India's Rubin Order Prices AI Power at $40M/MW • The Weighted Average

An Indian renewables group just placed one of Asia’s first binding orders for Nvidia’s newest rack-scale systems, and the deal’s most useful figure is one nobody printed. AM Intelligence, the AI infrastructure arm of the group founded by Greenko’s promoters, ordered 9,000 Vera Rubin NVL72 systems for first-quarter 2027 delivery into a Hyderabad AI factory whose first phase runs at 30 megawatts, the Economic Times reported. The company says it will bring 200 megawatts to market in the near term against capital expenditure exceeding $8 billion. Divide the two: roughly $40 million per megawatt of AI capacity, all-in.

What $40 million per megawatt buys, and what it assumes

That derived figure is the number to carry into any conversation about sovereign or regional AI capacity, because it converts a headline capex into a unit an operator can compare. A 200-megawatt build at $8 billion means every megawatt of delivered capacity — silicon, racks, liquid cooling, power interconnect, shell — costs about forty million dollars. Half that would be an aggressive claim; twice would suggest the buyer overpaid for grid access.

The compute side gives a second unit, and it also resolves an ambiguity in the coverage. The Hyderabad facility is engineered to deliver about 450 exaFLOPS of NVFP4 inference compute in its 30-megawatt first phase, per the company’s own account of the deployment. Nvidia’s published NVL72 specification rates a single rack at 3,600 PFLOPS of NVFP4 inference with 72 Rubin GPUs. Divide 450 exaFLOPS by 3.6 and you get 125 racks — which is exactly 9,000 GPUs at 72 per rack. So the “9,000 systems” in the headlines are GPUs, not racks, and the site’s density works out to about 240 kilowatts per rack and 15 exaFLOPS per megawatt at low precision.

That density is the tell. Nvidia claims Vera Rubin delivers up to 10x more tokens per megawatt than GB200 NVL72 and one-tenth the cost per million tokens for interactive agentic reasoning; AMI repeats the 10x token-cost figure. Both are vendor projections measured on a specific model and context length, but they set the yardstick the site’s economics will be judged against — and they explain why a power company, rather than a cloud, is the buyer.

Where the story stops being an equipment purchase is the parentage. AM Group’s other pillars are Greenko, with roughly 11 gigawatts of installed renewable capacity across solar, wind, hydro and storage, and AM Green in fuels. Chairman Anil Chalamalasetty framed the order as converting “electrons into value,” and beyond Hyderabad the group says it is developing 5 gigawatts of powered AI data centres across India, the US and Europe alongside a 1-gigawatt compute-as-a-service platform. This is a power company buying its way down the value chain, not a cloud provider buying its way into power.

Why the electron-to-token trade is the model to watch

The bet is straightforward and increasingly common: in a market where grid interconnect is the scarce input, whoever already owns firmed renewable capacity can build AI compute faster than whoever owns only capital. That inverts the usual sequencing, in which a data centre developer queues for power. It is the same logic driving the 32% compound capex growth required to reach $3 trillion in global data centre spending — the money exists; the megawatts are the bottleneck.

For operators outside India, the practical consequence is optionality on latency and jurisdiction. A 30-megawatt Rubin cluster serving Indian developers locally changes the calculus for anyone routing Indian traffic to Singapore or Virginia today, and a compute-as-a-service platform spanning India, the US, Finland and Malaysia is deliberately structured for data-residency shopping. If your compliance regime cares where inference happens, another supplier will exist in 2027 that does not today.

There is a supply-side reading too. Nvidia says the Vera Rubin platform has seven new chips in full production, with Taiwanese server makers shipping systems at scale. A binding order for 9,000 GPUs from a first-time AI operator, landing alongside national-scale deployments, tells you the allocation queue is open to buyers who bring their own megawatts. That is a meaningful change from the Blackwell cycle, when hyperscalers absorbed nearly everything and regional operators waited. If power is the currency that buys allocation, expect more utilities and renewable developers to appear on Nvidia’s customer list before expecting more clouds.

The counterpoints are serious. First, the entire order lands in Q1 2027 — a binding order is a promise, not capacity, and the 200-megawatt and 5-gigawatt figures are ambitions with no disclosed financing structure behind them. Second, $8 billion of capex from a group whose balance sheet was built on renewable project finance is a different risk profile than the same number from a hyperscaler; Greenko’s backers include GIC and ADIA, but AI infrastructure depreciates on a schedule renewables do not. Third, the 450-exaFLOPS and 10x figures are company claims measured in NVFP4, the low-precision format that flatters inference throughput.

The verdict: treat $40 million per megawatt as the benchmark to price rival announcements against, and treat AMI itself as a 2027 supplier, not a 2026 one. This paper has tracked the same discipline on the vendor side, where Nvidia’s server price increases concealed a much larger memory pass-through, and today’s lead argues that the efficiency of the silicon inside these racks is about to become a power-bill question. What would change the call: a disclosed financing package for the 200-megawatt tranche, or a competing operator publishing a lower cost per megawatt on comparable Rubin capacity.