Indian data center operator Yotta는 AI 컴퓨팅용 80,000개의 NVIDIA Vera Rubin (H200) GPU 배포를 계획 중입니다.

Yotta Data Services, backed by the Hiranandani Group and controlling an estimated 60 to 70 percent of India's installed GPU capacity, announced plans on September 11, 2026, to deploy 80,000 Nvidia Vera Rubin chips across two new facilities in a $12 billion expansion. The announcement, made at GFF Mumbai by CEO Sunil Gupta in an exclusive interview with MoneyControl, positions Yotta as the first Indian company—neither a subsidiary of a US hyperscaler nor backed by sovereign wealth—to assemble a GPU cluster comparable to the AI factories operated by Amazon, Microsoft, and Google. The disclosure comes weeks after Yotta filed plans to raise up to $1.5 billion through a public offering, placing investor scrutiny on the company's infrastructure strategy at a critical moment.

The expansion will span two facilities: 40,000 Vera Rubin chips at the D4 data center under construction in Greater Noida—a 120-megawatt facility—and 40,000 GB300 (Blackwell Ultra) chips at the 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 investment represents one of the largest single capital commitments by an Indian company in AI infrastructure, arriving as India's technology policy establishment explicitly treats advanced GPU capacity as a sovereign strategic asset.

Nvidia's Vera Rubin platform—named after the American astronomer whose galaxy rotation curve research confirmed dark matter's existence—succeeds the Blackwell Ultra GPU family. Each Rubin GPU is a dual-die chip fabricated on TSMC's 3-nanometer process, with full production beginning in June 2026 following the architecture announcement at CES 2026. Each Rubin GPU carries 288 gigabytes of HBM4 memory delivering 22 terabytes per second of bandwidth—2.75 times Blackwell Ultra's 8 terabytes per second in HBM3e. Single-chip FP4 inference performance reaches 50 petaflops, roughly 2.5 times Blackwell Ultra's per-chip compute. In the NVL72 rack configuration—72 Rubin GPUs and 36 Vera CPUs connected by NVLink 6, carrying 260 terabytes per second of all-to-all fabric bandwidth—a single rack delivers 3.6 exaflops of theoretical AI compute peak.

The performance gains demand proportional power infrastructure. A single VR200 NVL72 rack draws 190 to 230 kilowatts—compared to 132 to 142 kilowatts for Blackwell Ultra's NVL72 predecessor and roughly 40 kilowatts for Hopper-era racks—and requires 100 percent liquid cooling with no air-cooled fallback. Vera Rubin entered full production in June 2026, with partner deployments beginning in the second half of this year at AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius, and Nscale.

Yotta's combined facility capacity—120 megawatts at D4 and 80 megawatts at NM2, totaling 200 megawatts—warrants scrutiny against the hardware specifications. At 80,000 GPUs deployed in 72-GPU NVL72 racks, approximately 1,111 rack positions are required. At the VR200's rated 190 to 230 kilowatts per rack, these racks draw between 211 and 255 megawatts at full load before cooling overhead, with Supermicro implementations sized for approximately 227 kilowatts per rack. The 200 megawatt combined capacity sits at the mathematical lower bound of peak power requirements—and this figures total facility power, not IT load alone. A data center achieving a power usage effectiveness of 1.2, industry-leading for high-density liquid-cooled facilities, would allocate only 167 megawatts to IT equipment. The likeliest scenario is phased deployment: D4 commissioning targets May through August 2027, while NM2 buildout continues. Yotta has not disclosed a specific timeline for full 80,000-GPU deployment, though the 200 megawatt figure establishes a near-term ceiling rather than fixed capacity—both campuses sit within larger envelopes with scalability potential (Greater Noida to 250 megawatts; Navi Mumbai roadmapped to 2 gigawatts).

A second timeline risk concerns Nvidia's next-generation architecture, codenamed Kyber for the Rubin Ultra generation, specified at approximately 600 kilowatts per rack and requiring 800-volt DC power distribution infrastructure—a fundamental redesign from current AC distribution. Kyber targets the second half of 2027, meaning Yotta's current VR200 facilities at 190 to 230 kilowatts will require significant infrastructure upgrades or face obsolescence. Operators adopting Kyber first will achieve compute-per-megawatt advantages Yotta cannot match without additional capital, a calculus Yotta's IPO investors must price.

Nscale, a European cloud provider, announced in September 2026 a 100,000 Vera Rubin GPU cluster at its Barstow, Texas facility for second-half 2027 deployment. Yotta's 80,000 GPU plan, if executed, would represent roughly 80 percent of Nscale's target and rank among the largest announced GPU cluster deployments outside the United States.

India's context sharpens the significance beyond raw GPU counts. The country generates nearly 20 percent of the world's data but held just 3 percent of global data center capacity as of 2025. Total Indian data center capacity has reached approximately 1.7 gigawatts, with expansion projected significantly through 2030. Cumulative investment in India's data center sector has grown substantially through US hyperscaler commitments: 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, totaling more than $67 billion.

What distinguishes Yotta's announcement is operational identity. Google, Microsoft, and Amazon extend their global infrastructure into a new market; Yotta assembles sovereign compute capacity—hardware under Indian operational control, within India's legal jurisdiction, financed partly through Indian capital markets via the planned IPO.

Sunil Gupta has spent 2026 articulating sovereign AI infrastructure's meaning. At the AI Impact Summit in February 2026, he framed sovereignty not as isolation but as "strategic control"—ensuring "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."

Cellucci and Singh, 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 the third dimension: the physical substrate.

For enterprise customers using Yotta's infrastructure, the practical implication is data residency: AI workloads at Yotta's facilities remain within India's legal jurisdiction and do not route through US-controlled infrastructure. For Indian enterprises navigating the Digital Personal Data Protection Act of 2023—whose implementing rules were finalized [text truncates here]