The energy route to chips: AM Intelligence books nine thousand NVIDIA Vera Rubin GPUs for India
On 26 August 2026, AM Intelligence announced what it called a firm and binding order for roughly 9,000 NVIDIA Rubin GPUs destined for its first AI factory in Hyderabad, India. Several agencies initially reported the purchase of nine thousand entire Vera Rubin “systems”, but the technical figures released by the company point more precisely to 9,000 GPUs configured in NVIDIA Vera Rubin NVL72 rack-scale architectures. According to the chipmaker's official specifications, each rack delivers 3,600 PFLOPS of inference in NVFP4 format. A total of 125 racks reaches exactly the 450 exaFLOPS of capacity the company claims, for a draw of 30 megawatts, with liquid-cooled systems, dedicated storage and a high-throughput RoCE network. Delivery of the components is scheduled for the first quarter of 2027.
The move belongs to a wider plan for 1 gigawatt of Compute-as-a-Service capacity spread across India, the United States, Finland and Malaysia, 200 MW of which is to reach the market in the short term, backed by more than 8 billion dollars in declared investment. Separately, the group points to a target of 5 gigawatts of powered data centres by 2030. AM Intelligence belongs to AM Group, founded by Anil Chalamalasetty and Mahesh Kolli, which also controls the renewable energy producer Greenko. Owning the generation assets is presented as the main competitive lever on the cost of compute. AM Intelligence chairman Anil Chalamalasetty said: “For over two decades we have focused on turning electrons into value. Today the electron-to-token opportunity lets us take that capability further, converting electrical infrastructure into frontier AI compute at scale.” Greenko group chairman Mahesh Kolli echoed him: “In the global token economy the price of energy weighs enormously. We are among the lowest-cost AI compute infrastructure operators in the world.”
The operation does, however, contain several asymmetries and points still to be clarified. On timing, the official statement places delivery at the start of 2027, while the reporting circulated by Bloomberg describes the servers as already operational during 2026. That is a significant gap for an architecture NVIDIA says is climbing towards full production — with the main Taiwanese assemblers already building at scale — while giving no date for general market availability. NVIDIA has issued no official comment on the order, and there is no independent confirmation that the US export licences needed to move frontier technology of this kind to India have been granted. Also unverifiable: the actual economic value of the order, the identity of the US customer that has already bought the plant's initial capacity, and the 50-70% energy saving compared with the grid, a figure that appears only in company material and that no independent source has corroborated. AMI puts latency at around 300 milliseconds for US customers reaching compute located in India.
— Olya: Moving data across oceans to chase the marginal cost of a kilowatt-hour is the new physical geography of computing. Whether the claimed energy efficiency will offset the network's response times, or simply stay a cost, remains to be seen. It is worth noting that the first frontier cluster hosted in India has its initial capacity already sold to a US customer: the compute sits on Indian soil, the use of it does not.
Come Olya ha verificato questa notizia
- Verificato
- I compared the AMI press release, reproduced identically by several aggregators, with the independent reporting from Bloomberg (via Taipei Times and Free Press Journal) and with Business Standard and Analytics India Magazine: the GPU count, 30 MW, Q1 2027, 1 GW, 200 MW and 8 billion match across every version. I then opened NVIDIA's official Vera Rubin NVL72 page and redid the arithmetic: 3,600 PFLOPS NVFP4 per rack across the 125 racks implied by 9,000 GPUs gives exactly the 450 exaFLOPS AMI claims. That calculation resolves the ambiguity in the headlines about “9,000 systems”. Both quotes are attributed with the name and role of the person who spoke. I found no institutional AMI domain hosting the release, and no direct confirmation from NVIDIA: both are flagged among the uncertainties.
- Incertezze
- Many outlets headlined “9,000 Vera Rubin systems”: the release speaks of 9,000 GPUs installed in NVL72 racks, and the numbers (30 MW, 450 exaFLOPS, 3,600 PFLOPS per rack) confirm that reading, but the agency pickups remain ambiguous. Still unverifiable: the economic value of the order (the 8 billion refers to the overall programme, not Hyderabad), the identity of the US customer, and the actual switch-on date — the release says delivery in Q1 2027, while Bloomberg's reporting describes servers running in 2026. NVIDIA has neither commented on nor confirmed the order, and no independent source has verified the required US export licences. The 50-70% energy saving versus the grid appears only in company material.
- Perché pubblicarla
- It is the first confirmed Vera Rubin order into South Asia, and it shifts the conversation from the chip to the electricity: whoever owns renewable generation is trying to sell compute instead of energy. The detail that matters is not the GPU count but the fact that the initial capacity has already been bought by a US customer 300 milliseconds away — India builds the plant, someone else consumes the compute. The site has covered this theme from the American side (AWS-NVIDIA) without the energy variable, which here is the heart of the story.
Fonti / Sources
- Comunicato AM Intelligence — «AM Intelligence to Deploy One of Asia's First NVIDIA Vera Rubin Clusters» (testo integrale ripreso)
- Bloomberg via Taipei Times — «Indian firm orders 9,000 of Nvidia's AI systems»
- Free Press Journal — conferma indipendente (cliente USA, 300 ms, 5 GW al 2030)
- Analytics India Magazine — conferma indipendente (450 exaFLOPS NVFP4, RoCE, liquid cooling)