<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Julien Nauy on Medium]]></title>
        <description><![CDATA[Stories by Julien Nauy on Medium]]></description>
        <link>https://medium.com/@JulienNauy?source=rss-61170b42367d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*tu2UKc9s-OBJZCZrXQ0Egg.png</url>
            <title>Stories by Julien Nauy on Medium</title>
            <link>https://medium.com/@JulienNauy?source=rss-61170b42367d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 19:07:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@JulienNauy/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How Meta Green Cooling Is Entering the AI Liquid Cooling Market]]></title>
            <link>https://medium.com/@JulienNauy/how-meta-green-cooling-is-entering-the-ai-liquid-cooling-market-a37e242dd6f3?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/a37e242dd6f3</guid>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 11:38:58 GMT</pubDate>
            <atom:updated>2026-10-15T11:38:58.414Z</atom:updated>
            <content:encoded><![CDATA[<h3>From Wistron Orders to Data Center Construction: How Meta Green Cooling Is Entering the AI Liquid Cooling Market</h3><p>Once an AI server is connected to a liquid cooling loop, heat begins to flow away from its chips. But installation and maintenance challenges are only beginning. How pressure changes with temperature, where metal debris accumulates, and what takes over when equipment fails all affect whether a data center can keep running. For a liquid cooling supplier, delivery may extend well beyond shipping equipment from the factory.</p><p>Taiwan’s Meta Green Cooling (MGC) is building its business around this work. Founded in 2022, the company reported approximately NT$970 million in revenue for the first half of 2026. It supplies coolant distribution units and also integrates and installs data center cooling systems. Its listed customers include Wistron, Realtek and Global Unichip (GUC). The young company’s entry point is bringing equipment, piping, controls and on-site engineering together into a system it can deliver to the customer.</p><h3>One Leak Exposes a System-Wide Problem</h3><p>MGC President Clyde Chu recalls an early liquid cooling installation in which the team needed repeated inspections to identify the cause of a leak: liquid expanding as it warmed inside a closed loop, raising the pressure. The solutions included adding an expansion tank, monitoring temperature, pressure and flow, and arranging for backup equipment to take over when coolant delivery fell short.</p><p>Replacing a single component rarely resolves this kind of failure. The system design must allow for pressure changes, the controls must detect abnormalities promptly, and backup equipment must be ready to operate when needed. Even if each part works individually, engineers still have to establish how the parts respond together. The installation team’s accumulated value includes identifying where a failure originates and whether a repair will create a problem elsewhere.</p><p>This field experience grew out of MGC Chairman James Chen’s long career as an entrepreneur and executive in thermal management. In 1994, he founded Huashi, moving from distribution into CPU coolers and mounting clips. The business was later acquired by the US company Aavid, where he became Asia-Pacific president. In 2001, he founded Youju to make heat pipes and other thermal components. After the US company Vette acquired Youju in 2004, Chen served as Vette’s Asia president, while Chu was its chief technology officer.</p><p>At Vette, their work expanded from cooling electronic products to complete liquid cooling systems. As the company moved into liquid cooling technology used in IBM supercomputers, its Asian team took responsibility for system design and installation. Chen and Chu led engineers into university laboratories in different countries and the computing facilities of Taiwan’s Central Weather Bureau. They connected coolant piping and equipment, installed control software, and dealt with problems that emerged during operation. The leak, pressure monitoring and backup arrangements described above were part of that field experience.</p><p>After leaving Vette, Chen also joined Kenmec as general manager of its thermal division, leading a team that developed CDUs, liquid-cooled rear doors and liquid cooling components. His work continued to focus on liquid cooling product lines and the equipment needed to make them work as a system, before he founded MGC in 2022.</p><p>From CPU coolers and heat pipes to liquid cooling equipment and data center installation, Chen accumulated experience spanning components and on-site engineering. This helps explain why MGC could take on work beyond individual pieces of equipment early in its life. The company was young, but its core team had already dealt with the practical problems that arise when equipment is connected to different data centers.</p><h3>The CDU Connects Chip Cooling Loops to Facility Infrastructure</h3><p>A coolant distribution unit (CDU) sits at a critical interface in a liquid cooling system. In a liquid-to-liquid architecture, a heat exchanger transfers heat from the server-side coolant to the facility-side water loop while keeping the two fluids separate and regulating coolant supply conditions. Cold plates remove heat close to the chips; the CDU manages how that loop works with the facility’s cooling infrastructure.</p><figure><img alt="An in-row CDU assembly shown in an OCP document, with pumps, vertical piping and insulated components inside the frame" src="https://cdn-images-1.medium.com/max/689/1*pUj-Pp20QHZYd2md8ijy9g.jpeg" /></figure><p>Figure 1. An example of the internal assembly of an in-row CDU. Source: Berhanu Wondimu et al., OCP, <em>Liquid to Liquid CDU Test Methodology and Performance Rating</em>, Revision 1.0, p. 12, original Figure 4. Image unchanged, used under <a href="https://creativecommons.org/licenses/by-sa/4.0/">CC BY-SA 4.0</a>. <a href="https://www.opencompute.org/documents/ocp-wp-l-lcdu-test-methodology-performance-rating-r1-pdf">Original document</a>.</p><p>Chen says MGC initially made cold plates for ASUS motherboards. But with cold plates positioned close to expensive chips, the potential losses from a leak led the company to shift its focus to CDUs, which the team already knew well. The decision illustrates how a startup weighs its technical experience against the risks it can afford to carry. Moving into CDUs and system integration meant managing more connections and operating conditions between the cooling equipment and the customer’s site.</p><p>Materials and cleanliness can also change the scope of the work. Chen notes that customers have requested replacements for materials that could shed particles. Metal debris flushed out of cold plates can also accumulate along the piping, prompting MGC to develop filtration equipment. Once a problem moves with the coolant across equipment boundaries, the integrator needs to understand upstream components, the circulation loop and downstream loads. The value of a filter also depends on where it sits in the overall system and how it is maintained.</p><h3>Wistron Brings Engineering Work Beyond the Equipment</h3><p>Retrofitting existing data centers gave MGC another early opening. Chen recalls that when he first proposed liquid cooling to Realtek, the initial obstacle was the customer’s concern about bringing water into the server room. A retrofit must fit existing spaces and piping, and users need to accept a new way of operating the facility. Explaining equipment performance is only part of the job; site constraints and customer concerns also have to be addressed.</p><p>As projects increased, MGC’s work extended into data center water, electrical and network infrastructure, as well as coolant piping within racks. Chu compares turnkey construction to renovating a building: different trades need access to the site, and getting the sequence wrong can mean tearing out completed work and starting again. Customers are therefore buying construction coordination, problem-solving and responsibility for delivery. The staffing and judgment these tasks require are difficult to capture fully on a product specification sheet.</p><p>Wistron has played the roles of both investor and customer. MGC’s corporate history records Wistron’s investment in 2024 and lists Wistron and its US subsidiary WisLab among subsequent customers for data center infrastructure. Wealth Magazine also describes their collaboration on liquid cooling equipment and data center engineering. When server manufacturers need testing and production environments, a supplier’s work extends beyond the equipment itself into the layout of production lines and facilities.</p><p>MGC’s publicly described relationship with NVIDIA centers on product certification. MGC says its 1.3 MW TC1288 liquid-to-liquid CDU received NVIDIA certification in 2026. For a supplier, meeting platform requirements and deploying equipment at a customer’s site are successive tasks. The first makes a product eligible for use in the relevant systems; the second still requires work on each site’s piping, controls and construction arrangements.</p><h3>Large Suppliers Are Also Competing for the Whole Data Center</h3><p>MGC’s competitors are expanding their services too. Daniel Chang, Delta’s general manager of Taiwan sales, says AI data center planning must address power supply and distribution, liquid cooling, floor loading, fire protection, environmental sensing, and operations and maintenance together. Delta’s turnkey services cover design, procurement, construction, testing, acceptance, handover and maintenance. Large suppliers already have an incentive to take on these stages as a package.</p><p>Equipment comparisons are also becoming more systematic. The Open Compute Project’s CDU testing and performance document evaluates cooling capacity in relation to conditions such as water temperature, flow and pressure. Its co-authors include Chu and engineers from Intel and CoolIT Systems. As the industry develops common methods of comparison, MGC needs to translate its familiarity with equipment into successful delivery across different sites.</p><p>For MGC, the test of growth will lie in people and processes. More projects require more engineers who can diagnose site problems, as well as ways to turn senior staff’s experience into installation, testing and maintenance procedures. If orders grow faster than delivery capacity, the integration services that attract customers will also bring schedule and coordination pressures that the company must carry itself.</p><h3>Sources</h3><ul><li>Zhang Ruxian, “How a Four-Year-Old Cooling Startup Entered the Data Centers of Wistron and NVIDIA,” <em>Business Weekly</em>, issue 2031, pp. 54–56.</li><li>You Xiaoyan, “Meta Green Cooling Improves Cooling Efficiency and Drives an AI Data Center Revolution,” <em>Wealth Magazine</em>, issue 761.</li><li>Tang Ziqing, “<a href="https://www.mirrormedia.mg/story/20211201ind004">Intel Certification Makes It Asia’s Only Partner: How Kenmec Is Winning Business from Major US Companies</a>,” <em>Mirror Media</em>.</li><li>“<a href="https://www.netadmin.com.tw/netadmin/zh-tw/magazine/-Feature/8CD64A3BE58D49C6A6BCC2DBD47501A2">Turnkey Engineering Integrates High-Density Data Centers and Reduces Conversion Losses from Grid to Chip</a>,” <em>NetAdmin</em>, issue 247.</li><li>Meta Green Cooling, “<a href="https://mg-cooling.com/%E9%97%9C%E6%96%BC%E5%85%83%E9%88%A6/">About Us</a>” and “<a href="https://mg-cooling.com/%E6%8A%95%E8%B3%87%E4%BA%BA%E6%9C%8D%E5%8B%99/%E8%B2%A1%E5%8B%99%E8%B3%87%E8%A8%8A/">Financial Information</a>.”</li><li>Berhanu Wondimu, Edward Kung, David Zhou, Brad Zakaib and Clyde Chu, “<a href="https://www.opencompute.org/documents/ocp-wp-l-lcdu-test-methodology-performance-rating-r1-pdf">Liquid to Liquid CDU Test Methodology and Performance Rating</a>,” Open Compute Project, Revision 1.0, pp. 1–2, 5, 11–20.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a37e242dd6f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[TSMC Powers AI Chips. Why Does Taiwan Still Need Japanese Materials?]]></title>
            <link>https://medium.com/@JulienNauy/tsmc-powers-ai-chips-why-does-taiwan-still-need-japanese-materials-72b0064e10ba?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/72b0064e10ba</guid>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 09:21:10 GMT</pubDate>
            <atom:updated>2026-10-15T09:21:10.918Z</atom:updated>
            <content:encoded><![CDATA[<p>Trace Taiwan’s AI manufacturing capabilities upstream, and two companies come into view: Ajinomoto, known for seasonings, and Nittobo, whose origins lie in textiles. Their insulating films and specialty glass fibers go into chip substrates and high-speed circuit boards, helping determine what Taiwanese manufacturers can produce.</p><p>For readers in Europe and the United States accustomed to viewing the AI supply chain through Nvidia and TSMC, this relationship deserves a closer look. Once a chip design goes to Taiwan for manufacturing, its materials supply chain still extends to Japan. Taiwan’s manufacturing strengths depend on continued cooperation with upstream materials suppliers.</p><h3>Nan Ya’s Partnership Shows What Taiwan Needs</h3><p>Nan Ya Plastics is a major Taiwanese producer of electronic materials, with products spanning glass fiber cloth, copper foil, epoxy resins, and copper-clad laminates. Even with that range of capabilities, it chose to partner with Japan’s Nittobo to strengthen its position in high-end copper-clad laminates.</p><p>In November 2025, Nan Ya announced that Nittobo would provide a long-term supply of NER, its second-generation low-dielectric glass yarn. Nan Ya, in turn, would help weave specialty glass cloth for Nittobo. Under the plan announced by Nan Ya, the Taiwanese company is expected to weave 20% of the specialty glass cloth Nittobo supplies to the market by 2027.</p><p>This division of labor makes the relationship between Taiwan and Japan tangible. Nan Ya contributes weaving and volume manufacturing capacity while securing a critical yarn supply from Japan. As Taiwanese companies take on more advanced manufacturing, partnerships at the raw-material stage become more important. For buyers, establishing that a supplier has a factory in Taiwan is only the beginning: they also need to ask whose materials it uses.</p><h3>Ajinomoto’s Film Is Part of Taiwan’s Substrate Manufacturing Process</h3><p>ABF stands for Ajinomoto Build-up Film, an insulating material used in substrates. The ABF substrates produced by Taiwanese companies such as Unimicron, Nan Ya Printed Circuit Board, and Kinsus are products made by processing materials into multilayer circuitry. Film and substrate production occupy different stages of manufacturing.</p><p>Connections on a chip must pass through the substrate to reach external circuitry. Copper traces within the substrate need electrical insulation from one another, while fine circuitry must be built up layer by layer. Ajinomoto combines epoxy resin, a hardener, and inorganic microparticles in a film that provides insulation, heat resistance, and processability. Its surface also accommodates laser processing and copper plating.</p><figure><img alt="Ajinomoto technical illustration: surface microcircuits on the left; on the right, a semiconductor connects through a multilayer build-up substrate to the printed circuit board below." src="https://cdn-images-1.medium.com/max/1024/1*_9clMCj-5kZQP7BhL2bHwg.png" /></figure><p>Ajinomoto’s illustration shows how a multilayer build-up substrate connects a chip’s terminals to external circuitry. ABF provides interlayer insulation in this type of substrate. Source: Ajinomoto, “Ajinomoto Build-up Film (ABF).”</p><p>Ajinomoto identifies several development challenges, including uniformly dispersing organic and inorganic components that do not readily mix, and adapting material properties and processing conditions to customer requirements. Heat resistance, copper plating, and laser processing all require ongoing testing. This suggests that replacing a material involves a whole set of process conditions: can the new material work with the manufacturing methods, produce circuitry that meets requirements, and maintain quality?</p><p>That helps explain why Taiwanese substrate makers must also secure upstream supplies when they expand capacity. Unimicron chairman SC Chien has discussed tight material supplies at a shareholders’ meeting and traveled to Japan to visit major suppliers. Substrate manufacturers still need materials to arrive on time for their prepared factories and equipment to be put to use.</p><h3>Glass Fiber May Look Like Cloth, but Its Properties Determine Its Use</h3><p>Nittobo’s specialty glass fibers serve different purposes. T-glass emphasizes low thermal expansion and a high elastic modulus, making it suitable for high-density package substrates. NE-glass and the more advanced NER emphasize low-dielectric properties for high-speed signal transmission. The former addresses dimensional stability; the latter addresses electrical performance and transmission loss.</p><p>Glass fiber suppliers therefore need to deliver material properties suited to a particular application, with consistency maintained through subsequent processing. Grouping every product under the same “high-end glass fiber” label obscures the different problems each is designed to solve.</p><p>Tetsuya Kawasaki, chairman of Nittobo’s Chiayi operation, has explained that temperature, humidity, and operators affect production conditions. Even with a set of conditions supplied from Japan, the Taiwanese plant still needs to make adjustments on site. The Chiayi plant has also sent groups of employees to Japan for training. These efforts bring materials expertise into volume production, turning a formulation into products that can be supplied consistently.</p><p>Seen this way, Taiwan’s reliance on Japan takes several concrete forms: buying products from Japanese materials suppliers, using their technology to manufacture in Taiwan, and processing Japanese yarn into specialty glass cloth through partnerships. Nan Ya’s arrangement with Nittobo expands Taiwan’s manufacturing capabilities while linking the two companies’ materials and processes more closely.</p><h3>Western Buyers Need to Look Upstream of Taiwan’s Manufacturing</h3><p>European and American technology companies obtain chips, substrates, and servers through Taiwanese supply chains. In doing so, they also draw indirectly on capabilities accumulated by Japanese materials suppliers. From Ajinomoto’s insulating film to Nittobo’s glass fiber, material properties influence downstream processing and product delivery.</p><p>This adds a practical question to supply chain management: when downstream manufacturers expand capacity, can the supply, specifications, and qualification of the corresponding materials keep pace? Nan Ya and Nittobo’s arrangement combines yarn supply with access to more weaving capacity, illustrating how production can be coordinated across companies and national borders.</p><p>Taiwan’s role in AI manufacturing needs to be understood within this broader supply network. Alongside TSMC and Taiwan’s electronics industry, buyers need to see the Japanese materials companies supporting the supply chains in which they operate.</p><h3>Sources</h3><ul><li>Wei Sheng-feng, “Japan’s Advanced Materials Stocks Become Hidden Champions,” <em>Wealth Invest Weekly</em>, issue 2426.</li><li>Lin Yuan-qing, “Understanding the ABF Substrate Shortage,” <em>Wealth Magazine</em>, issue 771.</li><li>“The Story of AI Glass Fiber Cloth in Taiwan’s Pineapple Fields,” <em>Business Weekly</em>, issue 2006, highlights edition.</li><li>Nan Ya Plastics, “<a href="https://www.npc.com.tw/j2npc/zhtw/newsdetail/5JEFBU9NW1C">Nan Ya and Nittobo Form a Strategic Partnership in Specialty Glass Cloth to Meet Rapid Growth in AI Supply Chain Demand</a>,” November 28, 2025.</li><li>Ajinomoto Group, “<a href="https://www.ajinomoto.com/innovation/our_innovation/buildupfilm">Ajinomoto Build-up Film (ABF)</a>.”</li><li>Nittobo, “<a href="https://www.nittobo.co.jp/eng/business/electronicmaterials/index.htm">Electronic Materials Business</a>.”</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72b0064e10ba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Are Taiwan’s Suppliers Doing Before AI Chips Enter Mass Production?]]></title>
            <link>https://medium.com/@JulienNauy/what-are-taiwans-suppliers-doing-before-ai-chips-enter-mass-production-837e482b21cc?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/837e482b21cc</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[semiconductors]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 14:41:23 GMT</pubDate>
            <atom:updated>2026-10-12T14:41:23.725Z</atom:updated>
            <content:encoded><![CDATA[<h3>From WinWay to NVLink Fusion: What Are Taiwan’s Suppliers Doing Before AI Chips Enter Mass Production?</h3><p>Before an AI chip enters mass production, one practical question has to be answered: how will it be tested?</p><p>Even after the compute core and package have been designed, the test interface still has to work reliably with high-speed signals, large numbers of contacts, and heavy thermal loads. Unstable contact or signal distortion at the interface can make test results a poor reflection of the chip’s actual performance. These conditions shape test socket design and give suppliers a reason to join development before mass production begins.</p><p>From WinWay’s test interfaces to MediaTek’s collaboration with NVIDIA on custom chips, a common feature of these supplier relationships emerges: some suppliers establish their role during design and validation. Competition therefore extends to their ability to solve engineering problems together with customers, and to the systems their products will ultimately join.</p><h3>Why Do Test Sockets Need to Enter the Design Process Early?</h3><p>WinWay is a Taiwanese supplier of semiconductor test interfaces. A test socket connects a packaged chip to test equipment and must accommodate the chip’s contact layout, signal requirements, and thermal conditions during testing. Changes to the chip or package may require this interface to be redesigned as well.</p><p>WinWay chairman Mark Wang has said that testing has become an important part of product development, requiring early participation in specification design. WinWay’s engineering services cover simulation, mechanical design, and thermal solutions, and the company has built teams for signal integrity and power integrity simulation. For customers, this means test requirements can be discussed alongside package and product specifications, instead of being addressed only before shipment.</p><p>The solutions WinWay announced in 2024 included a 224 Gbps coaxial test socket and high-power thermal management products. Jason Chen, senior vice president of global business operations, said WinWay continued to work with IC design and custom-chip customers on joint development. These products and descriptions of collaboration illustrate an engineering role in which the test interface must address electrical, mechanical, and thermal requirements together.</p><figure><img alt="Front-view product photograph of WinWay’s coaxial spring-pin test socket, showing a silver metal frame, a dense array of openings, and the WinWay Tech. logo." src="https://cdn-images-1.medium.com/max/600/1*NkJOTCMm6kU3OIblNi4H2g.png" /></figure><p>Figure 1. WinWay’s coaxial spring-pin test socket, with its dense array of openings and metal frame visible from the front.</p><p>Image source: WinWay, <a href="https://www.winwayglobal.com/zh-TW/product-detail/Md84seNWgPSoVVb1?class=package">coaxial spring-pin test socket product page</a>, original product photograph.</p><p>Early engineering work also involves commercial arrangements. NRE, or non-recurring engineering, generally refers to a one-time engineering development fee for a specific project. It may cover design, simulation, or validation; the actual scope of work and subsequent supply depend on the individual arrangement. In analyzing a supply chain, the questions worth asking are when a supplier becomes involved, what problem it solves, and whether its work passes customer validation.</p><h3>How Does Joint Problem-Solving Change a Supplier’s Position?</h3><p>A test socket supplier that can offer design recommendations before the package is finalized has an opportunity to influence how the interface is implemented. Its work becomes connected to the customer’s development schedule: which signals need simulation, how contacts maintain reliable connections, and how heat is removed from the test area all have to match the product’s requirements.</p><p>This also offers a way to assess the depth of a collaboration. If changing suppliers means adapting the mechanical design, repeating electrical simulations, or confirming test conditions, procurement decisions must account for that engineering work and the time it requires. The size of the effect depends on whether the product is customized, whether the design can be transferred, and whether alternatives have already been validated.</p><p>Early involvement does not guarantee long-term orders. The package, power consumption, and signal requirements of the next chip generation may change, and existing solutions will need to be evaluated again. Taiwanese suppliers’ ability to remain in the development process still depends on the engineering results they deliver for each generation.</p><h3>MediaTek and NVIDIA: Which Platform Will Custom Chips Join?</h3><p>MediaTek’s collaboration with NVIDIA operates at another level. It addresses how custom compute chips connect to a larger AI system.</p><p>On August 31, 2026, the two companies announced an expanded partnership. NVIDIA’s announcement explicitly stated that it had invested in US$3.5 billion of convertible bonds issued by MediaTek. MediaTek would adopt the NVLink Fusion platform to help customers integrate custom compute chips into rack-scale AI systems connected by NVLink. The collaboration builds on an existing relationship: MediaTek was already listed among the initial partners when NVIDIA introduced NVLink Fusion in 2025.</p><p>Custom chips are commonly called ASICs, or application-specific integrated circuits. After designing their own compute cores, customers still have to address data interconnects, memory, packaging, and system integration. SerDes, the interface for transmitting and receiving high-speed serial data, is one of the engineering capabilities involved. MediaTek’s role along this path is to connect custom-chip design with the integration work that follows.</p><p>NVLink Fusion provides a route into NVIDIA systems. Johnny Shen, Alchip’s chairman and president, has argued that cloud providers would need considerable time to build an inter-chip interconnect platform from scratch. Adopting an existing solution can reduce development risk and accelerate progress.</p><p>For customers, this is a system-level choice: how much differentiation to retain in the compute core, and which interconnect and integration tasks to entrust to an existing platform. For MediaTek, the engineering value also extends to whether the chip can work with the surrounding technologies. Completing and validating that work creates the conditions for moving toward actual deployment.</p><figure><img alt="NVIDIA’s original NVLink Fusion deployment diagram, with three columns showing NVIDIA rack architecture, the addition of a custom CPU, and the addition of a custom accelerator. The custom CPU column is labeled NVLink-C2C, and the custom accelerator column is labeled NVLink." src="https://cdn-images-1.medium.com/max/1024/1*gFuCIyVpRLAlwWrt7vfE1Q.jpeg" /></figure><p>Figure 2. NVIDIA’s NVLink Fusion deployment configurations. The middle column connects a custom CPU through NVLink-C2C, while the right column integrates a custom accelerator through NVLink.</p><p>Image source: NVIDIA Technical Blog, <a href="https://developer.nvidia.com/blog/integrating-custom-compute-into-rack-scale-architecture-with-nvidia-nvlink-fusion">Integrating Semi-Custom Compute into Rack-Scale Architecture with NVIDIA NVLink Fusion</a>, original Figure 2.</p><h3>Capital, Joint Engineering, and Platforms Solve Different Problems</h3><p>Capital partnerships, joint engineering, and platform integration can work together within the same supplier relationship, while serving different purposes.</p><p>NVIDIA’s purchase of MediaTek’s convertible bonds is a financing collaboration that has already taken place. Its effects on products will depend on subsequent progress in design, validation, and manufacturing. Joint engineering brings suppliers into the work of adapting products, as illustrated by WinWay’s test interfaces. Platform integration addresses system compatibility, with MediaTek and NVIDIA’s NVLink Fusion collaboration offering another path.</p><h3>Taiwan’s Opportunity Depends on the Next Generation of Engineering Results</h3><p>WinWay and MediaTek play different roles, but they point to the same industry question: can suppliers turn a customer’s design into a solution that can be validated and integrated before mass production begins?</p><p>For Taiwanese companies, this capability can extend collaboration into earlier stages of product development. It may also increase dependence on a particular customer’s specifications or platform roadmap. Whether the engineering work can be reused across generations or support other customers will affect the long-term value of the relationship.</p><p>Three developments are worth watching: whether a solution completes customer validation; whether the next product generation continues to use or build on those capabilities; and whether the collaboration reaches actual mass production and adoption by multiple customers. Investment amounts and technology partnership announcements provide a starting point. These subsequent results reveal the supplier’s actual position.</p><p>Some of the competition in the AI supply chain is already taking place before procurement for mass production. For Taiwanese companies, the ability to solve problems involving signals, heat, packaging, and system interfaces at that stage will be an important clue to their industrial value.</p><h3>Sources</h3><ul><li>Wei Shengfeng, “Global Giants Play the Taiwan Card in the Battle for Resources and Market Position,” <em>Wealth Invest Weekly</em>, issue 2425.</li><li>Lin Yuanqing, “The World’s Largest Test Socket Supplier: Three WinWay Advantages Attract Major US Companies,” <em>Wealth Magazine</em>, issue 747.</li><li>Huang Yiyun, “What Is NVIDIA’s Latest AI Interconnect Technology, NVLink Fusion? Jensen Huang’s Decade of Preparation, with Alchip and MediaTek Named as Partners,” <em>CommonWealth Magazine</em>, included in issue 825. <a href="https://www.cw.com.tw/article/5135648">Article</a>.</li><li>WinWay, “WinWay Launches Four Advanced Testing Solutions for the AI Chip Era,” September 5, 2024. <a href="https://www.winwayglobal.com/zh-TW/news-detail/paPnysb9jdKwGpK1">Company announcement</a>.</li><li>NVIDIA, “NVIDIA Unveils NVLink Fusion for Industry to Build Semi-Custom AI Infrastructure With NVIDIA Partner Ecosystem,” May 18, 2025. <a href="https://nvidianews.nvidia.com/news/nvidia-nvlink-fusion-semi-custom-ai-infrastructure-partner-ecosystem">Company announcement</a>.</li><li>NVIDIA, “NVIDIA and MediaTek Deepen Long-Standing Partnership to Build AI Edge to Cloud Computing Platforms,” August 31, 2026. <a href="https://nvidianews.nvidia.com/news/nvidia-and-mediatek-deepen-long-standing-partnership-to-build-ai-edge-to-cloud-computing-platforms">Company announcement</a>.</li><li>WinWay, “Coaxial Spring-Pin Test Socket.” <a href="https://www.winwayglobal.com/zh-TW/product-detail/Md84seNWgPSoVVb1?class=package">Product information and original image</a>.</li><li>Joe DeLaere and Ivan Goldwasser, “Integrating Semi-Custom Compute into Rack-Scale Architecture with NVIDIA NVLink Fusion,” NVIDIA Technical Blog. <a href="https://developer.nvidia.com/blog/integrating-custom-compute-into-rack-scale-architecture-with-nvidia-nvlink-fusion">Technical article and original Figure 2</a>.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=837e482b21cc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Is AI Packaging Turning to the Toyota Group Supply Chain?]]></title>
            <link>https://medium.com/@JulienNauy/why-is-ai-packaging-turning-to-the-toyota-group-supply-chain-f2b2c16671ad?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/f2b2c16671ad</guid>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 12:41:56 GMT</pubDate>
            <atom:updated>2026-10-12T12:41:56.548Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why Is AI Packaging Turning to the Toyota Group Supply Chain? Eternal Precision Mechanics’ Cross-Industry Expansion Experiment</h3><p>Manufacturing challenges in AI hardware can bring two seemingly distant industries together. One is racing to expand chip packaging capacity; the other has precision manufacturing resources built for the automotive industry that remain underused. Taiwanese equipment maker Eternal Precision Mechanics is trying to connect these different demand cycles.</p><p>General manager Kazutoshi Iwata says the company’s order book extends into 2029. With its existing capacity, delivery of new orders could stretch into 2030. The company’s response is to seek additional assembly capacity in Japan’s automotive supply chain.</p><h3>Precision Manufacturing Resources in the Toyota Group Supply Chain</h3><p>Jian-he Lu (呂建和), a manager in Eternal Precision Mechanics’ general manager’s office, describes the partner as a wholly owned subsidiary of a Toyota Group member company. Headquartered in Toyota City, Aichi Prefecture, it designs and develops molds for automotive components and manufactures production equipment. According to Lu, the factory previously produced components for fuel cell vehicles and handled precision assembly. Slower-than-expected market adoption left capacity available for the collaboration.</p><p>Iwata says the contract manufacturing partnership began in June 2026. After roughly two months of staff training, assembly started in mid-August, with the first equipment shipments expected in October. He says the actual increase in capacity can be assessed more accurately once the operation has been running for some time.</p><h3>How Does AI Demand Create an Equipment Assembly Bottleneck?</h3><p>Eternal Precision Mechanics supplies vacuum lamination equipment, with technology and manufacturing capabilities that include its Japanese subsidiary, Nikko-Materials. A key market is integrated circuit (IC) substrates: advanced chips use the wiring in these substrates to connect signals and power to external circuit boards.</p><p>Ajinomoto Build-up Film (ABF) is an insulating film developed by Ajinomoto. When substrate manufacturers build multilayer wiring, they must bond this material evenly to the surface. Lamination equipment performs that bonding step. Materials suppliers, equipment makers, and substrate manufacturers each contribute distinct capabilities.</p><p>As AI chip packages grow larger and their wiring becomes more complex, substrate manufacturing must handle larger areas, more build-up layers, and stricter quality requirements. For Eternal Precision Mechanics, demand therefore comes from both customers’ capacity expansion and process upgrades. New factories can be planned and equipment orders placed, but equipment makers still have to assemble, test, and deliver each machine.</p><p>When several substrate manufacturers expand at the same time, their equipment suppliers’ assembly capacity can constrain the expansion schedule. This helps explain Eternal Precision Mechanics’ interest in the automotive supply chain. It needs manufacturing capacity it can draw on, and Japan’s established precision machinery clusters offer an option alongside building its own assembly lines.</p><h3>Transferring Precision Manufacturing Capabilities Requires Mastering New Quality Requirements</h3><p>Assigning people and factory space to new orders addresses only part of the task of reallocating capacity across industries. Eternal Precision Mechanics’ equipment must also enable customers to maintain consistent quality through repeated lamination cycles.</p><p>Vacuum lamination bonds film to surfaces with fine wiring and variations in height, reducing bubbles, wrinkles, and uneven thickness. The company’s product range includes three-stage systems that address advanced substrates’ requirements for flatness and process stability. These requirements carry the importance of machine assembly quality through to customers’ finished products.</p><figure><img alt="Front view of Eternal Precision Mechanics’ CVP-1700SP three-stage vacuum laminator against a white background, with an operator control panel on the right." src="https://cdn-images-1.medium.com/max/527/1*7miyYrTKHht7bKTRo19Cww.png" /></figure><p><em>Eternal Precision Mechanics’ CVP-1700SP three-stage vacuum laminator. Image: </em><a href="https://www.eternal-epm.com/%E7%94%A2%E5%93%81%E4%BB%8B%E7%B4%B9/"><em>Eternal Precision Mechanics</em></a><em>.</em></p><p>Iwata illustrates the cumulative effect with an example. If the pass rate for each lamination cycle stays at 98.5%, about 86% will pass after ten cycles. At 95% per cycle, only about 60% will pass after ten cycles. This illustrative calculation assumes a constant pass rate at each cycle and no rework, showing how repeated processing magnifies the difference.</p><p>Customers also care about reproducible process conditions. Discussing Kinsus’ K6 factory in Youshi, CEO Scott Chen says equipment must reach the required temperature, pressure, and other parameters before processing substrate panels. Kinsus also uses data to establish more consistent defect criteria with packaging companies.</p><p>Iwata notes that switching from a qualified equipment supplier requires customers to repeat qualification and accept the risk of fluctuations in production yield. As Eternal Precision Mechanics expands assembly capacity, it therefore needs to reproduce the equipment quality customers already know and can use consistently.</p><h3>Targets for Expanding Assembly Capacity in Japan</h3><p>The company’s August 2026 investor presentation outlines how Nikko-Materials is responding to demand through short-term factory leases and contract assembly personnel in Japan, along with cooperation with an automotive supply chain company. The plan is to raise maximum monthly assembly capacity in Japan from 18 machines to 28, increasing it quarter by quarter, with a target of mid-2027.</p><h3>Sources</h3><ul><li>Lin Yuanqing, “Eternal Precision Mechanics Taps AI Growth and Positions Itself for Advanced Packaging,” <em>Wealth Magazine</em>, no. 774, October 8, 2026.</li><li>Deng Kaiyuan, “Kinsus Raises PCB Yield Above 90%: What Changed in Its NT$25 Billion AI Production Line Investment?”, <em>CommonWealth Magazine</em>, no. 832. <a href="https://www.cw.com.tw/article/5137276">Article</a>.</li><li>Eternal Precision Mechanics, <a href="https://www.eternal-epm.com/%E7%94%A2%E5%93%81%E4%BB%8B%E7%B4%B9/">Product Introduction</a>, vacuum lamination equipment and model series.</li><li>Eternal Precision Mechanics, <a href="https://www.eternal-epm.com/wp-content/uploads/2026/08/20260826-%E6%B3%95%E4%BA%BA%E8%AA%AA%E6%98%8E%E6%9C%83%E7%B0%A1%E5%A0%B1_CH.pdf">Investor Presentation, August 26, 2026</a>, p. 18.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f2b2c16671ad" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Did Furukawa Choose Global Tek for a Major AI Liquid-Cooling Order?]]></title>
            <link>https://medium.com/@JulienNauy/why-did-furukawa-choose-global-tek-for-a-major-ai-liquid-cooling-order-ba5002fc89b0?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/ba5002fc89b0</guid>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 16:18:53 GMT</pubDate>
            <atom:updated>2026-10-05T16:18:53.319Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why Did Furukawa Choose Global Tek for a Major AI Liquid-Cooling Order? From Airbags to Server Cooling</h3><p>The race to build AI servers is bringing manufacturing capabilities once tucked away inside cars into data centers. As coolant flows through piping past chips, the precision, cleanliness, and sealing quality of metal components become part of what keeps the entire system running reliably.</p><p>Taiwan’s Global Tek Fabrication Co., Ltd. is one example. The manufacturer started with precision metal parts for automotive airbags and has gradually expanded into liquid-cooling quick disconnects, cold plates, and valve bodies. Business Today reports that Global Tek secured a Furukawa Group project for cooling tensor processing unit (TPU) servers and began shipments in mid-2026.</p><p>For international technology industry readers, the story is worth following because it shows how an automotive parts maker can apply existing manufacturing capabilities to AI infrastructure, and what new challenges it faces once it enters the market.</p><h3>There Is Room for Automotive Parts Makers Inside AI Racks</h3><p>To understand Global Tek’s opportunity, it helps to look at the components of a liquid-cooling system.</p><p>A cold plate sits close to the chip that needs cooling, allowing coolant flowing through its internal channels to carry heat away. Quick disconnects link piping and equipment, making the system easier to connect and disconnect while maintaining a seal. A coolant distribution unit (CDU) manages flow, pressure, and heat exchange within the circulation system. These components work together to move heat away from the chips.</p><p>For a precision manufacturer, the opportunity lies in the materials, shapes, and manufacturing requirements of these parts. At a seemingly minor connection, a problem with dimensional fit or sealing can expose the system to leaks. A component’s value therefore needs to be considered alongside the responsibility it carries within the system.</p><p>Global Tek’s entry point is the metal components and coordination across manufacturing processes that liquid-cooling products require. Furukawa Electric’s March 2026 announcement of plans to expand capacity for data-center water-cooling modules also shows that these products are becoming a business in which major industrial groups continue to invest. As system and module suppliers expand their product portfolios, a component supplier’s ability to turn design requirements into deliverable products takes on practical commercial importance.</p><h3>What Airbags Teach About Manufacturing Details</h3><p>Airbags and data-center cooling systems serve very different purposes, but both require careful attention to cleanliness and sealing during manufacturing.</p><p>Global Tek’s early products included precision metal parts for airbag gas generators. These products need to work reliably during a vehicle collision, with demanding requirements for internal cleanliness and protection against moisture from outside. The experience Global Tek accumulated across forming, machining, and welding taught it how to connect multiple production steps to meet a product’s quality requirements.</p><p>The new work in liquid cooling still revolves around materials, structures, machining, and sealing. Automotive experience provides a starting point for engineering judgment, but the operating conditions, customer specifications, and validation requirements of liquid-cooling components still need to be addressed individually. The ability to make a qualified automotive part does not, by itself, establish whether a different product will remain reliable over long-term use.</p><p>Global Tek also had a learning curve in liquid cooling. In 2022, President Yah-Hsing Huang accepted an invitation from a Taiwanese cooling company to help develop quick disconnects and related machining techniques. That collaboration did not reach large-scale production, but it gave the team liquid-cooling development experience. The company later took on cold-plate and heat-sink machining for other customers.</p><p>This progression shows how capabilities that span industries are often built gradually through actual projects. Existing technology gives a company a chance to tackle a problem, while new customer requirements push the team to fill gaps in its experience.</p><h3>Furukawa Needed Multiple Processes Connected Into a Solution</h3><p>According to Business Today, Furukawa approached more than ten suppliers to develop a cooling system integrating cold plates, quick disconnects, and valve bodies, but could not find a suitable partner because the required manufacturing processes were too complex. Global Tek offered a combination of stamping, forging, and machining capabilities, with work divided among different plants.</p><p>Huang summed up the challenge in one sentence: “The real challenge in liquid cooling is not machining, but the solution.”</p><p>Once a customer provides specifications, the manufacturing team must assess material properties, component geometry, and function before choosing suitable forming and subsequent machining processes. Huang gave the example of casting a more complex housing at the Xi’an plant before carrying out further machining. Such choices affect a product’s cost, lead time, and consistency in volume production.</p><p>Global Tek also brings value analysis and value engineering (VA/VE) into the design stage. For the supplier, this extends its role from carrying out machining to helping select the manufacturing approach earlier in development. For the customer, it creates an opportunity to resolve conflicts between design requirements and production conditions sooner.</p><p>At its April 2026 investor presentation, the company listed products including universal quick disconnects (UQDs), multi quick disconnects (MQDs), valve bodies, and cold plates, along with the division of manufacturing processes among its plants. These disclosures outline Global Tek’s liquid-cooling plans and help explain how its capabilities are distributed across multiple components and production steps.</p><figure><img alt="Global Tek’s investor presentation page on liquid-cooling products and manufacturing processes. The upper section lists product and cross-plant plans from 2022 to 2026; the lower section shows quick disconnects, piping, a valve-body cross-section, and cold-plate illustrations." src="https://cdn-images-1.medium.com/max/1024/1*-CES8K0jijwiwOWM7vgcKQ.png" /></figure><p><em>Global Tek’s plans for liquid-cooling components and manufacturing processes across multiple plants, including quick disconnects, valve bodies, and cold plates. The years shown follow the company’s presentation. Source: Global Tek Fabrication Co., Ltd., investor presentation, April 10, 2026, page 24.</em></p><h3>A Sample in Three Days: Manufacturing Has to Match a New Pace</h3><p>Capabilities across multiple processes address part of the question of how to make a product. Delivery speed raises another question.</p><p>Chen Guanfu, an assistant manager at Global Tek who leads the Furukawa project, said automotive projects typically start with complete drawings, go through repeated validation, and then enter volume production. A liquid-cooling customer, however, might submit a request today and expect a sample three days later. When there is not enough time to confirm every detail, engineers have to make initial judgments based on experience.</p><p>This puts pressure on the early stages of development. Teams must quickly determine which requirements are settled, which can be tested with samples, and which still need clarification with the customer. Fast sample delivery and reliable volume production are two successive tasks. Even after a customer receives a sample quickly, the product still needs to complete the applicable validation to support consistent shipments.</p><p>United Alloy-Tech, another Taiwanese company moving from automotive components into liquid cooling, has described a similar difference in pace. Chair Hu Baisha said automotive production follows long-term plans, while server projects demand rapid delivery and flexible adjustments to production lines. Xiong Dong, who leads the company’s liquid-cooling R&amp;D, said the company put at least 18 people on the project to deliver an immersion liquid-cooling rack sample within a month.</p><p>United Alloy-Tech and Global Tek make different products, but their experiences point to the same organizational requirement: entering a new market also calls for changes in development decisions, coordination among plants, and production-line scheduling. Existing machining equipment provides a foundation. Making that equipment work together at the delivery pace customers need depends on the team’s execution.</p><h3>Once You Have a Way In, How Do You Stay?</h3><p>Technical barriers create opportunities for precision manufacturers, but commercial relationships within the supply chain continue to change.</p><p>A liquid-cooling industry participant interviewed by Business Today cautioned that machining suppliers could still face pricing pressure as competition intensifies. An executive at another company that moved from automotive components into cooling pointed out that a supplier could lose business if a customer acquired a precision manufacturer.</p><p>System customers’ ability to manufacture components themselves is another relationship to watch. At a 2024 investor conference, Foxconn Chairman Young Liu said the company could manufacture everything from CDUs, quick disconnects, manifolds, and cold plates to racks. Such capabilities mean an integrator can both purchase components and supply them. A machining company’s competition may come from peers or from its own customers.</p><p>Global Tek’s subsequent performance can therefore be assessed through several questions. Do new projects lead to repeat orders? Can experience accumulated during development improve delivery in volume production? As the liquid-cooling business grows, can the company sustain the value of its manufacturing capabilities, rather than winning orders solely on price?</p><p>The answers will emerge through subsequent customer, shipment, and operating data. Entering the market shows that the company has found a new application for its capabilities. Continuing to participate in customers’ product development, while consistently managing quality, lead times, and costs, gives that application a chance to become a lasting business.</p><p>For Global Tek, the manufacturing experience gained from automotive parts has opened the door to liquid cooling. Its next task is to put those capabilities to work at the new pace of the AI supply chain, so that customers will still want to turn to it for their next project.</p><h3>Sources</h3><ul><li>曹悅華, “企業最前線：時碩靠金屬加工硬實力 拿日系液冷大單” (Global Tek Wins a Major Japanese Liquid-Cooling Order With Its Metalworking Capabilities), Business Today (今周刊), issue 1554, digest edition dated September 30, 2026; Hami subscription text edition. The article’s original publication date has not been verified.</li><li>羅之盈, “封面故事：精確實業靠防漏技術 搶AI液冷商機” (United Alloy-Tech Pursues AI Liquid-Cooling Opportunities With Leak-Prevention Technology), Business Today (今周刊), issue 1550, digest edition dated September 2, 2026; Hami text edition.</li><li>楊孟軒, <a href="https://www.cw.com.tw/article/5132831">“輝達GB200揭開散熱新時代！液冷價格躍升15倍，7家台廠已搶進供應鏈”</a> (Nvidia’s GB200 Opens a New Era in Cooling: Liquid-Cooling Prices Rise Fifteenfold as Seven Taiwanese Companies Enter the Supply Chain), CommonWealth Magazine (天下雜誌), issue 812, November 26, 2024.</li><li>Global Tek Fabrication Co., Ltd., <a href="https://www.globaltekww.com/uploads/legal_person_briefing/20260410.pdf">Investor Presentation, April 10, 2026</a>, pages 6 and 24.</li><li>Furukawa Electric, <a href="https://www.furukawaelectric.com/en/release/2026/kei_20260330.pdf">Notice Regarding Capital Expenditure (Acquisition of Fixed Assets) for Production of Heat-Dissipation/Cooling Products for Data Centers</a>, March 30, 2026.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ba5002fc89b0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Faster Military AI: Can Power and Cooling Keep Up?]]></title>
            <link>https://medium.com/@JulienNauy/faster-military-ai-can-power-and-cooling-keep-up-af71dd93549b?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/af71dd93549b</guid>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 14:05:04 GMT</pubDate>
            <atom:updated>2026-10-04T14:05:04.545Z</atom:updated>
            <content:encoded><![CDATA[<p>Installing a more powerful AI processor in existing military equipment raises a practical problem: the enclosure has not grown, and neither has the space available for power delivery and heat removal. New recognition, tracking, and data processing functions must draw power from the same equipment, which must also dissipate the heat they generate. Processor performance is only one part of the upgrade. Whether the platform can operate continuously also depends on its power and cooling systems.</p><p>Ron Gaw, president of Aegis Power Systems, notes that military platforms require more power as they take on AI target recognition, tracking, and real-time intelligence processing.</p><h3>More Computing Power in the Same Enclosure</h3><p>Vicor’s Matt Renola notes that some equipment retains the same physical dimensions while its power requirements increase. Such upgrades extend the challenge beyond processor selection to the entire platform: how can the power system deliver more power within the existing space, and how will the heat generated during conversion be removed?</p><p>This also changes how power modules should be evaluated. Smaller modules may free up space for other functions, but their benefits may not fully carry through to the complete system unless the heat removal path is addressed as well. Power density and cooling conditions need to be considered together when assessing an upgrade.</p><h3>High Efficiency Still Requires Heat Removal</h3><p>A conversion efficiency of 96.5% may look high, but at sufficiently high power levels, the remaining losses still create a cooling burden. Suppose a converter delivers 1,000 watts to its load at 96.5% efficiency. Its input power would be approximately 1,036 watts, leaving a difference of about 36 watts.</p><p>In this example, the approximately 36 watts represents the converter’s power loss. Heat generated by the processor and other loads must be counted separately. Estimating the cooling requirements of the complete system therefore requires accounting for both conversion losses and heat generated by the loads themselves.</p><p>For equipment housed in sealed enclosures, the lack of airflow makes heat removal more difficult. Higher efficiency can reduce conversion losses, but it cannot replace a design that conducts heat out of the enclosure. An efficiency percentage alone is insufficient to determine cooling requirements. How much power the equipment can deliver continuously in a particular environment still depends on its actual thermal conditions.</p><h3>Computing Moves Onto the Platform, and Power Demand Follows</h3><p>Vicor’s Scott Lee notes that limited communication between platforms is driving more processing capability, particularly AI processing, onto individual platforms. As work previously handled by other nodes moves onto the equipment itself, local power delivery and thermal management must support the additional load.</p><p>A DigiKey technical marketing engineer notes that engineers deploying single-board computers (SBCs) in edge AI systems face constraints involving memory, cooling space, and cost. An SBC integrates the main computing components onto a single board, helping industrial edge equipment become more compact. Space outside the board is still needed for power delivery, cooling, and other components.</p><p>Distributed deployment and modular power systems address different levels of the problem. Distributing troops or equipment requires coordinating missions and energy supplies across platforms. Modular power systems inside an enclosure focus on conversion and integration within a single piece of equipment. Even after the internal module design is complete, energy distribution between platforms still requires separate arrangements.</p><h3>New Components Create Design Opportunities, but System Integration Remains Essential</h3><p>Matt Renola says that the aerospace and defense market is using wide-bandgap power devices, including gallium nitride (GaN) and silicon carbide (SiC), to increase power density and reduce conduction and switching losses. For existing platforms, these improvements can create room for additional computing functions.</p><p>Texas Instruments (TI) notes that GaN’s conduction and switching loss characteristics can support higher switching frequencies while keeping losses under control, allowing some passive components to become smaller. Once component sizes are reduced, the complete system must still address load-generated heat, cooling, and packaging before those device advantages translate into usable space.</p><p>ROHM notes that realizing GaN’s high-speed switching capabilities requires compatible gate drivers and controllers. Adoption also requires consideration of cost, reliability, and application feasibility. Changing semiconductor materials is only one part of power system design; surrounding circuitry and integration conditions also matter.</p><figure><img alt="Diagram of a 650 V GaN power stage integrating a gate driver, EMI control, LDO, and thermal/power protection." src="https://cdn-images-1.medium.com/max/1024/1*ad2HToklyn1dm6Hwo-joxg.jpeg" /></figure><p>A GaN power stage integrates the power transistor, gate driver, and protection functions.</p><p>Source: ROHM, figure 4 in 新通訊, issue 301, “Small Size, High Power, and Low Losses: GaN Devices Advance Into Power Applications.”</p><p>For platform integrators, the key question is how these improvements translate into usable space and sustained operation. The interaction between components, power modules, and the enclosure must be assessed against actual load and cooling conditions. Each supplier provides one part of the solution, while the complete equipment must be evaluated under shared operating conditions.</p><h3>Assessing the Upgrade at Platform Level</h3><p>Ron Gaw believes that future power systems must report operating status in real time, provide backup capability, and rapidly route power to critical loads as needed. These capabilities broaden the evaluation from how much power a system can deliver to how it responds to an interruption and whether power can reach the required location in time.</p><p>EDATEC uses supercapacitors to provide power-loss protection, giving industrial SBCs time to shut down properly and reducing the risk of damage to data or storage devices. A safe shutdown and uninterrupted mission operation require different backup capabilities. When evaluating military platforms, the first step is to establish what must be preserved after a power interruption: data, equipment state, or the mission in progress.</p><p>Military AI upgrades should therefore be assessed in terms of the platform’s conditions, considering the power required by its loads, the paths for removing conversion heat, and the backup arrangements for power interruptions. The next useful evidence to look for is sustained operation within the same physical volume and under clearly specified cooling conditions. This provides a more meaningful basis for judging whether the complete equipment can support its new functions than comparing processor performance or power module specifications alone.</p><p>If a supplier lists only rated power and peak efficiency, readers still need to know whether those figures apply to the same load conditions. Clearly stated operating conditions make component specifications comparable with platform upgrade requirements and allow power delivery and cooling to become concrete criteria for evaluating the upgrade.</p><h3>References</h3><ul><li>新通訊, issue 308, “<a href="https://www.2cm.com.tw/2cm/zh-tw/magazine/-MarketTrend/075D2C4A18B243F6AB28008694701776">Military Power Demand Rises as System Design Moves Toward Greater Efficiency and Decentralization</a>.” The author is identified as Vicor’s Taiwan general manager. The issue date is October 1, 2026; the article’s individual first publication date has not been independently confirmed. This article draws on the archived text; no separate interviews were conducted with the industry representatives cited.</li><li>新電子, “<a href="https://www.mem.com.tw/%e5%8f%af%e9%9d%a0%e5%ba%a6%e8%88%87%e5%bd%88%e6%80%a7%e6%88%90%e9%97%9c%e9%8d%b5%e3%80%80%e5%b7%a5%e6%a5%adsbc%e5%8a%a0%e9%80%9f%e9%82%8a%e7%b7%a3%e9%81%8b%e7%ae%97/">Reliability and Flexibility Become Critical as Industrial SBCs Accelerate Edge Computing</a>.” The archived header records an online publication date of October 1, 2026, and identifies the article with issue 487. The author is identified as a DigiKey technical marketing engineer. Used for examples of integration constraints and power-loss protection in industrial edge equipment.</li><li>新通訊, issue 301, “<a href="https://www.2cm.com.tw/2cm/zh-tw/magazine/-Technology/0B62A944905B4FB2B528202FF5BD19E3">Small Size, High Power, and Low Losses: GaN Devices Advance Into Power Applications</a>.” March 2026 issue; contributed by ROHM. Used for GaN drive, control, and adoption requirements.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=af71dd93549b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Technology Taiwan Retired 20 Years Ago: Why Two-Stroke Engines Are Back in the Drone Supply Chain]]></title>
            <link>https://medium.com/@JulienNauy/a-technology-taiwan-retired-20-years-ago-why-two-stroke-engines-are-back-in-the-drone-supply-chain-cea1d1147a8e?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/cea1d1147a8e</guid>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[taiwan]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[drones]]></category>
            <category><![CDATA[defense-technology]]></category>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Fri, 04 Sep 2026 14:33:23 GMT</pubDate>
            <atom:updated>2026-09-04T14:33:23.801Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*BwjFlQ2G9Zy-7B1W.jpg" /></figure><p><em>President Lai Ching-te receives a briefing on Taiwan’s anti-radiation unmanned attack system and other drones developed by NCSIST on July 17, 2026. Image source: </em><a href="https://www.president.gov.tw/News/40197"><em>Office of the President, Republic of China (Taiwan)</em></a><em>.</em></p><p>Taiwan stopped producing high-emission, high-noise two-stroke motorcycle engines in 2004. More than 20 years later, a technology rejected by the civilian market is returning because of long-range attack drones.</p><p>The <em>Tianxia Zero Hour</em> podcast used the Shahed-136 attack drone as an example. One aircraft may cost only US$20,000 to US$50,000, yet it can force an opponent to use an interceptor missile worth more than US$1 million. In this kind of cost exchange, maximum engine efficiency, emissions, and service life are no longer the only design goals. Price, weight, range, and ease of production become more important.</p><p>This is not a step backward in technology. The mission has changed the definition of good technology.</p><h3>Long Range Makes Fuel Competitive Again</h3><p>Agricultural, photography, and short-range reconnaissance drones usually rely on batteries and electric motors. The calculation changes when an aircraft must carry a payload for hundreds of kilometers. Battery weight and endurance become serious constraints.</p><p>A fuel engine can add more than 100 kilometers of range with one liter of gasoline. The aircraft also becomes lighter as it burns fuel. This weight profile is very different from that of a battery, which remains on board for the entire flight.</p><p>Two-stroke engines fit this requirement well. They are mechanically simple, have a high power-to-weight ratio, and are relatively easy to maintain. They are also cheaper than rotary or four-stroke engines.</p><p>According to an assessment by Taiwan’s National Chung-Shan Institute of Science and Technology (NCSIST, 國家中山科學研究院) cited by the podcast, using a two-stroke engine in one attack-drone design could reduce the propulsion-system cost to about one-third of the cost of a four-stroke rotary engine.</p><p>Another example comes from a team at the National Yunlin University of Science and Technology (YunTech, 國立雲林科技大學). The team is developing a hybrid propulsion system for drones. Its target is eight hours of flight and a range of 400 to 500 kilometers. An engine generates electricity while electric motors provide propulsion. This design seeks a balance among range, payload, and controllability.</p><h3>The TPI Bearings Case: Even a Standard Bearing Needs New Operating Parameters</h3><p>Tung Pei Industrial Co., Ltd. (TPI Bearings, 東培工業) is Taiwan’s largest bearing manufacturer. Its entry into the drone supply chain did not begin with a mysterious new component. The company adapted ball bearings already used in engines to meet a different operating environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/0*V61lNOZwAflFRFF2.png" /></figure><p><em>A precision angular-contact ball bearing manufactured by TPI Bearings. Image source: </em><a href="https://www.tpibearings.co/zh-hant/catalog/standard/"><em>TPI Bearings, Standard Series</em></a><em>.</em></p><p>Motorcycle engines commonly use oil-bath lubrication. Drone engines require lubricants that can withstand higher temperatures. High heat and rotational speed also change the thermal expansion of bearings and nearby parts. Internal clearances must therefore be redesigned for the actual operating conditions.</p><p>The product may still be classified as a standard bearing. From an engineering perspective, however, it cannot simply be removed from a motorcycle and installed in an aircraft.</p><p>These products currently account for less than one-thousandth of TPI Bearings’ annual output of more than 100 million ball bearings. Orders are also unstable, and the company cannot confirm the products’ final applications.</p><p>The significance of this case is not that drone bearings have already become a new profit engine. It is that the market has given an established supply chain an entry point into a new industry.</p><h3>The Motorcycle Supply Chain’s Advantage Is Adaptation Speed</h3><p>The production value of Taiwan’s motorcycle-component industry has fallen by about 20 percent over the past four years. International manufacturers have shifted production to Southeast Asia, while low-cost Chinese products have added further pressure. Taiwanese suppliers need a new outlet for their manufacturing capacity.</p><p>Small-displacement engines, bearings, wiring harnesses, airframe structures, and precision machining are all required in drones. These needs overlap with capabilities that Taiwan’s motorcycle industry has developed over decades.</p><p>The individual parts may not have exceptionally high technical barriers. Taiwan’s advantage comes from independent engineering capabilities, a dense supplier network, and the ability to adjust small production runs quickly.</p><p>Some Japanese motorcycle manufacturers remain cautious because of their joint-venture relationships and the sensitivity of defense-related business. Taiwanese suppliers may therefore have more room to accept new orders. This is an advantage in supply-chain conversion, not a monopoly based on one component.</p><h3>International Orders Are Opening a Window</h3><p>This market is no longer limited to policy statements. According to <em>Wealth Magazine</em>, Taiwan exported about US$212 million worth of drones during the first half of 2026. That amount exceeded the combined total of the previous three years.</p><p>In June 2026 alone, Taiwan’s drone exports to Saudi Arabia reached US$47.18 million. These figures suggest that the geopolitical push for a non-China supply chain has started to produce actual purchasing orders.</p><p>NCSIST and other government bodies have also visited two-stroke motorcycle and agricultural-engine manufacturers with underused production capacity. They introduced these companies to overseas drone manufacturers, and some orders followed.</p><p>Procurement matchmaking by the American Institute in Taiwan (AIT, 美國在台協會) has also connected Taiwanese suppliers with US companies such as Shield AI. According to the report, the Taiwan share of Shield AI’s sourcing rose from less than 1 percent two years ago to 16 percent. The company plans to increase the share to 25 percent within two years.</p><p>For motorcycle-component manufacturers, this is a rare open door. An open door, however, is not a long-term competitive moat.</p><h3>Building a Prototype Is Easier Than Building a Production System</h3><p>Taiwan has engineers who can build an engine prototype. The more difficult task is assembling a team that can integrate mechanical design, thermal management, combustion, control systems, and aircraft engineering into a product that can be manufactured consistently.</p><p>YunTech’s engine project needed an advisory group of 15 professors. A single laboratory could not cover every area of knowledge required by the system.</p><p>Prototype production has also exposed problems involving capacity and cost. Another hybrid-engine team documented by <em>Business Weekly</em> needed a cylinder head produced through five-axis machining. One supplier quoted NT$1 million for the work. The team eventually found a solution that cost NT$180,000.</p><p>A successful prototype does not prove that delivery times, yield, traceability, and batch-to-batch consistency can satisfy a customer’s requirements. Taiwan’s opportunity is not based on proving that it can also build a two-stroke engine. The real task is shortening the path from prototype and testing to correction and volume production.</p><h3>A Non-China Supply Chain Is Only the Entry Ticket</h3><p>Taiwan’s immediate opportunity comes from efforts by the United States and its allies to remove Chinese components from drone supply chains. <em>CommonWealth Magazine</em> reported that more than 100 overseas companies had visited Taiwan to look for drone partners. Anduril Industries had also held discussions with seven or eight Taiwanese suppliers.</p><p>Taiwanese products that satisfy non-China manufacturing requirements may command a premium of about 20 percent. Buyers will still compare their prices with Chinese products. Geopolitical preference can open a market, but it cannot permanently remove price competition.</p><p>Hardware production is one of Taiwan’s strengths. Coretronic Intelligent Robotics Corporation (CIRC, 中光電智能機器人) and Thunder Tiger Corporation (雷虎科技) have both been expanding drone or motor capacity.</p><p>Important gaps remain. Taiwan lacks locally designed flight-control, communications, and satellite-navigation chips. It also lacks competitive flight-control and ground-control software.</p><p>In a combat environment, software and communications frequencies may need to be updated within 48 hours. A supplier that only provides mechanical parts and assembly will struggle to capture the higher-value portion of the product.</p><p>Two-stroke engines can help Taiwan secure an entry ticket. They cannot solve the gaps in flight control, communications, electronic warfare, and system integration.</p><h3>Why KYMCO and SYM Are Right to Be Cautious</h3><p>KYMCO (光陽工業股份有限公司) and Sanyang Motor Co., Ltd. (SYM, 三陽工業股份有限公司) both have the capability to develop complete engines. Their current approach is cautious. KYMCO’s position is straightforward: being technically capable of developing a product does not mean the investment is commercially justified.</p><p>That caution is understandable. Drone-related bearings account for less than one-thousandth of TPI Bearings’ production. A complete engine requires far more investment in development, testing, supply-chain management, and after-sales support than a standard bearing does.</p><p>Without predictable long-term purchasing volumes, a large manufacturer will carefully calculate the opportunity cost of converting production capacity. Greater technical capability can make a company more selective, not less.</p><p>Large motorcycle manufacturers and smaller component suppliers may therefore move at different speeds. Smaller companies can use existing production lines to serve low-volume orders. Larger manufacturers may wait until customers, specifications, and purchasing volumes become clearer.</p><p>This difference does not show that one group has greater foresight. The two groups simply face different investment thresholds.</p><p>The return of the two-stroke engine is not technological nostalgia. War has changed the definition of good technology. An engine rejected by the civilian market for its pollution has regained value in expendable drones because it is inexpensive, light, and suitable for long-range flight.</p><p>The real opportunity for Taiwan’s motorcycle supply chain does not lie in a few orders for standard components. It lies in converting independent manufacturing, rapid adaptation, and system-integration capabilities into a production base for the non-China drone supply chain.</p><h3>References</h3><ol><li><em>Tianxia Zero Hour</em> (天下零時差), “把台灣機車引擎送上天！消失20年的二行程技術，竟幫無人機成本砍7成” (“Sending Taiwan’s Motorcycle Engines into the Sky: How a Two-Stroke Technology Absent for 20 Years Cut Drone Costs by 70%”), September 4, 2026. This article used the full program transcript. <a href="https://open.spotify.com/episode/2JIntvjupFeRdaSzSCwff8">Public episode page</a>.</li><li>Lin Wei-chen (林洧楨), “雞舍倉庫誕生無人機引擎！直擊校園到山林裡的軍工全民運動” (“A Drone Engine Born in a Chicken-Shed Warehouse: Taiwan’s Grassroots Defense Movement from Campuses to Mountain Workshops”), <em>Business Weekly</em> (商業周刊), Issue 1976 Highlights Edition, September 25, 2025.</li><li>Chen Ya-chieh (陳雅潔), “打造下個護國神山 台灣無人機飛進全球國防領先群” (“Building the Next National Champion: Taiwan’s Drone Industry Enters the Global Defense Vanguard”), <em>Wealth Magazine</em> (財訊), Issue 769, July 30, 2026.</li><li>Wu Ching-fang (吳靜芳), “台灣無人機盛世來臨！6家台廠搶國際訂單，取代中國製造只差一個弱點？” (“Taiwan’s Drone Boom Arrives: Six Companies Chase Global Orders, but One Weakness Stands Between Them and Replacing China”), <em>CommonWealth Magazine</em> (天下雜誌), Issue 832, September 16, 2025.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cea1d1147a8e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[70 Racks Opened a Second Growth Curve: How Taiwan’s UAT Moved EV Battery Sealing Into AI Liquid…]]></title>
            <link>https://medium.com/@JulienNauy/70-racks-opened-a-second-growth-curve-how-taiwans-uat-moved-ev-battery-sealing-into-ai-liquid-21f33a952dab?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/21f33a952dab</guid>
            <category><![CDATA[data-center]]></category>
            <category><![CDATA[taiwan]]></category>
            <category><![CDATA[liquid-cooling]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[manufacturing]]></category>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Fri, 04 Sep 2026 13:10:38 GMT</pubDate>
            <atom:updated>2026-09-04T13:10:38.998Z</atom:updated>
            <content:encoded><![CDATA[<h3>70 Racks Opened a Second Growth Curve: How Taiwan’s UAT Moved EV Battery Sealing Into AI Liquid Cooling</h3><p>United Alloy-Tech did not invent a new cooling technology from scratch. It found a new market for the sealing, welding, surface-treatment, and large metal enclosure capabilities that it had already developed for electric vehicle battery housings.</p><p>That distinction matters. UAT’s move into AI liquid cooling may look like another auto-parts supplier chasing a popular theme. In practice, the company is transferring a mature manufacturing skill set into a market where failure is far more expensive and reliable production can command a higher price.</p><figure><img alt="Official rendering of a United Alloy-Tech manufacturing facility with the UAT logo, loading bays, trucks, cars, and surrounding greenery." src="https://cdn-images-1.medium.com/max/1024/1*MIcX4wquiVB3m6Ijhb7mwg.png" /></figure><p><em>An official rendering of a United Alloy-Tech manufacturing facility. Image: </em><a href="https://www.al-alloy.com.tw/about.php"><em>United Alloy-Tech</em></a><em>.</em></p><h3>UAT Is Not Selling “Cooling.” It Is Selling “No Leaks.”</h3><p>United Alloy-Tech Company (UAT, 精確實業) is applying its experience with airtight EV battery enclosures to liquid-cooling racks and manifolds. The products serve different industries, but they share three basic requirements. Liquids must not escape. Dimensions must remain within tolerance. Production must remain stable as volume increases.</p><p>This is the most important part of UAT’s transition. The company is not starting from zero. It is finding a new way to price manufacturing capabilities that it already owns. Automotive components provide an established revenue base. AI liquid cooling offers the possibility of faster growth and higher margins.</p><h3>Why AI Servers Are Bringing Liquid Cooling to the Foreground</h3><p>The power consumption and heat density of AI servers are rising quickly. Traditional air cooling is moving closer to its practical limits. A report on the Nvidia GB200 supply chain said that its heavy thermal load was pushing server racks toward broad adoption of liquid cooling. The report estimated that the related cooling components could cost 15 to 20 times as much as those in a traditional air-cooled rack.</p><p>This change is creating more than a market for one new component. It requires a complete fluid system.</p><p>Cold plates carry heat away from processors. Manifolds distribute coolant to individual computing nodes. Coolant distribution units, or CDUs, manage temperature, pressure, and flow. Quick disconnects allow technicians to service equipment without letting even a drop of liquid enter the server.</p><p>In a rack filled with expensive computing equipment, preventing leaks is not a minor manufacturing detail. It is a matter of reliability and risk control.</p><h3>The First Order Came From the Similarities Between Battery Housings and Server Racks</h3><p>In early 2025, UAT approached a cloud service provider. The customer did not immediately show much interest. Near the end of the meeting, however, it asked whether UAT would be willing to take on an immersion-cooling rack.</p><p>After examining the product, UAT’s team recognized several similarities to an EV battery enclosure. Both are large metal structures. Both require tight control over sealing, cleanliness, surface treatment, weight, and machining precision. UAT promised to deliver a sample within one month.</p><p>The company assigned 18 people to the project. The sample passed qualification on its first attempt. UAT then won an order for 70 racks and shipped them in the fourth quarter of 2025.</p><p>The significance of that first order went beyond its value. It showed that a new customer was willing to accept manufacturing processes originally developed for the automotive industry. UAT later said that another three or four customers had approached the company and that it had taken on more than 20 projects in total.</p><p>The speed is impressive, but speed alone does not explain the result. Delivering a sample in one month would have been a gamble without UAT’s existing experience in welding, flatness control, and leak prevention. Speed was the outcome. Transferable process know-how was the underlying reason.</p><h3>Blind-Mate Manifolds Are a More Difficult Mass-Production Test</h3><p>The manufacturing challenge becomes more demanding with blind-mate manifolds. Space inside a rack is tight, and engineers cannot directly see every connection point. The pipes must therefore connect accurately under limited visibility. Even a small welding deviation can create a coolant leak.</p><p>When Cooler Master Co., Ltd. (訊凱國際) sourced manifolds from UAT, it paid particular attention to the company’s control of precision-welding temperatures. It also valued UAT’s ability to integrate large equipment and maintain accuracy after moving a production line.</p><p>These requirements show why the opportunity is not simply about producing metal boxes. A qualified supplier must keep dimensions, welding quality, and sealing performance consistent across repeated production. That is where years of automotive manufacturing discipline can become commercially useful in a new market.</p><h3>This Is UAT’s Third Transfer of Its Core Capabilities</h3><p>UAT began as a bicycle-parts manufacturer. It later suffered consecutive losses as competition intensified. In 2017, Minth Group Limited (敏實集團) acquired a 44 percent stake. UAT then started moving into the automotive supply chain.</p><p>The company first produced battery-enclosure separators for Volkswagen Group (福斯集團). It later supplied battery enclosures and structural components to Chinese electric vehicle manufacturers. According to Business Weekly Taiwan, UAT’s revenue grew by about 12 times between 2017 and 2025, and the company returned to profitability.</p><p>This history changes how its liquid-cooling move should be understood. UAT is not becoming a technology company overnight. It has repeatedly moved its metal die-casting, machining, welding, and enclosure expertise into applications with greater value.</p><p>The move from bicycles to automobiles was one transfer. The move from conventional auto parts to EV battery structures was another. AI liquid cooling is the third.</p><h3>Taiwan’s Strength Is Manufacturing, but System Control Remains the Weak Point</h3><p>Taiwanese suppliers do not lead every part of the liquid-cooling supply chain. Cold plates and manifolds are closer to Taiwan’s established metalworking strengths. Companies such as Asia Vital Components Co., Ltd. (AVC, 奇鋐科技) and Hon Hai Technology Group (Foxconn, 鴻海科技集團) have advantages in these areas.</p><p>CDUs and quick disconnects present a different challenge. They require expertise in fluid control, long-term reliability, and on-site maintenance. European and American suppliers such as Vertiv Holdings Co. and Boyd Corporation held stronger positions in these areas during the market’s early development.</p><p>This division of labor creates both an opportunity and a limit for UAT. Its sealing and precision-manufacturing experience gives it a credible entry point. However, the companies that control more of the cooling system may have greater influence over specifications, customer relationships, and value distribution.</p><p>UAT’s first 70 racks proved that an automotive supplier could transfer its manufacturing capabilities into AI infrastructure. The next question is whether that entry point can become a durable position in the liquid-cooling supply chain. The answer will depend less on the excitement surrounding AI and more on UAT’s ability to deliver leak-free products with consistent quality at scale.</p><h3>References</h3><ol><li>羅之盈, “精確實業靠防漏技術搶AI液冷商機” [How UAT Uses Leak-Prevention Know-How to Enter AI Liquid Cooling], <em>Business Today</em> (今周刊), №1550, September 2, 2026.</li><li>高士閔, “從瀕倒自行車廠到AI液冷新星：精確8年營收翻12倍浴火重生記” [From a Failing Bicycle-Parts Maker to an AI Liquid-Cooling Contender: How UAT Increased Revenue Twelvefold in Eight Years], <em>Business Weekly Taiwan</em> (商業周刊), №2017, July 9, 2026.</li><li>楊孟軒, “輝達GB200揭開散熱新時代！液冷價格躍升15倍，7家台廠已搶進供應鏈” [Nvidia GB200 Opens a New Cooling Era: Liquid-Cooling Prices Rise Fifteenfold as Seven Taiwanese Suppliers Enter the Supply Chain], <em>CommonWealth Magazine</em> (天下雜誌), №812, November 26, 2024.</li><li>Ministry of Economic Affairs, Taiwan, “United Alloy-Tech Company,” official public company registry. Full public profile accessed September 4, 2026. <a href="https://findbiz.nat.gov.tw/fts/company/86406401?fhl=en">https://findbiz.nat.gov.tw/fts/company/86406401?fhl=en</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=21f33a952dab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Substrates Are Not All in Short Supply: Taiwan’s Real High-End ABF Race]]></title>
            <link>https://medium.com/@JulienNauy/ai-substrates-are-not-all-in-short-supply-taiwans-real-high-end-abf-race-04cdff72e30a?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/04cdff72e30a</guid>
            <category><![CDATA[taiwan]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[semiconductors]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 15:47:15 GMT</pubDate>
            <atom:updated>2026-09-02T16:23:56.998Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI Substrates Are Not All in Short Supply: How 4 Taiwanese Makers and 2 Equipment Suppliers Are Competing for the High-End ABF Boom</h3><blockquote><em>AI has not revived every part of the substrate market. The real shortage is in high-end ABF capacity for large, multilayer AI chips, and in the equipment and process control needed to produce those substrates at a stable yield.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/0*fbNjIOk5FCFJktm0.jpg" /><figcaption>Ajinomoto Build-up Film, the interlayer insulating material used to build fine circuitry in high-performance chip substrates. Image provided by Ajinomoto Group. Source: Ajinomoto Group.</figcaption></figure><p>The bottom line is simple: the current ABF substrate cycle is not a broad market recovery. It is a repricing of manufacturing difficulty.</p><p>Lower-end ABF substrates used in personal computers and smartphones may still face excess capacity. At the same time, GPUs, ASICs, and high-speed switching chips for AI servers require larger substrates, more layers, and lower signal loss. Industry estimates cited by <em>財訊 (Wealth Magazine)</em> suggest that an AI chip may need 2.6 to 3 times as much substrate area as a conventional chip. The layer count may also rise from 12 to more than 18.</p><p>This is not merely an increase in the number of substrates. Each unit consumes more material, takes more equipment time, and leaves less room for yield loss.</p><p>The most useful question is therefore not whether ABF is in short supply. The question is which companies can manufacture high-end products, improve yields, and deliver on time while both materials and equipment remain constrained.</p><h3>ABF Has Become Two Different Markets</h3><p>ABF stands for Ajinomoto Build-up Film. It is an interlayer insulating material used in package substrates for high-performance chips. It allows copper circuits to be built up layer by layer.</p><p>As a chip becomes larger and gains more input and output channels, the substrate must also grow. It needs more layers and finer circuitry. It must also control warpage when exposed to heat.</p><p>This has divided the ABF market into two distinct segments. Consumer electronics still drive demand for lower-end products. AI infrastructure drives demand for high-end substrates. The two segments share the same general name, but they serve different customers and require different materials, production processes, and yield levels.</p><p>Nominal capacity is also a poor measure of high-end supply. The number of production lines in a factory does not tell us how many qualified substrates the factory can deliver. Material allocations, equipment lead times, customer certification, and production yield determine effective capacity.</p><p>This explains why substrate makers can announce new plants while the market still expects high-end supply to remain tight through 2027 or 2028.</p><h3>Money Can Build a Factory, but It Cannot Instantly Buy Materials or Yield</h3><p>The first constraints are appearing in materials. High-end substrates require T-glass fabric with a low coefficient of thermal expansion, smoother copper foil, and higher-specification drill bits. ABF supplied by Ajinomoto Co., Inc. is another critical input. Wealth Magazine reports that Ajinomoto controls more than 90% of this market.</p><p>When supply becomes tight, substrate makers face more than a pricing problem. They also need to secure enough allocation to operate their new capacity.</p><p>Equipment creates a similar problem. Exposure, electroplating, lamination, drilling, and inspection systems take time to manufacture, install, and qualify. <em>天下雜誌 (CommonWealth Magazine)</em> reported that delivery times for some equipment had increased from 12 to 14 months. End customers have even become involved in coordinating material and equipment allocations so that every part of the supply chain expands on the same schedule.</p><p>Forecasts do not agree on the exact size of the future shortage. Some estimates cited by <em>Wealth Magazine</em> are more conservative, while other reports project a larger gap. The precise percentage should not be treated as certain. However, the reports point in the same direction: new demand may grow faster than qualified high-end capacity.</p><h3>Unimicron Is Near the Front, but Long-Term Contracts and Yield Still Matter</h3><p>Unimicron Technology Corp. (欣興電子) is one of the most direct beneficiaries of high-end ABF demand among Taiwan’s four major substrate makers. It has established an important position in GPU and ASIC substrates. Its existing high-end lines are reportedly close to full utilization.</p><p>In the near term, Unimicron plans to increase output by removing bottlenecks. New capacity will follow later.</p><p>Its advantage is not only scale. The company already participates in joint development with customers. When Intel Corporation changes the specifications of EMIB-T or another advanced packaging platform, the substrate supplier cannot wait until the chip design is complete. It must develop the substrate alongside the packaging solution.</p><p>That position also creates two risks. The first is yield. Early-stage yield for next-generation substrates still has substantial room for improvement. Nominal capacity will not immediately become shippable capacity.</p><p>The second risk comes from long-term agreements. These contracts improve order visibility, but they may limit pricing flexibility when supply becomes tight.</p><h3>Nan Ya PCB Is Betting on Size, While Kinsus Is Betting on Production Control</h3><p>Nan Ya Printed Circuit Board Corporation (南亞電路板) is concentrating its next investment phase on large, multilayer substrates. The company has announced capital spending worth tens of billions of New Taiwan dollars. Its targets include panel-level advanced packaging applications such as CoPoS, which is being developed by Taiwan Semiconductor Manufacturing Company, or TSMC.</p><p>The potential return comes from higher prices and higher technical barriers. The risk is equally clear. The investment is large, and mass production will begin later. Until the new factories contribute, the company must continue to remove bottlenecks from its existing lines.</p><p>A higher selling price becomes profit only when yield and utilization improve at the same time.</p><p>Kinsus Interconnect Technology Corp. (景碩科技) is following a different path. It is moving up from a secondary supplier position and has secured CPU and GPU substrate orders from US server customers. <em>Wealth Magazine</em> reports that Kinsus plans to increase monthly ABF capacity by about 25% by early 2027, with further expansion under consideration.</p><p>Kinsus learned a costly lesson in the past: winning an order does not guarantee successful mass production. It has therefore invested in sensors, production-line digitalization, and AI vision inspection at its K6 factory in Taoyuan.</p><p><em>CommonWealth Magazine</em> reported that the new factory improved yield from about 80% to more than 90%. This is important for more than the reduction in defects. It shows that high-end substrate competition has moved beyond buying machines. The central challenge is controlling the entire manufacturing system.</p><h3>Zhen Ding Is Betting on the Entire AI Interconnect System</h3><p>Zhen Ding Technology Holding Limited (臻鼎科技控股) produces flexible printed circuit boards, rigid boards, high-density interconnect boards, substrate-like PCBs, and IC substrates. Its exposure to AI demand is therefore different from that of the other three companies.</p><p>An AI server upgrade does not occur only underneath the chip. Motherboards require more layers. Optical modules are moving from 800G to 1.6T. Modified semi-additive processing, or mSAP, is also entering high-speed transmission boards.</p><p>Zhen Ding can seek growth in IC substrates, high-end rigid boards, and optical communication boards at the same time.</p><p>This broader portfolio can reduce dependence on one product cycle, but it makes execution more difficult. The company must divide large capital investments across several factories and product lines. Each new line must pass customer certification and improve yield.</p><p>The most useful measure is not how much Zhen Ding announces that it will invest. It is how much new capacity completes customer qualification and begins stable production.</p><h3>Equipment Suppliers May Get Paid Before the New Capacity Produces Revenue</h3><p>Substrate makers must build new factories, qualify products, and begin shipments before they receive the full benefit of expansion. Equipment suppliers may recognize revenue earlier because their machines must be ordered during the construction stage.</p><p>Eternal Precision Mechanics Co., Ltd., or EPM (長廣精機), is one of the clearest examples. Wealth Magazine reports that EPM has worked with Ajinomoto for more than 20 years and holds about 95% of the global market for vacuum lamination equipment used in ABF substrate production.</p><p>When Unimicron, Nan Ya PCB, Kinsus, and Zhen Ding expand high-end capacity at the same time, they all need lamination equipment.</p><p>EPM’s advantage is not limited to the mechanical hardware. Vacuum lamination must control film handling, pressure accuracy, air bubbles, and surface flatness. When material specifications change, machine parameters must change as well. The ability to develop processes with material suppliers and customers is more valuable than simply increasing machine output.</p><p>Ta Liang Technology Co., Ltd. (大量科技) addresses another bottleneck. Multilayer PCBs require more precise back drilling and tighter control of residual copper. Ta Liang integrates inner-layer measurement, CCD imaging, proprietary controllers, and back-drilling equipment.</p><p><em>Wealth Magazine</em> reports that Ta Liang is also a qualified supplier of residual-copper inspection equipment for PCBs associated with NVIDIA Corporation. It is extending the same capabilities into optical modules, co-packaged optics, and advanced packaging inspection.</p><p>Equipment suppliers do not need to predict which substrate maker will win the most business. They can benefit when the entire industry increases capital spending. However, this model still carries risk.</p><p>Equipment revenue can fluctuate with customer acceptance schedules. Customized systems also consume engineering resources. If every order requires extensive redesign, revenue growth may not produce the same rate of profit growth.</p><h3>Glass Core Substrates Will Not Immediately Replace ABF</h3><p>Glass core substrates are often described as a replacement for organic substrates, but this is unlikely to be a simple choice between two materials. Glass offers better flatness and dimensional stability. Its outer build-up layers may still use ABF or similar materials.</p><p>In the near term, glass, ceramic, and organic materials are more likely to coexist for different packaging applications.</p><p>For Taiwan’s supply chain, this means that equipment must handle more combinations of materials. Glass can crack during drilling, cutting, and lamination. Organic materials face warpage and thermal stress.</p><p>Equipment suppliers that can control coating, lamination, drilling, handling, and inspection may move closer to the customer’s core process. That position could be more defensible than supplying a single standard machine.</p><p>AI chips concentrate more computing power inside one package. They also concentrate supply-chain risk in the ability of substrate makers, material suppliers, and equipment companies to move in step.</p><p>The winners in this cycle may not be the companies that announce the largest expansion plans. They are more likely to be the companies that can align materials, equipment, yield, and customer qualification at the same time.</p><p>The true scarcity in high-end ABF is not factory space. It is a supply network that can deliver stable mass production.</p><h3>References</h3><ol><li>劉志明, “載板四雄、設備廠進入黃金成長期” (“Taiwan’s Four Leading Substrate Makers and Equipment Suppliers Enter a Golden Growth Period”), <em>財訊 (Wealth Magazine)</em>, Issue 771, August 27, 2026.</li><li>林苑卿, “完全解讀 ABF載板缺貨真相” (“Understanding the Real Cause of the ABF Substrate Shortage”), <em>財訊 (Wealth Magazine)</em>, Issue 771, August 27, 2026.</li><li>周佳蓉, “AI載板續享漲價紅利” (“AI Substrates Continue to Benefit from Price Increases”), <em>先探投資週刊 (Wealth Invest Weekly)</em>, Issue 2418, August 19, 2026.</li><li>鄧凱元, “景碩PCB良率提升九成以上 250億打造AI產線，改造了什麼？” (“Kinsus Raises PCB Yield Above 90% After Investing NT$25 Billion in an AI Production Line”), <em>天下雜誌 (CommonWealth Magazine)</em>, September 8, 2025.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04cdff72e30a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Taiwan Was Reportedly the First Country Hit by an Autonomous AI-Agent Cyberattack]]></title>
            <link>https://medium.com/@JulienNauy/taiwan-was-reportedly-the-first-country-hit-by-an-autonomous-ai-agent-cyberattack-1b57f58fdd7e?source=rss-61170b42367d------2</link>
            <guid isPermaLink="false">https://medium.com/p/1b57f58fdd7e</guid>
            <dc:creator><![CDATA[Julien Nauy]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 13:12:29 GMT</pubDate>
            <atom:updated>2026-09-01T14:58:05.157Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>The Bottom Line:</strong> A multi-agent attack system reportedly breached Taiwanese government networks over four days, compromised 85 accounts, and obtained more than 2,500 personnel records. If the identification of Taiwan is correct, this may be the first documented case of AI agents carrying out an end-to-end cyberattack against a national government.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fEYUlGHMzp82BoqVRLlvtw.jpeg" /><figcaption>Taiwan’s Ministry of Digital Affairs. Photo: 蕭文康. Aug. 13, 2026. Source: <a href="https://www.knews.com.tw/news/991E037CB70A9A852C3BECC3617CCF8E">Knews</a></figcaption></figure><p>The real concern is not that AI suddenly developed malicious intent. The change is much more practical. Cyberattacks are moving from a stage in which hackers use AI as a tool to one in which multiple AI agents can divide the work, exchange information, and execute much of the attack chain themselves.</p><p>Attackers can now operate at machine speed. Defenders may still need to wait for a meeting, management approval, or a reply from another department. That difference in speed is why this incident could mark an important turning point in global cybersecurity.</p><h3>Eight AI Agents Acting as an Automated Attack Team</h3><p>The operation reportedly took place from July 1 to July 4, 2026. The attackers used two open-source AI agent frameworks, Hermes and OpenClaw, to launch 12 waves of attacks. Up to eight agents were active at the same time.</p><p>These agents did not simply repeat a fixed set of instructions. They divided reconnaissance, password attacks, vulnerability research, and intrusion tasks among themselves. When one attack path failed, another agent searched online for information and developed a different approach.</p><p>The system eventually mapped 21 government systems, compromised 85 accounts, and obtained at least 2,564 personnel records. Its reconnaissance also extended to government suppliers, a nuclear safety agency, government email systems, and at least seven energy companies.</p><p>According to Dream Research Labs, the operators’ internal records were written in Simplified Chinese, while data in the targeted systems appeared in Traditional Chinese. Dream described the operators as Chinese-speaking, but it did not formally attribute the attack to a specific Chinese state or hacking group. The <em>Financial Times</em>, citing an unnamed person familiar with the matter, identified the targeted country as Taiwan.</p><p>Taiwan’s Ministry of Digital Affairs later confirmed that overseas hackers had used OpenClaw and other AI agents in attacks on government agencies. However, the ministry said the activity it observed began on July 20. That date does not match the early-July operation described by Dream, so the two disclosures should not automatically be treated as the same incident.</p><p>Earlier uses of AI in cyberattacks mainly involved helping a human write malware, find vulnerabilities, or prepare phishing messages. The more important change here is the level of delegation. A human operator can give the system a high-level objective, and multiple agents may then break it into smaller assignments. One agent conducts reconnaissance, another looks for vulnerabilities, another tests accounts, and another writes code. Their results can then be handed to other agents for the next stage.</p><h3>Why Taiwan Is an Obvious Early Target</h3><p>Taiwan combines three features that make it an especially valuable cyber target. It sits at the center of the global semiconductor and electronics supply chain. Its energy sector and critical infrastructure have strategic importance. It also operates in a highly sensitive geopolitical environment.</p><p>Government agencies, research institutions, technology companies, and suppliers in Taiwan hold information with both commercial and national-security value. A compromise at a Taiwanese supplier may also affect customers in the United States, Europe, and Japan through the supply chain.</p><p>This is why an attack on Taiwan should not be viewed as a local cybersecurity story. It is also a warning for every international company whose products, research, or operations depend on Taiwanese partners.</p><h3>AI Changes the Economics of an Attack</h3><p>AI does not create security vulnerabilities out of thin air. Attackers still need flawed software, leaked credentials, excessive access rights, or poorly configured systems.</p><p>Taiwanese cybersecurity company DEVCORE has made this point clearly. AI does not invent a completely new intrusion capability. It helps researchers and attackers find problems faster, analyze weaknesses, and operate at a larger scale.</p><p>In the past, a hacker might have scanned targets one by one, studied each system, tested several routes, and organized the findings manually. One person can now direct several agents working in parallel. An operation that once required a team of ten people may eventually require only one operator and a group of AI agents that do not need to rest.</p><p>The problem is not limited to attackers. Companies are also giving their own AI agents access to email, source code, customer information, cloud services, financial processes, and internal tools.</p><p>CyCraft Technology has described such agents as a new kind of “untrusted insider.” An agent may hold valid credentials, yet still take unexpected actions because of prompt injection, faulty reasoning, or malicious instructions.</p><p>Traditional cybersecurity asks whether a person should be allowed through the door. The age of AI agents requires an additional question: Does every action taken by this agent still serve the purpose for which access was originally granted?</p><p>This shift will require more than adding another layer of alerts. Organizations will need tighter limits on what agents can access, continuous checks on what they are doing, and a fast way to stop abnormal behavior. Human judgment will remain necessary, but a defense that moves only at human speed may no longer be enough.</p><h3>References</h3><h4>Sources in the Local Magazine Database</h4><ol><li>Madhumita Murgia, “台灣成首個被AI代理人攻擊國家 全球資安進入關鍵轉折點” [Taiwan Becomes the First Country Attacked by AI Agents as Global Cybersecurity Reaches a Critical Turning Point], <em>Business Weekly</em>, 2024 Highlights Edition, August 27, 2026. <a href="https://webreader.hamibook.com.tw/#/book/0100429961">https://webreader.hamibook.com.tw/#/book/0100429961</a></li><li>楊湛華, “台灣企業每週被駭7千次！AI駭客秒級攻擊，你的AI助理是資安最大破口” [Taiwanese Companies Face 7,000 Cyberattacks a Week: AI Hackers Attack in Seconds, and Your AI Assistant May Be the Biggest Security Gap], <em>CommonWealth Magazine</em>, Issue 846, April 13, 2026. <a href="https://www.cw.com.tw/article/5140589">https://www.cw.com.tw/article/5140589</a></li><li>楊喻斐, “AI正重寫資安產業遊戲規則” [AI Is Rewriting the Rules of the Cybersecurity Industry], <em>Wealth Magazine</em>, Issue 764. <a href="https://webreader.hamibook.com.tw/#/book/0100417149">https://webreader.hamibook.com.tw/#/book/0100417149</a></li><li>“代理式AI席捲企業 信任危機下的AI治理藍圖” [Agentic AI Is Sweeping Through Enterprises: A Governance Blueprint for the Crisis of Trust], <em>Micro-Electronics Magazine</em>, Issue 483, June 26, 2026. https://www.mem.com.tw/代理式ai席捲企業 信任危機下的ai治理藍圖/</li><li>“建構國家資安自動化治理 對齊平台化與AI判讀” [Building Automated National Cybersecurity Governance Through Platforms and AI Analysis], <em>NetAdmin</em>, Issue 242. <a href="https://www.netadmin.com.tw/netadmin/zh-tw/magazine/-Feature/D2CA6CCB2C274908A3705E426C6D2AD4">https://www.netadmin.com.tw/netadmin/zh-tw/magazine/-Feature/D2CA6CCB2C274908A3705E426C6D2AD4</a></li></ol><h4>External Sources Not in the Local Database</h4><ol><li>Tom Wilson, “China-linked hackers hit Taiwan in unprecedented ‘autonomous’ AI cyber attack,” <em>Financial Times</em>, August 12, 2026. <a href="https://www.ft.com/content/7d2ab3e0-9085-48f6-b38a-d90260d58795.">https://www.ft.com/content/7d2ab3e0-9085-48f6-b38a-d90260d58795.</a> The original FT article is behind a paywall and was not read through a subscriber account. The title and URL came from FT’s public page. The author attribution, unnamed-source attribution, and portions of the event description were available through public search excerpts, web archives, and open third-party quotations.</li><li>Dream Research Labs, “Inside a Multi-Agent AI Framework Used to Compromise Government Entities in Asia,” August 12, 2026. <a href="https://dreamgroup.com/blog/inside-a-multi-agent-ai-framework-used-to-compromise-government-entities-in-asia.">https://dreamgroup.com/blog/inside-a-multi-agent-ai-framework-used-to-compromise-government-entities-in-asia.</a> This is a freely accessible first-party research report. Dream did not identify Taiwan and did not release the complete underlying attack files for independent verification.</li><li>Administration for Cyber Security, Ministry of Digital Affairs, Taiwan, “境外駭客發動AI Agent攻擊政府機關 數發部啟動應變聯防強化資安防線” [Overseas Hackers Use AI Agents to Attack Government Agencies; MODA Launches a Coordinated Cybersecurity Response], August 13, 2026. <a href="https://moda.gov.tw/ACS/press/news/press/20394.">https://moda.gov.tw/ACS/press/news/press/20394.</a> This is a freely accessible first-party government statement. It does not include the figures reported by Dream and does not state whether the two disclosures concern the same incident.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1b57f58fdd7e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>