What Jensen Huang Said in Taiwan After Visiting Beijing

What Jensen Huang Said in Taiwan After Visiting Beijing

In mid-May, Jensen Huang accompanied President Donald Trump on his trip to China. On May 23, he landed at Taipei Songshan Airport and handed out Yakult drinks to reporters. The visit was timed ahead of Computex, but the questions on the ground were less about new products and more about China sales and the ongoing investigation into semiconductor smuggling.

Nvidia has to keep selling chips to Chinese customers while complying with U.S. export controls, all while working with its Taiwanese partners to produce next-generation chips. Even after you’ve built a good chip and taken the order, you still need export and import approvals and production readiness before any of it turns into actual revenue. What caught my attention on this trip was exactly what still stands between those two points.

Resuming China Sales and New Product Production in Taiwan

On May 13, the White House confirmed that Jensen Huang would accompany the delegation to China. He joined Air Force One in Alaska and headed for Beijing.

At the time, Nvidia’s main concern was resuming H200 sales in China. The US government had announced its policy allowing the exports back in December, and in January it set up a conditional, case-by-case review process. But getting US approval and actually delivering products to Chinese customers were two separate steps.

Even after the China trip, this issue wasn’t resolved right away.

At GTC in March, Jensen Huang said Nvidia was resuming manufacturing after taking orders for the H200 from Chinese customers. During his May visit to Taiwan, he repeated his desire to supply the Chinese market. But announcements about orders and approvals alone weren’t enough to conclude that shipments to China had actually normalized.

The US-China summit also touched on the Taiwan issue. Still, media coverage of the summit and forecasts from political figures alone can’t determine Taiwan’s future security situation. What Nvidia needed immediately was to align production schedules with its Taiwanese partners.

On May 23, in Taipei, Jensen Huang emphasized the scale of Vera Rubin’s production.

He predicted that Vera Rubin could become one of the largest product launches in the history of Taiwan’s supply chain. He also added that production partners would be extremely busy in the second half of the year.

China required the right conditions to resume sales, while Taiwan needed to be ready to produce the new product on schedule. Jensen Huang’s two visits show that Nvidia was working on both fronts at once.

US export review, China purchase approval, Taiwan investigation

Let me break down the problem each region needs to resolve.

The US has been restricting the performance and sales conditions of AI chips that can be exported to China. For the H200, it decided starting this January to review individual export license applications case by case. It checks whether US customer supply would be harmed, whether Chinese buyers have export compliance procedures in place, and so on.

Beyond export licenses, the H200 also carries a tariff burden. A US measure that took effect on January 15 imposes a 25% tariff on imports of certain advanced chips, including the H200, into the United States. There are exceptions for chips used in domestic data centers, among other things. Chips destined for export to China are required to undergo third-party performance testing in the US, so this tariff becomes an issue in the process of bringing chips into the US before shipping them onward to China. This is covered respectively in the White House announcement and the Commerce Department’s export review guidance.

Opinion within the US is split over allowing the H200. Michael Horowitz of the CFR, a foreign policy think tank, criticized the decision last December, arguing it could help China’s AI development and undermine the effectiveness of existing export controls. It’s a debate over how to reconcile Nvidia’s sales opportunities with America’s security goals.

China, for its part, is trying to secure the chips its companies need while also cultivating domestic semiconductor makers. That’s why a deal isn’t settled just because the US grants an export license.

Last December, reports emerged that Chinese authorities were considering a plan requiring H200 buyers to explain why domestic chips couldn’t meet their demand. This January, reports followed that some big tech firms had gotten purchase approval. At the time, Reuters reported that Alibaba, Tencent, and ByteDance had received purchase approval, while noting that uncertainty remained over the specific conditions.

In this process, Nvidia competes with Chinese companies like Huawei. Inference¹, where AI generates answers, and training², where models are built, require different computing conditions. Chinese customers also have to weigh performance, development environment, price, and procurement availability together. Not all training absolutely requires Nvidia chips, but for customers already using Nvidia hardware and software, a supply cutoff is a heavy burden.

Taiwan is where Nvidia’s core production partners — TSMC, server makers, and others — are concentrated. Mass-producing the next-generation Vera Rubin platform³ requires partners across chip manufacturing, packaging, and server assembly to be aligned and ready.

An investigation into illegal exports is also underway in Taiwan. Prosecutors in Keelung searched 12 related locations on May 20 and moved to take three people into custody on suspicion of forging documents to export Supermicro servers containing Nvidia chips. According to CNA’s May 22 report, some of the servers were shipped to Hong Kong, and investigators were looking into whether they were subsequently sent on to mainland China.

There was a separate related investigation in the US as well. The US Department of Justice announced on March 19 that it had charged three individuals, including a Supermicro co-founder. The indictment alleges that an intermediary company purchased roughly $2.5 billion worth of servers in 2024–2025, a substantial portion of which was illegally diverted to China. This is not a conviction, and Supermicro the company itself is not a defendant in this indictment.

US export review, Chinese purchase approval, and preventing illegal diversion during distribution are each separate matters that need to be verified. Approval from any single one of these doesn’t resolve the entire sales process.

Uncertainty Remains in Earnings Outlook and Production Costs

You can see how this situation affects the business by looking at the earnings outlook and the supply chain.

Nvidia’s Q2 FY2027 revenue guidance, issued on May 20, was $91 billion, with a margin of error of ±2%. This figure does not include China data center computing revenue. In other words, the company didn’t see a recovery in China sales as certain enough to build into its quarterly forecast. You can check this in Nvidia’s earnings release.

Transaction oversight has also become more important. In Taiwan, Jensen Huang stressed that partners must comply with export regulations, and he asked Supermicro to strengthen its compliance framework. Tracking where equipment goes after it’s sold and preventing illegal diversion is now being demanded alongside product supply itself. That said, it hasn’t been disclosed how much this investigation has raised Nvidia’s oversight costs.

Memory prices are another variable. In Taiwan, Jensen Huang noted that rising memory prices are pushing up costs for electronics like PCs and graphics cards, and he called for increased production. Securing enough memory in time — including the HBM4⁴ used in Rubin GPUs — is a critical condition for producing new products.

Competitors are also expanding their Taiwan supply chains. On May 21, AMD announced it would invest more than $10 billion in Taiwan’s industrial ecosystem. The plan includes expanded cooperation on advanced packaging and manufacturing. When considering Nvidia’s production capacity, it’s worth remembering that its partners are also handling demand from other customers at the same time. AMD’s announcement

Oswarld’s Lens

This time, I found myself more interested in Jensen Huang’s visit schedule and remarks than in the new product specs themselves.

There’s an issue I’ve watched play out repeatedly in setting technology strategy: when regulation hits a market with important customers, product competitiveness alone can’t protect a deal. You have to identify which products and customers are actually permitted, and also decide how to adjust production plans when deals get delayed.

Jensen Huang kept referencing the possibility of sales resuming in China while emphasizing expanded Vera Rubin production in Taiwan. To me, these two things need to be read together. Even if the resumption of China sales is delayed, orders from other markets still need to be fulfilled — and all the while, the relationship with Chinese customers has to be maintained too.

Nvidia’s strengths in technology and its development environment are undeniable. But I don’t judge long-term performance on that edge alone. I try to check, alongside it, whether customers can actually buy the product, whether partners can manufacture it on time, and what alternatives competitors roll out in the meantime. This visit made me think through that again.

Closing

When I assess the China sales resumption, I’ll be watching whether the approval announcement is actually followed by real shipments and revenue recognition. In Taiwan, what matters is the production schedule for the next-generation product and whether component supply can keep up.

I’ll be looking at these same two things during Jensen Huang’s keynote at Computex on June 1st. Beyond Vera Rubin’s performance specs, I’m curious about when it can actually reach customers, and what progress there’s been on the China business.

For disclosure: as I write this, my own position is NVDA, average purchase price $28.2, bought on February 8, 2023. I’m flagging this so you know I’m watching this as an invested shareholder, not just an outside observer.

The English draft is accurate and complete—no Hangul characters remain, all numbers are preserved, headings/footnotes/links/images match the source exactly. No corrections needed.

References & Further Reading

Primary sources

• Michael C. Horowitz, “The Consequences of Exporting Nvidia’s H200 Chips to China”, Council on Foreign Relations, 2025.12.9. : This piece analyzes the structural impact of H200 exports on U.S. security strategy. It’s the key evidentiary basis for today’s “U.S. Market” analysis.
• “Nvidia says its forecast for $200 billion CPU market includes China”, CNBC, 2026.5.23. : An article covering Jensen Huang’s remarks on the CPU market right after arriving in Taiwan.
• “Taiwan Seeks to Detain Three in AI Chip Smuggling Crackdown”, Bloomberg, 2026.5.21. : An article on Taiwan’s investigation into illegal AI server exports.
• “Nvidia CEO Urges Super Micro to Tighten Up Amid Taiwan Crackdown”, Bloomberg, 2026.5.23. : Jensen Huang’s comments on the Super Micro case, from an interview upon arriving in Taiwan.
• “Nvidia’s Jensen Huang predicts China will eventually open market for AI chips”, Taiwan News, 2026.5.19. : His outlook on the China market, from a Bloomberg TV interview right after his China trip.

Background

• KBS, “Trump’s China Trip Report Card: Taiwan as the Litmus Test”, 2026.5.19. : An article covering a Trump advisor’s remarks on Taiwan and concerns over the semiconductor supply chain.
• “Super Micro Co-Founder Arrested in $2.5B AI Chip Case”, 2026.3.20. : Useful for understanding the full context of the Super Micro smuggling case.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

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

2. Training: the process of adjusting an AI model’s parameters using data. The computation required varies with model and data scale. ↩

3. Vera Rubin platform: Nvidia’s next-generation AI data center platform, combining the Vera CPU, Rubin GPU, and networking equipment. It’s named after astronomer Vera Rubin. ↩

4. HBM4 (High Bandwidth Memory 4): a generation of HBM that stacks memory chips in multiple layers to provide wide data-transfer channels to GPUs and similar hardware. Not all AI workloads require HBM4. ↩