D4 is Yotta’s second Greater Noida building and the largest single facility record in Uttar Pradesh: 120 MW, planned to host 40,000 Nvidia Vera Rubin GPUs, with GPU deliveries expected March to June 2027 and the facility live between May and August 2027. The disclosure that matters most for this registry is buried in the coverage: facility construction is US$750 million, separately stated from the US$7.5 billion of Vera Rubin silicon. US$750 million for 120 MW is US$6.25 million per megawatt — ₹55.3 crore — and that is the fifth independent organisation to land inside the ₹51 to ₹60 crore band this registry has established from NTT, AirTrunk, CapitaLand and Colt.
1At a glance
2Seven hundred and fifty million dollars for 120 MW
| Programme | Rate per MW of IT | ₹ crore per MW | Basis |
|---|---|---|---|
| Colt DCS–RMZ | US$1.7bn for 250 MW | ₹60.1 crore | Two cities |
| CapitaLand Chennai 01 | ₹1,940 cr for 34 MW | ₹57.1 crore | One building |
| Yotta D4 | US$750m for 120 MW | ₹55.3 crore | Construction separated from silicon |
| AirTrunk India | US$6.0m per MW | ₹53.0 crore | Five sites |
| NTT NAV2 phase one | ₹5,500 cr for 107 MW | ₹51.4 crore | 54 acres |
| Band | — | ₹51.4 to ₹60.1 crore | A 17% spread over five operators |
Yotta D4 is the first disclosure in this registry where an operator has separated data centre construction from GPU procurement. That separation is what makes it usable, and it produces a number that sits almost exactly mid-band.
Five organisations, five disclosures, five countries of origin, one band. At this point the figure is not an estimate. Indian data centre construction costs about ₹51 to ₹60 crore per megawatt of IT load, and anyone quoting outside that range is either building something unusual or describing something else.
The AI capex headlines are dominated by silicon. D4 is US$7.5 billion of GPUs and US$750 million of building — a ten to one ratio.
That inverts the economics of the industry. A conventional colocation landlord invests about ₹55 crore per megawatt and sells it over twenty years. An AI operator invests ₹55 crore of building and ₹550 crore of chips per megawatt, and has to recover it in three to five years before the silicon is obsolete.
That is a fundamentally different business with a fundamentally different risk profile, and the Indian market currently reports both under the same word: data centre.
3Forty thousand GPUs and the power behind them
| Facility | Status or date | MW | GPUs | Chip |
|---|---|---|---|---|
| D2, Greater Noida | Live target Aug 2026 | 60 | 20,736 | Blackwell Ultra |
| D4, Greater Noida | 2027 | 120 | 40,000 | Vera Rubin |
| NM2, Navi Mumbai | Fit-out | 80 | 40,000 | GB300 |
| Programme | — | 260 | 100,736 | Mixed fleet |
260 MW and just over a hundred thousand GPUs across three facilities in two states. For scale, that is 10.7% of the entire 2,420.4 MW registry in one operator’s AI programme, of which 120 MW is counted here as an announced building.
The Vera Rubin choice is deliberate. Reports describe Yotta splitting generational bets — Vera Rubin in Noida, GB300 in Mumbai — so a single fab delay or architecture flaw does not wipe out the whole fleet. Vera Rubin carries 288 GB of HBM4 at about 22 TB/s and roughly 50 petaflops of FP4 inference per GPU, against Blackwell Ultra’s 8 TB/s and 15 petaflops.
That is a 3.3 times memory bandwidth improvement between generations, which is what matters for large-context inference and is the actual reason a new building is needed rather than a retrofit.
D4 is planned for live service in May to August 2027. Vera Rubin reached full production in June 2026. Nvidia’s roadmap has a successor cadence of roughly two years.
So a GPU installed in mid-2027 will be one generation old by mid-2029, and the building it sits in is designed to last twenty years. The mismatch between a three-year silicon cycle and a thirty-year building is the central unresolved economics of the AI data centre boom, in India and everywhere else.
No Indian operator has published a refresh or residual-value assumption. That is the single largest unknown in the sector and it is not in any capacity dataset.
4Uttar Pradesh, and what D4 does to it
| Measure | MW | Note |
|---|---|---|
| Uttar Pradesh counted today | 331.1 | 24 records |
| Plus Yotta D2 | 391.1 | +18% |
| Plus Yotta D4 | 511.1 | +54% on one announced building |
| NTT Noida 2 | 52.8 | Separate operator, Greater Noida |
| Iron Mountain NCR-1 | 0 | Capacity not disclosed |
| Yotta Greater Noida campus ceiling | 250 | Not counted |
| Registry total after D4 | 2,540.4 | From 2,420.4 |
Yotta alone would take Uttar Pradesh from 331.1 MW to 511.1 MW — a 54% increase on one operator’s two Greater Noida buildings, one of which does not exist yet.
That is the state-level consequence of the AI buildout: Uttar Pradesh becomes the second-largest data centre state in the registry on the strength of a single company’s announcements, ahead of Tamil Nadu’s 296.6 MW and Telangana’s 341.3 MW.
The registry counts D4’s 120 MW as additive because it is a discrete facility with a named location, a stated capacity and a construction start. It does not count the 250 MW campus ceiling, which would double it again.
5Procurement
6Limits and open questions
- 120 MW is an announced capacity. No construction stage, no phase schedule and no commissioning is published.
- IT versus gross basis is not stated for the 120 MW, and the registry counts it as stated.
- GPU counts conflict between sources — 40,000 Vera Rubin in one report, 50,000 in another. Neither is confirmed by Yotta directly.
- The facility may be a phase of the 250 MW campus claim. The relationship is not stated by any source.
- US$750 million is a reported construction figure, not an operator line item, and may exclude land and power infrastructure.
- US$7.5 billion is Vera Rubin procurement including networking, not facility capex, and is not counted anywhere.
- 120 MW is counted as a discrete facility with a named site and a construction start.
- What would close it. A construction start confirmation and a stated capacity basis.
7Sources
Sources: Moneycontrol report of 10 September 2026; ET Datacenters on the 95,000-GPU order; Tech Times on the 80,000-GPU programme; the Yotta Blackwell Ultra release of 18 February 2026; Hiranandani Group coverage; Data Center Dynamics; and StorageReview and Tech Insider on the Vera Rubin specification.