Yotta Vera Rubin in India: the 120 MW, May to August 2027 plan for GPU buyers

Summary. Yotta Data Services plans to put NVIDIA Vera Rubin GPUs into a new 120 MW data centre, D4, at its Greater Noida campus, with deliveries reported for March to June 2027 and go-live for May to August 2027. That is a plan given in interviews, not a signed order: neither Yotta nor NVIDIA has published a release on it. The GPU count has moved from 50,000 in late August to 40,000 in a 10 September interview. NVIDIA's side is firmer. Vera Rubin began ramping into full production on 31 May 2026 and has been generally available on Oracle Cloud Infrastructure since 27 October. NVIDIA sizes a 100 MW AI factory at 40,000 Rubin GPUs, which puts 50,000 beyond what 120 MW can hold. Yotta's announced pipeline is Blackwell: 20,736 B300 GPUs for its 60 MW D2 site, more than 10,000 of them committed to the IndiaAI Mission. Our verdict: plan 2027 work on capacity you can rent now (Yotta lists B200 bare metal at ₹473 per GPU-hour), treat Indian Vera Rubin as a second-half 2027 option, and sign nothing that depends on it without delivery, substitution and exit clauses.

About this article. For Indian CTOs, AI leads and procurement teams planning GPU capacity for 2027. On 7 November 2026 we checked Yotta's plans against its releases, media page, price list and the CEO interviews cited; NVIDIA's claims against its newsroom, product page and October 2026 datasheet; and the policy points against C-DAC's tender, PIB, PRS and the DPDP Rules (G.S.R. 846(E), 13 November 2025). Business Standard's 27 August interview page refused automated access, so we used a published summary of it. This is not investment, procurement, legal or tax advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

Status check: what is confirmed on 7 November 2026

Sources for the table: Yotta's release and press list, the CEO interviews, NVIDIA's newsroom and SEBI's filing list.

The August wording described an order still being placed, India's largest only if completed. By contrast, AM Intelligence (AMI), set up by Greenko's promoters, announced a firm and binding order on 25 August for about 9,000 Rubin GPUs in Vera Rubin NVL72 racks for a 30 MW AI factory in Hyderabad, with delivery in the first quarter of 2027. On 6 October it added two more binding orders for 20,000 Rubin GPUs across India and Malaysia.

Why the numbers keep changing

Within about two weeks Yotta's CEO gave three different mixes. Late August: 50,000 Vera Rubin plus 45,000 GB300. On 10 September to Moneycontrol: 40,000 Vera Rubin at D4 plus 40,000 GB300 at NM2, about $12 billion over the next 12 months. Around the same time to The Economic Times: 90,000 Vera Rubin and GB300 GPUs for $12 billion, bought through a tender and delivered in batches. The August account placed D4 in Delhi; Moneycontrol places it on the Greater Noida campus, where D1 and D2 already sit.

The totals move too. Yotta's February release set a roadmap to more than 80,000 NVIDIA GPUs by FY27, and its media page carries a 30 July 2026 headline of 85,000 GPUs by the end of FY27. An 11 October headline listed on the same page still describes the $12 billion order as being planned.

Read this as a procurement still forming, shaped by NVIDIA allocation, financing and which anchor customers sign first. For a buyer the count matters less than one question: when could a contracted block of Rubin capacity be yours, and at what price?

Does 50,000 Rubin GPUs fit in 120 MW?

Power, not floor space, limits how many Rubin GPUs a building can run. NVIDIA's Vera Rubin NVL72 page sizes a 100 MW AI factory at 40,000 Rubin GPUs, assuming its DSX platform with MaxLPS power management. That is 2.5 kW of site power per GPU, or 48,000 GPUs for 120 MW. On the same basis 50,000 GPUs need 125 MW. NVIDIA's October 2026 datasheet rates one NVL72 rack at 227 kW (MaxP), 187 kW (MaxQ) or 156 kW (MaxLPS). AMI's own order gives a second benchmark: about 9,000 Rubin GPUs in 30 MW, or 300 per MW.

Racks are rounded up to whole units of 72 GPUs, and the loads cover GPU racks only, before networking, storage and cooling. We do not know whether Yotta's 120 MW is IT load or total facility power. If it is total, a power usage effectiveness of 1.2, our assumption for a good liquid-cooled hall, leaves about 100 MW for IT.

The arithmetic points one way. The September figure of 40,000 fits a 120 MW building, comfortably at MaxLPS. The August figure of 50,000 would need deep power capping, a phased fit-out or more power than the headline. Ask Yotta which it is, because power-capped racks deliver less than peak throughput.

Timeline: from order to rentable capacity

The dates come from Yotta, NVIDIA, AMI and the interviews already cited.

Five risks to the date

Construction. On 10 September Gupta said D4 construction was starting. Going live by August 2027 leaves about 11 months for a 120 MW liquid-cooled hall, which is fast by any standard (our assessment, not Yotta's).

Order and financing. No signed order has been announced. Reuters reported that IPO proceeds would partly buy GPUs and that Yotta is exploring structures in which partners buy GPUs through special-purpose vehicles, share the revenue and hand over ownership after four to five years. Gupta told Moneycontrol that IPO money would mainly fund data centres and sovereign cloud, with GPUs financed through revenue-share contracts, leasing and investment vehicles. The draft IPO papers expected in October do not appear on SEBI's public list of filings, though a confidential pre-filing would not show there.

Anchor tenants first. Large blocks of the B300 cluster are already committed to NVIDIA's DGX Cloud and the IndiaAI Mission. Gupta told Reuters that global clients make up 75% to 80% of Yotta's customer base. Early Rubin capacity will probably go to large, long-term buyers before on-demand users.

Cooling. NVIDIA designs Vera Rubin NVL72 for 100% liquid cooling at a 45°C coolant inlet, so the hall has to be fully liquid-ready on day one.

The next generation. NVIDIA's roadmap puts Rubin Ultra systems in the second half of 2027, so D4 may open within months of a newer part. Our GTC 2027 buyer briefing covers that roadmap and the reported Kyber rack delay.

How Vera Rubin fits with Yotta's Blackwell supply

Most capacity Indian teams can use in the next nine months is Hopper and Blackwell. Yotta's announcements and its Shakti Cloud price list on 7 November 2026 show:

Yotta says data transfer is included in the bare-metal price and InfiniBand interconnect is charged separately. The page does not say how GST applies, so confirm tax treatment in any quote.

What does Rubin buy you over Blackwell? NVIDIA claims Vera Rubin NVL72 delivers one-tenth the cost per million tokens of GB200 NVL72 and up to 10x more tokens per megawatt, measured on Kimi-K2-Thinking with 32K input and 8K output tokens. It also claims mixture-of-experts training with one-quarter of the GPUs, for a 10-trillion-parameter model trained on 100 trillion tokens in a month. NVIDIA marks these projections as subject to change, and they compare against GB200, not the B300 or GB300 Yotta is installing, so the gap against Blackwell Ultra will be smaller. Each Rubin GPU carries 288 GB of HBM4 at 19.2 TB/s.

For rupee prices across Indian providers and the IndiaAI portal, see our IndiaAI 10,000-GPU empanelment guide.

Where IndiaAI compute fits

Yotta's February release committed more than 10,000 B300 GPUs from the D2 supercluster to the IndiaAI Mission, for foundation models, research institutions, startups and public platforms. That capacity reaches users through the IndiaAI compute portal, which has eligibility rules by user type and an approval process.

Separately, C-DAC's request for empanelment of 7 October 2026, for up to 10,000 AI compute units, covers H100, H200 and later or equivalent next-generation accelerators, with quantities and prices left to a Stage-2 procurement. Bids close on 13 November 2026. In principle an empanelled provider could offer Rubin in a later round. Nothing published says Yotta will put D4 capacity into IndiaAI, and Vera Rubin is not on the portal today.

The practical split for 2027:

• Eligible startups, researchers and government bodies should use the portal now and watch for next-generation rounds.

• Enterprises that are not eligible buy commercially. Rubin from Yotta will be a commercial contract.

What Indian GPU buyers should do now

Worked example: when would Rubin be worth paying more for?

No Indian Rubin price exists yet, so the useful sum is a break-even price. Take a team that needs 512 GPUs for 12 months of training.

Today's baseline on Blackwell. At Yotta's listed ₹473 per GPU-hour for 8x HGX B200 bare metal, which is also its one-year rate:

• GPU-hours: 512 × 8,760 = 4,485,120

• Cost: 4,485,120 × ₹473 = ₹2,12,14,61,760, about ₹212 crore, before GST and InfiniBand charges

The break-even rule. If a Rubin GPU finishes your job S times faster than a B200, the same job needs 1/S of the GPU-hours. Rubin is cheaper whenever its hourly price is below ₹473 × S.

The speed-ups are illustrative assumptions, not benchmarks. Your real S depends on model size, precision (NVFP4 versus FP8 or BF16), sequence length and how well your stack uses NVLink 6, so run your own training step on both before you sign. If Yotta quotes Rubin above ₹946 per GPU-hour, you need better than a 2x real-world gain over B200 just to break even. One reference point: Yotta's one-year B300 list price of ₹1,056 is already 2.23 times its one-year B200 rate (₹1,056 ÷ ₹473), so a B300 has to beat a B200 by more than that to win on this job.

Questions to put to Yotta, or any Indian GPU cloud, before you sign

Data residency, DPDP and the tax angle

For Indian personal data, an India-hosted GPU removes a cross-border transfer question that a foreign region reopens. Most obligations under the DPDP Rules start 18 months after their 13 November 2025 notification, on 13 May 2027, in the same window as D4's planned go-live. Our DPDP Rules 13 May 2027 checklist covers what has to be ready.

Tax policy is also pulling foreign demand into Indian halls, and Gupta told Reuters the February tax holiday had lifted overseas customers' confidence. The 1 February 2026 Budget proposed the holiday until 2047 for foreign companies serving cloud customers globally from Indian data centres, with Indian customers served through an Indian reseller. The Taxation and Other Laws (Amendment) Act, 2026, published in the Gazette on 17 August 2026, extended it to data centres an Indian company leases and operates, not only owns. Our piece on the India cloud tax holiday to 2047 explains what finance teams should prepare.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning team sizes training and inference in GPU-hours, runs like-for-like benchmarks on your own models across Indian GPU clouds and hyperscalers, and reviews delivery, power and exit terms in GPU contracts. Talk to us at /contact-us/.

Last updated: 7 November 2026.

References

1. Yotta to order 50,000 NVIDIA Vera Rubin GPUs for $7.5 billion, 28 August 2026 — Lapaas Voice

1. Yotta plans 40,000 Vera Rubin GPUs for new Noida data centre, eyes $12 billion investment, 10 September 2026 — Moneycontrol

1. Press releases — Yotta Data Services

1. NVIDIA Vera Rubin ramps into full production to power agentic AI factories worldwide, 31 May 2026 — NVIDIA Newsroom

1. NVIDIA and Oracle expand NVIDIA Vera Rubin to multi-gigawatt scale and bring NVIDIA Vera CPU to Oracle Cloud Infrastructure, 27 October 2026 — NVIDIA Newsroom

1. NVIDIA Vera Rubin NVL72 — NVIDIA

1. Yotta to deploy one of Asia's largest NVIDIA Blackwell HGX B300 superclusters with over 20,000 NVIDIA Blackwell Ultra GPUs, 18 February 2026 — Yotta Data Services

1. Shakti Cloud plans and pricing, checked 7 November 2026 — Yotta Shakti Cloud

1. Yotta targets Jan-March 2027 IPO, seeks up to $1.5 billion amid AI boom (Reuters), 2 September 2026 — The Economic Times

1. Yotta Data Services to order 90,000 Nvidia GPUs worth ₹1 lakh crore amid rising AI demand, 11 September 2026 — Kotak Neo

1. Public issues: draft offer documents filed with SEBI, checked 7 November 2026 — Securities and Exchange Board of India

1. AM Intelligence to deploy one of Asia's first NVIDIA Vera Rubin clusters, bringing frontier AI compute to India, 25 August 2026 — CXOToday

1. AM Intelligence expands NVIDIA Vera Rubin AI factory buildout in India and Malaysia, 6 October 2026 — Bernama

1. Yotta data centers in India — Yotta Data Services

1. Yotta media and news — Yotta Data Services

1. NVIDIA Vera Rubin datasheet, October 2026 — NVIDIA

1. NVIDIA sets conference call for third-quarter financial results, 28 October 2026 — NVIDIA Newsroom

1. GTC 2025 announcements and live updates, March 2025 — NVIDIA Blog

1. Eligibility criteria, IndiaAI compute portal — IndiaAI, MeitY

1. Request for empanelment of AI service providers for provisioning of AI compute capacity, C-DAC/GPU/01/2026, 7 October 2026 — C-DAC

1. Date change corrigendum, RFE C-DAC/GPU/01/2026 (bids close 13 November 2026) — C-DAC

1. Digital Personal Data Protection Rules, 2025, G.S.R. 846(E), 13 November 2025 — Ministry of Electronics and Information Technology

1. Tax holiday till 2047 to any foreign company that provides cloud services to customers globally by using data centre services from India, 1 February 2026 — Press Information Bureau

1. The Taxation and Other Laws (Amendment) Bill, 2026 — PRS Legislative Research

1. The Taxation and Other Laws (Amendment) Act, 2026 (No. 21 of 2026), Gazette of India Extraordinary, Part II Section 1, No. 37, 17 August 2026 — Ministry of Law and Justice (Legislative Department)