Context:

AM Intelligence, a Hyderabad-based AI-infrastructure firm, is pursuing up to $3 billion in debt (bonds and bank loans) to fund the purchase of Nvidia Rubin GPUs, aiming to deploy 29,000 chips for 100 MW of capacity and to expand to 300 MW within about 15 months with a multi‑billion‑dollar investment. The plan positions the company among leading Rubin GPU buyers outside North America and follows a broader strategy to build computing‑as‑a‑service data centers across India, Malaysia, the U.S., and Europe. The initiative reflects a broader wave of neocloud data centers in India targeting hyperscalers, leveraging advance payments and large-scale GPU deployments to capture first‑mover advantages in agentic AI. The outlook includes scaling to multi‑hundred megawatt capacity and shifting enterprise focus toward low‑latency production agents alongside foundational model work.

Dive Deeper:

• AM Intelligence, based in Hyderabad and owned by the founders of Greenko, is exploring up to $3 billion in debt financing via bonds and bank loans to acquire Nvidia's Rubin GPUs, with an estimated $300 million in advance payments anticipated from potential offtakers.

• The company markets itself as one of the largest committed Rubin GPU buyers outside North America and plans to purchase 29,000 Rubin chips to support a 100 MW computing capacity, with plans to grow to 300 MW within 15 months and a broader investment exceeding $20 billion.

• AM Intelligence intends to invest $6 billion to establish 100 MW of computing power split across data centers in India and Malaysia, signaling a cross-border expansion strategy for AI infrastructure.

• The deployment framework envisions 5 GW of AI data centers and 1 GW of compute‑as‑a‑service capacity globally, with a longer‑term target of 400 MW within two years, and underscores a shift toward computing‑as‑a‑service offerings for hyperscalers.

• Industry observers note Nvidia Rubin GPUs’ suitability for agentic AI, which could provide a first‑mover advantage in India, as enterprises move toward running low‑latency production agents beyond model training.