AM Intelligence (AMI) — the AI infrastructure arm of AM Group, founded by Greenko Group's promoters — has placed a binding order for 9,000 Nvidia Vera Rubin NVL72 rack-scale systems for a planned AI factory in Hyderabad, according to reports from Bloomberg via Taipei Times and BusinessWorld. The systems are scheduled for delivery in Q1 2027 and will power the first 30 megawatts of the facility. AMI plans to invest more than $8bn across AI data centres globally, building toward a 1GW compute-as-a-service platform across India, the US, Finland and Malaysia, with roughly 200MW near-term (BusinessWorld). The order lands squarely on a question the India Semiconductor Mission has not answered: the country can now buy world-class accelerators faster than it can make, package, power or cool them.
TL;DR
- AM Intelligence says it has ordered 9,000 Nvidia Vera Rubin systems for a Hyderabad AI factory, with about $8bn of committed capital.
- Vera Rubin is Nvidia's next-generation rack-scale platform, following Blackwell in the company's published data centre roadmap (Nvidia).
- The India Semiconductor Mission funds fabs and packaging, not GPU purchases, so this deal runs on a parallel track to ISM's ₹76,000 crore programme (ISM).
- The binding constraints are power, cooling water, transmission and skilled operations staff, not chip availability.
- Treat the headline figures as announced, not audited. Order books and installed capacity are different things.
- Last verified: 27 August 2026.
What exactly did AM Intelligence announce?
Three things: a location, a volume, and a timeline. The location is Hyderabad, which already hosts substantial hyperscaler and enterprise data centre capacity. The volume is 9,000 Nvidia Vera Rubin NVL72 rack-scale systems. The timeline is delivery in Q1 2027, powering the first 30MW of the facility, with ~200MW planned near-term and a 1GW platform across India, the US, Finland and Malaysia as the stated endpoint (BusinessWorld).
Rack-scale AI platforms still ship in allocation tranches, and a nine-thousand-unit order is almost certainly staged across multiple quarters rather than delivered as one shipment. AMI also claims the Vera Rubin generation could cut AI inference token costs by up to 10x compared with Grace Blackwell (BusinessWorld) — that figure is the operator's expectation, not a benchmarked result.
Why does the India Semiconductor Mission matter to a GPU order?
Because it is the piece that does not scale with a purchase order. The India Semiconductor Mission, approved with an outlay of ₹76,000 crore under the Ministry of Electronics and IT, targets fabrication, display fabs, compound semiconductors and assembly, testing, marking and packaging (ATMP/OSAT) capacity inside India (ISM, MeitY).
None of that produces a Vera Rubin accelerator. Leading-edge logic for these parts is fabricated at TSMC and packaged using advanced 2.5D techniques with HBM stacks sourced from SK hynix, Micron and Samsung. India's ISM-backed plants address mature nodes and back-end packaging, which is genuinely useful for power management, automotive and networking silicon, but it is not a substitute for the frontier supply chain.
So the honest reading is that AM Intelligence's order shows India buying its way into frontier compute while ISM works, on a much longer clock, towards making the less glamorous silicon that surrounds it. Both can be rational at once. Conflating them is where public commentary usually goes wrong. We covered the policy side of that gap in India's semiconductor mission and the 2026 governance agenda.
What is Nvidia Vera Rubin and why order it instead of Blackwell?
Vera Rubin is the generation Nvidia has positioned after Blackwell on its published data centre roadmap, pairing a new GPU architecture with a matching CPU and rack-scale interconnect (Nvidia). The practical argument for skipping a generation is lifecycle: a greenfield site that will not energise its first halls for several quarters is better off committing to the platform that will still be current when the building is finished, rather than filling it with hardware already mid-cycle.
The tradeoff is risk. Ordering an unreleased platform at volume means accepting the vendor's schedule, the vendor's power and cooling envelope, and the vendor's software maturity at launch. Sites that standardised on Blackwell have known thermals, known rack power and a settled software stack today. AM Intelligence is trading that certainty for a longer useful life.
Can Hyderabad actually power a facility of this size?
This is the part worth watching more closely than the GPU count. Rack-scale AI systems in this class are direct-liquid-cooled and draw well above the per-rack budgets legacy colocation halls were designed for, which means new build rather than retrofit, dedicated substations, and transmission capacity booked years ahead.
For a project of this scale the practical checklist is:
- Firm power allocation from the state utility, with a committed connection date, not a letter of intent.
- Transmission, since generation capacity is worthless without the lines to move it to the site.
- Cooling strategy — closed-loop liquid cooling to limit water draw, which matters in Telangana's summer.
- Redundancy and grid stability, because AI training clusters present sharp, synchronised load swings.
- Operations staffing, the least discussed constraint: liquid-cooled AI halls need technicians India is still training.
Progress on those five items, not the order size, is the signal that tells you whether the $8bn converts into working capacity.
How does this compare with India's other sovereign AI builds?
It is larger in capital terms than the publicly announced GPU expansions from established Indian operators, and it is structured differently. Existing players have generally grown capacity in tranches tied to customer demand, as with Yotta's sovereign AI GPU expansion. AM Intelligence has instead announced the endpoint first and is building towards it.
That approach has a clear failure mode. Announced-first capacity depends on demand arriving on schedule; if Indian enterprise and government AI workloads ramp more slowly than projected, the operator carries expensive idle silicon that depreciates on a short cycle. It also depends on talent, and India's deepest AI infrastructure operators are in demand globally, a pattern visible in the career of the Hyderabad-trained engineer now running compute at OpenAI.
What should Indian buyers and builders take from this?
If you procure AI compute in India, the useful consequence is optionality: a third large domestic supplier reduces dependence on a small number of operators and on offshore regions for data-resident workloads. Do not price that in yet. Ask for the energisation date, the cluster interconnect topology and the committed power draw per rack before you plan around it.
If you build hardware or manufacturing capacity, the lesson is where the addressable work sits. Frontier accelerators will be imported for the foreseeable future. Power electronics, busbars, cooling distribution units, cabling, racks and site construction will not be, and the same domestic-manufacturing logic already visible in schemes like the mobile phone manufacturing programme applies to data centre plant. That is the part of this $8bn most likely to be spent inside India.
There is also a governance point. When private capital moves faster than state programmes, the state's role shifts from funder to enabler: land, grid, clearances and skills. India has run that pattern before, most recently as public agencies handed hardware production to private firms, a shift covered in ISRO's exit from rocket manufacturing.
FAQ
Q: How many Nvidia Vera Rubin systems did AM Intelligence order?
A: The company states it has ordered 9,000 systems for its Hyderabad AI factory. Treat that as an order book figure rather than installed capacity, since delivery is almost certainly staged over multiple quarters.
Q: Is this project funded by the India Semiconductor Mission?
A: No. ISM's ₹76,000 crore outlay supports fabs, compound semiconductor plants and ATMP/OSAT packaging facilities, not GPU procurement or data centre construction. This is a private capital commitment on a parallel track.
Q: Why build in Hyderabad rather than Mumbai or Chennai?
A: Hyderabad combines existing data centre operations, available industrial land, state-level industrial power provisioning and a large local engineering talent pool. Coastal sites offer better subsea cable landing, which matters more for latency-sensitive serving than for training.
Q: Does Vera Rubin availability depend on TSMC?
A: Yes. Nvidia's leading-edge parts are fabricated and advanced-packaged outside India, with HBM memory from a small group of suppliers. Indian packaging capacity being built under ISM addresses different, more mature product categories.
Q: What is the biggest risk to the project?
A: Power and demand, in that order. Securing a firm grid connection and transmission on schedule is the hard engineering constraint; filling 9,000 systems with paying workloads on the projected timeline is the hard commercial one.
Q: Are the $8bn and 9,000-unit figures independently verified?
A: They come from the company's announcement of the project. Neither has been confirmed by audited filings or by vendor shipment disclosures, so they should be read as stated intent.
Corrections log: no corrections issued. If a figure here is superseded by an official filing or vendor disclosure, this article will be updated and the change noted in this section.
