Jensen Huang Finally Went to Japan — and the Deals He Signed Reveal How Nvidia Is Splitting Asia

In June, Jensen Huang toured Asia — South Korea, Taiwan — and did not stop in Japan. Japanese media gave the snub a name: “Japan passing.” For the world’s third-largest economy, home to Sony, Honda, Fanuc and the chemical suppliers without which advanced chips cannot be made, being skipped by the CEO whose company sits at the center of the AI boom was not a scheduling detail. It read as a verdict.

On July 15 and 16, Huang answered it. Two days in Tokyo produced a $6.2 billion government-backed national AI factory, robotics partnerships with Japan’s biggest industrial arms makers, a 30th-anniversary tribute to the company that once saved Nvidia from bankruptcy, and a genuinely strange amount of street food. Taken together, the announcements are more interesting than the apology tour framing suggests, because they quietly confirm something about how Nvidia now sees Asia: not as one market, but as three specialized suppliers with very different jobs.

What was actually signed

The centerpiece is Noetra Corp., a new consortium anchored by SoftBank, Sony, NEC and Honda, with roughly 44 companies and organizations participating as investors. Noetra will buy 27,500 Nvidia Rubin GPUs and 13,750 Vera CPUs — exactly 382 Vera Rubin NVL72 racks — to build a 140-megawatt AI factory. Backed by Japan’s Ministry of Economy, Trade and Industry (METI) under its FRONTia Project, and carrying a roughly $6.2 billion government commitment, Nvidia is calling it the world’s first national AI infrastructure built specifically for physical AI. Construction starts April 2027; full operation is targeted for June 2028.

Alongside it came the robotics layer. Nvidia announced partnerships with Fanuc and Yaskawa Electric — two of the handful of companies that dominate global industrial robot arms — while Fujitsu separately announced a physical-AI collaboration with Fanuc, Yaskawa and Kawasaki Heavy Industries, using Nvidia technology as the backbone. “With AI, robots will become smart, easily adaptable and accessible,” Huang said in Tokyo.

And then there was Sega. Huang’s Akihabara appearance marked 30 years of the Nvidia–Sega relationship, and the reason it matters is not nostalgia: in the mid-1990s, with Nvidia near bankruptcy, Sega put in a $5 million investment that kept the company alive. Framing the visit around that debt was a deft piece of diplomacy — it recast a make-up trip as a homecoming.

The three-way split

Here is the part worth sitting with. Read the Japan announcements next to what Nvidia has been doing in Korea and Taiwan, and a clear division of labor appears.

Taiwan makes the chips. TSMC fabricates the accelerators; that has not changed and is not being diversified away in any meaningful near-term sense.

Korea supplies the memory. Huang visited Korea in October last year and again in June, meeting Samsung and SK hynix leadership and, by all accounts, asking for one thing repeatedly: more HBM. Samsung is currently supplying HBM4 for Vera Rubin; SK hynix has deepened its long-term supply partnership. Korean coverage of the June visit summarized his message as, bluntly, “More HBM!”

Japan is being cast as the proving ground. Not the chip supplier, not the memory supplier — the place where AI gets bolted onto robots, factories and cars and actually has to work. Every major Japanese announcement points this way: a national AI factory explicitly for physical AI, partnerships with robot-arm makers rather than chipmakers, applications named as manufacturing, logistics, healthcare and telecom.

That framing explains the “Japan passing” anxiety better than any snub theory. The central axis of the AI hardware supply chain — fabrication in Taiwan, HBM in Korea — genuinely does run through Japan’s neighbors. Japan’s leverage was never going to be in that axis. What Tokyo secured this month is a different role, and arguably a smarter one given what it already owns: the world’s densest concentration of industrial robotics expertise, at exactly the moment the industry’s attention turns to physical AI.

What this does and doesn’t tell us

It would be easy to over-read a two-day trip. A few cautions worth holding onto.

The Noetra timeline is long. Construction begins in April 2027 and full operation lands in June 2028 — roughly two years out, on a Rubin-generation platform in an industry where the roadmap has been moving faster than most institutional buildouts. Government AI-infrastructure commitments have a way of being announced with more precision than they are delivered.

The robotics partnerships are, at this stage, partnerships. Fanuc and Yaskawa dominate industrial robot arms today, but as we’ve written before, the hard problem in physical AI is not the arm — it’s teaching it to handle situations nobody programmed, on far less real-world data than language models get. Announcements do not resolve that. Nvidia’s GR00T foundation models are the thing to actually watch, not the MOU count.

And the Korea comparison cuts both ways. Being the memory supplier to the AI boom has been extraordinarily lucrative for Samsung and SK hynix — but it is also, structurally, a component role in someone else’s stack. Japan’s physical-AI positioning is slower and less immediately profitable, and it is a bet on owning the application layer rather than a slice of the bill of materials. Which of those turns out to be the better place to stand is genuinely unsettled, and it is a more interesting question than who Huang visited first.

What would change the picture

Three things to watch rather than the headline deal values: whether Noetra’s April 2027 construction start actually holds, since a slip there tells you how much of the $6.2 billion is commitment versus intention; whether the Fanuc and Yaskawa collaborations produce a shipping product with measurable capability gains rather than a demo reel, which is the same bar physical AI has been failing to clear industry-wide; and whether Korea’s HBM position stays as unassailable as it currently looks, since the entire three-way division of labor rests on the assumption that no one else can supply that memory at scale.

This article is for informational purposes and is not financial or investment advice. Deal values, chip counts and timelines are drawn from public announcements and reporting current as of publication and may be revised; verify against primary sources before acting on them.

Sources

This article may contain affiliate links. Nothing here is financial or investment advice.