Blame Nvidia? Why Taiwan Faces Higher Reciprocal Tariffs Than Japan and South Korea
Liang-rong Chen
Hello everyone,
Fortunately, this newsletter goes out on Thursday—after Trump once again made a sudden U-turn and announced a 90-day pause in the implementation of reciprocal tariffs.
That gave me just enough time to adjust the angle of the original article. And then sit back to witness a historic sight: Taiwan’s stock market flipping from “more than a thousand stocks limit-down” to “more than a thousand stocks limit-up” within the span of a week.
Over the past week, nearly everyone and every company, has been inundated by information related to what may be the wildest trade war in history.
Still, I managed to carve out time to attend a rare opportunity: a talk by former U.S. Commerce Department’s Bureau of Industry and Security (BIS) export control officer, Mi-Yong Kim(金美英), who was invited to Taiwan by the local think tank DSET. She shared invaluable first-hand insights into the implementation of U.S. export controls on China.
Readers may recall that in our March 20 issue, we cited a U.S. think tank report analyzing the major loophole in the U.S. ban on Nvidia’s most advanced GPUs being exported to China—one that allowed Chinese AI companies like DeepSeek to continue acquiring GPUs in large quantities.
The Taiwanese semiconductor executives I interviewed found it puzzling: how could the U.S. government possibly not realize that Nvidia could rapidly release “downgraded” versions of its GPUs by tweaking technical specs post hoc? Some even speculated that Washington was playing a double game with Nvidia—deliberately leaving room for China to maneuver.
Mi-Yong Kim, who spent years at the BIS, told me the truth is much simpler: BIS simply lacks the technical expertise. It doesn’t understand the industry.
There are structural reasons for this. It’s extremely difficult for BIS to hire young engineers—not only are government salaries uncompetitive, but core departments like BIS require stringent security clearances. The majority of U.S. engineering graduates each year are international students, primarily of Asian descent, who often can’t pass those background checks. Over time, fewer and fewer apply at all.
This shortage of engineering talent has only worsened under Trump’s downsizing of the civil service. “Most engineers in government now are old and just waiting for retirement,” she said.
I think that shatters many readers’ (including my own) idealized view of America as an omnipotent Big Brother. And it also serves as a reminder: we can’t afford to be complacent about China’s future technological development under U.S. sanctions.
Now, back to this week’s newsletter.
Whenever I ran into slow news periods in the past, I often turned to customs import/export data for story ideas. By digging into product flows by 8-digit tariff codes—comparing monthly and yearly changes in exports to the U.S. and China against specific companies or industry segments—I often uncovered unique and interesting reporting angles.
That old trick came in handy during this reciprocal tariff saga.
It turns out that when you combine a few key terms from our previous Nvidia supply chain coverage—server motherboards, GPUs, HBM, and CoWoS—you uncover the main culprit behind Taiwan’s export surge to the U.S. over the past two years, which has pushed us into the top five U.S. trade surplus countries.
And judging by the red-hot growth of the AI sector, Taiwan’s trade surplus is poised to climb even higher over the next two years. In other words, if Trump decides to wage a four-year trade war during his next term, Taiwan’s sky-high reciprocal tariff rate may keep rising—and whatever mitigation tools we have, like importing more shale gas from the U.S., won’t be nearly enough.
There’s only one way to significantly and swiftly reduce Taiwan’s trade surplus with the U.S.
Find out in this issue.
When this issue of the newsletter goes out, it’s just past April 9 U.S. time—the original date the United States had set for implementing reciprocal tariffs on imports from various countries.
Just as the world braced for trade turmoil, President Trump made a dramatic U-turn, announcing that tariffs on countries other than China would be delayed by 90 days.
In other words, Taiwan’s 32% reciprocal tariff will remain hanging over exporters like a sword for the next three months.
Why Is Taiwan Hit Harder Than Japan and Korea?
Many in the industry were outraged. Taiwan’s industrial structure is quite similar to Japan’s and South Korea’s—so why are those two major trade rivals subject to reciprocal tariffs of 23% and 25%, while Taiwan faces a steep 32%?
That 7–9 percentage point gap is enough to put Taiwanese manufacturers—especially in sectors like machine tools and chemical raw materials that directly compete with Japan and Korea—on the edge of survival.
On the surface, the reason lies in the sharp increase in Taiwan’s trade surplus with the U.S. over the past few years. Under Trump’s crude formula for calculating reciprocal tariffs—Taiwan’s 2024 trade surplus divided by exports to the U.S., then halved—that’s the result you get.
The Real Driver: AI Servers, Nvidia, and Jensen Huang
But why has Taiwan’s trade surplus with the U.S. grown so dramatically in recent years?
For those of us tech reporters who follow Nvidia’s supply chain closely, the answer is obvious: we have Jensen Huang to thank—or blame—for this.
If you plug Taiwan’s 2022 trade data into the above “Trump formula,” the resulting reciprocal tariff would have been 26%—roughly in line with Japan and South Korea.
So what happened over the past two years?
Some readers might remember a viral YouTube clip from November 2022. Two young female singers were livestreaming outside the Taipei Expo Park when Jensen Huang unexpectedly walked into the frame. He even requested a song—a Lady Gaga hit.
The video captured Huang swaying to the music and playfully interacting with the singers. That moment was streamed live and later uploaded online. To date, it has racked up over 3.59 million views.
But what was Jensen Huang doing there?
He was actually attending an esports event held at the same venue. Back then, the man in the leather jacket was still known mainly as the “Godfather of GPUs,” and those seeking selfies or autographs were mostly hardcore gamers. The two singers didn’t recognize him at all. (One of Huang’s assistants introduced him as “Jensen,” to which one of the performers, a Taiwan-born Canadian singer, responded with a puzzled “…Jensen?”)
That was in the pre-generative-AI era. It was also the last time Huang appeared publicly in Taiwan. A month later, ChatGPT was launched—and the tech landscape shifted overnight.
Nvidia became the fastest-growing large tech company in history, and Huang was crowned the undisputed Emperor of AI.
Taiwan’s New Superstar Export: L6 Server Boards
Taiwan, meanwhile, found itself with a brand-new bestselling export to the U.S.—officially declared at customs as “other units of automatic data processing machines.”
In 2022, Taiwan exported $3.33 billion of these items to the U.S. By 2024, that figure had skyrocketed to $31.7 billion—accounting for 28.55% of Taiwan’s total exports to the U.S.
This seemingly new product category, which materialized almost out of nowhere over the span of just three years, is—according to officials from Taiwan’s Ministry of Economic Affairs—essentially server motherboards without key components installed.
In industry terms, these are “barebones” systems, or L6 units, which are shipped via air freight from Taoyuan International Airport to the U.S. There, at assembly plants, CPUs, GPUs, and other high-value parts are added.
According to Digitimes senior analyst Jim Hsiao(蕭聖倫), these L6 server boards are the primary force behind Taiwan’s AI server exports to the U.S.
As for server motherboards with CPUs and GPUs already installed—commonly referred to in the industry as L10 units—exports to the U.S. jumped from $3.47 billion in 2022 to $15.9 billion in 2024, exactly half the value of L6 exports.
Jim Hsiao notes that while the L10 numbers look substantial, they include the extremely expensive GPUs and CPUs, meaning the actual unit volume could be only one-tenth that of L6 shipments.
Why are so many of these units shipped to the U.S. instead of Mexico, which used to be the favored destination for Taiwanese manufacturers? (Taiwan’s L6 exports to Mexico account for less than one-tenth of those to the U.S.)
Hsiao explains that currently, Quanta is the only Taiwanese manufacturer with final assembly facilities in the U.S. But because local demand for AI data centers is enormous, large U.S.-based server integrators—like ZT Systems, acquired by AMD last year, as well as Google’s in-house server factories—are sourcing motherboards directly from Taiwan.
Shipments of GPUs and TPUs from Taiwan to the U.S.—classified as “other integrated circuits”—also soared, from $2.6 billion in 2022 to $6.7 billion in 2024.
Just these three product categories alone saw a combined increase of $44.9 billion in exports to the U.S. over two years.
Put differently: without this AI boom, Taiwan’s trade surplus with the U.S. would be less than 40% of what it is now.
Korean Memory, Taiwanese Surplus
This surge has also led to an unexpected side effect: South Korea overtook Japan for the first time last year to become Taiwan’s largest source of imports.
The reason? A dramatic spike in imports of one particular item: dynamic random-access memory (DRAM) integrated circuits.
Taiwan’s DRAM imports from South Korea jumped from $6.8 billion in 2023 to $17.7 billion.
Given the current slump in consumer demand for computers and smartphones—and consequently for standard DRAM—this astonishing growth clearly stems from a very specific segment: high-bandwidth memory (HBM) used in AI GPUs.
More precisely, the surge is due to Nvidia’s Hopper-series GPUs, which in 2023 consumed large quantities of HBM supplied by SK Hynix of South Korea. As a result, Hynix’s HBM revenue in 2024 is projected to grow more than fourfold year-on-year.
Amid fears over reciprocal tariffs, an intriguing pattern has emerged.
These ultra-expensive memory chips from South Korea—worth more than $10 billion—are shipped to TSMC fabs in Taiwan, where they are packaged with Nvidia and AMD GPUs or Google TPUs using CoWoS technology.
Roughly 70% of those finished products are then shipped to the United States. Through this form of triangular trade, Korean-made components become an invisible contributor to Taiwan’s trade surplus with the U.S.
The Tariff’s Future: Still Climbing?
Readers familiar with the AI industry must be thinking the same thing by now: TSMC CEO C.C. Wei(魏哲家) has said that the current AI boom is only just beginning, with accelerator revenue projected to grow at a compound annual growth rate of nearly 45% from 2024 to 2029.
Then won’t Taiwan’s U.S. trade surplus (and its painful side effect: the 32% reciprocal tariff) just keep climbing? As of now, that seems likely.
In January and February of this year, Taiwan’s exports to the U.S. of “other units of automatic data processing machines”—known in the industry as L6 server boards—rose 49% year-over-year. While that’s a far cry from the 200%-plus annual growth rates seen over the past two years, it’s still a strong showing.
If this growth rate holds through the year, this single product category alone could push Taiwan’s export-based tariff rate up by another 2 to 3 percentage points, assuming the reciprocal tariff formula is recalculated annually.
That slowdown in growth may point to something else: a shift in the AI supply chain toward Mexico.
During the same January–February period, Taiwan’s exports of “other units of automatic data processing machines” to Mexico reached $2.76 billion. In just two months, Taiwan exported more of these goods to Mexico than in the entire previous year.
This makes perfect sense. Taiwanese electronics shipped to Mexico are exempt from tariffs, whereas exports to the U.S. are now subject to a potential 32% tariff.
But here’s the contradiction: this shift runs directly counter to the Trump administration’s core manufacturing policy of reshoring production.
A recent report by Morgan Stanley on reciprocal tariffs and Taiwan’s electronics sector noted that among all AI server manufacturers, Quanta may be the worst positioned.
That’s because Quanta assembles all of its AI servers in the U.S., meaning its motherboards—shipped from Taiwan—face the full weight of the reciprocal tariff.
In contrast, Wistron and Wiwynn assemble their servers in Mexico, where components can be imported from Taiwan tariff-free.
“It’s totally irrational—it effectively penalizes U.S. manufacturing,” says Digitimes senior analyst Jim Hsiao. “There’s no doubt the U.S. government will adjust this policy.”
Will Motherboard Production Move to the U.S.?
What worries many in Taiwan is: Will the adjustment go so far that even L6 and L10 motherboards are relocated from Taoyuan to U.S.-based production? Could that leave TSMC’s GPUs and TPUs as Taiwan’s only major AI-related exports?
Hsiao says most companies are still in a wait-and-see mode—largely because Trump’s unpredictability makes it hard to plan.
According to Chip War author and Tufts University professor Chris Miller, rather than trying to follow Trump’s every move, people should focus on what’s consistent in his trade philosophy.
The key theme? “Bring manufacturing back to the U.S..”
In a speech this February, Vice President J.D. Vance put it more bluntly: “the Trump administration will ensure that the most powerful AI systems are built in the U.S. with American-designed and manufactured chips.”
That, so far, is the clearest policy stance the Trump administration has taken on AI hardware.
The second half of that statement—the chips themselves—is TSMC’s problem. That’s outside the scope of this article.
But the first half—the most powerful AI systems should be built in the U.S.—raises a question: Does that include the motherboard?
One Nvidia supplier thinks the answer is yes.
“Motherboards will likely be the next part of the AI server supply chain to relocate to the U.S.—driven by tariff pressure,” he says. “And four years from now (once Trump’s out of office), we’ll move it back.”
10-Second Recap:
Taiwan’s steep U.S. Reciprocal Tariff? Blame the Nvidia effect.
One “super product” saw exports to the U.S. surge nearly tenfold in just two years.
And there’s only one way to bring that trade surplus back down—fast.
Read more here:
Riding with Nvidia, Rising to Glory! But Why Do Taiwanese Firms Still Prefer Apple?
Why U.S. AI Chip Sanctions Aren’t Working? The Untold Secret from Nvidia’s GTC 25!
The Key Figure Behind TSMC’s U.S. Expansion: TSMC’s Shohei Ohtani, Ying-Lang Wang
Nice
Liang-rong, You are correct, not enough expertise for the technical matters of how close to the limit can an export license be obtained. Often in the technical community, the limit makes no sense at all.
At Motorola, then MIPS Technologies with respect to microprocessor performance export limits, as well as intellectual property laws, I found the Department of Commerce and Technology Liasons of other agencies to be competent and able to understand the technical nuance once presented.
Any technoloby company can do such presentations; it takes a similar one for customer executives to game changing deals.
What few companies don't realize, is these agencies are accessible. And like another other relationship; when you need the contact to understand you is not the time when you look up their contact info.
Proper Planning Prevents Poor Performance.
After taking companies to China, Taiwan, Israel multiple times, one learns: Plant, nurture, build, harvest. Underpromise and overdeliver.