can pair AI compute directly with clean generation and storage that Greenko-linked entities develop.
When will the Hyderabad AI factory be operational?
AM Intelligence has targeted the first quarter of 2027 for the site to come online, with roughly 200 MW of near-term capacity across its wider roadmap. Those dates depend on hardware delivery, construction and export licensing, none of which are fully locked in.
Does this mean India is building its own frontier AI models?
Not directly. AM Intelligence is selling capacity, and it lists sovereign-model developers among its target customers alongside cloud providers and AI labs. Whether Indian teams use that capacity to train competitive frontier models is a separate question that this order does not answer.
Primary source: Business Standard.
Sofia follows emerging technology, from AI and VR to IoT and blockchain, and translates the hype into plain language. She cares about what these tools mean for everyday users, not just the headlines.
