D2 is a 60 MW IT-load facility inside Yotta’s Greater Noida campus, and it hosts the largest single AI deployment recorded in this registry: 20,736 liquid-cooled Nvidia Blackwell Ultra GPUs, backed by more than US$2 billion in commitments and connected over 800 Gbps InfiniBand with 40 petabytes of parallel storage. More than 10,000 of those GPUs are committed to the government’s IndiaAI Mission. Nvidia is putting one of Asia-Pacific’s largest DGX Cloud clusters inside the building under a four-year engagement worth about US$1 billion. Sixty megawatts divided by 20,736 GPUs is 2.89 kW of data centre per GPU — and that ratio, not the megawatt, is the number that now defines an AI facility.
1At a glance
2Two point eight nine kilowatts per GPU
| Facility | Status | MW | GPUs | kW per GPU |
|---|---|---|---|---|
| D2, Greater Noida | Live | 60 | 20,736 | 2.89 |
| D4, Greater Noida | From 2027 | 120 | 40,000 | 3.00 |
| NM2, Navi Mumbai | Fit-out | 80 | 40,000 | 2.00 |
| Programme | — | 260 | 100,736 | 2.58 |
Every capacity figure in this registry before this record was a megawatt. Yotta has introduced a second unit, and it is more useful for AI: how much data centre a GPU needs.
The band is 2.0 to 3.0 kW per GPU across a Blackwell Ultra to Vera Rubin generation change. That is not a coincidence — an NVL72 rack draws roughly 140 kW on Blackwell Ultra and 120 to 130 kW on Vera Rubin for the same 72 GPUs, so the per-GPU figure is set by rack architecture rather than by chip efficiency.
A developer planning an AI data centre in India now has a defensible rule: 40,000 GPUs needs about 100 to 120 MW. That is what Yotta has published three times, on three facilities, without ever stating the ratio.
Compare a conventional colocation build, which sells racks. A 7 MW hall at 15 kW per rack is 467 racks. An AI hall at 140 kW per rack is 50 racks for the same power — nine times fewer cabinets, nine times the cooling per cabinet, and a completely different landlord business.
Any Indian capacity dataset that counts megawatts and stops there cannot tell those two buildings apart. This registry now records both.
3The IndiaAI Mission share
| Allocation | GPUs | Share |
|---|---|---|
| Total Blackwell Ultra deployment | 20,736 | 100% |
| Committed to the IndiaAI Mission | more than 10,000 | 48% or more |
| Available commercially and for DGX Cloud | up to 10,736 | up to 52% |
| Nvidia DGX Cloud cluster | Part of the remainder | Four-year engagement, about US$1 billion |
Nearly half of India’s largest AI cluster is committed to a government mission, and that single sentence explains both the political support for the Indian data centre buildout and the commercial risk in it.
The IndiaAI Mission is the sovereign foundation-model and public-AI programme. Allocating over ten thousand liquid-cooled Blackwell Ultra GPUs to it means Yotta is, in effect, the government’s compute supplier as well as a commercial operator.
The trade is straightforward: sovereign volume is contracted, predictable and creditworthy, and it anchors a US$2 billion capital commitment. The cost is that 48% of the flagship cluster is not available to the market — so the commercial GPU supply in India is materially smaller than the headline number implies.
For anyone pricing Indian AI compute: the private market competes for roughly 10,700 GPUs at D2, not 20,736.
4Two billion dollars, and what it buys
| Item | Basis | Reading |
|---|---|---|
| Committed capital | More than US$2 billion | Blackwell Ultra infrastructure deployment |
| What it includes | GPUs, liquid cooling, 800 Gbps networking, 40 PB storage | Not just the building |
| Facility capex | not broken out | Cannot compute a per-megawatt building cost |
| DGX Cloud revenue | About US$1 billion over four years | A customer contract, not an investment |
| Implied GPU cost | US$2bn ÷ 20,736 = US$96,450 per GPU | If the whole sum were silicon |
| For comparison, D4 | US$7.5bn for 50,000 Vera Rubin | US$150,000 per GPU |
| D4 facility construction | US$750 million | US$6.25m per MW = ₹55.3 crore |
The last row is the one that matters for this registry. A later Yotta disclosure separates the US$750 million of data centre construction from the GPU spend, and US$750 million for 120 MW is US$6.25 million per megawatt — ₹55.3 crore.
That is inside the ₹51.4 to ₹60.1 crore band this registry has established from NTT, AirTrunk, CapitaLand and Colt. A fifth independent organisation, disclosing construction separately, lands in the same place.
Which is the useful conclusion: Indian data centre construction costs about ₹51 to ₹60 crore per megawatt regardless of who builds it. What varies wildly is the silicon on top — from nothing at a colocation hall to ₹150,000 per GPU at an AI factory.
A US$2 billion AI announcement is mostly chips. A US$2 billion data centre announcement would be 350 MW of buildings. These are not the same product and they are reported in the same headlines.
This registry counts 60 MW at D2. It does not count dollars, GPUs or superclusters, because none of those is a facility.
That discipline is what keeps the national total comparable across an operator building a shell in Devanahalli and an operator installing ten thousand GPUs in Noida.
5Procurement
6Limits and open questions
- 60 MW is IT load as stated, and is what the registry counts. The 250 MW campus ceiling is not counted.
- US$2 billion is a deployment commitment including GPUs, cooling and networking — not facility capex, and not a valuation.
- “Live by August 2026” was a target stated in February 2026. No commissioning confirmation has been located.
- The facility’s own building cost is not disclosed, so no per-megawatt figure exists for D2 itself.
- “More than 10,000” GPUs to the IndiaAI Mission is not a precise number and the allocation mechanism is not published.
- 2.89 kW per GPU is a Gridvik division of published figures, not an operator statement.
- The DGX Cloud contract value is described as “over US$1 billion multi-year”, not a firm booked figure.
- What would close it. A commissioning date and the number of GPUs actually reserved by commercial customers.
7Sources
Sources: the Yotta Blackwell Ultra release of 18 February 2026; Hiranandani Group coverage of the same announcement; Data Center Dynamics on the US$2 billion deployment and the IndiaAI allocation; Manufacturing Today; Tech Times on the 80,000-GPU programme; and the Yotta data centre portfolio page.