Jensen Huang's "Dá Chain" has formed a closed loop.

Apple has its Apple Supply Chain, Tesla has its Tesla Supply Chain, and now, Jensen Huang has built the globally spanning "Jensen Chain" for NVIDIA.

Other industrial chains are built around product lines, but the Jensen Chain is different — it is structured around Jensen Huang's five-layer cake theory.

From the most fundamental energy sources, to chips, infrastructure, models, and finally end applications, everything is fully covered within the Jensen Chain.

The latest development is that after Jensen Huang arrived in Tokyo on July 15, in less than 24 hours, he signed a deal significant enough to reshape the whole of Japan, while also completing the closed loop of the Jensen Chain.

In short, Jensen Huang plans to gather all prominent Japanese industrial giants to jointly build an NVIDIA physical AI data center, and make physical AI the core industry of Japan's future.

This is exactly the interesting part of the Jensen Chain: none of the final products from each link in this chain are manufactured by NVIDIA itself, yet Jensen Huang aims to be the chief architect of the entire industrial chain.

So what exactly is the Jensen Chain like? Let's walk through it together.

1

Here is Jensen Huang's specific approach: NVIDIA will partner with the new Japanese company Noetra to build a Vera Rubin AI infrastructure, including 27,500 Rubin GPUs, 140MW of data center capacity, and adoption of the DSX architecture, Spectrum-X network, and BlueField DPU.

This is equivalent to 382 NVL72 rack-level systems, with a total GPU video memory of approximately 20.7TB (HBM4) and a maximum video memory bandwidth of around 1580TB/s. Calculated based on the performance of DeepSeek-V4 Pro (total parameters of 1.6 trillion, pre-trained data of 33 trillion tokens, 49 billion parameters activated per token), this facility can train a large model of the same scale in less than a day.

This project is funded by Japan's national treasury, with the first tranche of 387.3 billion yen (about 24 billion US dollars).

The equity structure mobilizes the full "national strength" of Japan.

Four core players — SoftBank, Sony, NEC, and Honda — each hold a 10% stake, with no single controlling shareholder. The remaining 40 participating enterprises cover almost all core industries in Japan, from manufacturing, automotive, electronics, and communications to finance.

The essence of Noetra is to integrate Japan's entire industrial strength into a single entity to collaborate with NVIDIA.

This facility will also serve as the computing power foundation for Japan's FRONTia initiative, used to train multimodal foundation models for robotics and industrial scenarios; its pre-trained weights are planned to be widely accessible to domestic developers and enterprises in Japan.

On the other hand, industrial robotics leaders including FANUC, Yaskawa, and Kawasaki have collectively joined the Cosmos Physical AI Alliance.

You might wonder: what's so special about that?

This development comes against a major industry backdrop. Over the past decades, Japanese robotics giants like FANUC, Yaskawa, and Kawasaki each had their own closed control systems. When factories installed robots from all three brands at the same time, joint debugging required three separate teams of engineers to modify their respective code, leaving industry standards fragmented.

Joining the Cosmos Alliance this time means the world's most powerful industrial robotics legion has unified its development standards onto NVIDIA's platform, just like how graphics card manufacturers all migrated to CUDA in the past.

The difference between the two is that CUDA previously controlled the computing entry point of the digital world, while Cosmos serves as the data and operation entry point of the physical world.

In the physical AI era, Jensen Huang has once again taken control of the underlying standards.

Japan's goal is to deploy 10 million AI robots across 18 industries by 2040, capturing 30% of the global AI robot market corresponding to an industrial scale of 133 billion US dollars.

The truly valuable part of this NVIDIA-Japan collaboration lies in Japan's decades of accumulated industrial data.

Internet large language models can source corpora from public web pages, but to develop physical AI, robots must learn the physical rules of the real world, such as force, velocity, collision, temperature, and material changes.

Japan hosts the world's most dense concentration of automotive factories, industrial robots, and precision manufacturing production lines, but this data has long been locked away in different enterprises, devices, and control systems — it cannot be directly interconnected, nor can it be easily used to train unified models.

What Jensen Huang aims to do is to convert all this scattered data onto NVIDIA's platform without requiring enterprises to hand over their core trade secrets, and then use Vera Rubin to train foundation models.

In this way, the narrative becomes that NVIDIA has unified Japan's robotics industry and helped it thrive.

2

If robots and applications are the top layer of the Jensen Chain, then the middle layer of the Jensen Chain is production and manufacturing, which is mainly executed by South Korea and China's Taiwan region.

Let's start with South Korea.

In June 2026, Jensen Huang traveled to Seoul and signed a binding agreement with SK Hynix centered on fourth-generation high-bandwidth memory (HBM4), valid through 2030.

In simple terms, NVIDIA paid SK Hynix in advance for HBM4 procurement, allowing the latter to have sufficient capital to build new factories ahead of schedule, while SK Hynix granted NVIDIA priority purchasing rights for HBM4.

Each NVIDIA Rubin GPU is equipped with 288GB of HBM4 memory, at a cost of 18.40 US dollars per GB of memory.

This means the HBM4 material cost for a single GPU is approximately 5,300 US dollars. An NVL72 rack holds 72 Rubin GPUs, with a total HBM4 capacity of 20,736GB, making the HBM4 cost per rack 382,000 US dollars.

For the aforementioned Noetra facility in Japan with 382 NVL72 racks, the total HBM4 material cost is approximately 146 million US dollars.

Worldwide, only three companies can manufacture HBM: SK Hynix, Samsung, and Micron.

On June 5, NVIDIA officially confirmed that all three have passed HBM4 certification for the Rubin platform.

But certification is one thing, and orders are another — the two are separate matters.

According to foreign media reports, 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%.

Why SK Hynix? Because it is the only manufacturer capable of mass shipping both HBM3E and HBM4 simultaneously.

In September 2025, SK Hynix took the lead in developing the world's first 12-layer HBM4; on July 14, 2026, it officially entered mass production.

Although Samsung is larger in scale than SK Hynix, it has long-standing unresolved HBM yield issues, resulting in a smaller shipment share than SK Hynix.

After joining the Jensen Chain, SK Hynix's fate has been completely rewritten.

When SK Hynix was only listed in South Korea, many US funds found it inconvenient to invest, and its valuation was easily determined according to South Korea's manufacturing and memory cycle stock metrics. Therefore, SK Hynix hoped to list on NASDAQ to have investors view it as an "AI infrastructure" company.

On July 9, SK Hynix issued additional ADRs on NASDAQ at a price of 149 US dollars per share, raising 26.5 billion US dollars — effectively completing a "secondary listing" in the United States.

This amount surpassed Alibaba's 25 billion US dollars US IPO in 2014, setting a new record for the largest fundraising by a foreign enterprise listing in the United States. The only larger offering before it was SpaceX's 75 billion US dollars fundraising a month earlier.

On its first day of trading on NASDAQ on July 10, its share price surged about 14%, rising over 18% at one point during the session, with a closing market value of approximately 1.22 trillion US dollars.

AI has brought SK Hynix tremendous profits.

In the first quarter of 2026, its revenue reached 52.6 trillion won, with an operating profit of 37.6 trillion won and an operating margin as high as 72%. Its stock price on the South Korean stock market has risen 634% in the past 12 months. The oversubscription ratio exceeded 7 times, with three institutions including Baillie Gifford and Coatue collectively placing intended subscriptions totaling 7 billion US dollars.

Besides NVIDIA, Google and Amazon are also SK Hynix's customers, queuing up to place orders for HBM3E and HBM4, and even proactively proposing to take stakes in SK Hynix's new production lines and fund the procurement of manufacturing equipment.

The chairman of the SK Group stated that although they plan to double their production capacity within five years, customer feedback consistently says "it's still not enough."

Beyond resolving HBM as the most critical raw material, the middle layer of the Jensen Chain also requires a central assembly workshop.

On May 28, Jensen Huang hosted a "Trillion Dollar Banquet" in Taipei, Taiwan region, attended by business leaders who control the economic lifeline of Taiwan region, including C.C. Wei of TSMC, Young Liu of Foxconn, Barry Lam of Quanta, Lin Hsien-ming of Wistron, Tung Tzu-hsien of Pegatron, and Rick Tsai of MediaTek.

A few days after the banquet, Jensen Huang announced plans to build NVIDIA's new headquarters in the Beitou-Shilin Technology Park in Taiwan region, codenamed Constellation.

He announced that the headquarters will break ground in 2026 and be occupied before 2030, while committing that NVIDIA's annual investment in Taiwan region will reach the hundreds of billions of US dollars level in the future.

Jensen Huang stated: "Chips come from here, packaging comes from here, systems are manufactured here, and AI supercomputers are born here."

A complete Vera Rubin unit has roughly 2 million components.

From chip fabrication, advanced packaging, PCB boards, high-speed connectors, liquid cooling heat dissipation modules, power management, optical communication modules, to rack structures, every link requires the coordination of a large number of factories.

NVIDIA has over 350 ODM and component factories in simultaneous production across 30 countries globally, 150 of which are concentrated in China's Taiwan region.

These 150 supply chain partners operate 25 main factories, supplying over 1 million components to NVIDIA each year. The supply chain scale of Vera Rubin is twice that of the previous generation Grace Blackwell.

The actual manufacturing work is carried out by the three major OEM giants in Taiwan region.

Data from TrendForce shows that Quanta, Wistron, and Foxconn together capture nearly 90% of the global AI server ODM OEM share.

Among them, Foxconn (Industrial FII) accounts for over 40% of NVIDIA's AI server OEM share, serving as the exclusive primary ODM for Vera Rubin full racks, with its capacity utilization rate approaching 95% — nearly operating at full load.

In the second quarter of 2026, Foxconn's revenue grew 39.8% year-on-year, and the proportion of AI server business in Foxconn's total business continues to rise.

Beyond OEM services, TSMC in Taiwan region can also provide chips with 3-nanometer process. As of now, over 70% of TSMC's CoWoS capacity is allocated to NVIDIA.

Such a complete production and assembly chain can only operate smoothly here in the whole world.

3

The United States is the cornerstone of the Jensen Chain. It is NVIDIA's largest customer base, and it is these customers who have made NVIDIA what it is today.

On September 22, 2025, OpenAI and NVIDIA announced plans to deploy at least 10GW of NVIDIA AI systems.

A single nuclear power unit has an installed capacity of about 1GW, so 10GW is equivalent to ten nuclear power plants operating at full capacity, all dedicated to providing computing power for AI.

This agreement plans to start deployment in the second half of 2026 and complete all deployments by the end of 2029.

SpaceX is also a super customer of NVIDIA.

SpaceX has already built two Colossus data centers. The first data center has deployed over 220,000 NVIDIA GPUs (including H100, H200, and about 30,000 GB200), and the second data center is even larger, with over 550,000 GPUs (mainly GB200 and GB300).

Combined, the two data centers spent approximately 18 billion US dollars just on GPU procurement. Elon Musk also leased all the computing power of Colossus No.1 to Anthropic, collecting 1.25 billion US dollars per month, with a contract running through May 2029 with a total value of about 400 billion US dollars.

Google leases part of the computing power of Colossus No.2 for 920 million US dollars per month.

Then there's Meta.

In April 2026, Meta raised its full-year capital expenditure guidance to 125-145 billion US dollars, almost doubling from 72.2 billion US dollars in 2025. Its largest infrastructure investment is the Hyperion supercomputing center in Louisiana.

When Hyperion was first announced in 2024, its expected computing power scale was 2GW. On July 13, 2026, Meta added a 40 billion US dollar investment to expand its computing power scale to 5GW. As a result, Hyperion has become the largest data center project in Meta's history.

Anthropic's largest order also comes from NVIDIA.

In November 2025, Microsoft and NVIDIA jointly invested in Anthropic, with Microsoft contributing up to 50 billion US dollars and NVIDIA up to 100 billion US dollars.

In exchange, Anthropic committed to purchasing a total of 300 billion US dollars worth of Azure computing power services, using up to 1GW of computing power from NVIDIA Grace Blackwell and Vera Rubin systems.

NVIDIA's full-year revenue for FY2026 reached 215.9 billion US dollars, a year-on-year increase of 65%, of which data center revenue was 193.7 billion US dollars, accounting for nearly 90% of total revenue. In the latest quarter, data center revenue surged 92% year-on-year.

It's easy to see that NVIDIA has made enormous profits here.

But the problem arises precisely here: the foundation of the Jensen Chain has begun to shake, as NVIDIA's most core group of customers are starting to develop their own chips.

On June 24, 2026, OpenAI partnered with Broadcom to officially launch its first self-developed AI inference chip, codenamed Jalapeño, which means "Mexican chili pepper" in Spanish.

From project initiation and design to successful tape-out, this chip took only 9 months.

OpenAI's AI models directly participated in chip architecture design, power consumption simulation, and reinforcement learning optimization, while Broadcom was responsible for manufacturing.

Broadcom CEO Hock Tan revealed, "Early tests show that compared to current mainstream AI chips, Jalapeño can reduce inference costs by approximately 50%."

Especially for OpenAI, which has over 800 million monthly active users, even a tiny reduction in the cost per token will ultimately result in astronomical savings.

The first batch of engineering samples has successfully run GPT-5.3, Codex, and Spark models in the laboratory at the frequency and power consumption standards for mass production. According to OpenAI's plans, small-scale deployment will begin at the end of 2026, rapid scaling in 2027, full large-scale mass production in 2028, with a maximum total power consumption of 10GW in the long term.

Anthropic has gone further than OpenAI, weaving a huge computing power network.

On one hand, it uses Amazon's Trainium chips to train and deploy Claude, on the other hand, it uses Google's TPUs, while not forgetting NVIDIA's GPUs.

Currently, Anthropic has deployed over 1 million Trainium 2 chips. Moreover, Anthropic's engineers are directly involved in the low-level kernel development of Amazon's Annapurna Labs chips.

Furthermore, Anthropic has expanded its partnership with Amazon, committing to investing over 100 billion US dollars on AWS in the next ten years to lock in up to 5GW of new computing power.

In April, Anthropic also signed agreements with Google and