Jensen Huang's Ten Days: Beijing to Taipei

Opening

Hello, dear reader. This is OZ Talking.

On May 15, Jensen Huang stepped off Air Force One alongside President Donald Trump. The location: Beijing. He was accompanying the first US-China summit in eight years and six months. Less than ten days later, on May 23, he arrived at Taipei Songshan Airport aboard his private jet — handing out bottles of Yakult to the reporters waiting for him.

One CEO visited the capitals of both sides of the US-China conflict within ten days. This wasn’t just a stop on his Computex itinerary. Here’s the bottom line upfront: Nvidia is now walking an unprecedented tightrope, managing three colliding markets simultaneously — the US (regulator), China (buyer), and Taiwan (producer) — and the cost of that balancing act is starting to show up in concrete numbers.

🌐 The Ten-Day Route: From Beijing to Taipei

Let’s first walk through what happened over the past ten days.

On May 13, the White House confirmed that Jensen Huang would accompany President Trump on his China trip. Huang joined Air Force One in Alaska and traveled with the president all the way to Beijing, where the summit with President Xi Jinping took place. The fact that a private-sector CEO rode along on a US president’s state visit is, in itself, a powerful political signal.

At this meeting, Trump delivered two important messages.

First, he formally reconfirmed the export of Nvidia’s H200 chips to China. The approval had been announced back in December, but actual shipments had stalled amid regulatory friction on both sides. At this year’s GTC 2026, Huang had said, “We’ve received purchase orders for H200 from Chinese customers and are resuming manufacturing.”

Second, there was a signal aimed at Taiwan. Reports emerged that President Xi had directly asked Trump how the United States would respond if China attacked Taiwan. Trump said he had discussed arms sales to Taiwan in detail, but that the decision would ultimately be his to make. According to Axios, one Trump adviser assessed that “this China trip signals a growing likelihood that Taiwan will be on the table within five years.”

Then, ten days later, on May 23, the first thing Huang said upon arriving in Taipei was this:

“Vera Rubin will be the largest product in the history of Taiwan’s supply chain.”

Taken separately, these two trips look like an ordinary diplomatic event and an ordinary tech event. But strung together into a single itinerary, they bring the structural dilemma Nvidia faces into sharp focus.

🔺 Three Markets, Three Dilemmas

The three markets Nvidia must manage simultaneously each want something entirely different.

The US government has been tightening semiconductor export controls on China step by step since 2022. Under the Biden administration, Nvidia was even forced to develop a separate, performance-capped chip for China alone: the H20. Trump’s H200 export approval partially reversed that trend — but the conditions attached are demanding.

Nvidia must pay the US government a fee equal to 25% of the relevant revenue and can only sell to customers approved by the Department of Commerce. Imposing this kind of fee on a domestic company’s overseas revenue is highly unusual. It’s less a tariff than a kind of “export licensing fee.” And Blackwell and the next-generation Rubin chips remain barred from export entirely.

There’s one more point worth noting. Michael Horowitz, a fellow at the Council on Foreign Relations (CFR), a US foreign-policy think tank, analyzed the H200 approval by observing that “the Trump administration is dismantling the very export control regime it built during its first term.” Even the US government itself can’t hold a consistent line between security concerns and industrial interests.

What’s interesting is China’s response. The US opened the door, and China is narrowing it from the inside.

According to the Financial Times, Chinese regulators are discussing an approval process requiring companies to justify H200 purchases by explaining why domestic chips can’t meet their needs. The public sector might be banned from buying Nvidia hardware altogether. Huang himself admitted on an earnings call that export controls have effectively handed the Chinese AI market over to Huawei.

Big tech companies like Alibaba, Tencent, and ByteDance have already shifted part of their inference¹​ workloads to Huawei’s Ascend chips. Large-scale model training²​ still requires Nvidia GPUs, though. Huang’s own remark that Nvidia’s China market share plunged from 95% at the start of 2025 sums up the situation.

Taiwan is the key production base for Nvidia’s chips. TSMC manufactures the Rubin GPU and Vera CPU on its 3nm process, and 100 to 150 Taiwanese partners participate in the Vera Rubin platform³​. Huang said “Taiwan’s supply chain will be very busy in the second half of the year.”

But Taiwan is also the front line against chip smuggling. Two days before Huang arrived, on May 21, Taiwanese prosecutors launched the island’s first-ever crackdown on semiconductor smuggling. They raided 12 sites and sought detention warrants for three people accused of illegally exporting Nvidia chips embedded in Super Micro servers to China, Hong Kong, and Macau using falsified documents.

This isn’t a minor incident. It connects directly to a case in which the US Department of Justice arrested a Super Micro co-founder this March over an alleged $2.5 billion (~₩3.5 trillion) AI chip smuggling scheme. According to the indictment, the smuggling ring systematically funneled Nvidia-powered servers into China through Southeast Asian intermediaries between 2024 and 2025.

Summed up in a line, the three markets’ demands look like this: the US says “don’t sell without a license,” China says “buy only when domestic won’t do,” and Taiwan says “don’t let it leak out through us.” The point where all three demands can hold true simultaneously is exceedingly narrow.

💰 The Cost of the Tightrope Is Starting to Show

This balancing act has already begun to carry concrete costs.

Let’s start with revenue uncertainty. Nvidia’s annual revenue guidance of $78 billion doesn’t include the recovery in H200 sales to China. Analysts believe an additional $3.5 to $4 billion (~₩5 to ₩5.6 trillion) in annual revenue is possible if the export framework functions normally, but the timing remains uncertain given China’s own restrictions and smuggling risk.

Next is compliance cost. Right after arriving in Taiwan, Huang addressed the Super Micro smuggling case, saying, “Nvidia thoroughly explains the rules to every partner,” but adding that “ultimately, Super Micro has to run its own company.” The implication shifts compliance responsibility onto the partner — but since it’s Nvidia’s chips being smuggled, Nvidia ultimately bears the brand risk. With Taiwan now cracking down as well, supply-chain monitoring costs can only keep rising.

There’s also a hidden threat: memory inflation. In Taiwan, Huang warned that “rising memory prices are pushing up the cost of nearly every electronic product, from PCs to graphics cards,” and urged memory suppliers to ramp up production quickly. The Vera Rubin platform uses next-generation HBM4⁴​ memory, and with memory prices already on an upward trend, the cost of building out this platform could climb further.

Finally, the competitive landscape is shifting. AMD announced a $10 billion (~₩14 trillion) investment in Taiwan. Huang shot back that “Nvidia has already invested more than that in its Taiwanese partners,” though without disclosing a public figure. In China, Huawei is rapidly eating into the market, and in Taiwan, AMD is now escalating an investment race.

Oz’s Lens

Honestly, what I find most interesting here isn’t the technical specs — it’s Huang’s itinerary itself.

In my years building tech strategy, I’ve seen this pattern countless times: a company’s core customer in one market becomes the regulatory target in another. Companies usually pick one of two paths — give up the market to comply with regulation, or lobby against the regulation to protect the market. Huang is doing both at once, right now.

He boards Air Force One to extract regulatory relief, then ten days later reassures the production base in Taipei by promising a product that will be “the largest in history,” while simultaneously signaling to Chinese buyers by saying “China is included in the $200 billion CPU market.” This isn’t a CEO doing sales — it’s practically private-sector diplomacy.

This tightrope walk is possible because Nvidia occupies an irreplaceable position in the AI chip market. But in my experience, “irreplaceable” has never been a permanent position. Huawei is rapidly filling in the Chinese market, and AMD has staked $10 billion on Taiwan’s supply chain. The cost of the tightrope keeps rising, while the time Nvidia can stand alone on it keeps shrinking.

Closing

To sum up:

Nvidia is walking a tightrope in the middle of three colliding markets — the US, China, and Taiwan — bearing concrete costs in the form of a 25% fee, smuggling risk, and memory inflation. The ten days from Beijing to Taipei aren’t a technology roadmap; they’re a microcosm of business strategy laid over geopolitics.

Huang is set to unveil the detailed specs of Vera Rubin at his Computex keynote on June 1. Rather than the technical announcement itself, I’d recommend watching what message he sends to each of the three markets on that stage. In my experience, the context of an announcement reveals more about strategy than the product specs do. Of course, don’t forget that I hold a position in $NVDA at an average cost of $28.2 (bought on February 8, 2023)…

References & Further Reading

Primary sources

• Michael C. Horowitz, “The Consequences of Exporting Nvidia’s H200 Chips to China”, Council on Foreign Relations, December 9, 2025. : An analysis of the structural impact of H200 exports on US security strategy. This is the core source behind today’s newsletter’s “US market” analysis.
• “Nvidia says its forecast for $200 billion CPU market includes China”, CNBC, May 23, 2026. : The most accurate summary of Huang’s remarks right after arriving in Taiwan.
• “Taiwan Seeks to Detain Three in AI Chip Smuggling Crackdown”, Bloomberg, May 21, 2026. : A detailed account of Taiwan’s first-ever semiconductor smuggling crackdown.
• “Nvidia CEO Urges Super Micro to Tighten Up Amid Taiwan Crackdown”, Bloomberg, May 23, 2026. : Huang’s arrival interview in Taiwan addressing the Super Micro case.
• “Nvidia’s Jensen Huang predicts China will eventually open market for AI chips”, Taiwan News, May 19, 2026. : His outlook on the China market from a Bloomberg TV interview right after the China trip.

Background

• KBS (Korea’s public broadcaster), “Grading Trump’s China Trip: Taiwan Is the Real Test”, May 19, 2026. : An article covering a Trump adviser’s remarks on Taiwan and concerns over the semiconductor supply chain.
• “Super Micro Co-Founder Arrested in $2.5B AI Chip Case”, March 20, 2026. : Helpful for understanding the full context of the Super Micro smuggling case.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Inference: The process by which an already-trained AI model produces an answer to a new input. For example, when you ask ChatGPT a question and it generates a response, that’s inference. ↩

2. Training: The process by which an AI model learns patterns from massive amounts of data. It requires far more computing resources than inference, often deploying thousands of high-performance GPUs. ↩

3. Vera Rubin platform: Nvidia’s next-generation AI data center platform, combining a CPU called Vera with a GPU called Rubin — named after astronomer Vera Rubin. It targets 5x the inference performance and 3.5x the training performance of the previous Blackwell platform, with mass production planned for the second half of 2026. ↩

4. HBM4 (High Bandwidth Memory 4): Ultra-high-speed memory stacked vertically next to an AI chip. It offers wider bandwidth than the previous HBM3, making it essential for large-scale AI computation — but its difficult manufacturing process has contributed to supply shortages and price increases. ↩