在OCP全球高峰會上，鴻海、緯穎、奇鋐、日月光等廠商推出的新系統，展現台灣企業如何打造更大規模AI部署所需的基礎設施。
台灣主要科技企業本週齊聚聖荷西的開放運算計畫（Open Compute Project，OCP）全球高峰會，展現全球AI基礎建設對台灣的高度依賴。各家展品涵蓋伺服器、散熱、電力電子與半導體封裝，共同回應機櫃密度每提升一級便隨之加劇的三大問題：電力必須以更低損耗送達處理器，熱量必須更快排出機櫃，資料則必須以銅線已無法負荷的速度在晶片之間傳輸。鴻海與和碩展出位於這些需求核心的完整機櫃，奇鋐、台達電、光寶科與日月光則呈現圍繞機櫃的散熱與供電設備。緯穎的現場光學展示指向下一步：以光直接將資料送入與送出晶片封裝的系統。
鴻海透過兩家子公司展現其布局廣度。鴻佰科技展出輝達Vera Rubin NVL72系統，鴻騰精密則供應周邊的連接、電力與散熱技術，包括高速光電連接器，以及額定功率高達480千瓦的液冷匯流排。匯流排負責在機櫃內分配電力，對其進行冷卻可因應功率需求攀升所產生的熱量。
和碩的Vera Rubin系統將72顆Rubin GPU與36顆Vera CPU整合於全液冷機櫃中。該公司同時強調其在機櫃出貨前所進行的模擬、散熱檢測與工廠測試，以便及早發現整合問題。
如此高密度的機櫃仰賴液冷，奇鋐的重點新品是一款六百萬瓦（6MW）的列間冷卻液分配裝置（CDU），負責管理伺服器設施的冷卻液循環與熱交換。其預測式控制可隨運算負載變化調整流量與壓力，在溫度迫使處理器降速之前先將熱量排除；模組化的泵浦、濾網與控制元件也讓維護更為簡便。
奇鋐正將液冷範圍從CPU與GPU延伸至記憶體、高速介面與供電元件。這項延伸之所以重要，是因為在高密度機櫃中，熱量會在其他部位不斷累積，僅冷卻主要處理器已不敷所需。
台達電與光寶科提出將供電與排熱相互銜接的互補方案。台達電的展品包括三百萬瓦（3 MW）的GoCool系統，結合冷卻液分配與預製管路，以及預計2027年初量產的Project Deschutes CDU。光寶科則推出來自投資夥伴DCX的2.6百萬瓦CDU，並搭配適用於輝達平台的電力轉換設備。這些產品共同為客戶提供更多安裝液冷機櫃所需的配套基礎設施。
兩家公司也強調正為輝達力推的800伏特直流架構做準備。以更高電壓配電可降低特定負載所需的電流，從而減少電力損耗與導體體積，但這項轉換需要電力轉換設備、保護機制與散熱系統協同調整，目前仍處於早期階段。
日月光則從供電鏈的另一端著手，也就是緊鄰處理器之處。其擴展後的PowerSiP平台將精巧的電壓調節模組置於處理器下方，縮短極大電流所行經的路徑。日月光表示，此舉可將寄生繞線損耗降低逾50%；這項數據指的是電力路徑上的損耗，而非伺服器整體耗電量。實際效益在於送達晶片的電力更多，以熱能形式浪費的電力更少。
緯穎、創意電子與Ayar Labs以現場共同封裝光學（CPO）展示聚焦連接議題。其系統將創意電子的晶片封裝與Ayar Labs的光學引擎及外部雷射光源結合，並同步監測溫度與訊號品質。將光學連接導入封裝內，提供了突破銅線頻寬與距離限制的途徑。穩定的訊號與可供查閱的運作數據，使這項合作從驗證元件可行，邁向打造資料中心團隊能夠實際管理的設備。
神達則專注於將上述能力整合為可部署的基礎設施。該公司展出與Tonomia合作開發的TonoForge模組化資料中心，並在神達伺服器上示範ZutaCore的兩相冷卻技術，利用介電液的相變在晶片處吸收熱量。神達同時強調開放韌體與管理軟體，讓營運者在安裝後擁有監控與管理系統的工具。
台灣廠商在OCP的強勢展出，顯示台灣企業能圍繞每一顆AI處理器提供的工程專業日益廣泛。伺服器製造商整合完整機櫃，專業廠商則分別解決電力損耗、散熱極限與資料傳輸問題。對客戶而言，價值在於穩定地讓算力上線，並讓昂貴的處理器持續發揮效能。對台灣產業而言，隨著AI設施日益龐大且複雜，這些需求正為整條供應鏈創造更多商機。
Taiwan Suppliers Tackle AI’s Power, Cooling, and Connectivity Challenges
At the OCP Global Summit, new systems from Foxconn, Wiwynn, AVC, ASE, and others show how Taiwan companies are engineering the infrastructure needed for larger AI deployments.
Taiwan’s leading technology firms have been out in force at this week’s Open Compute Project (OCP) Global Summit in San Jose, demonstrating how much of the global AI build-out now depends on them. Their exhibits spanned servers, cooling, power electronics, and semiconductor packaging, and together they addressed the three problems that grow with every increase in rack density. Power must reach the processors with less loss, heat must leave the rack faster, and data must move between chips at speeds copper can no longer sustain.
Foxconn and Pegatron showed the complete racks at the center of these demands, while AVC, Delta, Lite-On, and ASE presented the cooling and power-delivery equipment around them. Wiwynn’s live optical demonstration pointed to the next step, systems that use light to move data directly into and out of chip packages.
Foxconn illustrated its breadth through two subsidiaries. Ingrasys displayed Nvidia Vera Rubin NVL72 systems, and Foxconn Interconnect Technology supplied the connectivity, power, and cooling technologies surrounding them, including high-speed optical and electrical connections and a liquid-cooled busbar rated for up to 480 kilowatts. The busbar distributes electricity within the rack, and cooling it tackles the heat generated as power requirements climb.
Pegatron’s Vera Rubin system combines 72 Rubin GPUs and 36 Vera CPUs in a fully liquid-cooled rack. The company also highlighted the simulation, cooling checks, and factory tests it carries out to catch integration problems before racks ship.
Racks of this density rely on liquid cooling, and AVC’s headline introduction was a six-megawatt in-row coolant distribution unit, or CDU, which manages coolant circulation and heat transfer for server installations. Its predictive controls adjust flow and pressure as computing workloads change, removing heat before temperatures force processors to slow down, and modular pumps, filters, and controls simplify maintenance.
AVC is extending liquid cooling beyond CPUs and GPUs to memory, high-speed interfaces, and power-delivery components. The extension matters because cooling the main processors alone becomes insufficient as heat accumulates elsewhere in a densely packed rack.
Delta and Lite-On presented complementary approaches that link electricity supply with heat removal. Delta’s exhibits included its three-megawatt GoCool system, which combines coolant distribution with prefabricated piping, and the Project Deschutes CDU, scheduled for mass production in early 2027. Lite-On introduced a 2.6-megawatt CDU from investment partner DCX alongside power-conversion equipment for Nvidia platforms. Together these products give customers more of the supporting infrastructure needed to install liquid-cooled racks.
Both companies also highlighted their preparations for the 800-volt DC architecture that Nvidia has been promoting. Distributing power at higher voltage reduces the current needed for a given load, which limits electrical losses and the bulk of conductors, but the transition requires coordinated changes to conversion equipment, protection, and cooling, and it remains at an early stage.
ASE addressed power delivery at the opposite end of the chain, immediately beside the processor. Its expanded PowerSiP platform places compact voltage-regulator modules beneath processors, shortening the path travelled by very high currents. ASE says this can cut parasitic routing losses by more than 50 percent, a figure that concerns losses along the electrical path rather than total server consumption. The practical gain is more power delivered to the chip and less wasted as heat.
Wiwynn, GUC, and Ayar Labs turned to connectivity with a live co-packaged optics demonstration. Their system paired GUC’s chip package with Ayar Labs’ optical engines and external laser source while monitoring temperatures and signal quality. Bringing optical connections into the package offers a route past copper’s bandwidth and distance limits. Stable signals and accessible operating data move the collaboration beyond proving that components work towards building equipment that data center teams can manage.
MiTAC focused on assembling these capabilities into deployable infrastructure. It featured TonoForge modular data centers developed with Tonomia and demonstrated ZutaCore’s two-phase cooling on a MiTAC server, an approach that uses a dielectric fluid’s change of phase to absorb heat at the chips. MiTAC also emphasised open firmware and management software, giving operators the tools to monitor and control systems after installation.
The strong Taiwan presence at OCP showed the widening range of engineering expertise the island’s firms can supply around each AI processor. Server manufacturers integrate complete racks, while specialists address electrical losses, thermal limits, and data movement. For customers, the value lies in bringing capacity online reliably and keeping expensive processors productive. For Taiwan’s industry, those requirements create more business across the supply chain as AI installations grow larger and more complex.
