For the first time, an Indian company that is neither a subsidiary of a US hyperscaler nor backed by sovereign wealth has announced plans to assemble a GPU cluster large enough to be measured in the same breath as dedicated AI factories run by Amazon, Microsoft, and Google. Yotta Data Services, the Hiranandani Group-backed operator that already controls an estimated 60 to 70 percent of India’s installed GPU capacity, made the announcement at GFF Mumbai on September 11, 2026: it plans to deploy 80,000 Nvidia Vera Rubin chips across two new facilities in a $12 billion expansion. The disclosure arrives weeks after Yotta filed plans to raise up to $1.5 billion in a public offering — placing the financial credibility of this infrastructure bet under investor scrutiny at the same moment the company is asking the public markets to price it.

Two Campuses, One Claim

CEO Sunil Gupta disclosed the details in an exclusive interview with MoneyControl on the sidelines of the conference. The plan calls for deploying Nvidia’s latest GPU generations across two facilities: 40,000 Vera Rubin chips at Yotta’s D4 data center under construction in Greater Noida — a 120-megawatt facility — and 40,000 GB300 (Blackwell Ultra) chips at its NM2 campus in Navi Mumbai, an 80-megawatt site currently being fitted out.

“We are fitting out our next data centre in Mumbai, NM2, which is an 80-megawatt data centre,” Gupta said. “We are planning to put around 40,000 NVIDIA GB300 chips in it. We are also starting construction of our D4 data centre in the Greater Noida campus that will be a 120 MW data centre. That building will have 40,000 Vera Rubin chips.”

The $12 billion total investment represents one of the largest single capital commitments by an Indian company in AI infrastructure and arrives at a moment when the country’s technology policy establishment is explicitly treating advanced GPU capacity as a sovereign strategic asset.

What Vera Rubin Actually Is

Nvidia’s Vera Rubin platform — named after the American astronomer whose galaxy rotation curve research confirmed the existence of dark matter — is the successor to the Blackwell Ultra GPU family. Each Rubin GPU is a dual-die chip fabricated on TSMC’s 3-nanometer process. Nvidia’s Vera Rubin product page details the full architecture announced at CES 2026, which entered full production in June 2026.

Each Rubin GPU carries 288 gigabytes of HBM4 memory running at 22 terabytes per second of bandwidth — 2.75 times Blackwell Ultra’s bandwidth of 8 terabytes per second in HBM3e. A single GPU delivers 50 petaflops of FP4 inference performance, roughly 2.5 times the per-chip compute of Blackwell Ultra. In the NVL72 rack-scale configuration — 72 Rubin GPUs and 36 Vera CPUs bound together by NVLink 6, which carries 260 terabytes per second of total all-to-all fabric bandwidth — a single rack delivers 3.6 exaflops of theoretical AI compute peak.

The performance gains come with a commensurate power demand that is forcing a redesign of what “hyperscale data center” means in practice. A single VR200 NVL72 rack draws approximately 190 to 230 kilowatts — compared to 132 to 142 kilowatts for its Blackwell Ultra GB300 NVL72 predecessor and roughly 40 kilowatts for a Hopper-era rack — and requires 100 percent liquid cooling, with no air-cooled fallback. Nvidia entered full production on Vera Rubin in June 2026, with Vera Rubin partner deployments beginning in the second half of this year. Initial deployments have been confirmed at AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius, and Nscale.

What 200 Megawatts Actually Powers

The combined facility capacity Yotta has announced for these two GPU campuses — 120 megawatts at D4, 80 megawatts at NM2, totaling 200 megawatts — is a number worth examining against the hardware it is intended to house. At 80,000 GPUs operating in 72-GPU NVL72 racks, the deployment requires approximately 1,111 rack positions. At the VR200’s rated 190 to 230 kilowatts per rack, those racks draw between 211 and 255 megawatts of power at full load — before cooling overhead is added. Supermicro’s Vera Rubin NVL72 implementations are sized for approximately 227 kilowatts per rack.

This means the 200 megawatt combined capacity is at the mathematical lower bound of what full deployment at peak rack power requires — and that is the total facility power, not just IT load. A data center operating at a power usage effectiveness of 1.2 (industry-leading for high-density liquid-cooled facilities) would make only 167 megawatts of that available for IT equipment. The likeliest resolution is phased deployment: the D4 facility timeline puts go-live at May through August 2027, and the NM2 buildout is currently in progress. Yotta has not disclosed a specific commissioning timeline for the full 80,000-GPU count. What the 200 megawatt figure establishes is a near-term ceiling, not a fixed total capacity — both campuses are sited within larger envelopes (Greater Noida is scalable to 250 megawatts; Navi Mumbai campus has a roadmap to 2 gigawatts).

There is a second timeline risk embedded in the roadmap. Nvidia’s next-generation rack architecture, codenamed “Kyber” and targeting the Rubin Ultra generation, is specified at approximately 600 kilowatts per rack and requires 800-volt DC power distribution infrastructure — a fundamental redesign from the standard AC distribution in current VR200 facilities. Kyber is targeted for the second half of 2027. That means the facilities Yotta is building now for VR200 racks at 190 to 230 kilowatts will require significant power infrastructure upgrades — or will not support the generation after next. Operators that adopt Kyber first will offer a compute-per-megawatt advantage Yotta’s current buildout cannot match without additional capital expenditure. This is the upgrade-cycle math that Yotta’s IPO investors will need to price.

Where This Sits on the Global GPU Map

The scale of Yotta’s announced deployment places it in a small set of operators globally. Nscale, a European cloud provider, announced in September 2026 a 100,000 Vera Rubin GPU cluster at its Barstow, Texas facility — targeted for the second half of 2027. Yotta’s 80,000 GPU plan, if executed, would represent roughly 80 percent of Nscale’s targeted count and would be among the largest announced GPU cluster deployments outside the United States.

The broader India context is relevant to understanding why this matters beyond the raw GPU number. India generates nearly 20 percent of the world’s data but as of 2025 held just 3 percent of global capacity in data centers. India’s total data center capacity has reached approximately 1.7 gigawatts, with capacity projected to expand significantly through 2030. Cumulative investment in India’s data center sector has grown substantially, driven by Google, Microsoft, and AWS commitments totaling more than $67 billion — including Google’s $15 billion AI hub in Visakhapatnam, Microsoft’s $17.5 billion pledge through 2029, and AWS’s target of up to $35 billion by 2030.

What distinguishes Yotta’s announcement from those hyperscaler commitments is the operator’s identity. Google, Microsoft, and Amazon building data centers in India are extending their global infrastructure into a new market. Yotta building the same tier of infrastructure is an Indian domestic company assembling sovereign compute capacity — hardware under Indian operational control, within India’s legal jurisdiction, financed in part by Indian capital markets through the planned IPO.

Sovereign AI: What the Term Actually Means for Investors

Sunil Gupta has spent much of 2026 articulating what he means by sovereign AI infrastructure. At the AI Impact Summit in February 2026, he described sovereignty not as isolation but as “strategic control” — ensuring that “no single country or company can dictate a nation’s digital future.” In January 2026, he told Business Today that India’s five-year roadmap required “sovereign compute at scale, affordable power and creating deep talent and research ecosystems.”

Researchers who study sovereign AI infrastructure offer a more precise framing. Cellucci and Singh, in work cited across multiple 2026 academic analyses, define sovereign AI across four dimensions: training data sovereignty, model sovereignty, infrastructure sovereignty, and interaction sovereignty — with infrastructure sovereignty defined as maintaining “compute, storage, and inference infrastructure under national or institutional jurisdiction.” Yotta’s buildout addresses dimension three: the physical substrate.

The practical implication for enterprise customers using Yotta’s infrastructure is data residency: AI workloads processed at Yotta’s facilities stay within India’s legal jurisdiction and do not route through US-controlled infrastructure. For Indian enterprise customers navigating the Digital Personal Data Protection Act of 2023 — whose implementing rules were finalized in November 2025 — domestic processing offers a materially simpler compliance posture. For Yotta’s predominantly international customer base (75 to 80 percent of revenue comes from US and European clients), it offers low-latency South Asian inference without the sovereignty complications of routing through a foreign cloud. India’s data center tax holiday, a 20-year exemption for foreign firms using domestic data centers announced in the February 2026 budget, adds a financial incentive on top of the regulatory one.

The sovereign framing also has limits. Gupta himself acknowledged them at the AI Impact Summit: “We will always remain interconnected and interdependent across the world.” Every Vera Rubin GPU Yotta deploys is designed by Nvidia, fabricated by TSMC in Taiwan, and packaged using CoWoS advanced packaging technology — supply chains that are entirely outside Indian control. Export restrictions that cut off India from high-end GPU supply would immediately terminate Yotta’s sovereign compute ambitions regardless of where the data centers sit. Yotta’s sovereignty is infrastructure-layer sovereignty, not full-stack sovereignty. The CNAS Sovereign AI Index notes that nations in the highest tier of AI infrastructure development still depend on chips and toolchains built and controlled outside their borders.

Yotta’s Expansion History and the IPO Equation

Yotta has assembled its current GPU portfolio faster than virtually any company outside the US hyperscaler tier. The company began sourcing Nvidia GPUs in 2023 and in February 2026 unveiled its Blackwell Ultra supercluster — 20,736 liquid-cooled Nvidia Blackwell Ultra GPUs at its 60-megawatt D2 facility in Greater Noida, backed by more than $2 billion in investment. Nvidia simultaneously committed to establishing one of Asia-Pacific’s largest DGX Cloud clusters within that supercluster under a four-year contract worth approximately $1 billion — a validation from the world’s dominant AI chip supplier that treated Yotta as a tier-one deployment partner. The D2 facility was targeted to go live by August 2026.

In June 2026, Frost & Sullivan’s 2026 India award named Yotta its Indian Company of the Year in the AI infrastructure category, citing the company’s Tier IV-certified Navi Mumbai facility and its role supplying 75 percent of the IndiaAI Mission’s advanced GPU capacity — India’s government compute program, approved in March 2024 with ₹10,371 crore in government funding (approximately $1.09 billion), that is targeting 100,000 publicly accessible GPUs by December 2026.

In September 2026, ahead of the Global Fintech Fest announcement, Yotta’s Reuters IPO interview confirmed plans to raise up to $1.5 billion in a public offering on Indian exchanges, with a draft prospectus expected to be filed with the Securities and Exchange Board of India as early as October. The company has already raised $150 million in primary growth capital from high-net-worth individuals and family offices at a valuation of approximately ₹370 billion (approximately $3.87 billion), and Gupta said the IPO portion was expected to be smaller than originally planned because much of its fundraising target had already been met. The company is targeting a potential valuation of up to $6 billion at listing.

To manage balance-sheet pressure from GPU procurement costs, Yotta is also exploring financing structures in which partners purchase GPUs through special-purpose vehicles, share revenue generated by the chips, and eventually transfer ownership to Yotta after four to five years. This GPU revenue-share model defers capital requirements but creates future obligations not immediately visible on the company’s balance sheet — a financing structure that IPO investors and their advisors will need to model carefully when evaluating the prospectus.

Incident at a Customer’s Server

Context relevant to Yotta’s security posture: in July 2026, the ransomware group World Leaks claimed to have published approximately 858,000 files belonging to Reliance Group, including approximately 19,000 with potentially sensitive information connected to the Kudankulam Nuclear Power Plant in Tamil Nadu. Reliance Group confirmed a partial data breach on a server hosted by Yotta Data Services. Yotta said it had detected suspicious activity on May 29 involving the Reliance Infrastructure server, had immediately terminated it, and had prevented the suspected execution of ransomware. India’s CERT-In cybersecurity agency is examining the incident. The breach is a reminder that data center operators are accountable for the security posture of infrastructure they operate on behalf of customers — a responsibility that scales with the sensitivity of those customers’ workloads at 80,000 GPU cluster density.

Does India’s Grid Support What Yotta Is Building?

The power math above raised a question about Yotta’s 200-megawatt facility envelope. A separate question is whether India’s power grid can reliably supply it. India surpassed its 50 percent non-fossil fuel capacity goal five years ahead of schedule, and data center operators in India are increasingly investing in captive solar farms and battery energy storage systems to ensure 24-hour uptime. However, academic analysis of AI data center grid impacts notes that India’s incremental data center electricity demand is more likely to depend on existing fossil-dominated generation while renewable and transmission investment catches up — with coal continuing to supply the majority of grid electricity and grid reliability challenges leading operators to rely on backup generation and grid upgrades. Both the Greater Noida and Navi Mumbai campuses are purpose-built with dedicated power infrastructure; neither relies on standard commercial grid delivery without engineering backstops. But the constraint is real and is not resolved simply by announcing megawatt capacity figures.

What Yotta’s Bet Means for India’s AI Trajectory

The largest structural implication of Yotta’s announcement is geographic. Since frontier AI infrastructure emerged as a strategically significant asset class, overwhelmingly concentrated in the United States with a secondary cluster in China, the question has been which countries and which operators could assemble comparable compute capacity at comparable scale. A list of operators outside the US and China who have announced GPU clusters in the 80,000-GPU range is a very short list, and Yotta is on it. Unlike Japan’s government-backed sovereign AI programs or the Gulf states’ sovereign wealth-funded AI investment funds, Yotta’s buildout is anchored in private capital raised through commercial credit markets and, shortly, public equity. That means its viability is subject to market discipline in a way that state-funded competitors’ programs are not — which is both a structural strength (investor scrutiny enforces capital efficiency) and a structural vulnerability (if the IPO misses its target valuation or debt markets tighten, the $12 billion buildout timeline compresses).

For Nvidia, a confirmed 80,000-unit Vera Rubin order from an Indian independent operator — arriving at the start of the platform’s volume shipment ramp — is meaningful demand signal. Jensen Huang’s $1 trillion revenue projection through 2027 for Blackwell and Rubin combined requires broad adoption from operators willing to make high-capital early commitments. Orders of Yotta’s scale help demonstrate that appetite extends well beyond the hyperscaler tier.

Currency conversions in this article are approximate, based on an exchange rate of approximately ₹95.64 to the US dollar as of September 12, 2026, and are subject to change.

Frequently Asked Questions

What makes Vera Rubin different from the Blackwell Ultra GPUs Yotta deployed earlier in 2026?

The generational jump from Blackwell Ultra (GB300) to Vera Rubin involves three compounding improvements at the silicon level. Memory bandwidth per GPU increases from 8 terabytes per second (HBM3e) to 22 terabytes per second (HBM4) — a 2.75x improvement that directly determines how fast an AI model can access its weights and activations during inference. The NVLink 6 interconnect doubles total rack-level all-to-all fabric bandwidth from 130 to 260 terabytes per second, enabling larger models to run across multiple GPUs without network bandwidth becoming the bottleneck. And each Rubin GPU delivers approximately 50 petaflops of FP4 inference — roughly 2.5 times the per-chip compute of Blackwell Ultra. The tradeoff is power: at 190 to 230 kilowatts per NVL72 rack, versus 132 to 142 kilowatts for the Blackwell Ultra NVL72, the infrastructure demands per rack rise by approximately 50 to 70 percent. Vera Rubin also mandates 100 percent liquid cooling with no air-cooled alternative, which means any facility hosting it must have been engineered with direct-to-chip cooling loops from the start.

When is Yotta’s IPO, and what risks should investors watch?

Yotta plans to file its draft red herring prospectus (DRHP) with India’s Securities and Exchange Board (SEBI) as early as October 2026, targeting a public offering in the January to March 2027 window. The company is seeking up to $1.5 billion in proceeds, with uses including debt repayment, GPU purchases, and sovereign cloud expansion. The pre-IPO valuation was approximately ₹370 billion (roughly $3.87 billion); Yotta’s target listing valuation is up to $6 billion. Three risks deserve investor attention. First, GPU procurement risk: the $12 billion expansion depends on Nvidia GPU supply at a moment when TSMC CoWoS advanced packaging capacity is the binding bottleneck with lead times of 52 to 78 weeks — meaning delivery timelines are exposed to supply-chain disruptions. Second, upgrade-cycle obsolescence: Nvidia’s next-generation Kyber rack architecture (targeting 600 kilowatts per rack) is expected in the second half of 2027; facilities designed for today’s 190 to 230 kilowatt VR200 racks will require significant power infrastructure upgrades to support the following generation. Third, revenue concentration: approximately 75 to 80 percent of Yotta’s customer base is international, meaning a significant portion of IPO-era revenue depends on foreign enterprise spending in Indian data centers continuing at current pace. Full details are available in Yotta’s September 2026 IPO announcement and the company’s DRHP filing timeline.

What is “sovereign AI compute,” and does an Indian private company actually qualify?

Sovereign AI infrastructure, as defined by researchers at the Center for a New American Security and in academic work by Cellucci and Singh, refers to compute, storage, and inference infrastructure maintained “under national or institutional jurisdiction” — meaning it operates under the legal authority of the relevant state and is not subject to termination by a foreign government or corporation. Yotta’s infrastructure qualifies on that definition: the facilities are on Indian soil, under Indian corporate ownership, subject to Indian law. The limitation is that the hardware — every Nvidia Vera Rubin GPU — is designed in the United States, fabricated in Taiwan, and assembled using supply chains outside Indian control. True full-stack sovereignty would require domestic chip design and fabrication capacity India does not yet possess. What Yotta provides is infrastructure-layer sovereignty: the data stays in India, the operations are under Indian management, and the legal liability for the workloads runs through Indian law. That is meaningful — and it is also less than the full independence the word “sovereign” implies. The CNAS Sovereign AI Index tracks where nations stand on this spectrum globally.

Will the 200 megawatts of announced facility capacity be enough to run 80,000 Vera Rubin GPUs at full load?

Almost certainly not simultaneously, at full rack power — and that is likely deliberate. At 1,111 NVL72 racks (the number required for 80,000 GPUs at 72 GPUs per rack) running at 190 to 230 kilowatts each, total GPU load alone is 211 to 255 megawatts, before cooling and other facility overhead. Yotta’s announced 200 megawatts combined for both facilities is therefore at or below what full simultaneous deployment at maximum power would require. The likeliest resolution is phased commissioning: the D4 Greater Noida facility is not expected to go live until May through August 2027, and the NM2 Navi Mumbai campus is currently being fitted out. Yotta has not disclosed a specific commissioning timeline for the full 80,000-GPU count, and both campuses are sited within larger-capacity envelopes (250 megawatts scalable at Greater Noida; 2 gigawatts roadmap at Navi Mumbai) that can accommodate future power expansion. The 200 megawatt announcement describes current committed facility capacity, not final buildout ceiling. Supermicro’s Vera Rubin NVL72 specifications and Yotta’s campus scalability roadmap provide further technical context.