NVIDIA's Golden Decade: Is Taiwan Capturing Revenue or Profit?

Chapter 1: Mapping Taiwanese Manufacturers—150 Suppliers and Three ROE Engines

“Taiwan is not merely NVIDIA's supplier; Taiwan constitutes NVIDIA's production capacity.” This is not an exaggeration. NVIDIA has disclosed that Vera Rubin mass production involves hundreds of supply-chain partners, including 150 Taiwanese companies spanning more than 350 factories across 30 countries. Vera Rubin, Vera CPU, and RTX Spark are all produced by TSMC using its N3 process and CoWoS packaging. The Rubin GPU further adopts a 9.5-reticle-size CoWoS-L package—one of the industry's largest advanced-packaging specifications to date.

However, the statement that “Taiwan's supply chain benefits” is itself too broad. Looking only at gross margin can easily produce a misleading conclusion: assemblers have single-digit gross margins and appear to earn only low-margin contract-manufacturing income. The problem is the use of the wrong metric—gross margin merely reflects the shape of value capture, while ROE is the outcome actually delivered to shareholders. Wiwynn has maintained single-digit gross margins for years, yet its quarterly ROE reached 10.66% in 2026 Q1, implying approximately 42% on an annualized basis, higher than TSMC's; trailing-four-quarter EPS reached NT$298.32. Companies that quietly generate substantial returns cannot be properly measured using gross margin alone.

The correct framework is to classify Taiwan supply-chain beneficiaries into three ROE engines, irrespective of whether they operate upstream, midstream, or downstream:

Low gross margins do not imply low economic value. In the NVL72 generation, assembly is not merely low-value labor; it constitutes a technical barrier in its own right—Chapter 3 explains specifically why.

Zoom out, and this article is an extension of the "Taiwan Strategic Upgrade" series. That series makes one argument: Taiwan is not a contract-manufacturing island but an irreplaceable manufacturing ecosystem whose lead compounds year after year (see the "Taiwan's Hidden Champions" installment). NVIDIA's golden decade is the strongest accelerant that compounding has ever had—it is transforming Taiwan along three dimensions at once: quantitative change, as AI orders double or even quintuple the assemblers' revenue scale within four years; qualitative change, as the same companies evolve from notebook ODMs into rack-scale system integrators, fundamentally changing how they make money; and a change in scale, as Taiwan's stock market surpassed the United Kingdom in April 2026 at US$4.14 trillion to become the world's seventh largest, then overtook India barely five weeks later to reach fifth place globally at roughly US$4.95 trillion (the earlier series installment "Taiwan Is Already the World's Seventh-Largest Stock Market—and This Is Just the Beginning" delivered on its own subtitle within five weeks)—with this supply chain as the largest force behind it. The three ROE engines are the yardstick for measuring who, within this transformation, actually converts the wave into shareholder returns.

Chapter 2: The Margin Engine—Who Controls the CoWoS Bottleneck?

TSMC is the clearest representative of the margin engine: its gross margin reached 66.25% and net margin reached 50.51% in 2026 Q1, underpinned by the pricing power created by scarce CoWoS advanced-packaging capacity. TSMC reportedly targets monthly CoWoS capacity of approximately 125,000 wafers by the end of 2026, representing an expansion of 20% to 30% from the end of 2025. NVIDIA has reportedly secured more than 50% of incremental 2026 capacity, which was almost entirely reserved by NVIDIA and Broadcom before coming online. TrendForce estimates that the CoWoS supply-demand gap will narrow from approximately 20% to approximately 10% by the end of 2026.

Nevertheless, CoWoS is not solely a TSMC story. It is the final step in packaging integration, preceded by an entire supply chain in which each segment has its own bottlenecks:

Substrates, HBM, and testing—each layer faces genuine capacity constraints and merits inclusion in the "bottleneck" framework alongside TSMC's packaging lines. And note one thing about this table: aside from the three memory makers, every single segment sits in Taiwan, on the same island. A substrate shortage at Kinsus can be worked out in a same-day meeting; ASE's overflow capacity can be coordinated within the week. This is the cluster effect described in the "Taiwan's Hidden Champions" installment: supply-chain iteration inside the island runs on a cadence of weeks, while cross-time-zone coordination runs on months. NVIDIA's second moat is "supply-chain priority." That lock is anchored in Taiwan—but the key is not held by TSMC alone. It is held by the entire cluster.

Chapter 3: The Turnover Engine—Assemblers Quietly Generating Substantial Returns

This is the article's central analytical focus. Taiwan's five AI server assemblers—Wiwynn, Quanta, Hon Hai, Wistron, and Inventec—mostly operate with single-digit gross margins, yet their ROE levels differ dramatically. The figures below are taken from official financial statements filed through the Taiwan Stock Exchange's Market Observation Post System (MOPS) and calculated consistently as “net income after tax attributable to owners of the parent ÷ average beginning and ending equity.” Quarterly figures are not annualized:

All five companies’ figures are calculated on the same basis from official consolidated financial statements filed with MOPS.

Before reading the table, establish a baseline it does not show: revenue scale. Over the four years from 2021 to 2025, Wiwynn's revenue surged from roughly NT$192.6 billion to roughly NT$950.7 billion—nearly fivefold; Quanta's from roughly NT$1.13 trillion to roughly NT$2.12 trillion—nearly double; Wistron's from roughly NT$862.1 billion to roughly NT$2.19 trillion—about 2.5 times. This is the quantitative change the "Taiwan Strategic Upgrade" series describes: not a few extra orders, but an entire industry's revenue magnitude redefined by AI capital spending. What the ROE table must then answer is the more critical qualitative question: as scale doubled, did the efficiency of making money upgrade along with it? The answer: for some companies yes, for others no—and that is precisely the deepest dividing line among the five.

This table offers three observations worth incorporating into the investment framework:

First, Quanta and Wiwynn jointly validate the existence of an “AI plateau.” Both assemblers' ROE rose markedly in 2024–2025 from a lower base—Quanta moved from its previous norm of 17%–22% to a 29%–32% plateau; Wiwynn's increase was even more dramatic, surging from 29.73% in 2023 to 48.03% in 2025, the highest among the five and the only company approaching 50%. This is not a one-off fluctuation but a structural upgrade in engine efficiency: AI racks carry high average selling prices, lifting net margins at the same turnover rate. Both companies maintained elevated quarterly figures in 2026, indicating that the plateau has not yet receded.

Second, Hon Hai has the weakest engine among the five. Hon Hai's full-year ROE remained within a range of 9.65%–11.08%, while its quarterly ROE fell to just 2.81% and 3.25% in 2026—the gap versus Wiwynn over the same periods was not merely “twofold,” but three to four times. Its gross margin was also among the lowest of the five, at 6.04%–6.30%, but the real drag on ROE is neither leverage nor insufficient asset turnover; it is an excessively thin net margin. A three-factor DuPont analysis makes this clearer:

Third, Wistron and Inventec are the mid-sized and smallest players, respectively. Wistron's ROE trajectory moved in the same direction as Wiwynn's and Quanta's—recovering from its 2023 low, but by a much more moderate degree. Full-year 2025 ROE reached 17.66%, between Hon Hai and the higher levels of Wiwynn and Quanta. Inventec presents a completely different story: its full-year ROE remained within a stable 10.16%–11.79% range, without the sharp increases seen at Wiwynn and Quanta. Its returns are closest to Hon Hai's, though modestly higher. Inventec has positioned itself in the Google TPU ASIC segment—it is the only Taiwanese supplier selected to assemble Google TPU L6 motherboards and is reportedly expanding toward L10 and L11 rack integration. Its estimated AI server market share is only 5%–7%, compared with approximately 40% for Hon Hai, approximately 25%–30% for Quanta, and approximately 8%–10% combined for Wistron and Wiwynn, making it the smallest of the five. However, AI revenue is rapidly increasing as a share of the company's total revenue and is estimated to exceed 50% for the first time in full-year 2026; previously, notebook contract manufacturing was its core business. This is a “small but rapidly transforming” story, distinct from the other four companies, whose core businesses are already centered on AI servers.

Technical Barriers in Assembly

“Assembly” in the NVL72 generation is not simply fastening screws. Liquid-cooling integration requires the design, assembly, and leak testing of full-rack liquid-cooling loops—a single drop of leakage can destroy an entire rack of GPUs. The NVLink spine involves backplane assembly and testing for the interconnection of 72 GPUs, with yield directly determining profitability. L10 and L11 system-level design requires ODMs not only to manufacture products but also to participate in mechanical, thermal, and power-system design. Hon Hai's Ingrasys and Quanta's QCT already possess capabilities comparable to those of branded vendors. The supplier that first raises yield on a new platform secures the first wave of high-value orders. Customer concentration itself demonstrates the moat: customized racks for Microsoft and AWS are concentrated at Wiwynn, while Meta and Google allocate theirs to Quanta. Hyperscalers do not switch assemblers casually; switching costs are real. This is the mechanism behind “quietly generating substantial returns”: a turnover engine combined with technical barriers means that low gross margins and high ROE are not contradictory, but intrinsic to the business model. Hon Hai, Quanta, and Wistron are expected to ramp production in the second half, with large-scale shipments beginning as early as 2026 Q3.

Chapter 4: The Expansion Engine—Rising Content Value in Power and Thermal Management

The third engine consists of companies whose content value and gross margin rise simultaneously. Delta Electronics is the representative example: its gross margin reached 37.0% in 2026 Q1, while full-year 2025 ROE reached 22.5%, making it a primary beneficiary of rising power-system content value. NVIDIA's own technical blog has established 800V HVDC as the next-generation power architecture, and both Delta Electronics and Lite-On are official partners for GTC 2026 and Computex 2026. Market estimates—not official figures—indicate that Delta Electronics has secured approximately 60%–65% of NVIDIA's 800VDC ecosystem, compared with approximately 20% for Vertiv, with mass production scheduled for 2026 Q3. Lite-On, meanwhile, is jointly developing a solution with Wolfspeed and targets mass production in November 2026.

Chapter 5: Constellation—An Expansion, Not a First Entry

NVIDIA has operated in Taiwan for nearly 30 years. Its earliest base was in the Neihu Technology Park, which has long served as an operating center for AI software R&D and technical support. At the end of 2025, NVIDIA first relocated its R&D center from Neihu to floors 3 through 17 of the Ruentex Yucheng Office Building in Nangang under a 10-year lease, accommodating more than 1,000 employees—this is an interim location pending Constellation's opening. During his May 2025 visit to Taiwan, Jensen Huang personally explained the reason for the expansion: the Neihu office had become so crowded that “if one employee wanted to sit down, another had to stand up.” The joke directly confirms that Constellation is an expansion necessitated by insufficient space at the existing site, not NVIDIA's first entry into Taiwan. According to official statements and Taiwanese media reports, the Neihu, Nangang, and Constellation sites will coexist as a technology corridor; Constellation will not replace the first two.

The description “first overseas headquarters” is not incorrect, but its scope must be clearly defined—it refers to NVIDIA's first purpose-built overseas headquarters campus, not its first location in Taiwan.

The construction timeline requires particular clarification: Jensen Huang attended a ceremonial groundbreaking event on May 27, 2026, but that did not constitute the start of actual construction. Taipei's Construction Management Office stated at the time that NVIDIA had not yet applied for a building permit, which is legally required before construction can begin. Some media reports suggested that construction was targeted for June or July, but as of August 2026, no source had confirmed that physical construction had begun. More recent reports even indicated that actual construction might not start until the end of 2026 at the earliest. The official phased timetable for the NT$40 billion investment has not yet been disclosed.

Taiwan's Own AI Infrastructure: Both Supplier and Customer

Hon Hai is using the NVIDIA Blackwell architecture to build Taiwan's fastest AI supercomputer in Kaohsiung. The first phase became operational in mid-2025, with full deployment scheduled for 2026. The National Science and Technology Council's 2027 technology budget reached a record NT$176.8 billion, with additional funding for the “Ten Major New AI Infrastructure Projects” and sovereign AI. TAIDE is reportedly adopting NVIDIA's latest AI chips. Taiwan AI Cloud, an ASUS subsidiary, focuses on healthcare and national sovereign-AI projects; it entered the Emerging Stock Market in May 2026 and plans to file for a main-board listing by year-end.

Chapter 6: Going Global—Capital Allocation on the World Stage

July 2026, Fort Worth, Texas. Wistron's D1 AI smart factory officially came online, and off its production line came the first NVIDIA GB300 Grace Blackwell Ultra superchip ever manufactured on U.S. soil—with preparations for the next-generation Vera Rubin platform already underway inside the same plant. For the first time, "Made in USA" appeared on the origin label of NVIDIA's most advanced product. The company that made it happen was Taiwanese.

How should this be characterized? The reflexive reaction has always been anxiety: is Taiwan's capacity being moved away? But shift the coordinate system, and this is the standard playbook of a multinational corporation—American companies allocate capacity globally, EU companies allocate capacity globally; capital goes where the customers are, where policy requires, where power and land make sense. Taiwan's five assemblers have already reached multinational scale—combined 2025 revenue of roughly NT$14.05 trillion. Building plants in America is not a retreat; it is global capital allocation—the first time Taiwanese companies are genuinely competing on the world stage as multinationals in their own right.

Nor is this a one-company story; the whole plate is shifting. Hon Hai is making an additional US$569 million investment in Wisconsin, adding 1,400 jobs over four years and doubling its scale there by 2030; its Houston, Texas subsidiary, Foxconn Assembly LLC, is investing an additional US$295 million to build AI servers for North American customers and will put NVIDIA Isaac GR00T N humanoid robots onto its production lines; the company guides full-year 2026 capital expenditure up more than 30% YoY. Quanta's president has publicly stated the company will add three U.S. plants by the end of 2026 to support the Vera Rubin ramp; subsidiary QCT has signed a 10-year, US$61.71 million lease on a California facility, and the company projects full-year capital expenditure of NT$30 billion. Of the five assemblers, the three largest are all putting down roots in America.

More importantly, what this wave of expansion carries abroad is no longer just hardware capacity. Hon Hai's Ingrasys and Quanta's QCT do own-brand-grade system design, not bare contract work; Delta Electronics and Lite-On participate directly in NVIDIA's 800VDC next-generation power-architecture ecosystem, staking positions in an industry standard; Hon Hai is building Taiwan's fastest AI supercomputer in Kaohsiung on the Blackwell architecture, and ASUS subsidiary Taiwan Web Service is carrying national-level sovereign AI projects—that is compute as a service; the Houston plant will put NVIDIA Isaac GR00T N humanoid robots onto its production lines, training next-generation manufacturing capability. Hardware, software, services, industry standards—Taiwan's supply chain is progressively building the next generation of competitiveness, and that is the full meaning of playing on the world stage.

Will Taiwan itself be diluted, then? A key number from the "Taiwan's Hidden Champions" installment can settle that anxiety first: TSMC's U.S. fab cost nearly 50% more to build than its Taiwan fabs—not because of wages, but because America lacks the surrounding ecosystem. That is the price of tacit knowledge: factories can be moved; clusters cannot. The real question posed by this migration is not "where is the production line" but "can the yield knowledge and iteration speed leave the island"—and so far, the answer is that the capability now running in Wistron's Texas plant was grown inside Taiwan's cluster and then transplanted. However far the away games go, the home stadium and the training ground remain in Taiwan.

For investors, the reading needs an update. The old reading treated migration as a risk; the new reading treats it as a necessary stage in Taiwanese companies' upgrade into genuine multinationals. The financial tuition is real—depreciation and labor costs at U.S. plants will compress already-thin gross margins, a number visible in every quarterly report from here on—but what it buys is delivery closer to the customer, diversified policy risk, and the pricing confidence that comes with a global footprint. The Taiwan-level question remains: the jobs and tax base that stay in Taiwan do not move with the orders—but a team that plays in the World Cup brings home far more than gate receipts.

Will the Taiwan Supply Chain's Golden Decade Be a Golden Decade for Taiwan, or for Taiwanese Companies?

Finally, place this article back into the narrative it belongs to. Taiwan's irreplaceability is not static; it deepens at a compounding rate—every passing year raises the cost and time required to replicate Taiwan a little further. NVIDIA's golden decade is feeding all three dimensions of that compounding at once: order magnitudes doubling is the quantitative change, assembly upgrading into systems integration is the qualitative change, and Taiwan's stock market reaching fifth in the world is the change in scale. Constellation is NVIDIA doubling down on this ecosystem, not placing a new bet; Wistron's Texas plant is the first true stress test—testing not whether Taiwanese companies can replicate production lines abroad (that is now proven), but whether the cluster that makes everything work will dilute along with the orders.

On that question, this article's answer leans toward "the golden decade of Taiwan": this is not dilution—it is frontier deployment of Taiwanese technology and intellectual property. Yield knowledge, rack-scale system design, a seat at the table on power-architecture standards—the most valuable assets are owned by Taiwanese companies and embodied in engineering teams grown in Taiwan. The Texas, Wisconsin, and California plants deploy technology that Taiwan owns to the frontier closest to customers and policy—just as American multinationals keep design in California and put manufacturing around the world. What each overseas plant earns back is profit for Taiwanese shareholders, bargaining power for Taiwanese headquarters, and a path along which "the golden decade of Taiwanese companies" and "the golden decade of Taiwan" converge into the same thing. Taiwan no longer exports only products—it exports the technology itself. That is the true compounding of the golden decade.