Jensen Huang has constructed what insiders call the "Jensen Chain," a vertically integrated industrial ecosystem that controls the foundational standards for AI infrastructure without NVIDIA manufacturing most of the final products. Unlike traditional supply chains built around specific product lines, this network spans from energy sources through semiconductor manufacturing, data center infrastructure, AI model development, and real-world applications like robotics. The architecture mirrors how graphics card makers once unified around NVIDIA's CUDA computing platform, except this time the stakes involve controlling the emerging physical AI industry.

What Is the Jensen Chain and How Does It Work?

The Jensen Chain operates across five distinct layers, each handled by specialized partners rather than NVIDIA itself. At the foundation sits energy infrastructure, followed by chip manufacturing, then data center systems, AI model training, and finally end-user applications. This structure allows Huang to function as the chief architect of an entire industrial ecosystem without directly producing most components. The strategy proved particularly effective in Japan, where Huang arrived on July 15, 2026, and within 24 hours signed a deal that would reshape the nation's AI infrastructure.

Japan's new initiative centers on a company called Noetra, which will build the Vera Rubin AI infrastructure with support from Japan's national treasury. The first funding tranche totals 387.3 billion yen, approximately 24 billion US dollars. The equity structure mobilizes Japan's entire industrial base, with four core players each holding 10% stakes: SoftBank, Sony, NEC, and Honda. The remaining 40 participating enterprises span manufacturing, automotive, electronics, communications, and finance.

How Does the Vera Rubin Data Center Compare to Other AI Infrastructure?

The Vera Rubin facility represents one of the most ambitious AI infrastructure projects globally. It will contain 27,500 Rubin GPUs (graphics processing units) with 140 megawatts of data center capacity. This translates to approximately 382 NVL72 rack-level systems, with a total GPU video memory of roughly 20.7 terabytes using HBM4 (high-bandwidth memory, fourth generation) and maximum video memory bandwidth around 1,580 terabytes per second. Based on the performance characteristics of DeepSeek-V4 Pro, a model with 1.6 trillion total parameters and 33 trillion tokens of training data, this facility could train a comparable large-scale model in less than a day.

The infrastructure will serve as the computing foundation for Japan's FRONTia initiative, which aims to train multimodal foundation models for robotics and industrial applications. Pre-trained weights from these models are planned to be widely accessible to domestic developers and enterprises throughout Japan, creating a shared foundation for the nation's AI robotics industry.

Why Are Japanese Robotics Companies Unifying Around NVIDIA's Standards?

For decades, Japanese robotics giants FANUC, Yaskawa, and Kawasaki operated with separate, incompatible control systems. When factories installed robots from multiple manufacturers simultaneously, joint debugging required three separate engineering teams to modify code for each system, leaving industry standards fragmented. By joining the Cosmos Physical AI Alliance, these companies have unified their development standards onto NVIDIA's platform. This mirrors the historical shift when graphics card manufacturers converged on NVIDIA's CUDA computing standard, except Cosmos serves as the data and operation entry point for the physical world rather than the digital computing layer.

Japan's ambition extends far beyond current robotics capabilities. The nation aims to deploy 10 million AI robots across 18 industries by 2040, targeting 30% of the global AI robot market. This corresponds to an industrial scale of approximately 133 billion US dollars. The robotics unification represents a critical step toward achieving this goal by eliminating technical fragmentation.

What Makes Japan's Industrial Data Valuable for Physical AI?

The true strategic value of the NVIDIA-Japan collaboration lies in Japan's decades of accumulated industrial data. Internet-based large language models can source training material from public web pages, but physical AI systems must learn the actual rules governing the physical world: force, velocity, collision, temperature, and material changes. Japan hosts the world's densest concentration of automotive factories, industrial robots, and precision manufacturing production lines. However, this data has remained locked within individual enterprises, devices, and control systems, unable to be directly interconnected or easily used to train unified models.

Jensen Huang's strategy involves converting all this scattered data onto NVIDIA's platform without requiring enterprises to surrender their core trade secrets. The Vera Rubin facility will then use this consolidated data to train foundation models that can be shared across Japan's industrial base. This approach allows NVIDIA to position itself as the unifying force that helps Japan's robotics industry thrive while simultaneously establishing itself as the underlying standard for physical AI development.

How Does NVIDIA Secure Its Supply Chain for Advanced Chips?

The middle layer of the Jensen Chain involves production and manufacturing, primarily executed by South Korea and Taiwan. In June 2026, Huang traveled to Seoul and signed a binding agreement with SK Hynix focused on HBM4 memory, valid through 2030. Under this arrangement, NVIDIA paid SK Hynix in advance for HBM4 procurement, providing the manufacturer with sufficient capital to build new factories ahead of schedule. In return, SK Hynix granted NVIDIA priority purchasing rights for HBM4 supplies.

Each NVIDIA Rubin GPU requires 288 gigabytes of HBM4 memory at a cost of 18.40 US dollars per gigabyte, making the memory cost per GPU approximately 5,300 US dollars. An NVL72 rack containing 72 Rubin GPUs has a total HBM4 capacity of 20,736 gigabytes, resulting in an HBM4 cost per rack of 382,000 US dollars. For the Noetra facility with 382 NVL72 racks, the total HBM4 material cost reaches approximately 146 million US dollars.

Steps to Understanding NVIDIA's Competitive Advantage in Memory Supply

• Manufacturing Concentration: Only three companies globally can manufacture HBM: SK Hynix, Samsung, and Micron. On June 5, 2026, NVIDIA officially confirmed that all three passed HBM4 certification for the Rubin platform, but certification differs significantly from actual order allocation.
• Market Share Dominance: SK Hynix secured approximately 70% of NVIDIA's first batch of HBM4 orders, far exceeding the industry's previous estimate of 50%. Samsung and Micron split the remaining 30%, demonstrating NVIDIA's preference for SK Hynix's manufacturing capabilities.
• Production Capability: SK Hynix is the only manufacturer capable of mass shipping both HBM3E and HBM4 simultaneously. In September 2025, SK Hynix developed the world's first 12-layer HBM4, and on July 14, 2026, it officially entered mass production, giving it a significant timing advantage over competitors.
• Yield and Scale Issues: Although Samsung operates at larger scale than SK Hynix, it has long-standing unresolved HBM yield issues, resulting in smaller shipment volumes. This manufacturing challenge has allowed SK Hynix to capture the dominant position in NVIDIA's supply chain.

SK Hynix's position within the Jensen Chain has been completely rewritten. When the company was listed only in South Korea, many US investment funds found it inconvenient to invest, and its valuation was determined primarily by South Korea's manufacturing and memory cycle stock metrics. SK Hynix sought a NASDAQ listing to be viewed as an "AI infrastructure" company rather than a traditional memory manufacturer.

On July 9, 2026, SK Hynix issued additional American Depositary Receipts (ADRs) on NASDAQ at 149 US dollars per share, raising 26.5 billion US dollars and effectively completing a secondary listing in the United States. This amount surpassed Alibaba's 25 billion US dollars US initial public offering in 2014, setting a new record for the largest fundraising by a foreign enterprise listing in the United States. Only SpaceX's 75 billion US dollars fundraising a month earlier exceeded this amount. On its first day of trading on NASDAQ on July 10, SK Hynix's share price surged approximately 14%, demonstrating strong investor confidence in its role within NVIDIA's supply chain.

The Jensen Chain represents a fundamental shift in how technology ecosystems organize themselves. Rather than controlling every step of production, NVIDIA has positioned itself as the orchestrator of an entire industrial network, establishing the underlying standards and architectural principles that govern how the world builds physical AI systems. By securing Japan's industrial robotics sector, controlling memory supply through strategic partnerships, and unifying fragmented manufacturing standards, Huang has created a self-reinforcing system where NVIDIA's dominance becomes increasingly difficult to challenge.