AINEWS
The Hyderabad firm, backed by Greenko's renewable energy, is using an early allocation of Nvidia's next gen platform to sell low cost AI compute to US and Asian customers from a 1 gigawatt India campus.
AM Intelligence, a Hyderabad-based AI infrastructure company most readers have never heard of, has placed a binding order for 9,000 Nvidia Vera Rubin systems, one of the first Asian commitments to the chipmaker's next-generation AI compute platform, which Nvidia announced at GTC 2025 and hasn't shipped at scale. The order is the first concrete test of whether India can sell AI compute to global customers at materially lower cost by stacking cheap renewable power, sub-300-millisecond transcontinental fiber latency, and an early allocation of hardware that Microsoft, Google, and Amazon have already locked up through multiyear backlogs.
The Hyderabad company sits inside the Greenko Group, one of India's largest renewable-energy producers, and that relationship is the structural reason the order is plausible. Greenko's renewable-energy portfolio lets AM Intelligence pitch itself as "one of the lowest-cost AI computation infrastructure players globally," according to Greenko founder and president Mahesh Kolli in on-record comments to Business Standard. The pitch has already produced a binding pre-sale of initial capacity to an unnamed US customer under NDA, and Kolli told Economic Times that prospective buyers are "hunting desperately" for AI compute.
The Vera Rubin platform is Nvidia's successor to the Blackwell generation that currently dominates AI training clusters. Vera Rubin is the name of the late astronomer who helped establish the existence of dark matter; Nvidia uses it as the codename for its full rack-scale system, including the Rubin GPU, Vera CPU, NVLink 6 switch, and ConnectX-9 SuperNIC. Nvidia says Vera Rubin is engineered for training and running trillion-parameter and agentic AI models, and the platform is the same one Japan committed to in late 2025 for a national robotics foundation model, with a first data center targeted for June 2028. AM Intelligence is targeting its first gigawatt by 2027 (the 2027 target is stated elsewhere in the Business Standard reporting on the order).
The 9,000-system order is the first phase of an $8 billion plan to stand up roughly 1 gigawatt of computing capacity at a southern India site, large enough to host multiple hyperscale data center buildings. AM Intelligence says its target customers are major cloud-service providers, AI labs, and organizations building homegrown Indian AI models, with planned markets in India, the United States, Finland, and Malaysia. The geographic mix matters: it positions India not as a domestic-only compute buyer but as a low-cost exporter of AI compute, routing transcontinental traffic through undersea cables where Kolli says round-trip latency runs around 300 milliseconds.
That latency figure is Kolli's own framing, isn't a third-party measurement, and is the load-bearing technical claim behind the cost-arbitrage thesis. A 300-millisecond round trip is acceptable for inference workloads and for training pipelines that are not tightly synchronous, but it is materially worse than the single-digit-millisecond latency buyers get when their GPUs sit in the same rack as the model they are querying. The bet is that the price gap is wide enough to absorb the latency penalty for a meaningful slice of the AI compute market.
The other load-bearing claim is that AM Intelligence can actually get the systems. Nvidia's next-generation AI chips are in multiyear backlogs, with Microsoft, Google, Amazon, and the Japanese consortium all competing for allocation. That AM Intelligence is booking a 9,000-unit binding order while those buyers are still waiting on Blackwell-class hardware is either a sign of Nvidia's willingness to seed new geographies, an undisclosed customer relationship, or a marketing claim that may not translate into shipped systems on the company's 2027 timeline. CryptoBriefing and other outlets have reported the order, but no independent reporting has confirmed delivery dates or financing.
Kolli told Business Standard that the funding plan is a mix of debt and equity, and asserted that "capex to cash flow in this business is very short," meaning the company expects to start generating operating cash before the full $8 billion build is deployed. The full quote was cut off in the visible source text, so the specific working-capital and payback assumptions behind that framing aren't on the public record. The structural case is that selling pre-sold AI compute against long-dated customer contracts can produce cash flow faster than a typical real-estate or power-infrastructure build, but the case hasn't been independently validated.
What to watch: whether AM Intelligence names its anchor US customer before the 2027 target date, whether Nvidia confirms Vera Rubin delivery slots publicly, and whether the 1-gigawatt southern India site clears power-transmission and permitting milestones on a timeline that matches the chip delivery schedule. If those three converge, the order moves from a kickoff signal to a working template for low-cost AI compute exporting. If any of them slip, the renewable-arbitrage thesis has a much harder first test case.