NVIDIA in Taiwan: The World's Most Valuable Company Does Not Make a Single Chip Itself

30-second overview: NVIDIA is the most valuable company on earth. On October 29, 2025, its market value broke US$5 trillion¹. Yet it does not own a single wafer fab. Every AI chip is manufactured at TSMC; every AI server is assembled by Foxconn, Quanta, and Wistron; Taiwan handles 90% of the world's outsourced AI server manufacturing². This dependence runs so deep that NVIDIA is already TSMC's largest customer, accounting for 19% of revenue³, and so deep that NVIDIA's chip architecture is now being determined in reverse by Taiwan's packaging yield⁴. The problem is that holding someone else's lifeline and sharing in the gains from it are two different things: NVIDIA's gross margin is 75%, while Taiwanese ODMs have only 5% to 8%⁵. This article is about how this unequal relationship grew into what it is today.

Jensen Huang speaking at Computex Taipei in 2016. Since 2023, he has returned to this trade show almost every year to announce NVIDIA's latest AI chips, while seated below him is the entire Taiwanese supply chain that makes those chips for him. Photo: NVIDIA Taiwan, 2016.

May 19, 2025, Taipei Nangang Exhibition Center. Jensen Huang, wearing his signature leather jacket, stepped onto the Computex main stage. Behind him, a huge backdrop lit up into a wall of logos: AAEON, Accton, Delta Electronics, Gigabyte, Quanta, Wistron, Wiwynn, Foxconn, MediaTek, TSMC, UMC... one after another, finally freezing into the marks of 55 Taiwanese companies⁶. Including the thank-you video shown at the venue, the Taiwanese companies named totaled 122⁶.

That was the first time Taiwanese people "saw" their entire industry publicly named, all at once, by an American company.

The pride was real. But that wall carried an unspoken question: every logo on it was working for this American company, while the real power lay in the hands of the person projecting the wall, not on the wall itself.

The full COMPUTEX keynote delivered by Jensen Huang on May 19, 2025, from NVIDIA's official channel. This was the speech that revealed the wall of logos from 55 Taiwanese companies and announced that NVIDIA would locate its overseas headquarters in Taipei.

A Company That Manufactures Nothing Became the Most Valuable on Earth

NVIDIA is the fabless semiconductor model taken to its extreme. It designs chips, but it does not build fabs, buy lithography machines, or produce a single wafer. An empire worth US$5 trillion has no wafer fab of its own.

It has outsourced manufacturing wholesale to an island across the Pacific.

Micrograph of the NVIDIA Ampere GA102 die, produced on TSMC's 8-nanometer process. NVIDIA designed it, but every dense line etched across this surface was carved inside Taiwanese fabs. Photo: Fritzchens Fritz, CC0.

Its most profitable chips, including H200, Blackwell, and the forthcoming Rubin, all depend on TSMC's 3-nanometer and 4-nanometer processes⁷. In its own annual report to the U.S. Securities and Exchange Commission, NVIDIA acknowledged this concentration in black and white: the company's supply chain is mainly concentrated in the Asia-Pacific region, and it uses foundries such as TSMC to produce its semiconductor wafers⁸. This sentence appears in the company's legal filing to the SEC. In other words, NVIDIA itself identifies Taiwan as its largest source of geopolitical risk.

Making the chip is not enough. For a GPU to become something that can compute, it must first undergo advanced packaging and then be installed in a server. TSMC's CoWoS packaging is now a global bottleneck. NVIDIA alone consumes roughly 60% of CoWoS capacity, with Taiwanese media estimates even higher, at 70%⁹. Packaged chips are then handed to Taiwanese manufacturers for assembly: Foxconn's Ingrasys makes GB200 NVL72 rack systems, with E.Sun Securities Investment Consulting estimating that it holds more than 40% of AI rack assembly market share¹⁰; Quanta makes cloud servers and has more than half the share among the top 50 data centers¹¹; Wistron's new AI plant in Zhubei has been entirely booked by NVIDIA orders¹².

Taiwan accounts for 90% of global outsourced AI server manufacturing; if U.S.-brand suppliers are included, the figure is 100%². This means that virtually every physical machine used for AI computation on earth has passed through Taiwanese hands.

Stack these numbers together and the central contradiction emerges: the most valuable company manufactures nothing because its entire physical body is lodged in Taiwan. Taiwan is the switch it cannot turn off.

But holding someone's lifeline and taking money from that person are two different things. The next wall is hidden behind those 55 logos.

Behind the Logo Wall Is the Bottom of the Smile Curve

Manufacturing has an old "smile curve": high at both ends, low in the middle. The two ends that control brands, design, and technology earn thick profits; the middle segment, responsible for assembly and manufacturing, earns the thinnest margins. Taiwan stands precisely in that middle segment.

NVIDIA's full-year FY2025 gross margin was 75.0%, a figure it reported to the SEC itself⁵. During the same period, the Taiwanese companies assembling servers for it had gross margins like these: Foxconn 6.18%, Quanta 4.78% (the lowest in nearly 15 quarters), Wistron 5.21%, and Wiwynn 7.2%¹³. NVIDIA's gross margin is roughly twelve times Foxconn's and sixteen times Quanta's.

This chart contains a counterintuitive fact. Among Taiwanese companies, those closer to pure "assembly" have thinner margins; those closer to "technology" have higher margins. Delta Electronics, which makes power and cooling systems, has 37%. Unimicron, which makes ABF substrates, is estimated at 21.3%¹⁴. The difference is not "whether it is a Taiwanese company," but "where it stands on the curve." The assembly segment is thin no matter who does it.

📝 Curator's note: Morgan Stanley calculated an even sharper figure in May 2026: ODMs' assembly value-added gross margin at the complete-system level falls from 2.7% for the previous-generation GB300 rack to 1.9% for the next-generation VR200¹⁵. In other words, every time NVIDIA releases a stronger new chip, Taiwanese manufacturers must front more capital while earning a lower margin. The further this supply chain advances, the flatter the bottom of the curve is pressed. The logos can look glorious on the wall while gross margins remain thin at the bottom of the curve; both can be true at the same time.

Pride and cost tear at each other along the same chain. Because of the AI boom in Taiwanese stocks, Taiwan's economic growth in 2025 was about 7.37%, its fastest in 15 years and among the world's leading performers¹⁶. Yet Min-Hua Chiang, a scholar of Taiwan's economy, points to a cold number: most Taiwanese people did not feel the benefits of this thriving economy¹⁶. The highest-earning 10% took 48% of Taiwan's total income, while the bottom 50% received only 12%¹⁶. On a per-capita basis, the top 10% earned twenty times as much as the bottom 50%.

A June 2026 commentary in The Reporter made this K-shaped divergence more concrete: the "directly employed population" along the main growth chain of AI, semiconductors, and electronics supply chains accounts for less than 10% of total employment¹⁷. A person working in food service has total monthly pay of NT$38,484, only 34.6% of the figure in electronic components manufacturing¹⁷. The AI dividend is real. It is simply concentrated in capital and a small group of engineers, while most people watch the wave from outside it.

This is the first meaning of "being indispensable does not mean calling the shots": Taiwan holds the switch, yet receives that 5%.

Taiwan Once Caught It, and Nearly Sank It

To understand how this unequal relationship began, we need to return to the late 1990s, when NVIDIA was still racing against death.

NVIDIA was founded in 1993 and came close to collapse several times in its early years. When it launched the RIVA 128 graphics chip in August 1997, the company had "only one month of payroll left"¹⁸. In those days, Jensen Huang opened every monthly company meeting with the same English maxim: our company is thirty days from going out of business¹⁸. The sentence later became part of NVIDIA's internal creed, but it was never a Chinese sentence.

What truly pulled NVIDIA out of that crisis was a US$5 million payment from SEGA, not Taiwan¹⁹. This must be made clear first, because the claim that "Taiwan saved NVIDIA" is often told too romantically.

Taiwan's role was something else: the manufacturing lifeline. Around 1996, 32-year-old Jensen Huang wrote a letter to TSMC founder Morris Chang, asking whether TSMC could manufacture chips for NVIDIA²⁰. TSMC's Y.J. Mii later recalled in 2025 that this "far-reaching collaboration began at a critical moment, in 1997," when "TSMC founder Morris Chang personally contacted NVIDIA founder Jensen Huang in response to NVIDIA's request for foundry services"²¹. In 1998, the two sides formally signed an agreement, and TSMC became NVIDIA's main wafer foundry²⁰. The division of labor, "design in Silicon Valley, manufacture through TSMC," attached NVIDIA's physical body to this island from then on.

💡 Did you know: One widely circulated version says that when Jensen Huang received Morris Chang's phone call, he excitedly shouted to the people around him to quiet down because Morris Chang was calling²². This scene is second-hand retelling, and the tone may not be precise, but it captures something real: that small company on the verge of failure saw a phone call from TSMC as a lifeline.

But that rope almost became a noose. In 1998, an error in one of TSMC's chemical processes caused a large batch of NVIDIA chips to be scrapped, nearly dragging the company down again²³. So the more honest formulation is this: Taiwan was not NVIDIA's "bankruptcy savior"; Taiwan was its "manufacturing lifeline." That lifeline bound both sides. Taiwan once caught it, and Taiwan once nearly sank it. Symbiosis was never a one-way favor.

The next part of the story is more familiar. In 2006, NVIDIA launched CUDA, a decision almost everyone considered irrational at the time. In 2012, graduate student Alex Krizhevsky used two NVIDIA GTX 580 graphics cards in a bedroom at his parents' house to train AlexNet, cutting the ImageNet image-recognition error rate from 26% to 15.3%²⁴. That moment proved that the GPU was the engine of deep learning. In 2022, ChatGPT detonated global hunger for computing power, and NVIDIA's market value began rising like a rocket.

From a small company thirty days from collapse to history's first US$5 trillion company, every chip along the way was made in Taiwan.

Jensen Huang holding up the next-generation Blackwell GPU at CES in 2025. From a founder who had to say every month that "we are thirty days from going out of business" to the person holding the most sought-after chip in the world, that chip still has to be made in Taiwan. Photo: Pronoia, CC0.

Does Irreplaceability Have an Expiration Date

This leads to a question that must be faced honestly: is Taiwan's "irreplaceability" permanent, or does it have a time boundary?

In the short term, the boundary is so hard that there is almost no opening. From 2025 to 2027, NVIDIA's most advanced AI GPUs, from fabrication to final packaging, are 100% pinned to TSMC's CoWoS-L lines inside Taiwan²⁵. TSMC controls roughly 90% to 92% of the world's advanced processes below 5 nanometers, and its advanced packaging capacity exceeds that of all competitors combined²⁶. Research by National Taiwan University of Science and Technology professor Yuntsai Chou states it plainly: diversifying TSMC foundry production is not feasible in the short term; building a new leading-edge fab takes three to four years and more than US$10 billion²⁷.

The strongest engineering evidence is not in any report, but in NVIDIA's own product design. The next-generation Rubin Ultra was originally planned to use a "four-die" package. But in April 2026, TrendForce noted that the four-die package would expand the package area to 7.5 to 8 times the reticle limit, "severely dragging down yield and cost," so the design "is now shifting to a dual-die architecture"⁴. This sentence deserves to be read slowly: the physical limit of Taiwan's packaging yield is determining, in reverse, what NVIDIA's chips should look like. Even the world's strongest chip-design company has to redesign around Taiwan's yield. This is already a bottleneck at the physical layer, hard enough to leave no room for negotiation.

But a "hard short-term boundary" is not the same as "forever." Taiwan has painful precedents.

In 2002, Taiwan promoted the "Two Trillion, Twin Star" industrial policy. Flat panels and DRAM memory were also once treated as irreplaceable national lifelines. What happened? Because Taiwan lacked core technologies and R&D spending was only 6%, far below the 10% to 21% levels in South Korea, Japan, the United States, and Europe, the two industries were hollowed out by South Korea and China. A later United Daily News retrospective was cutting: "The once-glorious flat-panel makers and semiconductor memory makers suffered heavy losses a few years later because of oversupply in the international market, and netizens mocked them as 'Maoshan priests' (a pun on gross margins of three to four), meaning their product gross margins were only 3% to 4%; the 'Two Trillion, Twin Star industries' became the 'Two Trillion, Heartbreak industries.'"²⁸

⚠️ How this time differs from flat panels and DRAM: Flat panels and DRAM were hollowed out because Taiwan did not control core technologies and anyone could catch up. The moat around AI this time appears much deeper: TSMC does control process IP, and CoWoS packaging yield and the stickiness of the entire ecosystem cannot be replicated in three to five years by throwing money at them. But this is not a reason for Taiwan to rest easy. SMIC has claimed mass production at 5 nanometers; although its yield is only one-third of TSMC's and its cost is 50% higher, leaving it roughly five years behind²⁹. But five years is not forever in the technology industry. Treating "being deeply depended upon" as "being forever safe" is exactly the mistake made by the people of 2002.

After 2028, cracks begin to appear. In December 2025, NVIDIA spent about US$5 billion to take a stake in Intel. Its real purpose was to "ensure priority access to Intel's advanced packaging capacity in the United States," expected for use in the 2028 Feynman architecture³⁰. TSMC's Arizona packaging facility AP1 is expected to enter mass production in 2028³¹. Taiwan's Powertech has developed PiFO packaging as a counterpart to CoWoS-L, with production costs about 30% lower, and "multiple U.S. AI chipmakers" have rushed to seek its help³². These are all real signs of loosening, but none has yet entered the main supply chain for NVIDIA's most advanced GPUs.

One figure best captures the subtlety of the current situation: TrendForce forecasts that the CoWoS supply-demand gap will shrink from about 20% now to around 10% by the end of 2026³³. But the gap is shrinking because TSMC itself is expanding capacity, not because substitute suppliers are filling it³³. In other words, up to today, this bottleneck can still be solved only by Taiwan itself.

Irreplaceability is an engineering fact, but it has a hard boundary written somewhere near 2028. Taiwan's bargaining chip has a shelf life.

NT$4.434 Billion: A City Clears Land for a Trillion-Dollar Company

If gross margin is an abstract scale of power, what happened in the Beitou-Shilin Technology Park in Taipei in the second half of 2025 is one of its sharpest slices.

The story begins in 2021. The Taipei City Government put two Beitou-Shilin Technology Park plots, T17 and T18, totaling 3.89 hectares, up for sale as 50-year superficies rights. Shin Kong Life was the only bidder and won with NT$4.4 billion³⁴. For the next three years, the land sat idle and overgrown.

In May 2025, Jensen Huang announced at Computex that NVIDIA's overseas headquarters, "Constellation," was intended for the Beitou-Shilin Technology Park³⁵. But a problem emerged: the superficies rights were still held by Shin Kong Life, and publicly tendered superficies rights could not be directly transferred. NVIDIA, Shin Kong Life, and the Taipei City Government were stuck for five months over this land³⁶.

The final solution was for the Taipei City Government to pay Shin Kong Life to leave. On November 12, 2025, the Taipei City Council unanimously approved a termination payment of NT$4.434 billion, to be paid by the city government to Shin Kong Life in order to take back the land³⁷. The full figure was NT$4,434,064,085³⁷.

The termination bill submitted by Shin Kong Life set off an uproar in the council. Kuomintang councilor Yu Shu-hui wrote after receiving the bill: "When I saw that in Shin Kong Life's eight-page bill, even weed removal, environmental maintenance, logo adjustment, and scrivener fees were all being charged back to the Taipei City Government, I could only force a bitter smile. Isn't environmental maintenance something the tenant was supposed to do in the first place? Even the logo adjustment cost for the Taishin-Shin Kong Life merger has to be paid by the Taipei City Government? I nearly fainted... sigh, three helpless sighs."³⁸ But she still voted in favor, adding a line that captured the mood in the chamber: "Shin Kong Life's bill is unfathomable, but the greater good comes first"³⁸.

That day, the council witnessed a rare scene: the blue, green, and white parties were unusually harmonious, and the Democratic Progressive Party caucus even shouted, "Support NVIDIA, sign the contract as soon as possible"³⁹. One parcel of land and one foreign company made parties that usually bite at one another speak with one voice.

📝 Curator's note: It is worth pausing to look at this power structure. To let a trillion-dollar company land in Taipei, a city's government and council used public funds, crossed party lines, overcame every obstacle, and cleared a parcel of land still occupied by the previous tenant. NVIDIA did not pay this NT$4.434 billion; Taipei taxpayers fronted the money, with only Shin Kong Life's own costs and already paid taxes, totaling NT$1.441 billion, to be assumed in aggregate by NVIDIA⁴⁰. "Being indispensable does not mean calling the shots" takes its most concrete form here: when you need someone too badly to stay, you will pay bills for them that were never yours to pay.

Is the Headquarters a Signboard or a Root

So what has NVIDIA given Taiwan? This has to be viewed in two parts, otherwise it is easy to misjudge.

One part is the "signboard": the Constellation headquarters, modeled after the "starship" form of NVIDIA's U.S. headquarters, will accommodate about 4,000 people, begin construction in 2026, and open only in 2030. As of today, construction has not begun⁴¹. Looking only at this, questioning whether the headquarters is merely a PR operation is not unreasonable.

The other part is the "root that had already been planted." NVIDIA did not come to Taiwan only in 2025. It already had an office in Taipei's Neihu District, with about 1,800 existing employees, according to media estimates rather than official figures⁴². In 2021, it received Ministry of Economic Affairs approval for its "AI Innovation R&D Center Project," with total investment of NT$24.3 billion and government subsidies of NT$6.7 billion, and a plan to hire 1,000 new R&D staff from 2022 to 2027⁴³. In November 2025, it established "Taiwan NVIDIA Classic Co., Ltd.," increasing capital from NT$1 billion to NT$3.3 billion. This is an independent legal entity that can pay taxes and hold assets on its own⁴⁴.

So the truth of the "headquarters landing" lies in between: the substantive R&D roots are real, and the tax-paying subsidiary has been established; but the most visible Constellation headquarters remains a blueprint. Neither half should be told alone.

The Island Drained by This Supply Chain

Beyond the halo and the bill lies another account, paid by everyone living on this island: water, electricity, air, and housing they cannot afford.

TSMC's Taichung fab. The physical bodies of each NVIDIA GPU generation take shape inside fabs like this, and the water and electricity these fabs draw away are another bill this island is paying. Photo: Briáxis F. Mendes, CC BY-SA 4.0.

Start with electricity. In 2023, TSMC used 24.775 billion kWh of electricity, accounting for 8.96% of Taiwan's total electricity consumption⁴⁵. S&P predicts that by 2030, TSMC's electricity use may reach 23.7% of Taiwan's total⁴⁵. In other words, by then, nearly one in every four kWh of electricity in Taiwan may be consumed by this one company.

Carbon emissions rise with it. Greenpeace's April 2025 report, The Dark Side of the Chip Boom, calculated that electricity use for global AI chip manufacturing surged from 218 GWh to 984 GWh, an annual increase of more than 3.5 times; Taiwan alone saw electricity use soar to 375.8 GWh, "as high as 38% of the global total"⁴⁶. Because TSMC relies heavily on fossil fuels, its AI chip manufacturing produced 185,700 metric tons of CO2 equivalent, and Greenpeace directly named it the "carbon emissions champion" of AI chip manufacturing⁴⁶.

What about NVIDIA itself? Greenpeace gave it an F. The report wrote that NVIDIA's "supply chain emissions nearly doubled over the past three years, rising from 3.51 million metric tons in 2022 to 6.91 million metric tons in 2024," while what it was doing was "merely shifting the supply chain's carbon emissions and pollution onto other parts of the world"⁴⁶. In other words, the valuation and halo sit under NVIDIA's name, while the carbon emissions and pollution remain in Taiwan's sky.

Water, too. TSMC uses more than 200,000 metric tons of water per day; the 81,000 metric tons supplied each day by four reclaimed-water plants in Tainan go "almost entirely to TSMC"⁴⁷. The cost falls on farmland: farmland in Chianan was twice left unirrigated, in 2021 and 2023, to divert water to the semiconductor industry⁴⁸. Making one 12-inch wafer requires 8,327 liters of water⁴⁹, and on this island, some farmers' fields once stopped cultivation so chips could have water.

Then there is housing. After NVIDIA announced that it would establish its headquarters in the Beitou-Shilin Technology Park, home prices in the area moved first. Taipei's Department of Economic Development estimates that the park's resident working population may eventually reach 60,000, but 591 statistics show only about 1,476 homes available for sale, with residential land accounting for only 13.8% of the district⁵⁰. When 60,000 people compete for 1,476 homes, the direction of prices is easy to imagine: reports say new projects in the core Beitou-Shilin Technology Park area have already broken NT$1.5 million per ping⁵¹.

Two things must be separated here to avoid misplacing the anxiety. According to actual-price registration, the highest 2025 residential building transaction in Shilin District was NT$570,400 per square meter, equivalent to about NT$1.88 million per ping, at No. 39 Jihe Road, with a total price of NT$447 million⁵². But those sky-high transactions were luxury homes in Tianmu and urban Shilin, which are different from the Beitou-Shilin Technology Park itself; the figure of new projects breaking NT$1.5 million per ping in Beitou-Shilin Technology Park is another number reported by the media. The two figures should not be conflated. But both point to the same feeling: growth is happening on my street, yet I cannot afford to buy there.

⚠️ "The company has not even broken ground, so why are prices rising?" Local residents' backlash is real. One industry source said privately that because the park's everyday amenities are poor, "many employees are unwilling to move here with the company... and even after moving here, one-third of employees leave because transportation is inconvenient"⁵³. Anonymous opinion on PTT was blunter: the TSMC theme had been used to exhaustion, and now it was simply being swapped out for the NVIDIA theme⁵⁴. A headquarters that has not broken ground has already pushed up local housing prices. This is the version of K-shaped divergence closest to the dinner table. Hope must be honest: AI has indeed made Taiwan visible to the world, and the drained water and electricity and unaffordable housing are also real. Both have to be written down.

Taiwan Is Not the Only One That Cannot Leave It

Pull the camera back, and a larger fact appears: in the relationship between NVIDIA and Taiwan, "cannot leave" is mutual, even multidirectional. Even the other side of the Taiwan Strait is caught in this structure.

The Computex floor at Taipei Nangang Exhibition Center. Every June, buyers from around the world pour into this hall for things made by Taiwan's supply chain. Taiwan has never been the only one unable to leave it. Photo: NVIDIA Taiwan, CC BY 2.0.

This is the idea of the "silicon shield": Taiwan controls chips the whole world needs, and this irreplaceability becomes a layer of protection. But the silicon shield has always been two-sided. It is both a talisman and a powder keg tied to the island. Scholars call these two sides the "Silicon Shield" and the "Silicon Trap": the same concentration can deter aggression, and it can also become an incentive for aggression or a single point of failure⁵⁵.

An even sharper debate was stirred in 2021 by an article from the U.S. Army War College that proposed an extreme "scorched earth" or "broken nest" strategy: if China invades Taiwan, Taiwan should not hesitate to destroy its own semiconductor industry to make the island profitless to the aggressor. The counterarguments were equally forceful: even if such a move successfully deterred China in the short term, the economic self-harm might merely delay aggression until China can produce semiconductors domestically; and Taiwanese people themselves are unlikely to see such industrial sabotage as being in their own interest⁵⁶. This is not a judgment this article will make for Taiwan, but it truly hangs behind every discussion of whether Taiwan has leverage.

The silicon shield is even being diluted by TSMC itself. To diversify geopolitical risk, TSMC is investing US$165 billion in U.S. expansion⁵⁷. The August 2025 headline in MIT Technology Review was "Taiwan's silicon shield could be weakening"⁵⁸. The concern is that moving capacity overseas will dilute Taiwan's domestic leverage and make countries such as the United States feel that Taiwan is no longer so worth defending⁵⁸. But Bonnie Glaser of the German Marshall Fund also reminds us that this ecosystem is not so easy to move: the ecosystem Taiwan created is truly unique, the result of its talent pipeline, culture, and laws, and cannot easily be replicated anywhere⁵⁹. China technology researcher Paul Triolo put it more directly: when it comes to leading-edge manufacturing, Arizona is nowhere near that level, and never will be⁶⁰.

The clearest illustration of how asymmetrical this dependence is came in one political moment.

On May 29, 2024, Jensen Huang publicly said in Taiwan: "Taiwan is one of the most important countries in the world."⁶¹ A few days later, on June 2, he spoke at National Taiwan University, describing Taiwan as "an unsung hero, yet a pillar of the world"⁶².

Eighteen days later, Chen Binhua, spokesperson for China's Taiwan Affairs Office, responded: "Regarding this extremely erroneous statement, people and netizens in China have already expressed strong dissatisfaction. Taiwan has never been a country... I hope he catches up on his lessons."⁶³

But what is more intriguing is something else. CNA noted at the time that Chinese financial media had published many reports on Huang's Taiwan visit, yet "did not appear to mention Huang's statement that 'Taiwan is an important country,' seeming to omit a sensitive issue usually regarded as 'the most important of the important'"⁶⁴. In other words, officials protested sternly, while financial media chose to mute the matter.

📝 Curator's note: This muting reveals where power really sits. China needs NVIDIA's chips, so even after Jensen Huang said the thing Beijing finds most unacceptable, Chinese media chose not to report it or amplify it, fearing damage to the relationship with this "AI godfather." A widely circulated line says: China needs NVIDIA, but NVIDIA does not need China⁶⁵. In this relationship, even the vast market across the Taiwan Strait has, to some degree, had its neck caught by an American company's supply chain. This is the island of Taiwan's strange position: the entire world, including the side that most wants to change its status, cannot do without the chips made here. But "the whole world cannot leave you" and "you are therefore safe and get to decide" remain two different things. This article does not draw a political conclusion for Taiwan, but the tension itself is worth every reader weighing for themselves.

Being Indispensable Does Not Mean Calling the Shots

Return to that wall of 55 company logos.

Every name on the wall is real. They are the physical body of the AI revolution on earth. Without them, the US$5 trillion NVIDIA could not ship a single chip. This irreplaceability is an engineering fact, not rhetoric. Taiwan has every reason to be proud of it.

But after looking across the whole chain, the halo, valuation, and decision-making power fall into the hands of the person projecting the wall; the 5% gross margins, drained water and electricity, unaffordable housing prices, and war risk placed on the island fall on the names on the wall. Taiwan holds the switch the world cannot turn off, but that does not mean it gets to call the shots. And this leverage also has a shelf life written somewhere near 2028.

Taiwan is not standing still. In 2025, Lai Ching-te proposed making Taiwan one of the "world's top five computing centers" and developing "sovereign AI"⁶⁶; Foxconn is building a national supercomputer in Kaohsiung using 10,000 Blackwell chips⁶⁷; the Executive Yuan's "AI New Ten Major Construction Projects" plans to invest more than NT$100 billion, with a target of NT$15 trillion in output value⁶⁸. This is an attempt to grow "compute for ourselves" out of "manufacturing for others": to climb one step upward from the bottom of the smile curve.

But the road is still long. Taiwan's own language model, TAIDE, has been described as being like a "high school student," while international giants are already at the "graduate student" level⁶⁹. The South Korean government purchased 260,000 GPUs in one move, while Taiwan is still going back and forth over one parcel of land and one termination payment⁷⁰. From catching one phone call from Morris Chang to catching the world's computing power, Taiwan has spent nearly thirty years getting onto the wall. But being on the wall and taking back the pen are two different things.

That wall will keep glowing. At the next Computex, Jensen Huang's backdrop will show even more logos. In 2026, he disclosed that NVIDIA's annual spending in Taiwan had reached about US$150 billion, compared with only US$10 billion to US$15 billion five years earlier⁷¹. The question "Is Taiwan important?" was answered long ago. Taiwan has to answer the harder one: when the whole world cannot do without what you make, how do you slowly turn "cannot do without" into "gets to decide"?

The names on the wall are multiplying. Whether the hand holding the pen can become Taiwan's own is a question Taiwan has only just begun to reach for.

Further Reading:

• Jensen Huang: From a Boy Who Cleaned Toilets to the Leather-Jacketed Leader of a US$5 Trillion Empire — The personal life story of NVIDIA's founder. This article only touches on it lightly; his Tainan roots and upbringing are here.
• Semiconductor Industry — Why Taiwan became the center of global chip manufacturing; the supply chain discussed in this article has fuller context here.
• Taiwanese Enterprise: TSMC — The "sacred mountain protecting the nation" that manufactures every NVIDIA chip, and the other side of what it drains.
• Morris Chang: The Recipient of That Letter and the Foundry Empire He Built — The person who received Jensen Huang's letter in 1996, and the founder of TSMC.
• Computex: How Taipei's Computer Show Became the Opening Ceremony of Global AI — The stage where that logo wall lit up, and the annual home court of Taiwan's technology industry.
• AI Artificial Intelligence Industry — From manufacturing NVIDIA chips to building an AI ecosystem, Taiwan's position in the AI wave.
• Taiwan's Artificial Intelligence Development and Future Strategy — Sovereign AI, TAIDE, and Taiwan's national ambition to climb upward from contract manufacturing.
• Taiwanese Enterprise: Foxconn Precision — The contract-manufacturing giant assembling 40% of the world's AI racks, and the largest pair of hands at the bottom of the smile curve.

Image Sources

• Jensen Huang at Computex Taipei — Photo: NVIDIA Taiwan, 2016, CC BY 2.0 (hero, Jensen Huang speaking on the Computex stage)
• NVIDIA Ampere GA102 GPU die — Photo: Fritzchens Fritz, CC0 (micrograph of chip die)
• Jensen Huang holding RTX Blackwell at CES 2025 — Photo: Pronoia, CC0
• TSMC factory in Taichung — Photo: Briáxis F. Mendes, CC BY-SA 4.0
• Computex Taipei at Taipei Nangang Exhibition Center — Photo: NVIDIA Taiwan, 2015, CC BY 2.0
• Video: NVIDIA CEO Jensen Huang Keynote at COMPUTEX 2025 — NVIDIA official YouTube channel

References