免責聲明： 本文內容僅供參考，不構成任何投資建議。投資人應獨立評估風險並自負盈虧。

Subscribe now

寫在最前面，本文想做的是成為第一次了解此領域的朋友的一塊敲門磚。主要是想從一些方向說明主要參與者的差異。公司往後的營運方向、管理層的執行力、資產負債和股權具體狀況，甚至是高管股權激勵和地緣政治，都會改變投資敘事和績效。這篇文章沒辦法直接帶各位到具體的投資決定，這也不是本文的目的。

Neocloud 在目前市場環境有被放大檢視的趨勢，資本運作如增發、舉債、投資等等，市場上的風吹草動都會造成短期股價的劇烈波動，有些是飛刀，有些是機會，要不要接和往後拿不拿得住，非常仰賴投資者對產業和公司自身情況的了解，充分準備才能讓風險變成機遇。

隨著生成式人工智慧模型參數規模的呈指數級擴張，全球運算基礎設施的瓶頸已從早期的「先進封裝與晶片短缺」顯著轉移至「電力供應與實體資料中心容量的稀缺」。傳統的超大型雲端服務供應商（Hyperscalers，如 AWS、Microsoft Azure、Google Cloud）受限於龐大且僵化的傳統架構、電網互連的長期延遲，以及針對通用運算設計的高昂定價結構，無法完全滿足人工智慧原生企業與前沿模型實驗室對高密度、低延遲算力的急迫需求。

這一結構性的市場真空，為新興的人工智慧專用雲端服務商（Neoclouds）創造了擴張契機。

商業模式與戰略路徑的底層邏輯分化

雖然 IREN、CoreWeave 與 Nebius 皆致力於提供高效能的 GPU 雲端運算資源，但其資產結構、技術基因與擴張策略卻呈現出不同的風貌。這些底層設計的差異，決定了它們當前的毛利率表現，更影響了其在未來宏觀經濟波動與硬體迭代週期中的生存韌性。

IREN 是一家具備比特幣挖礦原生背景的企業，目前正全面轉型為人工智慧雲端服務商。其核心商業模式建立在「重資產與垂直整合」的基礎上。與多數依賴第三方主機代管（Colocation）的同業不同，IREN 選擇直接擁有土地、專屬變電站、高壓電網連接權以及自行設計建造的資料中心基礎設施。

這種戰略優勢在當前全美電網極度壅塞的環境下被放大，而且使 IREN 免受第三方資料中心營運商（如 Equinix 或 Digital Realty）不斷攀升的租金影響，更賦予其對電力調度與冷卻系統設計的絕對控制權，這對於部署功耗極高的下一代液冷 GPU（如 NVIDIA GB300）至關重要。

為了彌補在軟體與企業級服務上的短板，IREN 於 2026 年中斥資約 6.25 億美元，以全股票交易的方式收購了開源雲端軟體供應商 Mirantis。這項收購為 IREN 帶來了 k0rdent AI 平台、深厚的 Kubernetes 容器編排技術，以及超過 1,500 家企業客戶基礎，讓 IREN 能從單純的「裸機（Bare Metal）」運算力供應商，進化為提供完整平台即服務（PaaS）的綜合型人工智慧雲端整合商。

然而，其劣勢同樣明顯。重資產模式意味著極高的前期資本支出，且在從 ASIC 礦機過渡至 GPU 雲端的過程中，短期內必須承受挖礦硬體退役帶來的非現金減值損失，這導致其 GAAP 淨利潤在轉型時承受巨大壓力。

相對於 IREN 的重資產策略，CoreWeave 是目前全球規模最大、市場估值最高的 Neocloud，他也同樣擁有加密貨幣（以太坊）挖礦背景，但已徹底剝離該業務。CoreWeave 採取了極致的「輕資產（Asset-light）」與「高槓桿」雙軌並行的擴張策略。透過租賃第三方資料中心的空間，CoreWeave 能夠以超越所有競爭對手的速度在全球部署算力，並成功抓住了 NVIDIA H100 晶片的第一波爆發紅利。

其戰略優勢在先發規模效應與 NVIDIA 的股權綁定，累積了高達 1,040 億美元的積壓訂單（Backlog），並在 2026 年第二季實現了 25.8 億美元的營收，年增率高達 112%。但這種極速擴張的背後潛藏著流動性與債務風險。為了支撐龐大的 GPU 採購，CoreWeave 大量使用資產抵押貸款（Asset-backed loans），截至 2026 年中，其總負債已膨脹至 720 億美元，其中包含超過 350 億美元的債務本金，以及 795 億美元的未折現租賃付款。在輕資產模式下，CoreWeave 必須完全依賴外部電網與合作夥伴的供電進度。而在基礎設施緊缺的宏觀背景下，這也成為履行千億訂單的最大執行風險。

Nebius Group 源自俄羅斯科技巨頭 Yandex 的國際業務剝離，總部設於荷蘭阿姆斯特丹，是一家由軟體工程師主導的全端（Full-stack）人工智慧基礎設施公司。Nebius 的核心優勢在於其深厚的研發底蘊與強健的資產負債表。該公司擁有超過 1,000 名演算法與雲端工程師，能提供涵蓋虛擬機器、受管 Kubernetes 集群、受管資料庫及 MLflow 等高級雲端原生工具，其軟體附加價值使其在定價上享有溢價空間。

財務方面，截至 2026 年第二季，Nebius 帳上擁有約 80 億美元的現金及現金等價物，幾乎沒有長期債務負擔。此外，NVIDIA 和對 CoreWeave 依樣也對其進行了 20 億美元的戰略注資，目標在 2030 年底前共同部署超過 5 GW 的超大規模容量。

儘管具備資金與技術雙重優勢，Nebius 面臨的挑戰在於其資本消耗速度極快，2026 年上半年的資本支出高達 81 億美元。市場對其最大的隱憂在於，這種驚人的燒錢速度是否能在未來數個季度內，順利轉化為相匹配的經常性營收增長。

Subscribe now

核心基礎設施的博弈：電力儲備與電網互連

在當前的人工智慧基礎設施競賽中，晶片供應鏈的瓶頸已逐漸緩解，取而代之的是「電力（Power）」與「變電設施」的極度稀缺。美國兩大電網系統——覆蓋東海岸資料中心走廊的 PJM Interconnection，以及德州的 ERCOT——目前均處於歷史性的緊繃狀態。

宏觀數據顯示，PJM 電網在針對 2027-2028 交付年度的容量拍賣中，首次未能滿足其 20% Installed Reserve Margin 目標，出現高達 6,625 MW 的容量缺口，也導致容量價格飆升至 333.44 美元/MW-day 的歷史新高。同時 Gartner 發布警告，預計到 2027 年，高達 40% 的人工智慧資料中心將面臨電力供應限制的硬約束。

另一方面，在德州 ERCOT 電網，Interconnection Queue 的規模已爆增至 474 GW，相當於德州歷史峰值用電量（85.5 GW）的五倍以上。由於電網面臨崩潰風險，德州州長於 2026 年 8 月下令對 ERCOT 佇列中的所有資料中心專案進行嚴格審計，實質上暫停了新的資料中心並網開發。

在這種電力稀缺性溢價急遽攀升的背景下，已獲批准的並網電力與實體資料中心控制權，成為決定 Neocloud 企業估值乘數的核心變數。

IREN 因為其重資產與擁有獨立電網的優勢，戰略護城河正在快速拓寬。其位於德州 Childress 總容量高達 750 MW 的自營園區，已不受近期 ERCOT 暫停審查政策的影響。2026 年中，IREN 成功向微軟（Microsoft）交付了 Horizon 1 專案，一座 50 MW 的直達晶片液冷（Direct-to-chip liquid-cooled）資料中心，為價值 97 億美元、總計 200 MW 的微軟多年期合約拉開了序幕。該合約明確指定使用 NVIDIA 尚未全面鋪貨的 GB300 晶片，證明了 IREN 的基礎設施已完全達標超大型企業對下一代高密度液冷技術的嚴苛要求。

CoreWeave 雖然在 2026 年第二季達到了驚人的 1.5 GW 活躍電力，並將年底目標上調至 1.85 GW，但因為對第三方主機代管的嚴重依賴，使其暴露在極高的 PJM 與 ERCOT 電網審批延遲風險中。當新世代 GPU 對單機櫃功率密度（Rack Density）的要求從 15-20 kW 躍升至 100 kW 以上時，多數傳統第三方資料中心可能無法在不進行徹底重建的情況下滿足散熱需求，對其實現千億美元積壓訂單成為重要隱患。

財務預測、資本支出與客戶預付款模型

財務結構的穩健度決定 Neocloud 在行業洗牌中是否能存活的關鍵。GPU 叢集的採購需要天文數字的資本支出（Capex），而如何為這些資本支出進行融資，在業界衍生出兩條不同的路徑：「依賴高息債務槓桿」與「獲取客戶預付款及股權支持」。

在營收與年度經常性營收（ARR）方面，CoreWeave 在 2026 年第二季，其營收達到 25.75 億美元，年增 112%，並將全年營收指引上調至 124 億至 132 億美元以及高達 1,040 億美元的積壓訂單（另有 Q3 初新增的 250 億美元承諾），這些支撐了其年底 185 億至 195 億美元的 ARR 目標。

IREN 則在簽署了包含微軟（97 億美元）、NVIDIA（34 億美元）以及多家領先人工智慧開發者（28 億美元）的長約後，將 2026 年底的 AI 雲端 ARR 目標上調至超過 40 億美元，其中約 85% 的目標營收已簽約鎖定。

Nebius 方面，核心 AI 業務在 2026 年第二季實現了高達 514% 的年增長，總營收達 5.82 億美元，當季 ARR 提升至 30 億美元。

第二層次的洞察顯示，在 Neocloud 產業鏈中，「客戶預付款（Customer Prepayments）」具備遠優於高收益債券的護城河效應。IREN 在最近與企業客戶簽署的合約中，成功要求客戶預先支付約 45% 的 GPU 資本支出。以微軟的 97 億美元合約為例，合約包含了 20% 的整體價值預付款（約 19.4 億美元），結合針對該專案的 36.5 億美元債務融資，IREN 幾乎無需動用自身股權資本，便覆蓋了高達 96% 的硬體設備採購成本（總計 58.1 億美元）。

Nebius 同樣利用其軟體優勢與客戶黏性，在 Q2 的新交易中獲得了高達 70% 的預付款支持。這種將資本開支壓力轉嫁給下游客戶的模式，極大地降低了企業自身的淨資金需求與財務槓桿風險。

相比之下，CoreWeave 走的是 GPU 為抵押品的資產支持貸款（Asset-backed loans）。分析師指出，這些貸款的貸款價值比（LTV）通常高達 60-70%。儘管 Q2 實現了高營收，但單季高達 6.4 億美元的利息支出（去年同期僅為 2.67 億美元），以及高達 57.4 億美元的負自由現金流，正在快速吞噬其營運利潤。這種高度依賴外部債權融資、用短期債務支撐長期資產支出的模式，讓華爾街擔憂其於 2026 至 2028 年間到來的「再融資之牆（Refinancing Wall）」。

GPU 折舊政策與經濟現實的錯配

人工智慧基礎設施領域目前最受爭議的會計處理，在於 GPU 設備的折舊年限（Depreciation Period）。根據歷史經驗，NVIDIA 晶片架構的迭代週期大約維持在 1 到 2 年之間（例如 Ampere 轉向 Hopper，隨後迎來 Blackwell，並預計於 2026 年推出 Rubin）。然而，硬體在實體經濟中的實際壽命（Economic Life）與財報上的會計壽命（Accounting Life）之間存在錯配，扭曲了企業的獲利能力指標與資產品質。

三家企業的折舊政策：

• CoreWeave：採用長達 6 年的直線折舊法。

• IREN：未明確公告固定年限，但其實務操作與設備租賃合約普遍暗示 3 至 5 年的生命週期。

• Nebius：採用相對保守的 4 年折舊法。

會計差異首先體現在「利潤率的幻覺（Margin Illusion）」。CoreWeave 在 2026 年第二季報告了 15.1 億美元的調整後 EBITDA，其利潤率高達 59%。然而，透過將 GPU 折舊年限拉長至 6 年，CoreWeave 大幅壓低了每年的折舊費用攤提，推高了帳面上的營運利潤。

相較之下，Nebius 採用 4 年折舊期，意味著其必須在更短的時間內吸收相同硬體的成本，導致其 Q2 的調整後 EBITDA 利潤率僅為 41%。如果將 Nebius 的折舊年限人為調整至 6 年，其利潤率表現將不亞於 CoreWeave；反之，若 CoreWeave 改用 4 年折舊，其帳面淨損將擴大。

更深層的威脅在於「資產減值與抵押品崩潰風險（Impairment & Collateral Risk）」，這可能會觸發企業破產的系統性風險。當 2026 年末至 2027 年 Blackwell 與後續 Rubin 架構全面普及，舊有 H100 GPU 的次級市場價值與算力租金必將面臨斷崖式下跌。對於折舊期長達 6 年的 CoreWeave 而言，當這些資產進入第 3 或第 4 年時，財報上仍將保留著極高的未折舊資產價值（Book Value）。一旦這些舊款 GPU 的現金流創造能力低於其帳面價值，公司將會被迫認列巨額的資產減值損失（Impairment charges）。

更何況 CoreWeave 的 350 億美元有息債務中有絕大部分是以這些 GPU 叢集作為抵押品進行借貸的。如果 H100 的市場價值崩潰，抵押品價值低於未償還的債務餘額，這極有可能觸發融資銀行的 Margin Call，或者導致其違反債務契約條款。因激進折舊政策掩蓋的抵押品減值危機，很可能是 Neocloud 產業最大的灰犀牛。

Subscribe now

雲端定價權與單位經濟效益

GPU 的租賃定價直接決定了 Neocloud 的營收天花板與回本週期。根據 2026 年年中的市場實時數據，新一代硬體（如 B200）以及主流硬體（如 H100）的定價呈現分化。

CoreWeave 目前的高定價能力建立在其絕對領先的算力規模與先發優勢上。但隨著 IREN 等重資產玩家的龐大產能陸續上線，並透過整合 Mirantis k0rdent 軟體平台補足了管理層的短板，IREN 將有能力在提供同等軟體體驗的前提下，利用電力成本優勢持續壓低市場平均租金。屆時，背負高額利息負擔、必須維持高毛利才能生存的 CoreWeave 將面臨毛利壓縮（Margin Compression）危機。這也解釋了為何在長期合約報價中，IREN 能以約 1,500 萬美元/MW（甚至微軟合約約 970 萬美元/MW）的價格換取高達 45% 的預付款，犧牲部分價格以換取絕對的資金安全。

投資分析框架

樂觀情境（Bull Case）：「電力霸權」確立與超大型企業外溢效應

此情境的觸發條件在於宏觀基礎設施的極度惡化。德州 ERCOT 以電網安全為由，徹底凍結新資料中心的互連審批，同時 PJM 因容量缺口正式向部分既有資料中心發出限電或斷電警告。AWS、Microsoft 等雲端巨頭發現自身新建的資料中心根本無法獲得電力，導致巨量的人工智慧算力訂單無處安放，最終強烈溢出至具備現成電力儲備的 Neoclouds。

• IREN： 作為市場上極少數同時擁有自有土地、專屬變電站與高達 5 GW 簽約電力的垂直整合商，IREN 成為整個北美科技圈最稀缺的戰略資產。微軟等超大型企業為了確保算力供應，不僅追加百億美元級別的長期合約，甚至可能引發對 IREN 的直接收購。憑藉高達 45% 的客戶預付款機制，IREN 以幾乎零債務槓桿的姿態完成擴張，市場給予其基礎設施極高的稀缺性溢價，推動 EV/ARR 乘數強勢擴張至 8.0x 以上，股價將迎來「戴維斯雙擊（Davis Double Play）」。

• Nebius： NVIDIA 先前承諾的 20 億美元注資全面發揮產業鏈協同效應，Nebius 憑藉在歐洲樞紐的 5 GW 擴張計畫，順利避開北美電網混戰，成為歐洲無可爭議的 AI 主權雲端霸主，股價突破估值重力持續攀升。

• CoreWeave： 受益於整個算力基礎設施板塊 Beta 值的全面抬升，市場資金寬鬆，使其租賃資產的盈利能力短暫超越高昂的利息成本，成功跨越迫在眉睫的再融資之牆，實現估值的進一步擴張。

悲觀情境（Bear Case）：折舊陷阱引爆與再融資之牆（Refinancing Wall）坍塌

此情境的觸發點在於技術迭代與需求錯配。2027 年 NVIDIA Blackwell 與 Rubin 架構產能大量湧入市場，同時 AI 基礎模型訓練的演算法效率大幅提升，導致對原始算力的絕對需求增長短暫降溫。市場供過於求，舊款 H100 的隨選租金從目前的每小時 4-6 美元暴跌至 1-2 美元。

• CoreWeave (最大受害者)： 公司面臨「資產負債表危機」。長達 6 年的非理性折舊政策使其面臨數十億美元的資產減值損失。GPU 租金的暴跌導致其作為抵押品的硬體價值快速縮水，引發資產支持貸款（Asset-backed loans）的違約條款與追加保證金要求。在流動性枯竭與資本市場極度避險的環境下，CoreWeave 無法進行再融資，高達數百億的租賃與債務將其推向破產重整的邊緣。

• IREN： 由於整體市場算力租金下滑，其高達 40 億美元的 ARR 目標確定無法達成，營收面臨衰退。然而，受惠於其重資產策略與依賴客戶預付款的保守融資模式，IREN 幾乎沒有剛性的龐大債息壓力，公司生存無虞。面對資本市場冷卻導致的股權融資困難，擴張步伐被迫大幅放緩，遠期 EV/ARR 乘數被嚴重壓縮至 3.5x 至 4.5x 的區間。

• Nebius： 過去 4 年的保守折舊政策此時發揮了巨大的安全墊作用。儘管前期的帳面利潤率看似不如對手，但在產業寒冬中免於毀滅性的資產減值。更重要的是，高達 80 億美元的現金儲備使其成為市場上唯一的流動性提供者與行業整合者。Nebius 能夠趁勢以極低的折價，收購諸如 CoreWeave 等陷入困境同業的優質硬體資產與客戶合約，進一步擴大市佔率。

結論

IREN 目前提供了當前行業內較佳的風險報酬比。其垂直整合的重資產模式，在全美電網互連日益艱難、電力稀缺性溢價飆升的當下，賦予了戰略定價權。透過精準併購 Mirantis 補齊了長久以來的軟體短板，並善用高比例的客戶預付款大幅降低融資風險，IREN 建立了一道難以被輕資產同業複製的堅固護城河。(註：不少投資者對管理層薪酬結構有疑慮，這在市況不好時可能會是額外的隱憂)

CoreWeave 則呈現出「高風險高報酬」特徵。千億美元的積壓訂單背後有高達 350 億美元的巨額債務與激進的折舊政策，在未來一至兩年的硬體換代期時，投資者必須對其再融資能力保持最高級別的警惕。

Nebius Group 則是軟體實力與現金流的結合體，健康的財務體質賦予了極高的容錯率，但當前高昂的估值倍數反映其輕資產的特性，也要求管理層展現無懈可擊的執行力，以證明其資本支出的長期轉換效率。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

免責聲明： 本文內容僅供參考，不構成任何投資建議。投資人應獨立評估風險並自負盈虧。

Subscribe now

前面這一篇文章我們從一個很簡單的問題開始：

為什麼 AI 需要越來越多的光？

當 GPU 數量不斷增加、資料交換速度不斷提升之後，傳統的銅線開始碰到物理極限，所以我們逐漸讓光訊號取代依靠銅傳輸的電訊號。

也就是資料傳輸的路徑逐漸改變：

從 GPU → Copper → Optical Module → Fiber

走向：

GPU / Switch ASIC → Optical Engine → Fiber

而 CPO 的核心，就是進一步把 optical engine 往 switch ASIC / compute package 靠近，縮短高速電訊號在封裝內或板級的傳輸距離。

現在我們可以把問題再延伸一步：

當 AI 對光學傳輸的需求越來越高，Scale-out/Scale-up/Scale-across 帶動傳輸的流量越變越大，從 800G 走向 1.6T 以上，這部分的資本支出會流向誰？

800G → 1.6T：真正改變的不只有光模組數量

800G 時代，一個典型架構可以理解為：

8 × 100G = 800G

到了 1.6T，目前主流方向則是：

8 × 200G = 1.6T

未來更高速度則會進一步走向更高的單 lane 速率。

這裡其實可以看到兩種不同的產業升級方式：

第一種是：需要更多 lanes / 更多 components

第二種則是：lane 數量不變，但每個 component 的速度、複雜度與製造要求提升。

而後者正是產業升級的重要方向。

因此，1.6T 的產業故事不只是：

「需要更多光模組。」

而開始變成：

「需要更多能夠穩定量產高速 optical components 的產能。」

目前來看，Lumentum（LITE）已經有 200G/lane EML，並將其用於 1.6T pluggable transceiver；Coherent（COHR）也早已推出 200G InP EML，並將其用於 1.6T transceiver。Lumentum 在 OFC 2026 更展示了採用其 200G EML 的 NVIDIA 1.6T 2×DR4 OSFP。

當單一 lane 從 100G 提升到 200G，整條 signal chain 都必須跟著升級。

包括：

• EML / Laser

• Photodiode

• Driver

• DSP

• Optical Engine

• Packaging

• Thermal management

Lumentum 早在 2024 年就展示 200G/lane InP photodiode，並搭配其量產中的 200G EML，定位於下一代 800G / 1.6T optical transceivers。

到了 2026 年，Lumentum 已展示使用其 200G EML 的 NVIDIA 1.6T 2×DR4 OSFP，而其 1.6T 模組採用 8×約 212.5G PAM4 lanes 的架構。

這代表 1.6T 的升級不是一個單純的產品換代，而是整條 optical chain 同時提高規格。

而且這裡要注意一個重要區別：AI deployment 增加會帶動 optical ports 數量增加；但從 800G → 1.6T 的升級，首先代表的是每一個 optical port 能夠傳輸更多資料。

因此，AI CapEx 同時會帶來兩個方向的需求：

Port count ↑

以及：

Bandwidth per port ↑

前者帶來 volume，後者則提高每一個 port 裡面高階 optical components 的價值。這是第一層 Value Migration。

從「Volume」走向「Bottleneck」

這也是為什麼 Lumentum（LITE）值得研究。

但這裡需要先釐清一個概念：

Bottleneck 不代表「只有一家供應商會做」。

LITE 並不是唯一一家能夠提供高速 EML / laser 的公司。後面會談到的 COHR，同樣具備 200G EML、InP laser、photodiode 等能力，甚至已經展示更高速度的 EML 技術。

真正值得研究的是：

當 optical lane 從 100G 提升到 200G，哪一個環節最難跟著擴大產能？

比「誰會做 EML」更重要是誰有更好的量產能力。

因為從 100G → 200G，並不是簡單地把原本的元件速度提高一倍。

高速 EML / laser 同時面臨：

• 更高的 modulation speed

• 更嚴格的 wavelength control

• 更高的 optical power

• 更低的 error rate

• 更高的 manufacturing yield 要求

• 更嚴格的 reliability requirements

而且這些元件往往需要較長時間的製程、良率爬坡與客戶 qualification。

換句話說：

需求可以在一兩年內快速增加，但高速光電元件的有效供給能力，不一定能以同樣速度增加。

這才是 Bottleneck 的真正來源。

假設 800G 時代，每個 transceiver 主要使用 100G/lane EML。

當市場進入 1.6T，開始大量轉向 200G/lane 時，需求不只是從：800G → 1.6T，而是同時發生：100G/lane → 200G/lane，這意味著整個產業必須重新建立更高速 EML / laser 的有效產能。

Transceiver 廠可能可以很快擴充 assembly capacity，但高速 EML 的產能、良率或 qualification 跟不上，那麼限制整個產業出貨的就不再是 transceiver assembly，而可能變成上游的 EML / laser。

這時候真正稀缺的不是「光模組」，而是光模組裡最難擴產的高速元件。

當 200G/lane 成為 1.6T 的主流架構，那麼高速 EML / laser 的產能、良率與 qualification 可能成為整個 optical supply chain 的限制因素，而 LITE 是其中值得追蹤的主要供應商之一。而在 bottleneck 這個位置會發生下面的傳導：

Capacity/Utilization/Yield → ASP → Gross Margin

總結它的投資邏輯：

AI bandwidth ↑ → optical speed ↑ → high-speed component demand ↑ → supply constraint ↑ → bottleneck value ↑

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

如果 LITE 和 COHR 都做 EML，COHR 的故事是什麼？

如果只看 200G EML：

LITE ≈ COHR

但如果把視野拉大，就會看到兩家公司策略上的差異。

COHR 的產品組合和應用更加廣泛，也可以說它的 AI optical exposure 並不像 LITE 那麼集中在部分高速 connectivity components。

它同時布局：

• InP

• EML

• Laser

• Photodiode

• VCSEL

• Silicon Photonics

• Optical Engine

• CPO

例如在 OFC 2026，COHR 展示了多種 1.6T transceiver architectures，包括 Silicon Photonics PIC、high-power InP CW laser、200G InP EML 與 200G VCSEL；同時也展示 400G/lane Differential EML 與 Silicon Photonics 路徑，對應下一代 3.2T architectures。

在 CPO 方面，COHR 也展示了 Silicon Photonics + External Laser Source、VCSEL-based CPO，以及 400G/lane InP modulator 等不同架構。

差異比較像是：

LITE：

High-Speed Optical Content + Manufacturing Scale

也就是在高速 EML、laser 等關鍵元件上的技術與量產能力。

COHR：

Photonics Technology Breadth + Architecture Optionality

也就是跨越不同 photonics technologies 與 architectures 的平台能力。

當然，這並不代表 LITE 只能做 EML。公司也已經展示了 1.6T 中使用 UHP laser + Silicon Photonics engine 的架構，因此 LITE 同樣可以參與部分 Silicon Photonics / CPO 路徑。

更精確的比較應該是：

LITE 的投資邏輯更偏向「高速 optical components 的技術與量產能力」；COHR 則更偏向「跨不同 photonics architectures 的選擇權」。

假設未來 1.6T 最終主要採用：

200G EML

LITE 和 COHR 都能受惠。

如果未來更多採用：

Silicon Photonics + External Laser

兩家公司仍然都有參與機會，但受惠方式不同。

如果未來短距離 AI interconnect 更多採用：

200G VCSEL

COHR 也有對應產品。

如果未來 CPO 大量部署，COHR 同樣有 Silicon Photonics、InP laser、VCSEL 與 CPO architecture 的布局，較可以從不同技術路徑受惠。

所以 COHR 的核心價值，相較於 LITE 傾向的：

「某一個高速元件可能成為 bottleneck。」

更像是：

「無論下一代 optical architecture 最後選擇哪一條技術路線，我都有機會提供其中的重要元件。」

1.6T 時代出現了三種不同的受惠方式

總結一下前面的內容，其實提到了三種不同的 AI Optical thesis 受惠方式：

第一種：Volume Play

代表：AAOI、InnoLight、Eoptolink

邏輯很簡單：

AI CapEx ↑ → Network Ports ↑ → 800G / 1.6T Transceivers ↑ → Revenue ↑

它們直接受惠於更多的光。

第二種：Bottleneck Component Play

代表：LITE

邏輯是：

AI bandwidth ↑ → 200G/lane ↑ → EML / Laser complexity ↑ → Capacity becomes constrained → Component suppliers gain leverage

它受惠的不只是「更多的光」，而是：

更高速、更難製造、供給更容易受限的光。

因此，真正要驗證的是 capacity、yield、ASP、utilization 和 gross margin 是否開始反映供需緊張。

第三種：Photonics Platform Play

代表：COHR

邏輯是：

1.6T → EML / SiPh / VCSEL / Laser → Optical Engine → CPO → Multiple architecture opportunities

光學架構和元件本身的升級，它都有機會受惠。

特別提醒光通訊概念近期 (mid-2026) 在市場上有很高的熱度，較高的估值代表著較低的容錯率，同時高額的資本支出可能帶來舉債和股權稀釋的風險。研究公司不是為了馬上投資，是為了在機會來臨時能知道這是機會。

1.6T 與 CPO 是兩個不同階段，或者說是兩個不同的投資概念

這兩者容易被放在一起討論，但其實解決的是不同問題。

800G → 1.6T

核心問題：

How much data can each optical lane carry?

因此價值增加主要集中在：

EML

Laser

Photodiode

Driver / DSP

Optical Engine

也就是：

提高每一個 optical lane 的 bandwidth。

CPO

核心問題：

How close can optics move to the compute?

因此價值轉向：

Optical Engine

Silicon Photonics

External Laser Source

Advanced Packaging

CPO Architecture

也就是：

縮短高速電傳輸路徑，降低功耗與訊號完整性問題，讓 optics 更靠近 switch ASIC / compute。

所以：

1.6T 主要是 speed / bandwidth upgrade；CPO 則是 architecture / packaging upgrade。

兩者可以同時發生，都和光有關，但其實是兩個概念。

最後整理幾個未來幾年的 AI Optical 追蹤框架，在投資時可以根據公司財報和市場消息，驗證投資趨勢是否真的正在發生。

① 1.6T 到底有多快放量？

觀察：

200G/lane adoption

1.6T transceiver shipments

Hyperscaler deployment

當 1.6T 快速放量：LITE / COHR / transceiver vendors 都會受惠。

更重要的觀察是：這是單純的 volume growth，還是同時伴隨 ASP / mix / margin improvement？如果有才代表 Value Migration 正在發生。

② Bottleneck 到底在哪裡？

觀察：

Capacity

ASP

Utilization

Yield

Gross margin

供需失衡正好代表 Bottleneck component suppliers 的價值捕獲能力提高，從毛利趨勢最容易判斷這個節點是否是 Bottleneck。

③ CPO 真正意義上的量產？

很多公司都在說 CPO，但真正重要的問題不是：「誰展示了 CPO？」，而是：「誰的 CPO 真正開始量產，並產生實質營收？」

這會是 Architecture Shift 真正發生的證據。

一旦 CPO 從 demonstration 進入 production，資本支出的受益者可能會進一步從傳統 pluggable transceiver，向：Optical Engine → Silicon Photonics → Laser → Advanced Packaging → CPO 重新分配。

因此，未來真正值得追蹤的，不只是 AI 需要多少光，而是：

AI 每增加一單位 bandwidth，價值究竟流向 optical chain 的哪一個環節？

這才是 AI 「光」投資真正值得研究的問題。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

免責聲明： 本文內容僅供參考，不構成任何投資建議。投資人應獨立評估風險並自負盈虧。

Subscribe now

本文著重在解釋公司未來的策略和發展方向。除了Elon Musk本人的能力和執行力，公司現有的技術和路徑很值得更深入討論。在傳統估值體系都在說股價高估，公司眼中的未來或許可以幫助理解市場為了什麼買單。

特別註明，文章中的時間點完全根據公司發布的內容，這很可能是官方較樂觀的推演，同時在商業化落地的過程有很多被簡化的地方。願景很美，過程也可能很曲折，需要大家自行判斷。

以下是公司最新財報提到的公司發展三大方向：

1. Starship V3：從試飛向每日軌道運輸轉型，把「上太空」這件事變成一條全年無休的物流輸送帶

依目前的市場版圖，現在要把東西送上太空的門檻依然極高，SpaceX靠著獵鷹（Falcon）火箭已經吃下全球超過80%的發射噸位。2025年SpaceX一年送往軌道的質量超過2,200噸，占全球超過80%；這也是目前全球其他發射業者難以抗衡的規模。之後透過極致的完全回收再利用，公司的長期目標是將年度運力從現在的幾千噸，拉到100萬、甚至1,000萬噸以上。

喊出這種指數級別的目標，其原因是近期的進展已經讓Starship最重要的幾個技術風險逐步降低。

Flight 13成功部署20顆V3 Starlink衛星，並成功讓單具Raptor引擎在太空中重新點火。同時，Elon Musk表示，根據目前取得的數據與目視檢查，他認為隔熱盾（Heat Shield）問題已經解決；他也特別強調，這並不代表未來不會繼續改善隔熱盾，而是目前看不到達成「完整且快速回收（full and rapid reusability）」的技術障礙。這是非常重要的進展，但仍需要後續飛行與實際快速重複使用來驗證。

接下來的12個月，塔架捕獲與每日一飛是公司的重要技術目標。等待監管機關放行，Flight 14將挑戰用發射塔直接捕獲第二級的星艦船體（Ship）。目前公司預計最早在8月底進行Flight 14，並嘗試第一次以發射塔捕獲Ship。

如果未來能夠完成第一級與第二級火箭的「雙重捕獲」，將逐步消除對傳統降落支架與海上回收流程的依賴，可以減輕部分系統重量並大幅縮短兩次發射間的整備時間，朝全面且快速重複使用（full and rapid reusability）的目標前進。

最終的藍圖，是在12個月內實現「每日一飛（Daily Flight）」。這將不僅僅是發射次數的疊加，SpaceX正在打造的，是一條連接地球與太空的物流輸送帶。

SpaceX在SEC文件中指出，Falcon 9可將約23公噸送入LEO，而Starship V3的目標載荷能力為100公噸；未來版本則可能進一步提高到200公噸。Starship的運力躍升是支撐下一代Starlink大規模部署的重要物理途徑，其龐大的有效載荷空間將解鎖前所未有的軌道頻寬。

Subscribe now

2. Starlink 2.0：重塑全球電信與AI聯網的底層邏輯

透過V3衛星的推進和EchoStar頻譜的整合，SpaceX想打造「全方位連網基礎設施」，直接鎖定全球數千億美元的電信與行動通訊市場。

從產業佈局的角度來看，這裡面有三個戰略轉折點：

1. V3衛星的「百倍乘數效應」

算力與運力的疊加：10倍的單體頻寬提升（V3 vs. V2）乘上10倍的發射量，理論上創造的是總頻寬最高約100倍的爆發性成長。SpaceX官方資料顯示，V3單顆衛星具備約1 Tbps下行容量，相較V2約提升10倍；Elon Musk則表示V3衛星的部署量也將增加10倍以上。

目前公司已在Flight 13首次部署20顆V3衛星，但這批衛星是測試任務的一部分，並未進入長期營運軌道。Flight 14則預計首次將V3 Starlink衛星部署到實際營運軌道。

2. Starlink Mobile：傳統電信商要擴充網路，就得砸大量資本建基地台；Starlink走的是基礎設施極致簡化的路線

SpaceX在IPO文件中表示，2027年將部署下一代V2 Mobile衛星，並結合EchoStar頻譜與新的5G協定，大幅提升Starlink Mobile的容量。

另一方面，SpaceX也計畫利用Starlink現有地面基礎設施，在原本安裝Starlink接收盤的設備上增加小型蜂巢式基地台（small cellular base stations / femtocells），形成低成本、分散式的地面網路，再與衛星網路結合。也就是說，利用既有安裝位置與設備，降低新增行動網路基礎設施的成本。

背後的邏輯已不僅僅是去填補偏遠地區的訊號死角，而是進一步切入都會區與次都會區，挑戰全球電信巨頭的核心利潤。但這部分仍需要大量頻譜、地面設備與監管許可，目前應視為SpaceX的擴張計畫，而非已驗證的商業模式。

3. AI與自動化時代鋪設「機器聯網」

隨著AI運算、人形機器人跟自動化設備的全面爆發，未來對「低延遲、高頻寬」網路的需求，主力可能逐步從現在的「手機消費」轉移到「機器間的生產協作」。

3. 20 GW基建與「星載AI」的戰略邏輯

此策略的核心在於SpaceX想把航太等級的「熱管理」與「系統工程」經驗，直接拿來對傳統資料中心進行降維打擊。對那些習慣處理火箭極端高溫與能量密度的工程師來說，在地球上蓋機房，可能像是「大聯盟選手去打小聯盟」一般。

目前AI基礎設施最大的限制之一，是算力需求、電力、散熱以及高階晶片供應之間的失衡。SpaceX採取的資本支出（Capex）策略是「基礎設施（BOS）先行」。

目前SpaceX的AI運算容量約為1.4 GW。公司希望到2027年底建立一系列累計約20 GW的電力與冷卻基礎設施；但Elon Musk也表示15 GW可能更為現實。

公司的目的在確保即便供應鏈交貨延遲，SpaceX仍能提前準備足夠的供電與散熱能力，讓基礎設施等GPU，而不是讓昂貴的晶片躺在倉庫裡等機房。公司目前的AI基礎設施投資也已經顯著加速，Q2單季AI相關資本支出約158億美元。

同時SpaceX全面採用NVIDIA的AI硬體，並與NVIDIA合作設計Starmind AI1衛星。Starmind將搭載NVIDIA Rubin GPU與Vera CPU，提供資料中心等級的太空運算能力，將Vera Rubin架構從地面延伸至軌道。

2027年即將啟動的「Starmind AI衛星」，本質上就是把資料中心級的AI算力搬到軌道。太空環境確實可以利用大型太陽能板提供能源，也能透過輻射散熱進行熱管理，但簡單理解為「天然低溫且零衰減的太陽能」可能過於理想化；太空中的熱管理本身仍是一項重要工程挑戰。

SpaceX把造火箭時累積的散熱、電力管理與系統整合能力平移到AI基礎設施。這種垂直整合不僅可能降低建置與營運成本，也可能將資本支出的回收期壓縮到一年以內，但這目前仍是公司的經濟模型與目標，需要依後續的執行狀況來證實。

600億美元Cursor收購案，目的在於加速SpaceX的AI企業業務與工程軟體能力。核心邏輯不只是讓強大的算力最終落地為生產力，還希望透過Cursor的coding與knowledge-work能力，進一步強化Grok與SpaceX內部工程流程，形成更完整的AI閉環。SpaceX在最新財報中也直接將Cursor交易定位為「加速AI企業機會」。

最後，公司提到了1兆美元營收目標從2031年前移至2030年，同時也說明關於月球基建計畫。由於時程較遠，暫不評論過多。需要特別注意的是，這個1兆美元的目標高度依賴Starlink、AI基礎設施與Starship等多個業務同時實現大規模商業化，因此目前更適合作為SpaceX的長期願景與估值框架，而不是單純的財務預測。

總結

整體而言，公司的發展方向和敘事邏輯蠻清晰的，但正如每個段落都有點到，從故事到事實都需要一連串的驗證，譬如發射業務並不是Flight 14發射成功、雙重捕獲成功，就可以直接接上每日發射這個目標，其中省略了非常多技術節點。Starlink Mobile也不僅僅是把更多更強的衛星發射上去就能商業化，監管批准和頻譜整合都是不確定性。總而言之，公司給的時間線個人感覺相當激進。

在目前極高的估值下，市場可能提前反映Starship快速重複使用、Starlink大規模部署，以及AI基礎設施高速擴張等未來成長，因此任何重要里程碑的延期，都可能讓原本被提前定價的收入與現金流往後移，進而壓縮估值。此外，AI基礎設施也有一個需要密切觀察的風險：SpaceX正在以極高的資本支出速度建置算力與電力容量，未來AI相關營收與長期合約能否同步成長，將決定這些Capex究竟是在提前建立稀缺產能，還是存在資本效率問題。

SpaceX的安全邊際在於公司並不是一個早期新創，Starlink與既有發射業務已經建立穩定且快速成長的收入基礎，因此並不需要Starship、Starlink Mobile、AI基礎設施三條曲線同時成功，才能支撐公司的基本價值。這也使SpaceX的投資邏輯更接近一個「具有現金流支撐的多重成長選擇權」。對投資人而言需要思考與追蹤的，不僅是管理層提出的目標是否準時完成，而是在每個新聞和驗證節點給出自己的理解和判斷，時程儘管延後，公司的技術成熟度、部署速度、合約收入與資本效率是否有在逐步提升。倉位管理在投資這類型需要定價成長選擇權的公司尤為重要。控制好倉位，才能在波動和雜訊來臨時，增加持股和心態的穩定。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

免責聲明： 本文內容僅供參考，不構成任何投資建議。投資人應獨立評估風險並自負盈虧。

Subscribe now

大公司的財報季告一段落。在資本支出持續擴大的背景下，每家公司因為市場定位與既有優勢不同，也發展出不同的AI成長與變現路徑。

問題在於，如果無法理解這些差異，很容易把所有AI投資都看成同一件事。現在看到自由現金流下降，就會和市場一樣擔心投資是否能得到回報。之後當市場風向或關注點改變時，也更容易因為缺乏清楚的基本面框架，犯下隨波逐流的錯誤。

這也是我希望透過文章幫助大家釐清的地方。此篇我想特別說明的是雲端工作流，以及目前主流三大公司Amazon AWS、Microsoft Azure與Google Cloud的差異。

需要特別說明的是，企業雲端與AI工作流只是它們龐大商業版圖的一部分。每一家公司的整體業務，都值得單獨用一篇甚至數篇文章深入分析。本文著重的是我觀察到許多投資人目前最容易混淆的問題：

三家公司都在提供AI算力、模型與Agent平台，但它們分別依靠什麼競爭優勢？是如何進入企業工作流的？

AWS：提供企業AI工作負載的執行環境

Amazon在最新一季財報中表示，AWS的AI業務與自研晶片運算業務，各自都已超過250億美元的Annual Revenue Run Rate，而且仍以三位數速度成長。AWS本季營收達422億美元、年增37%，營業利益達166億美元，代表企業AI已經轉化為相當可觀的雲端收入與獲利。

AWS眾所皆知是企業雲端發展最早、市占率最高，基礎設施與服務種類也最成熟的參與者之一。

今天以一家零售商為例。假設它想使用AI Agent，將異常訂單處理自動化，通常會需要以下流程配合：

• 收集訂單、客戶、庫存及物流資料

• 整理資料，並建立身分、權限與安全規則

• 選擇模型，建立Agent與決策流程

• 將Agent提供給員工或客戶使用

• 讓Agent呼叫企業系統並執行動作

• 記錄結果，供後續稽核與改善使用

當一筆高價值客戶訂單延遲時，Agent不只是回答「訂單目前在哪裡」，還可能需要查詢其他倉庫的庫存、讀取改單政策、計算額外物流成本、取得主管批准，最後修改訂單並通知客戶。

從AWS的角度出發，許多企業原本就已經把應用程式、資料庫、檔案、API與安全架構建構在AWS平台之中。因此，企業不必先把資料和系統搬到另一個環境，便能在原有後端架構上加入AI。

這能顯著降低AI與既有後端系統的整合成本，也是AWS的第一項優勢。

之後，這家零售商可以透過Amazon Bedrock使用不同的基礎模型，再利用自己的文件、資料庫與企業規則建立RAG或Agent。

AWS刻意採取多模型策略，讓企業依照成本、速度、準確度、資料主權與合規要求選擇不同模型，而不必完全綁定在單一模型上。這說明AWS不一定需要擁有市場上最強的消費端AI模型；只要企業最終仍在AWS上執行訓練、推論、資料查詢與工作流程，Amazon就能持續收費。

簡而言之，AWS提供的是這家零售商AI工作負載的執行環境。

員工在實際操作時，可能是在自家客服系統、網站或其他應用程式中與Agent互動，甚至不會直接看到AWS的品牌或介面。但背後的模型推論、資料庫查詢、文件搜尋、儲存、API呼叫、網路傳輸、安全與監控，都發生在AWS上。

因此，每當Agent完成一次工作，背後都可能帶動多項AWS服務的使用量。AWS賣的不是一個獨立的聊天機器人，而是Agent執行每一個步驟時所需要的完整雲端環境。

這也解釋了為什麼Amazon的AI收入快速成長，卻仍需要同步擴大資本支出，讓自由現金流承受壓力。AWS必須先建置資料中心、採購晶片、取得電力，之後才能等待客戶逐步使用這些產能。截至第二季的過去十二個月，Amazon合併營運現金流年增33%至1,614億美元，但自由現金流由去年同期的正182億美元，下降至負76億美元。公司指出，主因是房地產與設備採購淨額年增661億美元，而這項增幅主要反映AI投資。

所以，AWS目前的狀況是：

AI相關收入已經快速成長，但Amazon擴建AI基礎設施所投入的現金，現階段仍高於新增業務所帶來的即期現金回收

這是一種典型的基礎設施商業模式。Amazon需要先支付龐大資本支出，再透過未來數年的算力使用量、資料服務和長期客戶合約回收投資。

自研晶片如何強化這個飛輪？

AWS積極把Trainium、Inferentia等自研晶片導入AI訓練與推論服務。當AWS擁有足夠龐大的客戶與實際工作負載，就可以形成：

更多企業AI需求
→ 增加Trainium使用量
→ 累積更多真實工作負載與軟體優化經驗
→ 改善晶片效能與成本
→ 提供更有競爭力的價格
→ 吸引更多工作負載回到AWS

Amazon最新財報也提到，Anthropic與OpenAI已對Trainium做出多年、數GW等級的使用承諾，顯示自研晶片不再只是AWS內部的降本工具，也開始成為客戶選擇AWS算力的重要產品。這使AWS的自研晶片並非獨立存在，而是與既有雲端規模形成正向循環。

AWS的核心優勢可以濃縮成一句話：

企業可以使用不同模型、建立不同應用，但只要資料、運算與執行仍留在AWS，Amazon就能從整個工作負載持續變現。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Microsoft：掌握員工入口，再把AI需求導向Azure

從Microsoft的角度出發，它在既有企業工作流上的優勢非常明確，就是掌握大量員工每天工作的前端環境。員工本來就已經在Teams、Outlook、Excel、Dynamics、Power Platform或GitHub之中工作。

如果同一家零售商想使用AI Agent處理異常訂單，它不一定需要另外建立一個新的操作介面，而可以直接把Agent放進員工既有的工作流程中。

例如，員工只要在Teams或Dynamics中輸入：「找出目前系統判斷異常的訂單，整理延遲原因、受影響的高價值客戶，以及可以採取的處理方案。」

Copilot或Agent便可以：

• 從Dynamics取得客戶與訂單資料

• 查詢ERP或庫存系統

• 讀取SharePoint中的公司政策

• 分析Excel報表

• 將結果整理成Excel表格

• 草擬一封Outlook郵件

• 在主管批准後，呼叫系統執行改單

Microsoft最大的優勢在於，員工不必主動前往一個新的AI平台：

AI已經存在於工作實際發生的地方。

這家零售商可以使用Copilot Studio建立Agent，連接知識來源、工具、其他Agent與企業應用，並設計執行順序、核准流程與權限。Microsoft Foundry則負責更複雜的模型選擇、Agent開發、資料連接、評估、監控與治理。

整體架構大致可以分成兩層：

• Copilot、Teams、Dynamics與GitHub負責員工入口及工作流程

• Azure、Foundry、Fabric與模型基礎設施負責背後的運算、資料與AI能力

Agent的使用不只可能提高Microsoft 365、Dynamics或Copilot的軟體收入，也會增加Azure的模型推論、資料庫與運算用量。Microsoft便能從同一條工作流收取多層收入。

AWS主要從後端使用量回收投資；Microsoft則有機會先從每位員工的軟體授權與Copilot方案收費，再從背後的Azure消耗取得第二層收入。

最新一季Microsoft Cloud收入達593億美元、年增27%；Azure在2026財年全年收入首次超過1,000億美元、年增41%；Microsoft 365 Copilot付費席位超過3,000萬。這反映AI需求同時推動Azure算力與Microsoft第一方AI應用。

然而，這種模式同樣需要龐大的前置投資。Microsoft最新一季資本支出達410億美元，其中約三分之二用於壽命較短的CPU與GPU，其餘則是可使用多年的資料中心、土地與其他長期資產。當季營運現金流仍年增30%至554億美元，自由現金流為196億美元，顯示公司目前仍有足夠的軟體訂閱、雲端帳款與企業合約現金流承接AI投資。

Microsoft在企業工作流的AI路徑是：

一套由員工工作入口、企業資料、Agent執行能力與Azure基礎設施共同組成的AI化工作環境。

Microsoft的核心優勢是，企業不必先思考如何把AI推廣給員工。因為Teams、Outlook、Excel、Dynamics與GitHub本來就在員工面前，AI只需要被加入既有工具中。

不過，Microsoft未來真正需要證明的，也不只是有多少企業購買Copilot，而是企業是否願意持續增加使用量，以及席位收入、Agent收入與Azure消耗，能否長期高於GPU折舊、能源和模型推論成本。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Google：先理解與推理，再把企業資料轉化為行動

Google在企業工作流中沒有AWS同等規模的既有雲端存量，也沒有Microsoft在大型企業辦公軟體與工作入口上的控制力。

因此，它需要透過另一種方式建立差異化：

把模型、資料分析、搜尋與自研算力整合起來，讓企業不只自動執行既有流程，也能從資料中找出原本不容易被發現的問題與決策機會。

仍以訂單異常為例。

Google Cloud可以協助零售商：

• 分析大量歷史訂單與物流資料；

• 找出特定倉庫、供應商或路線的異常模式；

• 預測哪些訂單可能在未來幾天延遲；

• 判斷哪些客戶與營收受到的影響最大；

• 比較改倉、補貨或調整配送方式的成本；

• 讓Agent根據分析結果主動提出處理方案。

它不只是處理「已經被系統標記為異常」的訂單，而是試圖更早回答：

哪些訂單可能即將出現異常？原因是什麼？現在應該採取什麼行動？

這比較像是在現有工作流程之上，替企業多增加一層分析、預測與決策能力。

在這套架構中，Google會使用：

• BigQuery負責整合與管理企業資料

• Gemini負責理解問題、多模態推理、生成與規劃

• Vertex AI與Agent平台負責建立、部署Agent

• TPU與GPU基礎設施負責訓練和推論

• Gemini Enterprise及Workspace則提供員工搜尋資訊、使用AI助理的介面

如果這家零售商原本已經在使用Google Workspace，員工可以透過Gemini Enterprise進行企業搜尋、使用AI助理或啟動Agent。Google同樣提供低程式碼與無程式碼工具，讓業務團隊建立較簡單的工作流程。

然而，相較Microsoft，Google在大型企業核心工作介面中的控制力通常仍然較弱；相較AWS，Google既有後端工作負載與服務廣度也較小。

因此，Google可能更需要依靠以下因素吸引企業：

• Gemini的模型品質

• BigQuery的資料分析能力

• Google累積的搜尋技術

• TPU的成本與效能

• 模型、資料和Agent平台之間的垂直整合

Google Cloud的主要賣點比較像是：

我們可以把企業分散的資料整理成AI能理解的商業脈絡，再由Gemini和Agent把資料轉化為分析、預測、決策與執行。

這套模式的潛在價值很高，因為它偏向產出與結果導向。然而，問題一直都是：Google能否把技術優勢轉化為足夠規模的企業合約、工作流與現金流。

最新財報顯示，Google Cloud第二季收入達248億美元、年增82%，其中除了GCP、AI解決方案與AI基礎設施需求外，也包含首次認列的TPU系統銷售收入。公司表示，即使排除TPU系統銷售，Cloud收入成長仍明顯加速。

這種成長同樣需要非常高的資本支出。Alphabet第二季資本支出達449億美元，高於當季391億美元的營運現金流，使自由現金流降至負59億美元。公司也將2026年全年資本支出預估提高至1,950億至2,050億美元。

再次補充一下，這裡討論的單純是Google Cloud的AI路徑，只是其商業版圖其中一部分。Google這套模式最需要證明兩件事：

第一，Gemini、TPU與BigQuery的整合，能否提供企業明顯優於其他模型與資料平台的效果。

第二，Google能否把技術與資料優勢，進一步轉化為長期企業合約、員工採用與大規模Agent部署，而不只是AI基礎設施及模型API收入。

總結

下面分類主要代表的是三家公司最具差異化的切入點與變現優勢，不是它們能力的絕對邊界；實際上，三家公司都正在向基礎設施、資料平台、Agent開發與員工入口的其他環節擴張。

AWS：企業AI工作負載的執行環境

AWS比較像提供建造AI工廠所需要的一切。

它不必控制員工最終看到的介面，也不一定需要擁有最強的模型。只要Agent背後的運算、資料、儲存、網路與系統呼叫都在AWS上進行，Amazon便能持續收費。

它最依賴的是：

基礎設施廣度、既有客戶存量，以及自研晶片帶來的成本與供給優勢。

目前AWS的AI收入已經快速成長，但自由現金流轉負，也反映基礎設施投資必須先於收入回收。我們可以持續觀察新增產能的利用率、Trainium的採用，以及AI工作負載能否持續帶動資料庫、儲存與其他AWS服務。

Microsoft：AI化的企業工作入口與工作流程

Microsoft把AI放進員工每天工作的辦公室。

企業可以在Teams、Outlook、Excel、Dynamics與GitHub中直接使用Copilot或Agent，再由Azure、Foundry與Fabric提供背後的模型、資料與運算。

它最依賴的是：

企業軟體分發、工作流程控制力，以及從席位收入延伸至使用量收入的能力。

Microsoft目前仍維持正向且龐大的自由現金流，顯示既有軟體訂閱與Azure業務可以承接資本支出。但隨著GPU、資料中心與推論成本快速上升，未來仍需要證明Copilot與Agent能提高每位客戶收入，而不只是增加運算成本。

Google：企業資料的分析、預測與決策層

Google則試圖成為企業資料之上的搜尋、分析與智慧層。

它希望結合BigQuery、Gemini、TPU與Agent平台，讓企業不只是加速既有流程，也能從資料中發現異常、預測結果並主動採取行動。

它最依賴的是：

模型、資料、搜尋與自研算力垂直整合後，能否創造足夠明顯的技術與經濟優勢。

以各公司公布的雲端業務成長率來看，Google Cloud本季成長最快，但Alphabet當季自由現金流已轉負，反映公司正在提前建設以滿足未來需求。我們需要追蹤的，是這些Cloud合約與AI需求能否逐步轉化為收入、獲利及持續的Agent使用量。

Subscribe now

免責聲明： 本文內容僅供參考，不構成任何投資建議。投資人應獨立評估風險並自負盈虧。

Subscribe now

我想很多人都聽過CPO（Co-Packaged Optics，共同封裝光學），甚至已經開始投資CPO相關的概念和股票。

但我發現，很多人即使知道一些細節，例如這家公司是做交換器的、那家公司著重光學引擎，卻不一定真正理解背後最基本的原理。

或許在「CPO是未來趨勢、總體市場將大幅成長」的背景下，這暫時不會造成太大問題。但如果未來某些技術路線發生改變，或從新聞中聽到新的產品進度、客戶採用和量產消息，投資人可能很難判斷這些變化對持股與原有投資邏輯究竟代表什麼。

因此，我想用較淺白的方式，幫助讀者理解CPO的一些基本原理和概念。之後也會分享CPO的其他細分市場，讓大家比較有完整的脈絡。

資料離開GPU之後，去了哪裡？

要理解CPO，我們先不要從矽光子、雷射或封裝技術開始。

最簡單的方法，就是跟著一筆資料，看它如何從一顆GPU或ASIC，走到另一顆GPU或目的地晶片。

當一顆GPU完成一部分模型運算，需要把結果送給另一顆GPU繼續處理接下來的任務時，資料離開晶片後，會先變成高速電訊號。這些電訊號經過晶片封裝與基板，再沿著PCB上的銅線向外傳送。

如果目的地很近，電訊號可以直接完成傳輸。

但當傳輸速度愈來愈快、距離愈來愈長，問題便開始出現。電訊號在銅線中傳輸時會逐漸衰減、失真，也更容易受到雜訊與其他訊號干擾。

為了讓資料繼續可靠地前進，系統會使用SerDes，將晶片內部的平行資料轉換成適合高速傳輸的序列訊號，並在接收端將它還原；系統也可能加入訊號補償、Retimer或DSP，重新恢復時脈、整形訊號，並修正傳輸過程中產生的失真。

這些技術可以維持訊號品質，代價則是更高的功耗、成本與設計複雜度。

因此，資料通常不會全程使用銅線傳輸。

它走到光模組或光學引擎之後，會由電訊號轉換成光訊號，再透過光纖送到遠端。到達另一端後，光訊號重新轉成電訊號，最後交給另一顆GPU、ASIC或其他晶片繼續處理。

整條路徑就是：

GPU／ASIC產生電訊號
→ 經過封裝與PCB銅線
→ 電轉光
→ 光纖傳輸
→ 光轉電
→ 另一顆GPU／ASIC

真正的問題是：

電訊號要走多遠，才轉換成光？

CPO解決的正是這個問題。

它的目的，就是把電轉光的位置搬得更靠近晶片，讓最難處理、最耗電的高速電訊號少走一段路。

傳統光模組、LPO與CPO有什麼差別？

傳統可插拔光模組、LPO與CPO都在做電訊號與光訊號之間的轉換。

三者最核心的差別，在於：

電訊號要走到哪裡，才轉成光？

傳統可插拔光模組

在傳統架構中，資料從交換器ASIC出發後，需要先沿著PCB上的銅線走到設備前面板，再進入可插拔光模組。

光模組內的DSP會負責訊號等化、時脈恢復與其他數位訊號處理，再驅動光學元件，把電訊號轉換成光訊號。

這種設計最大的優點是成熟、標準化，而且容易維修。光模組如果損壞，可以直接從前面板拔掉並更換。

但當交換器頻寬與單通道傳輸速度愈來愈高，交換器ASIC到前面板光模組之間的這段電訊號路徑，會消耗愈來愈多功率，也需要更複雜的訊號處理。

LPO

LPO是Linear Pluggable Optics，也就是線性可插拔光學。

在LPO架構中，電訊號仍然需要從交換器ASIC沿著PCB走到前面板，因此光模組仍然是可插拔的。

它和傳統光模組的主要差異，是省略傳統光模組中的完整DSP或Retimer訊號處理路徑，改由交換器ASIC的SerDes直接驅動線性光學元件。

LPO的優點是降低DSP帶來的功耗、成本與延遲。

但它並沒有完全解決電訊號需要走到前面板的問題，而且對交換器ASIC的SerDes品質、PCB通道、光學元件以及整體系統的共同設計要求更高。

換句話說，LPO不是把光學搬到晶片旁邊，而是嘗試讓電訊號走到前面板時，不再需要經過一套完整的DSP處理。

交換器CPO

交換器CPO則更進一步。

它不再讓高速電訊號一路走到前面板，而是把光學引擎直接放到交換器ASIC旁邊，讓資料離開交換器ASIC後，只經過一小段電訊號路徑，就完成電轉光。

接下來，資料主要透過光纖走到前面板，再連接到外部設備。

這種架構的優點，是可以縮短高速電訊號的傳輸距離，降低功耗，並提高交換器的頻寬密度。

但代價是封裝、散熱、測試與維修都會變得更複雜。

原本光模組壞掉，可以直接從前面板拔掉。現在光學引擎與交換器ASIC高度整合，一旦出現問題，維修成本與系統設計難度都可能上升。

雷射光源、光纖連接方式、封裝良率與現場維修模式，也都需要重新設計。

這也是為什麼CPO不會立刻取代所有可插拔光模組。

它最可能先出現在頻寬、功耗與前面板密度壓力最大的高階AI網路，再逐步向其他市場擴張。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

如果光學不只進入交換器，還能去哪裡？

交換器CPO解決的，是交換器ASIC到前面板光模組之間的問題。

但AI系統還有另一個更靠近運算端的瓶頸：GPU與GPU之間的scale-up互連。

Scale-up的目的，是讓多顆GPU或XPU協同運作，看起來像一個更大的運算系統。這類網路通常要求更高的頻寬、更低的延遲，以及更緊密的記憶體和運算資源協同。

目前，這些連接主要依賴高速電訊號，例如NVLink或其他銅線互連。只要GPU位於同一塊板子、同一台伺服器，或距離非常接近的機架內，銅線仍然非常有效率。

但銅線的傳輸距離有限。

當系統設計者希望將scale-up網路從單一伺服器或單一機架，擴展到更多機架時，電訊號就必須走得更遠。這不只會提高功耗，也會增加訊號完整性、散熱與系統設計的難度。

而XPU Optical I/O或XPU CPO的構想，就是把光學再往前移一步。

資料離開XPU後，只經過一小段高速電訊號路徑，便轉換成光，接著透過光纖直接連接其他XPU，或先進入scale-up switch及其他fabric元件，再送往另一個機架中的XPU。

路徑因此可能變成：

XPU
→ 超短電訊號
→ XPU旁的光學引擎
→ 光纖
→ Scale-up switch或另一顆XPU

這裡需要注意，XPU使用光學I/O並不代表系統不再需要交換器。光學主要改變的是資料傳輸的媒介與電轉光的位置。資料仍可能需要透過scale-up switch、fabric switch或其他網路元件，決定應該送往哪一顆XPU。

如果這個架構成熟，scale-up網路便能降低對銅線傳輸距離的依賴，更多機架與更多加速器也能被納入同一個高速互連域。但這也是比交換器CPO更困難的技術。

交換器本來就是負責資料轉送的網路元件，因此把光學整合進交換器，有相對清楚的產品邏輯。

XPU則是功耗與熱密度極高的運算晶片。把光學引擎、封裝與運算晶片放得更靠近，會帶來更大的散熱、良率、測試與可靠性挑戰。

實際產品還必須處理雷射光源、封裝內光纖連接、光學元件壽命，以及某個元件損壞後如何更換等問題。

因此，XPU Optical I/O代表更大的潛在市場，也代表更長的商業化時間與更高的不確定性。

三家公司都走向光學，但出發點不同

理解Broadcom、NVIDIA和Marvell的差異，可以先想像一座正在擴建的AI工廠。

工廠裡已經有愈來愈多GPU與AI加速器，每一台機器都能快速運算，但它們之間必須持續交換資料。

當機器數量愈來愈多，真正限制工廠效率的，不再只是單一GPU跑得多快，而是資料能不能快速、穩定地在所有機器之間移動。

三家公司都看到了這個問題，也都認為光學互連會成為重要的解決方案。但它們站在AI工廠裡的不同位置，看到的機會也不一樣。

Broadcom (AVGO)

Broadcom原本就是資料中心交換器ASIC市場的領先者，因此它的發展邏輯很直接：

先利用已經掌握的交換器市場，讓CPO在交換器端落地，再把同一套SerDes、矽光子與封裝能力延伸到客製化XPU。

Broadcom的優勢，是它本來就控制交換器ASIC與高速SerDes，也已經擁有大型雲端客戶和客製化ASIC業務。

對Broadcom而言，CPO不是重新建立一個市場，而是在原本已經掌握的平台上，增加更多光學內容。

NVIDIA (NVDA)

NVIDIA站的位置不同。

它是從整座AI工廠開始思考。

對NVIDIA而言，光學本身不是終點。真正的目標，是讓更多GPU可以組成更大的叢集，同時降低網路功耗並提高可靠性。

因此，NVIDIA將矽光子與CPO導入Quantum-X InfiniBand和Spectrum-X Ethernet交換器。NVIDIA發展光學交換器，不只是為了銷售一個新的光學零組件，而是為了讓GPU、NVLink、InfiniBand、Ethernet交換器和軟體平台一起變得更有價值。

網路規模愈大、GPU數量愈多，NVIDIA整套AI平台的價值就愈高。

Marvell (MRVL)

Marvell的出發點，則是高速DSP、PAM4、矽光子與Inphi累積的電光互連技術。

它希望同時參與交換器CPO和XPU Optical I/O：

在交換器旁整合光學；

在客製化XPU旁整合光學；

最後再透過收購Celestial AI取得的Photonic Fabric技術，把更多運算資源連接成更大的scale-up網路。

三家公司最白話的差異是：

Broadcom：以交換器CPO為近期落地基礎，同時將光學能力逐漸延伸至客製化XPU。

NVIDIA：用光學交換器擴大GPU叢集，強化整套AI平台的價值。

Marvell：從高速光互連與客製化晶片能力出發，同時向交換器、XPU與scale-up fabric延伸。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Broadcom的故事，是先控制交通樞紐，再延伸到運算端

Broadcom的Tomahawk與Jericho交換器ASIC，原本就處於AI網路中的重要位置。

在傳統架構中，Broadcom主要負責交換器ASIC。當資料離開交換器後，電光轉換通常由外部可插拔光模組完成。

CPO改變了這個價值分配。

如果光學引擎直接放到交換器ASIC旁邊，Broadcom就不再只是出售交換器ASIC，而是可以連同資料離開交換器後的部分光學傳輸功能一起提供。

這讓Broadcom從原本的：

交換器ASIC供應商

進一步變成：

交換器ASIC＋SerDes＋光學引擎＋CPO封裝平台供應商

而這條路徑最重要的地方，是它已經非常接近實際商業化，而且目前的產品進度相對明確。

Broadcom本來就有交換器客戶，也控制交換器ASIC和高速SerDes。因此，它不必等待其他交換器晶片公司採用自己的光學元件，而是可以直接在自己的交換器平台上完成整合。

Broadcom已經推出多代交換器CPO平台，並將光學引擎與Tomahawk交換器ASIC共同封裝。最新一代產品也持續提高交換器頻寬與單一光學引擎的傳輸能力。

這也是Broadcom與許多單純提供光學零組件公司的主要差異。它不只是在等待CPO市場形成，而是本身就掌握一個可以導入CPO的交換器平台。

同時，Broadcom也參與大型Hyperscaler的客製化AI加速器設計。這代表它有機會在設計XPU時，就把光學I/O納入整體架構。

Broadcom已經展示將6.4T光學引擎連接到客製化加速器的技術。資料離開客製化XPU後，只需要經過很短的電訊號路徑，便可以進入光學引擎，再透過光纖建立更大範圍的scale-up fabric。

不過，投資人仍然需要區分：

技術展示、客戶設計導入，以及大規模量產，是三個不同階段。

Broadcom已經證明它具備把光學能力從交換器延伸到客製化加速器的技術路徑，但實際的收入貢獻，仍取決於客戶採用、產品驗證和量產時間。

Broadcom的CPO投資命題因此可以分成兩段：

第一段，是較接近落地的交換器CPO，觀察它能否提高每台交換器的內容價值，並逐步取代部分原本由外部可插拔光模組提供的功能。

第二段，則是XPU CPO，觀察Broadcom能否利用客製化ASIC業務，把光學進一步帶進Hyperscaler的AI加速器與scale-up系統。

Marvell的故事，是先掌握光互連，再把拼圖一路向系統各處推進

Marvell的主要出發點和Broadcom不同。

它並不是以市場領先的交換器ASIC平台作為CPO商業化的核心，而是先累積高速DSP、PAM4、矽光子與電光互連能力，再逐步向交換器與XPU整合。

不過，這並不代表Marvell沒有交換器產品或交換器CPO技術。

Marvell本身擁有Teralynx交換器ASIC，也早已發展交換器CPO相關方案。

真正的差別是，Broadcom可以依靠龐大的既有交換器平台推動CPO，而Marvell更依賴光學技術、客製化晶片能力與不同互連架構的彈性，爭取進入下一代AI系統。

在AI產業開始大量討論CPO以前，資料中心就已經需要處理高速電訊號。當訊號速度愈來愈快，如何維持訊號品質、進行PAM4處理，並可靠地完成電光轉換，早已是一個重要市場。Marvell收購Inphi後，進一步強化了PAM4 DSP、coherent DSP、高速類比與電光互連能力。

這代表Marvell先掌握的是：

如何讓高速資料可靠地在電訊號與光訊號之間轉換。

接著，Marvell開始把這些能力從傳統可插拔光模組，往晶片封裝旁移動。

這套光學能力可以分成兩條主要路徑。

交換器CPO

Marvell可以將矽光子光學引擎放到交換器ASIC旁，縮短高速電訊號的傳輸距離。

在基本架構上，這與Broadcom的交換器CPO相似。

Marvell也能結合自己的交換器晶片、SerDes、DSP與矽光子技術，提供交換器CPO方案。不過，Marvell目前並不具備Broadcom在交換器ASIC市場的相同平台優勢，因此能否擴大交換器CPO業務，仍取決於客戶採用與整體生態系合作。

客製化XPU CPO

Marvell也擁有客製化ASIC能力，因此可以把光學引擎整合到客戶的AI加速器旁，讓資料更早轉換成光，再透過光纖及scale-up switch建立更大的scale-up網路。

Marvell已經展示6.4T 3D SiPho光學引擎，並公布把光學引擎、XPU運算晶片、HBM與其他chiplet整合在同一基板上的客製化加速器架構。

但和Broadcom一樣，投資人仍需要區分技術展示、客戶設計導入與大規模量產三個不同階段。

目前這些產品證明Marvell已經擁有進入XPU CPO市場的技術能力，但不代表相關收入已經大規模發生。

Photonic Fabric：從光學引擎走向完整scale-up網路

Marvell的故事還有下一個階段：

透過收購Celestial AI取得的Photonic Fabric。

Marvell已於2026年2月完成Celestial AI的收購。Celestial AI的Photonic Fabric是一套針對AI scale-up互連設計的光學技術平台，目標是支援封裝、系統與機架層級的高速連接。

這一步的想像，不只是「在XPU旁邊放一個光學引擎」。它希望透過光學chiplet與互連技術，把客製化XPU、scale-up switch、記憶體和其他運算資源連接起來。可以把它想成，Marvell希望從提供高速道路中的零件，逐步參與設計整個交通網路。

如果Photonic Fabric成功進入Hyperscaler的scale-up架構，Marvell參與的價值就可能不再局限於單一DSP或光學引擎，而是擴展到整套AI加速器互連系統。

但這也是Marvell投資邏輯中不確定性最高的部分。

Photonic Fabric仍需要經過客戶驗證、系統整合與量產爬坡，才能證明它不只是技術上可行，也能轉化成具有規模的商業收入。

Marvell的主要投資觀察可以整理成一句話：

從高速光互連出發，透過Inphi、自研矽光子、客製化ASIC與Celestial AI的Photonic Fabric，觀察其應用深度及廣度能否逐步向交換器CPO和XPU互連兩端擴張。

Broadcom與Marvell：同樣投資CPO，實際押注並不相同

對於看好CPO是未來發展趨勢，並考慮投資Broadcom和Marvell的投資人來說，兩家公司CPO投資邏輯最本質的差異是：

Broadcom是在既有交換器與客製化ASIC平台上增加光學內容；Marvell則是在利用高速光互連、矽光子與客製化晶片能力，爭取成為下一代AI互連平台的一部分。

以目前的產品進度與市場地位來看，Broadcom的確定性相對較高。它可以先依靠交換器CPO提高單一系統的內容價值，再等待XPU CPO逐步商業化。即使XPU光學I/O的發展速度比預期慢，Broadcom仍有交換器ASIC、客製化ASIC與現有AI網路業務作為基礎。

不過，由於Broadcom公司體量較大，單一CPO趨勢對整體營收與估值的影響可能相對有限。

Marvell提供的則是另一種路徑。

它在可插拔DSP、LPO、交換器CPO、客製化XPU與Photonic Fabric等不同架構上都有布局。如果其中一項或多項技術進入大規模採用，對Marvell營收的影響可能更明顯。但相對而言，它也伴隨更高的技術、客戶採用與量產時間風險。

由於XPU CPO與Photonic Fabric仍在早期階段，投資人應持續關注三項訊號：

第一，是否出現具名或可信度較高的客戶採用。

第二，產品是否從技術展示進入正式設計導入與量產。

第三，CPO相關收入是否開始對公司營收與毛利率產生可衡量的影響。

在這些訊號更清楚以前，Broadcom比較接近：在既有平台上增加光學內容價值。

Marvell則比較接近：以較高的不確定性，換取成為下一代AI互連平台一部分的成長彈性。

CPO並不是單一產品，也不是所有公司都在押注同一條技術路線。

當光學逐步從前面板走向交換器，再從交換器走向XPU，價值鏈中的受益者、競爭優勢與商業化時間都可能發生變化。

真正重要的，不只是知道哪些公司「有CPO」。

而是理解：光學正在往哪裡移動、誰控制那個位置，以及這項技術何時能從展示轉化成實際收入。

Subscribe now

如果你喜歡這種從產業發展、商業邏輯到個別公司細節的分析，歡迎訂閱，不錯過之後的文章。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

CPU為什麼在AI時代重新變得重要？

過去幾年，關於AI基礎建設的討論幾乎都圍繞著GPU。這並不難理解，因為不論是模型訓練或大規模推論，最主要的矩陣運算仍然由GPU與其他加速器負責完成。

但隨著AI從單純的模型訓練，逐漸轉向由AI Agent主導的推論與任務執行，CPU在整套系統裡的重要性也開始上升。

CPU重要的原因在於，AI Agent需要拆解任務、查詢資料庫、執行程式碼、操作外部工具、檢查結果，再決定下一步，並反覆執行。更簡單來說，agentic workflow就是由CPU負責準備和安排工作，GPU負責模型運算。一個任務能否有效率地完成，不再是由GPU說了算。這也說明為什麼NVIDIA近來更強調Vera CPU，而不只是Rubin GPU。

對資料中心營運商來說，真正重要的也不會只是單顆晶片的峰值效能，而是整個機櫃每瓦能完成多少任務，以及GPU能否維持高使用率。這使CPU重新成為AI系統設計中重要的一環，而不單單只是GPU旁邊的輔助晶片。

Intel與AMD的x86內部競爭

既然CPU重要，接下來的問題就是：現有市場的競爭格局會如何改變？

在x86市場裡，Intel長期掌握著個人電腦與伺服器市場的主導地位。除了處理器本身，Intel還擁有完整的OEM上下游關係、企業認證、軟體最佳化與伺服器平台生態系。值得注意的是，對企業和伺服器客戶來說，由於轉換平台需要考慮驗證、穩定性與相容性等問題，轉換成本也會較高。這說明一旦客戶願意轉換，背後通常不是短期價格促銷，而是產品差距已經大到足以抵銷遷移成本。

所以AMD能夠蠶食Intel的市占，並不只是因為價格較低。更重要的是，AMD在Zen架構、chiplet設計與台積電製程的配合下，持續提高核心數、記憶體頻寬與每瓦效能。對雲端服務商與大型資料中心來說，這些差異會直接影響一台伺服器或一個機櫃可以承載多少工作負載。這也是AMD EPYC能夠逐步進入原本由Intel高度主導的伺服器市場的重要原因。

根據Mercury Research的資料，AMD在2026年第一季已取得約33.2%的x86伺服器CPU出貨市占，營收市占則達到46.2%。營收市占高於出貨市占，代表AMD已經不只是依靠低價搶量，也逐漸進入價值較高的伺服器配置。

不過，Intel最新一季的表現也顯示，市場並不是單向發展。Intel第二季Data Center and AI (DCAI) 業務營收年增59%至約63億美元，伺服器CPU需求明顯加速。

這裡需要特別區分兩種可能性。

第一種是整體市場快速擴張。AI推論與Agentic AI增加了CPU需求，即使Intel沒有奪回市占，也可能因為市場變大而錄得強勁成長。但這也說明AMD同樣受益，而且程度可能更大。

第二種才是Intel產品競爭力改善，使客戶重新增加Xeon採購。

目前僅靠單季營收，還無法完全判斷是哪一種情況。後續追蹤AMD的財報，或許會有答案。

就目前x86市場的發展來看，AMD持續搶占高價值市占，同時AMD與Intel也都受惠於AI帶動的整體市場回升，這兩個方向是相對明確的。

Arm-based CPU為何會興起？

既然Intel與AMD能提供成熟的伺服器CPU，為什麼Arm-based CPU仍然開始興起？

Arm最重要的優勢，並不是單純的「RISC比x86省電」，而是它提供了一套可以授權與客製化的架構。

Intel與AMD主要向市場出售通用CPU。客戶必須從兩家公司提供的產品中選擇最適合的產品。Arm的模式則是讓AWS、Google、Microsoft、Apple與NVIDIA等公司，可以根據自己的需求設計CPU，再將CPU與自身的軟體、網路、加速器與資料中心架構進行整合。

AWS Graviton是一個典型案例。

AWS需要針對EC2上的常見工作負載，例如Web server、微服務、開源資料庫、快取與容器環境進行最佳化，因此可以單純依照這些需求設計晶片。AWS宣稱，Graviton執行個體相較於可比較的x86 EC2執行個體，價格最多可降低20%，能源使用最多可降低60%。

雖然這是AWS在特定比較條件下的官方說法，但這其實還說明Graviton的重要性不只在於價格和效能，也代表AWS可以藉此同時控制晶片、雲端服務、定價方式與客戶入口。

Google的Axion與Microsoft的Cobalt也沿著相同方向發展。Google表示，Axion可針對Java、MySQL等通用工作負載提供更高的價格效能；Microsoft也已將自研的Cobalt 100部署到多個Azure區域。

更重要的是，雲端服務商還可以把部分架構差異隱藏在託管服務之下。

假設客戶直接管理虛擬機器，從x86轉移到Arm可能需要重新編譯程式、替換不相容的函式庫，並重新測試效能。但若客戶使用的是託管資料庫、容器服務或Serverless平台，底層CPU由雲端服務商管理，客戶甚至不一定會明確感受到架構差異。這降低了Arm的導入門檻，也改變了CPU市場的銷售方式。

不過，Arm也不是在所有工作負載上都更有優勢。

一些高度依賴x86指令集、AVX-512向量運算、專有軟體或既有企業系統的工作負載，仍可能在Intel或AMD平台上表現得更好。企業若擁有大量舊程式、商業軟體授權與特定硬體最佳化，遷移成本很可能高於節省下來的運算成本。

小結，Arm-based CPU比較容易在雲端原生、開源軟體、容器化與大規模水平擴展的工作負載中成長，而在許多現有工作環境下，仍無法直接取代x86伺服器。

Subscribe now

開始免費訂閱，不錯過後續分析與更多有價值的內容。

x86與Arm的市場細分與長期方向

從長期來看，x86與Arm不會在所有市場正面交戰。兩者更可能逐漸形成不同的市場定位。

x86走向通用性與相容性

x86最大的優勢仍然是龐大的既有軟體與硬體生態系。

大量企業應用、資料庫、虛擬化平台與內部系統，都是在x86環境下開發與驗證。對很多客戶來說，CPU價格只是總成本的一部分。系統穩定性、認證、技術支援與遷移風險往往更重要。

因此，x86未來仍可能在以下領域保持優勢：

• 需要廣泛軟體相容性的通用伺服器

• 高效能運算與特定向量運算工作負載

• 必須同時支援大量不同客戶軟體的公有雲服務

• AI伺服器中的通用主機CPU需求

目前由AI帶動的需求提升會持續發酵，進而讓AMD和Intel同時受惠。而AMD只要能維持核心密度、每瓦效能與產品路線圖的執行力，就可能持續搶占Intel的市場份額，取得更大的增長。

Intel仍然擁有規模、生態系與客戶關係優勢。而它需要證明的不只是單季需求反彈，而是能否穩定推出具有競爭力的產品，並改善製程、供貨與平台執行。

Arm走向客製化與垂直整合

Arm未來成長的領域，是能夠針對特定工作負載設計整套系統的市場。

這包括：

• AWS、Google與Microsoft的自研雲端CPU

• NVIDIA等通用晶片公司的CPU與加速器整合

• Apple、Qualcomm與其他PC廠商的Arm PC

• 網路、儲存與邊緣運算

• 汽車、機器人與嵌入式系統

• 需要高度能源效率的大規模水平擴展服務

在這些市場裡，CPU不會是一個獨立採購的零件，而是整個平台架構設計的一部分。

例如，NVIDIA設計Vera CPU，不是為了單獨與Xeon或EPYC競爭，而是希望CPU、GPU、NVLink、網路與軟體能共同構成一個最佳化的AI機櫃。當所有環節都經過共同設計，就算Vera CPU單獨比較時未必在所有效能指標上領先，整體機櫃產出仍可能為NVIDIA帶來更大的優勢。

這同時帶出「系統控制權」逐漸重要的原因。

Intel與AMD出售CPU，主要賺取晶片本身的價值。AWS、Google、Microsoft與NVIDIA則可能透過Arm架構，把CPU變成強化自身雲端或AI平台的工具。即使自研CPU本身沒有產生獨立營收，只要它能降低資料中心成本、提高GPU利用率或增加客戶黏著度，就能創造更大的平台價值。

幾個未來觀察指標

未來判斷Intel、AMD與Arm-based CPU的競爭格局，關注新產品發布或單一效能測試可能不再那麼重要。個人認為，以下的長期指標更值得關注，可以讓你知道發展方向是否如預期，在趨勢改變時能採取行動。

Arm雲端執行個體的實際使用率

Arm在雲端服務商的執行個體占總運算時數、客戶支出與新增工作負載的比例是否持續提升。

軟體原生支援與遷移成本

主流作業系統、資料庫、容器、開發工具與商業軟體是否能原生支援Arm。愈多工作負載能在不修改程式的情況下遷移，Arm的市場空間就會愈大。

CPU與GPU或加速器的綁定程度

若CPU與GPU之間的記憶體、互連與軟體排程愈來愈緊密，市場競爭單位就會從單顆CPU轉向整個AI平台。

這對Intel與AMD既是機會，也是風險。AMD可以利用同時擁有EPYC與Instinct的優勢發展整機系統；Intel則需要讓Xeon、加速器、網路與先進封裝形成完整平台。另一方面，NVIDIA也能透過Vera與Rubin的綁定，降低AI系統對x86 CPU的依賴。

AMD能否繼續取得Intel的高價值伺服器市占

AMD的伺服器CPU營收市占已經明顯高於出貨市占。未來若這個差距持續擴大，代表AMD不只是搶到更多CPU數量，也進入更多高階與高價值配置。

Intel能否在產品與製造執行上縮小差距

Intel真正的問題不是缺乏客戶或生態系，而是過去產品與製程的執行不夠穩定。若Intel能夠落實產品路線圖、改善供應並提供具有競爭力的產品，就仍有機會利用既有的生態系優勢守住市場。

Hyperscaler自研CPU是否擴展至外部企業工作負載

自研CPU先在雲端服務商內部使用並不困難，更大的考驗是，外部企業客戶是否願意把資料庫、企業應用與關鍵系統部署到這些自研CPU上。

如果Arm架構只侷限於雲端服務商的內部或特定服務，那麼它主要影響的是Intel與AMD對hyperscaler的出貨。但如果Arm執行個體更多被外部企業廣泛採用，那它改變的就會是整體伺服器CPU市場。

現階段結論

目前市場主流觀察如下面三張圖所示，但市場是動態的，當發現趨勢加速或轉向，或許就會是獲取增額報酬的機會。

中文圖:

Subscribe now

如果你喜歡這種從產業發展、商業邏輯到個別公司細節的分析，歡迎訂閱，不錯過之後的文章。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

Disclaimer: For informational purposes only. Not financial advice. Invest at your own risk.

Subscribe now

This article looks beyond Tesla’s delivery rebound to identify the financial and operating signals that will show whether its AI story is developing a sustainable commercial foundation. 這篇文章將帶你跳脫交付量的表面反彈，掌握哪些財務與營運訊號，才能真正判斷特斯拉的 AI 故事是否開始具備可持續的商業基礎。中文版請見文末。

Tesla’s second-quarter earnings matter less as a verdict on delivery recovery than as a test of whether that recovery produced durable economics that can fund the company’s AI ambitions.

The delivery numbers were strong: Tesla reported 480,126 vehicle deliveries in Q2, up from 358,023 in Q1, alongside production of 451,758 vehicles. It also said energy-storage deployments reached a record 13.5 GWh. That is enough to show the business can still move volume at scale. It is not enough to answer the more important investor question: are Tesla’s existing businesses generating the margins, cash flow, and operating evidence needed to support a valuation that increasingly leans on robotaxis, autonomy, and robotics?

That is the real lens for this quarter.

Tesla is no longer being judged mainly as an EV manufacturer. The story investors are being asked to underwrite is much bigger: autonomous transport, a robotaxi network, Optimus, and eventually a broader physical AI platform built on Tesla’s software, hardware, and manufacturing stack. But the capital supporting that transition still comes from far more ordinary places, mainly cars, energy storage, and related services.

So the key issue is not whether Tesla is an auto company or an AI company. It is whether the businesses that exist today are healthy enough to finance the ones that are still mostly future tense.

That is why the delivery rebound is not the real headline. Tesla has long warned that deliveries are not a direct read-through to quarterly financial performance. Average selling prices, incentives, leasing mix, foreign exchange, production costs, and inventory movements all matter.

A good quarter is not simply one in which more vehicles left the lot. It is one in which higher volume translated into better economics. Because Q2 delivery volume is already known, the market reaction to earnings is likely to come from the quality of that volume, not the volume itself.

The most useful place to start is automotive margin quality. Investors should look past headline delivery growth and focus on revenue per vehicle, gross margin excluding regulatory credits, and free cash flow.

Revenue per vehicle is a pricing-and-mix signal: it helps show whether Tesla protected economics or leaned harder on discounts and lease support to move units. Gross margin excluding regulatory credits is a cleaner operating measure because it says more about the underlying vehicle business than a volatile revenue line that is not directly tied to core unit economics. Free cash flow is the funding test. If Tesla wants investors to keep backing a larger AI story, it has to show that the current business is still producing real cash to support it.

Q1 is part of what makes this quarter harder to read. Tesla’s total automotive gross margin improved year over year, but management also pointed to one-time warranty and tariff-related benefits. That means Q2 has to separate what was structural from what was temporary.

Subscribe for research that looks beyond the headline—into the business, the expectations, and what the market may be missing.

Subscribe now

If margins stay resilient without those benefits doing most of the work, the recovery starts to look more durable. If they slip once those temporary supports fade, then Q2 may end up looking more like a rebound in activity than an improvement in business quality. That distinction matters because higher deliveries can still disappoint if average selling prices fell, incentives rose, or incremental gross profit was absorbed elsewhere.

The reason this matters so much is that Tesla’s AI narrative is getting expensive. Even if investors stay optimistic about robotaxis and Optimus, those projects do not fund themselves. Faster progress in AI could be positive for long-term value while still pressuring near-term profits and cash flow.

If Tesla is spending more on compute, engineering talent, deployment operations, and commercialization before meaningful new revenue arrives, the burden on the legacy business gets heavier, not lighter. That tightens the chain of cause and effect in this report: automotive and energy cash flow fund AI investment, and if that base weakens, the AI story becomes less self-financing.

That is also why the robotaxi discussion matters less as spectacle than as operating proof. Investors should be careful not to confuse an expansion story with a business model. Announcing a larger service area or new geography may sound important, but geography alone does not tell you whether the economics work.

What matters more is whether Tesla discloses the kind of metrics that indicate commercial viability: how many vehicles are actually giving paid rides, how many trips each vehicle completes per day, how long riders wait, how often remote support intervenes, what weather or road-type restrictions remain, and how quickly the service is moving from pilot status toward something repeatable and scalable.

That distinction between software capability and business viability is crucial. An autonomous driving system and a profitable robotaxi network are not the same thing. Even if the software can safely handle most trips, Tesla still has to solve the operating layer: charging, cleaning, maintenance, repositioning, customer support, insurance, and demand imbalances across time and geography.

The bottleneck may turn out to be operations rather than intelligence. That is why robotaxi operating data matters so much. Without it, investors are left evaluating technological promise without enough evidence about the business wrapped around it.

Energy may be more important in that picture than many investors appreciate. Tesla’s energy-storage deployments rose to 13.5 GWh in Q2 from 8.8 GWh in Q1, suggesting the segment could become a larger contributor to revenue and gross profit. If energy margins remain strong, or at least healthy after any normalization, that business could become a financial bridge between Tesla’s current earnings power and its future autonomy and robotics ambitions.

In other words, if automotive economics hold and energy scales profitably, Tesla has a more credible way to fund AI expansion without leaning entirely on investor belief.

There is also a more subtle risk on the earnings call: insufficient disclosure. Tesla may talk confidently about learning curves, fleet growth, and future timelines without providing the operating data investors need to value progress properly. That would leave bulls and bears free to tell very different stories from the same presentation.

For a company whose valuation increasingly reflects autonomy and robotics expectations, the absence of measurable operating detail is not a side issue. It is a valuation risk.

Tesla does not need to prove the full robotaxi or Optimus thesis in one quarter. That would be an unreasonable standard. But it does need to prove something more concrete than improved deliveries.

The strongest result would show resilient automotive margins without excessive incentive pressure, healthy revenue per vehicle, solid free cash flow, meaningful support from energy, and at least some measurable evidence that autonomy is moving from demonstration toward commerce.

The weakest result would be the opposite: high deliveries, softer pricing, thinner margins once temporary benefits fade, rising expenses, and little real disclosure on robotaxi operations.

That is the real test of this quarter. Tesla’s AI opportunity may still be substantial, but the market now needs evidence that today’s businesses are strong enough to fund that story and that the AI businesses are beginning to produce real-world proof in return for the capital they are consuming.

Subscribe now

Subscribe for free and never miss future analysis and other valuable insights.

特斯拉第二季財報的重要性，與其說是檢驗交付量是否回升，不如說是要確認這波反彈能否轉化為可持續的經濟效益，並為公司持續投入 AI 願景提供足夠支撐。

交付數據確實亮眼。特斯拉第二季共交付 480,126 輛汽車，明顯高於第一季的 358,023 輛，產量則為 451,758 輛。公司同時宣布，能源儲存產品部署量達到創紀錄的 13.5 GWh。這些數字足以證明，特斯拉仍具備大規模生產與銷售產品的能力。

但它們仍無法回答投資人更重要的問題：特斯拉現有業務是否正在創造足夠的利潤、現金流與實際營運成果，支撐一個愈來愈依賴 Robotaxi、自動駕駛與機器人願景的估值？

這才是本季財報真正值得關注的角度。

如今，市場已不再主要將特斯拉視為一家電動車製造商。投資人被要求相信的故事遠比汽車更大，包括自動駕駛運輸、Robotaxi 網路、Optimus 人形機器人，以及最終建立在特斯拉軟體、硬體與製造能力之上的實體 AI 平台。

然而，支撐這場轉型所需的資金，目前仍主要來自相對傳統的業務，包括汽車、能源儲存與相關服務。

因此，關鍵問題並不是特斯拉究竟是一家汽車公司，還是一家 AI 公司，而是今天已經存在的業務是否足夠健康，能夠為那些目前仍大多停留在未來藍圖中的業務提供資金。

這也是為什麼交付量反彈並不是真正的重點。特斯拉長期以來一直提醒投資人，交付量並不能直接反映單季財務表現。平均售價、折扣與促銷、租賃占比、匯率、生產成本和庫存變化，都會影響最終結果。

一個真正好的季度，不只是有更多車輛交到消費者手中，而是更高的銷量能否轉化為更好的獲利與現金流。由於第二季交付數據早已公布，市場對財報的反應，更可能取決於這些銷量的「品質」，而不是銷量本身。

首先值得關注的，是汽車業務的利潤率品質。投資人應該跳脫表面的交付成長，將焦點放在單車營收、排除監管積分收入後的毛利率，以及自由現金流。

單車營收反映的是產品定價與銷售組合。透過這項指標，投資人可以判斷特斯拉究竟守住了每輛車的經濟效益，還是更依賴降價、優惠或租賃方案來推動銷量。

排除監管積分收入後的毛利率，則更能反映汽車本業的真實營運狀況。監管積分收入波動較大，也無法直接代表核心汽車業務的單位經濟效益。

自由現金流則是最直接的資金能力測試。如果特斯拉希望投資人持續支持更宏大的 AI 故事，就必須證明現有業務仍能產生足夠的實際現金，支撐後續投資。

Subscribe now

第一季的表現，也讓第二季更難解讀。特斯拉整體汽車毛利率較去年同期改善，但管理層同時指出，其中包含一次性的保固與關稅相關正面影響。因此，第二季的關鍵之一，就是分辨哪些改善具有結構性，哪些只是短期因素。

如果在這些一次性效益減弱後，利潤率仍能維持韌性，這波復甦才更可能具有持續性。

反過來說，如果暫時性支撐消退後，利潤率便隨之下滑，那麼第二季可能只代表營運活動回溫，而不是真正的商業品質改善。

這項區別相當重要。即使交付量增加，如果平均售價下降、促銷成本上升，或新增的毛利被其他支出吸收，最終財報仍可能令市場失望。

之所以如此重要，是因為特斯拉的 AI 故事正變得愈來愈昂貴。

即使投資人仍看好 Robotaxi 與 Optimus 的長期潛力，這些計畫也不會自行產生資金。AI 發展加速，或許有利於長期價值，卻也可能在短期內壓縮獲利與現金流。

如果特斯拉在新業務尚未帶來明顯收入之前，就必須持續增加對運算資源、工程人才、服務部署與商業化的投入，那麼現有業務承擔的壓力不會減輕，反而會進一步加重。

這也讓本季財報背後的因果關係更加清楚：汽車與能源業務產生的現金流負責支撐 AI 投資；一旦這個基礎轉弱，特斯拉的 AI 發展就會更難依靠自身資金推進。

這也是為什麼投資人應該將 Robotaxi 視為一場營運能力的考驗，而不只是一次技術展示。

投資人不應把服務範圍擴大，直接解讀為商業模式已經成立。宣布擴大服務區域或進入新的城市，聽起來或許很重要，但地理覆蓋範圍本身並不能證明這項業務具有可行的經濟效益。

真正重要的，是特斯拉是否願意公布能夠反映商業可行性的數據，例如實際提供付費乘車服務的車輛數量、每輛車每天完成的行程數、乘客等待時間、遠端支援介入頻率、仍存在的天候或道路限制，以及服務正以多快的速度從試點階段走向可複製、可擴張的營運模式。

軟體能力與商業可行性之間，存在一項非常重要的差別。

一套能夠運作的自動駕駛系統，並不等同於一個可以獲利的 Robotaxi 網路。即使軟體已能安全處理大多數行程，特斯拉仍必須解決整個營運層面的問題，包括充電、清潔、維修、車輛調度、客戶服務、保險，以及不同時間與地區之間的供需不平衡。

最終的瓶頸，未必是人工智慧本身，也可能是營運能力。

這正是 Robotaxi 營運數據如此重要的原因。缺乏這些資訊時，投資人只能評估技術潛力，卻沒有足夠證據判斷圍繞在技術之外的商業模式是否可行。

在這個架構下，能源業務的重要性可能高於許多投資人的想像。

特斯拉第二季能源儲存部署量，從第一季的 8.8 GWh 增加至 13.5 GWh，顯示這項業務未來可能對營收與毛利做出更大貢獻。

如果能源業務的毛利率能持續保持強勁，或即使經歷正常化後仍維持健康水準，它就有機會成為連接特斯拉當前獲利能力與未來自動駕駛、機器人願景之間的一座財務橋梁。

換句話說，如果汽車業務的經濟效益能夠守住，同時能源業務也能在維持獲利的前提下持續擴張，特斯拉就會擁有更可信的 AI 資金來源，而不必完全依賴投資人對未來的信念。

財報電話會議還存在另一項較不明顯的風險：資訊揭露不足。

特斯拉可能會充滿信心地談論學習曲線、車隊成長與未來時程，卻未提供投資人真正需要的營運數據，以合理評估這些進展究竟具有多少價值。

如此一來，多方與空方仍然可以從同一場簡報中，解讀出完全不同的故事。

對一家估值愈來愈反映自動駕駛與機器人預期的公司來說，缺乏可量化的營運細節並不是次要問題，而是一項實質的估值風險。

特斯拉當然不需要在一個季度內證明完整的 Robotaxi 或 Optimus 投資論點。這樣的要求並不合理。

但它確實必須拿出一些比交付量改善更具體的成果。

最理想的財報結果，應該包括汽車利潤率保持韌性，且沒有過度依賴促銷；單車營收維持健康；自由現金流穩健；能源業務提供實質支撐；同時至少提出部分可量化的證據，顯示自動駕駛正從技術展示逐步走向商業化。

最疲弱的結果則恰好相反：交付量很高，但售價走弱；暫時性利多消退後，利潤率隨之下降；費用持續增加；同時公司仍未提供 Robotaxi 營運的實質資訊。

這才是本季財報真正的考驗。

特斯拉的 AI 機會或許仍然非常龐大，但市場現在需要看到更具體的證據：今天的業務是否足夠強健，能夠持續為這個故事提供資金；而 AI 業務在消耗大量資本的同時，是否也開始交出可衡量的真實營運成果。

Subscribe now

Disclaimer: For informational purposes only. Not financial advice. Invest at your own risk.

Subscribe now

Investors often talk about CoWoS as if it were one vertically integrated packaging business that TSMC should simply scale and own end to end. That framing misses the real economic choice. The mistake is not just technical. It can lead investors to overvalue packaging revenue itself and undervalue the role packaging plays in protecting TSMC’s much larger wafer economics.

TSMC is not trying to maximize every dollar of packaging sales. It is deciding which parts of advanced packaging are strategic enough to keep, and which parts can be outsourced without giving up the value that matters most.

That distinction matters because CoWoS is increasingly central to AI infrastructure. In simple terms, CoWoS is TSMC’s advanced packaging technology for combining high-performance logic chips with high-bandwidth memory, or HBM, in a tightly integrated package. HBM matters because AI accelerators need to move enormous amounts of data quickly and with reasonable power efficiency.

In other words, packaging is no longer a minor back-end step. It is part of what makes modern AI systems possible.

Even so, CoWoS should not be treated as one uniform activity. At a high level, it can be simplified into two broad stages. First is Chip-on-Wafer, or CoW, where the logic dies and HBM are integrated on an advanced structure with very dense connections. Then comes the downstream wafer-on-substrate, assembly, and test work, where that integrated unit is attached to the package substrate, finished, and qualified for shipment.

Both stages are important. But they do not carry the same strategic weight, and they do not have the same economics.

That is why TSMC’s current model makes sense. It keeps CoW in-house, where the value of the leading-edge wafer is most exposed and the integration work is most sensitive, while relying more on ASE, SPIL, Amkor, and other partners for downstream scale. This is not just a temporary capacity release valve. It reflects a deliberate choice about control, capital efficiency, and where differentiated returns are most likely to sit.

CoW is the stage TSMC has the strongest reason to own because it sits closest to the leading-edge logic die. It determines how the processor connects to HBM, and that affects bandwidth, power efficiency, thermals, and overall package yield. If something goes wrong here, the damage is not limited to a cheap downstream step. It can compromise multiple very expensive dies and memory stacks at once.

That makes CoW much more than routine manufacturing execution.

This is also why CoW matters to the customer relationship. For many AI customers, the foundry decision is increasingly tied not only to transistor performance, but also to whether the full package can be delivered at scale with acceptable yield and system performance. In that sense, CoW helps protect wafer revenue, not just packaging revenue. It sits close to the technical boundary where packaging stops being a support function and starts becoming part of the product architecture.

That broader strategic role is why TSMC places CoWoS inside its 3DFabric platform rather than treating it as a stand-alone service. 3DFabric matters because it turns packaging from an isolated manufacturing step into part of a broader customer lock-in system that includes advanced nodes, chip integration, memory coordination, and design enablement.

For investors, the implication is straightforward: some of CoWoS’s economic value does not show up neatly as packaging margin. It shows up in stronger wafer pull-through, deeper customer dependence, and a platform that is harder to replace.

The downstream stages have different economics. Wafer-on-substrate, assembly, and testing are still demanding, especially as AI packages become larger, hotter, and more complex. But they are also more operations-heavy and more scale-driven. They sit closer to the kind of high-volume manufacturing OSATs have spent decades optimizing.

That does not make them low value in an absolute sense. It makes them different from CoW in a way that matters for capital allocation.

TSMC could build more of that downstream capacity itself. The better question is whether it should. Every additional facility brings not just capital spending, but also depreciation, labor, maintenance, and utilization risk. Those costs persist even if demand normalizes or package designs evolve.

For TSMC, the opportunity cost is unusually high because the alternative uses of capital are unusually attractive. Advanced process technology, future device architectures, SoIC, and more CoW capacity all compete for the same resources.

This is where the outsourcing debate often gets muddled. TSMC has not publicly framed OSAT cooperation as an admission that downstream packaging has structurally weak margins. Management’s stated rationale has been capacity: avoiding packaging bottlenecks that could limit customer growth and, in turn, TSMC’s own wafer sales.

That should be taken at face value. But it is also reasonable to infer that margin and return on capital still matter indirectly. Even if downstream packaging becomes more profitable, that still does not mean it deserves the same capital priority as the businesses that anchor TSMC’s long-term returns.

The key point is that TSMC does not need advanced packaging to be its best standalone business. It needs advanced packaging to support its best overall business. Those are not the same thing. A packaging step can be economically attractive for an OSAT and still be a lower-priority use of capital for TSMC.

Once viewed through that lens, outsourcing looks less like a compromise and more like a rational division of labor.

Seen that way, the model is controlled orchestration. TSMC keeps control over the CoWoS platform, the CoW integration, the qualification flow, and the yield-sensitive know-how that protects wafer value and customer stickiness. OSAT partners absorb more of the downstream assets and operating burden needed to scale output.

That can let TSMC expand the overall supply chain faster without internalizing every fixed cost.

The partnership evidence fits that interpretation, though the line between fact and inference matters. TSMC has entered arrangements with OSAT partners such as Amkor and works with ecosystem participants such as ASE within the broader 3DFabric framework. Those facts show coordinated capacity building and ecosystem participation.

The further conclusion is an inference: TSMC appears to be positioning itself as the platform owner while letting partners handle more of the downstream manufacturing load. That reading fits the structure of these relationships, but it is still an interpretation rather than management’s explicit wording.

For investors, this leads to two practical conclusions.

Interested in this topic? Subscribe for free and keep reading.

Subscribe now

Disclaimer: For informational purposes only. Not financial advice. Invest at your own risk.

Subscribe now

The AI infrastructure race is usually described with a familiar set of numbers: GPU orders, secured megawatts, and planned capital spending.

Those figures matter. But they are still mostly input metrics.

For investors, the more important question is how much of that spending can actually become usable, billable AI compute.

That is where deployable capacity matters.

In plain English, deployable capacity is the share of infrastructure that can support the required hardware and workload now, under real constraints of power delivery, cooling, networking, backup systems, and physical layout.

Announced capacity tells you what management intends to build. Installed capacity tells you what has been built. Deployable capacity gets closest to what can actually be monetized.

That distinction sounds subtle until you apply it to a real facility.

A data center can still show remaining power on paper and yet be unable to host the next AI cluster. The issue may not be total electricity in the abstract. It may be that the available capacity is scattered across separate halls, rows, electrical paths, or cooling zones, while the workload needs a large, contiguous block of power, cooling, space, and network connectivity in one place.

That is why deployable capacity is a better economic lens than headline megawatts alone.

The relevant question is not how much infrastructure exists in theory. It is how much of that infrastructure can be turned into compute customers can use and providers can bill for. That conversion from installed assets to live AI capacity is where capital efficiency shows up.

A useful framework for thinking about this comes from a 2026 Stanford University and Microsoft Azure Research paper on AI data-center power architecture and deployable capacity.

The paper is best read as a simulation framework calibrated with real operating assumptions, not as proof that Azure or the broader industry is currently suffering from severe underutilization. That distinction matters. Its value is conceptual: it offers a sharper way to evaluate how infrastructure design holds up as AI hardware, density, and deployment patterns change.

Interested in this topic? Subscribe for free and keep reading.

Subscribe now

Disclaimer: For informational purposes only. Not financial advice. Invest at your own risk.

Subscribe now

Most AI companies are trying to solve two hard problems at once: how to reach users at scale and how to pay for the compute required to serve them. Google starts from a different position. It already has enormous distribution through Search, YouTube, Android, Chrome, Workspace, and Cloud. And the business most exposed to AI disruption, Search, is also one of the businesses financing the response.

That is why Google’s AI story is bigger than Gemini alone. The more important idea is a full-stack flywheel: profitable products generate cash; that cash funds data centers, models, and custom chips; those investments can improve products and lower AI costs; and stronger usage, subscriptions, and cloud revenue help finance the next round.

If that loop works, Google’s edge will not come from one benchmark win or one polished demo. It will come from owning more of the system than most rivals do.

That is the promise. The harder question is whether this becomes a durable business advantage or simply a more expensive way to defend existing businesses.

Search is both the risk and the funding source

A lot of the discussion around Google and AI starts with the threat to Search. That concern is real. If AI changes how people look for information, compare options, or complete tasks, then the business that has long anchored Google’s economics is directly exposed.

But that framing is incomplete. Search is not just the business under pressure. It is also the engine paying for Google’s adaptation. The same company that has to absorb the risk also has the cash flow to fund infrastructure, engineering talent, model development, and the hardware needed to keep improving those systems.

That matters because AI is not just a software contest. It is a capital problem. The cost of staying competitive rises when products need more compute, more specialized infrastructure, and faster iteration. In that environment, companies with both distribution and financial capacity start from a much stronger position than companies that have one but not the other.

So the useful question is not simply whether AI hurts Search. It is whether Google can move commercial intent into an AI-shaped interface without breaking the economics that made Search so valuable in the first place. The full test sits beneath the product experience: query volume, click behavior, ad placement, advertiser returns, revenue per query, and the cost of serving that query still all have to work.

That distinction is easy to miss in product conversations. Users may prefer direct answers, conversational search, or AI-generated summaries. But if those features raise compute costs faster than they improve monetization, engagement alone will not save the business. The real upside comes only if AI improves the product while preserving the revenue and margin structure that funds the rest of Google.

Gemini gets most of the attention because it is the visible part of Google’s AI story. The more economically important piece may be TPU.

Google does not need custom chips to dominate the broader chip market for them to matter. It only needs them to improve the internal economics of its own system. If TPUs lower inference cost, use power more efficiently, or reduce dependence on outside suppliers, Google gets more room to serve AI inside products where demand is large and margins matter.

That matters most in high-volume workloads. AI Overviews in Search, conversational responses, and assistant features in Workspace all consume compute. At Google scale, a lower cost per response is not a technical footnote. It can be the difference between an AI feature that supports the business and one that quietly dilutes margins.

That is why TPU looks more important as infrastructure leverage than as a standalone story. If Google can make intelligence cheaper to serve, it can push AI into higher-volume products without letting the cost structure spiral. That does not guarantee success, but it changes the economics of deploying AI across the stack.

Seen that way, TPU is not mainly a side business. It is part of the cost foundation underneath the broader flywheel, which leads to Gemini’s real job: turning that cheaper, more available intelligence into better product economics.

None of this makes Gemini unimportant. It clarifies where it matters.

Subscribe for research that looks beyond the headline—into the business, the expectations, and what the market may be missing.

Subscribe now

The key question is not whether Gemini is impressive in isolation. It is whether Gemini makes Google’s products more valuable in ways that improve monetization, retention, or margin. In Search, that could mean handling more complex queries, improving commercial discovery, or helping users complete tasks that advertisers care about.

It may not need to reshape every query. It needs to work where the economics justify it: where better answers or easier task completion create enough user and advertiser value to cover the higher cost to serve.

Outside Search, Workspace may be the cleaner monetization path. If AI meaningfully improves Gmail, Docs, Sheets, or Meet, Google may gain pricing power, stronger retention, or broader enterprise adoption. That is a more direct business signal than consumer chatbot engagement alone. A useful assistant embedded in daily workflows can become sticky in a way stand-alone AI usage often does not.

Android and Chrome matter differently. They are valuable primarily as distribution layers. If AI assistants become a meaningful interface for finding information, completing tasks, or navigating commerce, Google wants Gemini present where people already spend time. But distribution alone does not create loyalty. Defaults matter, but people can still switch if another product feels more useful, faster, or more trustworthy.

That is why product quality still matters. Google’s advantage, if it has one, is not that defaults last forever. It is that it can pair distribution with infrastructure and models. The next question is whether that shared stack can also be monetized outside Google’s own surfaces.

That is where Cloud gets more interesting. If Search is the main earnings engine and Workspace is the clearest paid AI upsell, Cloud is the main external revenue lever.

Google has already invested heavily in data centers, networking, servers, storage, and custom AI hardware to support its own services. Through Google Cloud, it can try to monetize that same foundation in several ways: selling model access, managed AI tools, and the compute capacity companies need to build on top of them.

That reuse matters because shared infrastructure only becomes a flywheel if it stays busy on attractive terms. High-utilization capacity serving both Google products and enterprise customers can improve returns. Underused capacity, or AI demand won through heavy pricing pressure, does the opposite: it turns expensive infrastructure into lower-return inventory.

This is what makes Google’s full-stack position so interesting. Search monetizes intent. Workspace offers a direct paid productivity use case. Cloud commercializes the stack externally. Android and Chrome help distribute AI into daily behavior. Not many companies control that many layers at once.

But control is not the same as advantage. The unresolved question is whether Google’s breadth actually raises returns across those layers, or whether competition and capital intensity absorb most of the benefit before it shows up in the numbers.

That is the clean skeptical case: AI becomes a capital-spending treadmill rather than a compounding flywheel.

The risks are straightforward. AI interactions may remain structurally more expensive than traditional search. Monetization may lag behind usage. Cloud customers may capture much of the value through pricing pressure. Custom silicon carries design risk and the possibility of underused capacity. And each new model cycle may demand another round of heavy spending before the last one has earned an adequate return.

In that world, Google still has to invest. Standing still is not an option. But forced investment is not the same thing as attractive investment.

That is why this thesis should be judged through business mechanics, not model hype. Over the next few years, the important signals are whether Search can hold revenue quality as AI usage grows, whether Workspace shows real paid AI adoption or pricing power, whether Cloud turns AI demand into durable growth with decent margins, whether cost per AI interaction improves, and whether capex intensity eventually translates into stronger free cash flow.

The bottom line is simple. Google’s real AI bet is not just Gemini. It is the idea that a profitable, distributed, infrastructure-heavy company can use AI to reinforce its economics across multiple businesses.

The final test is simpler still: after this investment cycle, does the system produce better returns on capital and more free cash flow, or does it mostly demand another round of spending?

Subscribe now

Subscribe for free and never miss future analysis and other valuable insights.

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

免責聲明： 本文內容僅供參考，不構成任何投資建議。投資人應獨立評估風險並自負盈虧。

Subscribe now

如果過去兩年 AI 半導體產業有一個幾乎無法避開的關鍵字，很多人的答案大概都是 CoWoS。

這很合理。隨著 NVIDIA、AMD，以及大型雲端業者持續推進新一代 AI 晶片，市場對先進封裝的關注快速升高，CoWoS 也自然成為討論焦點。對投資人來說，最直覺的研究框架就是：AI 晶片需求很強，而 CoWoS 是影響出貨節奏的重要環節，所以接下來要看的是台積電如何擴產、什麼時候能擴得更快。

但我在研究這條線時，慢慢把問題往後推了一步。

如果 CoWoS 是今天的瓶頸，那 CoWoS 自己的瓶頸又是什麼？

我認為，這個問題比單純追蹤 CoWoS 擴產更重要。因為當市場資金與注意力都集中在同一個環節時，真正有研究價值的，往往是下一個還沒有被充分定價的限制因素。

把一顆 AI GPU 或 ASIC 的封裝系統拆開來看，它不只有 CoWoS。除了矽中介層（interposer，負責在 GPU 與 HBM 之間做高密度連接）之外，還有 ABF 載板（把整個封裝往下接到系統板，並承接供電與高速訊號），以及最後連接到系統端的 PCB。也就是說，就算先進封裝產能持續增加，整體出貨能不能順暢往下走，仍然取決於其他關鍵零組件能不能一起擴產與升級。

我的判斷是，未來市場焦點很可能會逐漸從「封裝做不做得出來」，轉向「高階載板跟不跟得上」。

這也是我開始看欣興 (3037.TW)、南電 (8046.TW)、景碩 (3189.TW) 這幾家公司的原因。

很多時候，市場會把這些公司統稱為「CoWoS 概念股」，但這種分類其實太粗了。真正值得區分的，從來不是有沒有站上題材，而是誰有能力持續承接下一代 AI 封裝對載板規格的要求。當封裝難度提升到某個程度之後，供應鏈的差異不再只是有沒有產能，而是有沒有能力把高難度產品穩定量產出來。

而 AI 的演進方向，正好在把這個問題推向前台。

不論是更高算力、更大的運算模組、更高的 HBM 容量與頻寬，還是更高功耗，這些趨勢最後都會反映在封裝尺寸變大、結構更複雜、熱與機械應力更難控制。對載板來說，這通常意味著更高層數、更細線路、更嚴格的翹曲控制，以及更高的量產良率要求。

換句話說，下一階段的競爭核心，未必只是「誰能供貨」，而會越來越接近「誰能做出最難的那塊載板」。所以我會先看兩個訊號：AI 晶片的封裝尺寸是否持續放大，以及高階 ABF 載板的 ASP 是否開始反映規格升級。

如果從這個角度切入，我會先把欣興放在第一順位追蹤，而不是直接下結論說它已經是答案。

原因不是因為它能被簡單歸類為 AI 題材受惠股，而是以目前公開訊號來看，它比較接近我想驗證的那個投資假設：當 AI 封裝持續升級時，哪些公司有機會不只吃到需求成長，還能吃到規格升級帶來的價值提升。

如果把比較點收斂成三個，我最在意的是高階製程與量產能力、與 AI 平台的承接深度，以及資本支出是否持續往先進載板傾斜。南電、景碩當然也可能受惠，但如果主線是「高規格升級」，而不只是「跟著補量」，我目前會先把重心放在欣興身上。

短期觀察：未來 1 到 4 季的高階 ABF

我並不是否定 CoWoS 的重要性，也不是跳過當前最重要的產業現實，而是把研究視角延伸到 CoWoS 之後，看看哪個環節更可能接棒成為新的瓶頸。如果未來幾季封裝持續放量，而高規格載板的供給、良率或價格開始出現更明顯變化，同時欣興這類供應商的 CAPEX 也持續往先進載板集中，那整個投資敘事就可能從「先進封裝擴產」逐漸轉向「先進載板升級」。

不少人會把 T-Glass 和 Glass Core 混淆。T-Glass 說的是材料層的優化方向，用來改善大型 ABF 載板在高溫與大尺寸條件下的翹曲與可靠性問題；它代表的不是架構革命，而是既有高階載板規格繼續往上走。從投資角度來看，這件事的重點不是誰講了最多新故事，而是誰能在現有體系裡，把更高規格的產品穩定做出來。

Subscribe now

長期觀察：未來 2 到 5 年的 Glass Core 路線

如果把時間拉長，我認為更值得提早研究的是 Glass Core。

這不是同一層級的變化。Glass Core 指的是以玻璃核心取代現有有機核心的載板架構，關心的是尺寸穩定性、訊號表現與更高 I/O 密度的上限；T-Glass 則仍屬於現有 ABF 體系內的材料與製程優化。

Glass Core 之所以值得追蹤，不是因為它明天就會大規模取代 ABF，而是因為一旦 AI 晶片持續往更大尺寸、更高 I/O、更高訊號要求發展，現有有機載板架構終究可能面臨新的物理極限。到那個時候，能不能提早卡位下一代技術路線，就可能重新定義誰能留在高階供應鏈核心。

我的想法很明確：至少在未來 1 到 4 季，我不認為 Glass Core 會成為主要投資主線；放到更長的 2 到 5 年，我才會把它當成需要持續追蹤的技術分水嶺。現在如果把所有投資邏輯都壓在 Glass Core 上，我會覺得太早。但這不代表它不重要。越是離商業化還有距離的技術，越值得在市場共識形成之前開始建立觀察框架。

而欣興吸引我的地方，就在於它的投資故事不只停留在今天。短期我看的是它能不能在高階 ABF 這條較接近當下變現能力的主線上持續受益；長期我看的是，如果公司對玻璃基板、Glass Core，甚至更上游的先進封裝協同技術持續投入，它是否有機會在下一次封裝架構升級時，依然保有技術位置。

當然，這裡也必須保留一個反向思考。即使 CoWoS 擴產持續推進，也不代表高階載板瓶頸會立刻、線性地接棒。中間可能還會受到客戶導入節奏、良率爬坡、產品世代切換，甚至整體 AI 資本支出波動影響。

也因此，我接下來不會只看「AI 很熱」這種總體敘事，而會更具體地觀察幾個指標：AI 晶片的封裝尺寸是否持續放大、高階 ABF 載板的 ASP 是否有結構性提升、欣興的資本支出是否持續往先進載板傾斜，以及公司未來是否揭露更多與 Glass Core、TGV 或其他先進載板技術有關的布局。

現階段的答案看來是成立的，但未來是否依舊成立會影響投資邏輯。

這篇文章真正想說的，不是「欣興是不是 AI 概念股」這種表層的問題，而是另一個更核心的思考：如果今天市場的瓶頸是 CoWoS，那未來的瓶頸，會不會逐步轉向高階載板，甚至更長期的 Glass Core 技術？

市場總是先追逐眼前最明顯的瓶頸。但很多時候，真正值得研究的，不是今天已經很擁擠的答案，而是下一個瓶頸還沒被大多數人看見之前，誰已經開始準備了。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

Disclaimer: For informational purposes only. Not financial advice. Invest at your own risk.

Subscribe now

If you are looking at Rocket Lab (RKLB) as a long-term investment, the first thing to update is your mental model. This is no longer just a company launching a small rocket in the shadow of SpaceX.

Rocket Lab reported $602 million in revenue in 2025 and ended that year with $1.85 billion of backlog. Electron, Rocket Lab’s small orbital launch vehicle, still has a clear role. Customers that care about schedule control, mission-specific orbits, and not being one payload among many on a larger rideshare flight still value dedicated launch. Not every customer wants the lowest possible cost if it means giving up flexibility.

That launch business still matters strategically. Electron gives Rocket Lab a proven operational foothold, and HASTE, the company’s suborbital test variant used for hypersonic technology work, extends that foothold into a defense niche that is becoming more important for the United States and its allies.

Even if small launch never becomes a huge market, it still matters as a capability and as a customer relationship engine.

Launch is no longer the whole story. Rocket Lab now operates through two interdependent businesses, Launch Services and Space Systems, and that second segment is doing much of the work in turning the company from a niche launcher into a broader contractor.

Space Systems includes full spacecraft as well as components such as solar products, reaction wheels, star trackers, separation systems, radios, and flight software. In plain English, Rocket Lab is increasingly selling more of the mission stack, not just the ride to orbit.

That is strategically important because components, software, and spacecraft manufacturing can embed Rocket Lab more deeply inside customer programs and widen the revenue base over time. The company is trying to build the satellite, supply more of the subsystems inside it, support the software around it, and eventually launch it on its own rocket.

Rocket Lab now has enough evidence behind the narrative that investors should judge it as a real operating business, not just an ambitious space story.

The company reported $602 million of revenue in 2025, up 38%, and ended the year with $1.85 billion of backlog. In the first quarter of 2026, Rocket Lab reported record revenue of $200.3 million, up 63.5% year over year, with GAAP gross margin of 38.2% and backlog rising to roughly $2.2 billion.

Those numbers do not prove economic maturity, but they do suggest that the company has moved beyond pure promise.

The clearest example is the Space Development Agency award to design and manufacture 18 missile-warning and tracking satellites, a contract Rocket Lab said was worth about $816 million. One award does not prove the company has fully arrived as a prime contractor, and investors should resist making too much of a single win.

Still, it is a meaningful signal. It suggests Rocket Lab is being trusted with larger and more strategically important work.

That shift in role is important. Prime status usually means deeper customer relationships, more control over the program, a wider revenue opportunity, and a stronger position in future competitions. It also means the bar on execution rises. Being trusted with more of the mission is attractive only if the company can deliver without losing discipline.

The acquisitions fit the same pattern. The Mynaric deal expands Rocket Lab’s position in optical communications, and the planned Motiv Space Systems acquisition would add robotics and motion-control capabilities. The strategic logic is easy to see: bring more critical subsystems in-house, reduce dependence on outside suppliers, and offer customers a broader package.

But this is also where the story gets less simple. Vertical integration is not automatically a virtue. It can improve control and potentially strengthen margins, but it can also create managerial sprawl, integration risk, and higher capital intensity.

A company that tries to do more has to prove it can do more without becoming less coherent. Investors should recognize the upside without treating expansion itself as proof of success.

Neutron Is the Unlock, and the Biggest Risk

If one program ties the entire Rocket Lab narrative together, it is Neutron.

Electron gives Rocket Lab a credible place in small launch, but small launch has natural limits. Payloads are smaller, mission opportunities are narrower, and low-cost rideshare options cap pricing power. A well-run small-launch business can still be useful and strategically valuable. It is much harder to argue that small launch alone supports a platform valuation of major scale.

Neutron, Rocket Lab’s planned medium-lift rocket, is supposed to change that. The company has said the vehicle is designed to carry up to 13,000 kilograms to low Earth orbit and return its first stage for reuse. If it delivers that capability, Rocket Lab moves into a much larger market that includes constellation deployment, national-security missions, cargo resupply, and potentially deeper-space applications.

The strategic significance goes beyond launch revenue itself. A successful Neutron would make Rocket Lab one of the few Western companies outside SpaceX with a credible medium-lift capability. That matters because many customers do not want to rely on a single provider, even one as dominant as SpaceX.

Governments care about supply resilience. Satellite operators care about schedule certainty and bargaining power. Defense customers care about assured access to space. If Neutron becomes reliable, Rocket Lab may win business not because it is always the cheapest option, but because a viable alternative to SpaceX has real strategic value.

There is also a more ambitious version of the thesis. A medium-lift rocket does not just support third-party launches. It can support services or infrastructure Rocket Lab might one day operate itself. That possibility helps explain why some investors see the company less as a contractor in the making and more as an integrated space platform.

Still, none of that works without execution. Neutron is the biggest technical, operational, and financial uncertainty in the story. Rocket Lab still has to qualify the Archimedes engine, complete integration, validate its reuse design choices, and successfully fly an entirely new system.

Timing is the next layer of risk. As of the company’s May 2026 update, management was still targeting a debut later in 2026, but first launches often slip. A modest delay would not necessarily break the thesis. Repeated delays, major redesigns, or an early failure would be more serious because they would hit cash use, customer confidence, and the timeline for turning Neutron from a development program into a real commercial asset.

Rocket Lab’s financial trajectory is improving, but investors should be careful not to confuse rapid growth with proven economics.

The company is still loss-making because it is funding Neutron, expanding production, integrating acquisitions, and taking on larger programs that can create quarter-to-quarter volatility. For the second quarter of 2026, management guided to revenue of $225 million to $240 million, GAAP gross margin of 33% to 35%, and an adjusted EBITDA loss of $20 million to $26 million.

The message is straightforward: scale is arriving, but profitability is not yet durable.

The key financial question is narrower than the headline growth story. Rocket Lab can clearly grow. What still has to be proven is whether that growth can turn into attractive free cash flow after the capital demands of Neutron, production expansion, and vertical integration.

The most useful operating signals are therefore not flashy headline beats. Investors should watch whether Space Systems margins improve as scale builds, whether Electron and HASTE cadence produce better unit economics rather than just more activity, and whether Neutron moves from development drag toward a repeatable business without overwhelming cash consumption.

Until those things become clearer, profitability remains something the market expects rather than something the business has fully earned.

A Good Business Can Still Be a Difficult Stock

This is where Rocket Lab becomes harder to own, even for investors who admire the company.

At roughly $81 per share in July 2026, Rocket Lab’s market capitalization was around $49 billion despite continued accounting losses. On the company’s 2025 revenue base, that is roughly an 80-times-sales valuation.

Whether one thinks that multiple is justified or not, it changes the nature of the bet. At a lower valuation, investors could have done well if Rocket Lab simply survived, executed competently, and kept expanding into adjacent areas. Here, much more has to go right.

Neutron likely has to work. Defense and satellite programs likely have to keep scaling. Vertical integration likely has to improve the economics enough to support a bigger platform narrative. The burden of proof is simply much higher now.

Traditional valuation tools are not especially satisfying here, which is why a qualitative end-state lens is more useful than pretending to have precision. Earnings-based metrics do not help much while profits remain negative. Revenue multiples are easier to cite, but they can mislead because Rocket Lab’s revenue mix spans launch, components, spacecraft manufacturing, and still-unproven future businesses.

That leaves a more practical question: is Rocket Lab becoming a very capable contractor, or a genuinely integrated platform with launch, spacecraft, components, and eventually higher-value services? The current stock price leans toward one of the more ambitious answers.

That is why admiration and investment judgment need to stay separate. Rocket Lab has already proven more than most early-stage space companies ever will. It has a reliable launch vehicle, meaningful demand, a large backlog, expanding defense relationships, and growing spacecraft capabilities. This is not a slide-deck company.

But none of that automatically makes the stock attractive at any price.

The cleanest way to approach the name is to watch a short list of milestones without romanticizing them: Neutron’s path to first flight and early execution, whether Space Systems converts scale into better margins, whether Electron and HASTE show real operating leverage, and whether cash consumption stays under control.

If those signals keep improving, the platform thesis becomes easier to underwrite. If Neutron slips repeatedly, margins remain messy, or cash demands keep rising, the stock can disappoint even if the company continues to win business.

That is the bottom line. Rocket Lab no longer needs investors to believe it is a clever niche player with upside. It needs investors to decide whether it can become a scaled, durable, integrated space company, and whether that future is still ahead of the stock or already embedded in it.

Subscribe now

Subscribe for free and never miss future analysis and other valuable insights.

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

免責聲明： 本文內容僅供參考，不構成任何投資建議。投資人應獨立評估風險並自負盈虧。

Subscribe now

談到上銀科技（2049.TW），市場最熟悉的標籤，仍是線性滑軌、滾珠螺桿，以及與工具機景氣高度連動的製造業循環股。但我認為，上銀現在真正值得看的，不是它能不能被歸進最純的人形機器人概念股，而是它是否有機會更早在半導體設備、物流機器人與專用自動化領域，成為把 AI 決策轉成真實動作的 Physical AI 次系統供應商。

這個轉型還沒有被完全證明，但已開始反映在營收占比與獲利數據上。

這也是我會繼續追蹤它的原因。上銀不是那種只靠題材敘事就能一路把故事講下去的公司；反而因為底子夠傳統，更適合用基本面去檢查，轉型到底有沒有真的發生。對投資人來說，重點不是「機器人熱不熱」，而是市場重估需要看到什麼：產品組合是否往新業務傾斜、獲利結構是否同步改善，以及公司能不能從賣零件，走到拿到重複性的次系統訂單。

Physical AI 之所以值得談，關鍵從來不只是讓 AI 理解環境，而是把運算結果變成精準、穩定、可重複執行的實體動作。這一步沒有大模型那麼吸睛，卻是自動化真正落地的核心。上銀長期累積的線性傳動、精密定位與機電整合能力，剛好卡在這個轉換環節。

從產品來看，它也早已不只是滑軌和螺桿，而是延伸到單軸及多軸機器人、諧波減速機、末端效應器、晶圓機器手臂、晶圓尋邊器與 Load Port。也因此，它的機會未必需要等通用型人形機器人大量商業化，反而更可能先在對精度、穩定性與可靠性要求更高的場景落地。若這個角色升級真的開始發生，最早反映的地方，通常不會是口號，而會是產品組合與獲利結構。

先看基本面。從2025年營收結構來看，上銀雖然開始出現轉型跡象，但仍未脫離傳統製造循環股的本質。最明顯的改變，是機器人業務占營收比重由2024年的7%提高至約10%，反映晶圓搬運、自動化設備與專用型機器人開始對營收產生更實質的貢獻。

把時間拉到 2026 年第一季，公司單季營收為 63.79 億元，年增 9%；其中機器人營收由一年前的 5.11 億元增至 7.47 億元，年增約 46%，明顯快於整體營收成長，營收占比也提高到 12%。傳統產品占比雖然仍高，但已沒有再往上集中。

這組數字至少說明一件事：目前的成長，不只是傳統景氣回升，機器人與自動化相關業務也開始在營收結構裡變得更有分量。

更重要的是，這個組合變化目前沒有伴隨明顯的獲利犧牲。2026 年第一季毛利率由一年前的 29.8% 升至 31.9%，營業利益率也由 7.2% 提高到 9.9%。至少就已揭露的結果來看，機器人占比上升，並沒有把整體獲利往下拖，反而與獲利能力改善同步出現；同季淨利與每股盈餘也都較去年同期成長。這是現階段最有價值的訊號，因為它代表新業務至少還沒有稀釋整體體質。

但這裡也正是最需要克制的地方。這些數據足以說明轉型開始進入可觀察階段，卻還不足以證明轉型已經完成。公司並未單獨揭露晶圓搬運、物流機器人或半導體次系統的毛利率，因此整體毛利改善，也可能部分來自產能利用率、匯率、價格或景氣循環回升。機器人營收增加，本身也不等於高價值次系統已經建立出可持續的競爭優勢。

如果這些新業務還停留在出貨增加，而沒有真正走進客戶的設備設計、模組整合與認證流程，那麼它的產業位置，就未必已經完成本質升級。

所以真正值得看的，是下一層的二階效應。標準滑軌與螺桿，終究還是比較容易受價格競爭與景氣波動影響；但如果上銀能從標準零件供應商，逐步走向晶圓移載系統、智慧傳動模組與完整運動次系統的提供者，它賣的就不只是元件，而是更深地參與客戶設備架構。

這種升級若成立，最能證明的不是一次性出貨，而是半導體與自動化設備端出現重複訂單、設計導入與認證延續。那代表的才是單機價值提高、客戶黏著度上升，以及替換成本變高。

因此，上銀的投資看點，與其說是押注人形機器人何時全面爆發，不如說是觀察它能不能先在更務實、需求更明確的場景裡，完成自己的角色升級。接下來我會聚焦三件事：機器人營收占比是否持續提升、半導體次系統能否形成重複訂單，以及毛利率與資本報酬能否隨著產品組合升級而持續改善。

這三項若能一起往上，市場才有理由把它從傳統傳動元件廠，重新定義成更高價值的智慧運動與機器人次系統供應商。

市場現在看到的，也許仍是一家等待工具機復甦的公司；但更值得研究的，可能是它是否正在變成連接 AI 決策與實體動作的關鍵一環。這個轉型還沒有被完全證明，所以不該提早下結論；但它也已經不只是題材故事，因為部分變化，開始能被營收占比、獲利率與訂單品質逐季驗證。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

Disclaimer: For informational purposes only. Not financial advice. Invest at your own risk.

Subscribe now

Most investors look at Unity (U) through two familiar lenses: whether Vector can keep driving ad growth and whether Unity 6 can stabilize the core game-engine business. That makes sense, but it may also be too narrow. The more interesting question is whether Unity has a longer-term option in industrial digital twins and Physical AI that the market is barely recognizing.

That possibility is easy to miss because the visible numbers point elsewhere. In Unity’s first-quarter 2026 outlook, the company projected Strategic Grow revenue of $302 million to $306 million, versus Strategic Create revenue of $154 million to $158 million. Management has also been clear that Vector is a major driver of improved growth and profitability.

The problem is that Unity does not separately disclose Industry revenue, growth, or profitability. Those activities sit inside Create, which means investors can track Vector quarter to quarter but have much less visibility into Unity’s industrial exposure.

That does not prove the market is overlooking a valuable business. But it does create an asymmetry. The easiest part of Unity to measure gets most of the attention, while the part that could matter in a different computing cycle remains largely buried.

That buried exposure matters because Physical AI adoption, if it expands, will likely require more than chips, models, and simulation infrastructure. It will also require software that connects industrial data, turns it into usable interfaces, and helps people interact with increasingly autonomous systems. Unity may already touch part of that layer, even if the business is still too opaque and too early to value with confidence.

This is where first-order and second-order thinking matter. First-order thinking says Physical AI should benefit the companies selling chips, robotics models, sensors, and simulation infrastructure. That is the obvious layer, and NVIDIA is already pushing Omniverse as part of the industrial digital twin and robotics stack.

Second-order thinking asks what happens after that infrastructure exists. If factories, warehouses, vehicles, and other physical systems become more software-defined and AI-assisted, companies will need more than training environments. Engineers, operators, designers, and maintenance teams will need applications that let them visualize systems, inspect predictions, test scenarios, supervise autonomy, and step in when necessary.

In other words, Physical AI may need an application layer between intelligent systems and human users.

That is where Unity could become relevant. More precisely, Unity appears to have some of the building blocks for that role today, not a proven end-state position. It can turn CAD, BIM, and other 3D assets into interactive applications and deploy them across PCs, phones, browsers, XR headsets, vehicle displays, and embedded devices. If digital twins evolve from static visualizations into operating interfaces, that capability set becomes more strategically important.

The bull case here is not displacement. Unity does not need to beat NVIDIA or replace engineering platforms like Siemens to matter. The more credible version of the thesis is complementarity: infrastructure players provide the computing, simulation, and models, while Unity could help provide visualization, application development, cross-platform deployment, and the human-facing interface layer.

For investors, that means the right question is not whether this sounds interesting in theory. It is what would make it investable. The main proof points are fairly clear: measurable Industry traction, stronger strategic validation through partnerships or customer wins, and clearer evidence that Unity’s product stack is becoming more capable for industrial workflows.

That is why this should still be treated as optionality, not a base-case assumption. Unity needs stronger robotics integrations, better industrial data connectors, clearer OpenUSD support, more deterministic simulation capabilities, and real proof that customers are deploying these workflows at scale. Its transfer of digital twin professional services to Capgemini may improve focus, but it also increases dependence on partners for enterprise execution.

So the investment takeaway is straightforward. Unity should not be valued today as a Physical AI winner. There is not enough evidence for that.

But investors may also be making a mistake if they view the company only as a game engine plus a recovering ad business. The right posture is skeptical but open-minded: watch for measurable Industry traction, stronger partnerships, and clearer product validation.

If those signals appear, Unity’s industrial role could move from an interesting idea to something the market has to price. Until then, it remains what the best optionality often looks like early on: real enough to watch, but still too unproven to underwrite.

Subscribe now

Subscribe for free and never miss future analysis and other valuable insights.

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

The market is still talking about AI infrastructure as if the whole story can be reduced to one question: who wins the GPU race? That framing made sense during the first phase of the generative AI buildout, when access to NVIDIA accelerators was the constraint that seemed to matter most. It makes less sense now. The more important shift is not that AI is moving beyond GPUs altogether. It is that value is starting to migrate upward, from the standalone accelerator to the architecture wrapped around it.

That distinction matters for anyone trying to understand where capital, margin, and strategic leverage will sit over the next five years. The industry is moving toward a more heterogeneous AI stack, one in which training, inference, host compute, packaging, and memory do not carry equal importance across workloads. Raw accelerator performance still matters. NVIDIA still matters enormously. But the center of gravity is changing. If you only watch the GPU leaderboard, you will miss the next set of bottlenecks: training-versus-inference economics, host-CPU orchestration, memory bandwidth, packaging capacity, and the degree of vertical integration behind the system.

What follows is the more useful frame: training is likely to remain structurally favorable to NVIDIA for longer than many people expect, while inference is opening the door to specialized silicon. At the same time, agentic workloads are turning inference into a system orchestration problem, which raises the strategic importance of CPUs, host architecture, and memory bandwidth. In that world, the real contest is not just accelerator versus accelerator. It is merchant silicon versus vertically integrated systems.

The market is not breaking NVIDIA’s hold. It is splitting around it.

A lot of commentary treats the AI chip market as if it is headed toward a simple post-monopoly future. That is too neat. What is actually happening looks more like a bifurcation.

Training remains a fortress business. The reasons go well beyond installed market share. Training large frontier models rewards a full-stack platform: mature software tooling, compiler support, interconnect, operational familiarity, and the ability to deploy at scale with limited risk. That combination is still extremely difficult to replicate. Even when competitors narrow the performance gap on paper, the switching cost inside production training environments remains high. That is why NVIDIA’s position in training continues to look stronger than headlines about an imminent collapse suggest.

Inference is different. Once a workload becomes large, stable, and predictable enough, the economic logic changes. The buyer starts caring less about maximum flexibility and more about cost per query, power efficiency, thermal limits, latency, and total system utilization. That is where custom ASICs become more attractive. The important point is not that all inference will leave GPUs. It will not. General-purpose GPUs will remain valuable for many inference workloads, especially where demand is spiky, models change quickly, or developers need a common platform across use cases. But the share of inference that can justify specialization is growing.

That transition is not driven by demand alone. It is also being enabled by supply. For several years, advanced packaging capacity acted as a hidden governor on how fast custom AI silicon could scale. Advanced packaging is the assembly step that binds compute die, memory, and interconnect into one high-performance module. If you could not secure enough of that capacity, a chip design could be ready and still fail to ship in volume. As more packaging and memory-integration capacity comes online, hyperscalers gain more room to move inference onto purpose-built silicon. In other words, the next inference wave is not just a software or model story. It is an industrialization story.

That is why the cleanest interpretation is not “the monopoly is breaking.” It is “the monopoly is concentrating where it is strongest, while adjacent value pools begin to open.” NVIDIA’s moat in training can deepen even as inference becomes more plural.

Agentic AI makes inference less of a compute event and more of a coordination problem.

The most underappreciated shift in AI infrastructure is that inference is becoming less like a single-pass transaction and more like a multi-step workflow.

A conventional chatbot query already involves more than matrix multiplication, but agentic systems make this much more explicit. A single request might trigger a planner model, a retrieval step, a tool call, a second model pass, and a verification loop before the user sees an answer. That turns inference into an orchestration problem. The challenge is no longer only how fast the accelerator can process tokens. It is how efficiently the system can manage state, move data, schedule work, and keep the pipeline full without creating new bottlenecks.

That does not mean CPUs suddenly matter more than networking, storage, or software. All of those layers matter. But the host processor sits directly in the path of this orchestration. It handles much of the control-plane work that agentic systems generate in abundance: coordinating tool calls, tracking intermediate state, queueing follow-on tasks, and feeding the accelerator the next job with minimal idle time.

This is why older assumptions about CPU-to-GPU balance look less stable than they did during the first wave of model serving. In heavily accelerator-centric designs, the CPU could be treated as a supporting actor. In more agentic environments, that becomes harder. The exact ratio will vary by workload, architecture, and deployment philosophy, so it is unwise to treat any single number as universal. But the directional shift is clear: host compute is moving back toward the critical path.

That has two implications. First, system throughput may increasingly be capped by orchestration quality rather than pure accelerator abundance. Second, the economic value of the “boring” parts of the server starts to rise. If a deployment bottlenecks on host coordination, memory movement, or software overhead, then adding more accelerator horsepower does not fully solve the problem. It simply moves the queue.

If the CPU matters more, the architecture of that CPU matters more too.

Once the host processor becomes strategically important again, the question is not just how many CPU cores exist in the system. It is who designs them, what architecture they use, and how tightly they are integrated with the rest of the stack.

This is where ARM’s position strengthens, especially inside hyperscaler and custom-silicon deployments. In those environments, custom ARM server CPUs offer a combination that is hard for merchant x86 vendors to match: better power efficiency, tighter workload specialization, and more direct control over system integration. The value proposition is not ideological. It is operational. If a company is already designing, or tightly co-developing, its own accelerator, it becomes increasingly rational to pair that accelerator with a host CPU optimized for the same orchestration model, thermal envelope, and software path.

That does not mean x86 disappears. It still has major advantages in enterprise compatibility, broad software support, and standard server deployments. It may remain the practical default in large parts of the market for years. But the most strategically important AI buildouts are not representative of the whole server market. They sit inside hyperscalers and other scaled operators with the capital, software talent, and fleet control to optimize vertically. In those environments, merchant x86 faces a structural problem: it is competing not just against another CPU, but against a full-stack optimization loop.

That is the context in which ARM can take disproportionate share in the AI ASIC segment even if x86 remains relevant elsewhere. The core issue is not architectural purity. It is ownership. When orchestration becomes central, the hyperscaler or platform owner wants tighter control over the host in its AI-specific buildouts.

Memory and packaging are moving to the center of the story.

The first phase of AI infrastructure investment trained the market to focus on compute density. The next phase will force more attention onto memory and packaging.

For many inference workloads, especially interactive ones, the decisive metric is not only total throughput. It is responsiveness. Time-to-first-token, the delay before a model produces the first visible token for the user, depends heavily on how quickly the system can access model state and move data through memory. As models grow more complex and agentic flows generate more intermediate work, memory bandwidth becomes more than a component specification. It becomes a direct constraint on latency and perceived product quality.

That is why high-bandwidth memory, or HBM, has moved from premium feature to strategic dependency. HBM is the stacked memory placed close enough to the accelerator to feed it data at very high speed. If compute improves faster than memory delivery, the accelerator spends more time waiting. The problem compounds when systems are designed around larger contexts, more concurrent sessions, and more specialized inference silicon. At that point, whoever controls HBM supply, advanced packaging, and the ability to integrate them at scale controls a meaningful part of the industry’s economics.

This is also why advanced packaging deserves more investor attention than it usually gets. Technologies such as CoWoS, which links logic and HBM in a tightly integrated package, are not just manufacturing details. They are the physical mechanism that allows logic, memory, and interconnect to work as one high-performance system. Custom AI chips need that packaging layer to ship at scale, so when packaging is constrained, roadmap ambition gets capped. When packaging expands, the range of economically viable system designs expands with it.

The important takeaway is that the bottleneck is moving outward from the chip core. Performance is increasingly determined by how well the system is assembled, not merely by how fast the accelerator is in isolation.

Google is the clearest case study in where this market is heading.

Google matters here not because it is the only company pursuing custom silicon, but because it shows the logic in unusually clear form. It offers a relatively visible roadmap, deep stack integration, and a clean enough separation between training and inference design choices to make the architectural direction easier to see than it is at many peers.

Its TPU program reflects a world in which large operators no longer treat AI compute as a simple vendor procurement exercise. They treat it as a strategic stack. That means differentiating between training and inference needs, shaping silicon accordingly, and pushing for more control over design, packaging, sourcing, and deployment economics. The exact roadmap details will evolve, and not every reported specification should be treated as settled fact until products are fully in market. But the direction is unmistakable: deeper vertical integration, sharper workload specialization, and a more aggressive effort to reclaim margin and reduce dependence on external platforms.

Google is not alone. AWS has shown the same instinct with its Trainium, Inferentia, and Graviton strategy. Meta has similar motivations around fleet economics and infrastructure control. Microsoft has every reason to move in the same direction where it can. The broader pattern is that hyperscalers increasingly want to own more of the stack whenever the volume justifies it.

This is the key strategic point: custom silicon is not only about beating a merchant GPU on benchmark efficiency. It is about changing who captures value across the system. Once the hyperscaler controls the accelerator, the host CPU, more of the software path, and portions of the supply chain, the margin pool shifts.

The post-2028 question is not whether GPUs matter. It is where durable leverage lives.

The most plausible post-2028 equilibrium is not a world in which GPUs vanish. It is a world in which GPU dominance becomes more segmented.

Training may remain concentrated around a small number of platforms, with NVIDIA continuing to hold the strongest position. Inference, meanwhile, is likely to become more fragmented and more specialized, especially inside the largest cloud and consumer internet platforms. If that happens, the architecture around the accelerator becomes the real battleground: host CPUs, memory bandwidth, packaging capacity, interconnect, and orchestration software.

That is why the long-term pressure on x86 is real even if the transition is uneven. If hyperscaler AI deployments increasingly converge on custom accelerators plus custom ARM hosts, then the merchant CPU loses not only unit share but strategic relevance in the fastest-growing AI-specific part of the market. It can still remain important across enterprise and standard server deployments. It just may no longer sit at the center of value creation.

The simplest way to say it is this: the AI compute market is not merely moving beyond NVIDIA GPUs, and it is not merely searching for the next accelerator winner. It is moving toward system-level competition. The companies that win from here may not be the ones with the single best chip. They may be the ones that control the orchestration layer around the chip.

That is the shift to watch. The GPU era is not ending. But the era in which the GPU alone explained where value would accrue is.

Subscribe now

Subscribe for free and never miss future analysis and other valuable insights.

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

免責聲明： 本文內容僅供參考，不構成任何投資建議。投資人應獨立評估風險並自負盈虧。

Subscribe now

市場談南亞科 (2408.TW)，最常見也最省事的說法是：AI起飛、HBM吃緊、記憶體全面受惠，所以南亞科也是AI概念股。

這樣講不能說全錯，但它把重點講得太平，也容易把真正的獲利來源看偏。若要判斷南亞科這一輪到底只是短期行情，還是有更長一些的延續性，真正該看的不是它有沒有直接站上HBM主戰場，而是AI需求如何改變整個DRAM產業的資源配置。

更精確地說，南亞科眼前最有價值的投資邏輯，不是「它成了下一個HBM贏家」，而是「當大廠把更多資本、產能與先進製程轉去追HBM時，傳統DRAM供給反而變得更緊」。

這個差別很重要。因為它決定了你該用題材框架看它，還是用供需框架看它；也決定了你接下來要追蹤的是新產品故事，還是價格、供給與製程執行力。

南亞科技本身是台灣主要的DRAM製造商，產品應用涵蓋個人電腦、伺服器、消費電子、車用與工業設備。相較於三星、SK海力士與美光，南亞科規模較小，也不是HBM主要供應商。

這裡必須多保留一句：如果沒有更細的產品組合與營收占比資料，我們很難精確量化它對DDR4、DDR5或伺服器DRAM的敏感度。所以比較穩妥的說法，不是「AI一定直接大幅改寫南亞科的價值」，而是「AI帶來的產業重分配，可能讓南亞科在既有產品線上拿到比過去更強的定價能力」。

這也是市場最容易低估的第二層效應。AI伺服器需要大量HBM，也需要伺服器DRAM；當全球三大記憶體廠把更多資源優先留給HBM與高階產品，原本供應DDR4、DDR5及其他傳統DRAM的能力就會受到排擠。

需求不需要全面失控，只要供給收縮夠明顯，價格就可能快速上去。對南亞科來說，這比「有沒有直接做HBM」更重要，因為它解釋了公司為什麼能在產業地位沒有根本改變的前提下，突然出現極強的獲利彈性。

數據與產業證據

先看公司端。從目前揭露的財務表現來看，這一輪改善更像價格驅動，而不是銷量擴張。

根據公司資料，南亞科2026年第二季營收為825.49億元，季增68.2%；同期間DRAM平均售價季增超過六成，但銷售量大致持平。這個組合非常關鍵，因為它代表公司這一波不是靠「賣更多」賺錢，而是靠「賣更貴」賺錢。

這不是文字上的細節，而是判讀獲利品質的核心。價格驅動的獲利爆發，通常對損益表有很強的槓桿效果。毛利率升至79.5%，單季每股盈餘達14.66元；2026年上半年累計每股盈餘為23.38元，和一年前仍在虧損週期時形成鮮明對比。

這些數字當然足以讓市場興奮，但它們同時也提醒一件事：南亞科此刻的獲利，高度依賴價格環境。當價格傳導效果這麼強，景氣上來時會非常亮眼；反過來，若價格回落，修正速度也可能比市場想像更快。

再把視角從公司端切到產業端，邏輯會更完整。公司財報只能告訴我們「價格確實在漲，而且獲利對價格非常敏感」；但要回答「為什麼在漲、能漲多久」，還是得回到整個DRAM市場的供需變化。

產業面的訊號，確實支持供給偏緊的判讀。TrendForce指出，2026年第一季傳統DRAM合約價季增約93%至98%；南亞科2026年6月營收達293.88億元，月增6.21%、年增621.34%，也顯示報價上行仍在往實際營收轉化。

比較審慎的解讀應該是：這些資料支持AI需求升溫與HBM資源排擠，正在強化傳統DRAM的議價環境；但它們本身不能單獨證明所有漲價都來自同一個因素，更不能直接把「價格動能可能延續到2027年」當成已被驗證的結論。

較合理的說法是，如果大廠持續把產能與資本優先配置到HBM，而終端補庫存也沒有快速結束，那麼傳統DRAM偏緊的局面就有機會延續更久。

風險與驗證點

但也正因為多頭故事不難理解，現在更需要防範的，其實是把多頭講得太直線。DRAM終究仍是高度循環產業。價格快速上漲時，供應商會想擴產，客戶也會傾向提前拉貨、建立庫存，整個鏈條都可能因為對緊缺的預期而自我強化。

這種時候，帳面數字通常最好看，市場也最容易把短期高點誤判成長期常態。若未來新產能釋出、終端需求不如預期，或雲端業者放慢AI資本支出，報價與毛利率都可能顯著回落。接近八成的毛利率，至少在目前，沒有理由被直接視為可長期維持的新基準。

所以，討論南亞科不能只停在「這一季賺很多」，而要往前再問一句：這筆景氣紅利最後會留下什麼？

真正的中長期考驗，在於公司能不能把這一輪由價格帶來的現金流，轉化成更穩固的技術、成本與產品競爭力。公司正推動自主1B製程、16Gb DDR5與低功耗產品，也在發展更高容量的伺服器與邊緣AI記憶體；但這些方向目前更像是能力建設路線，而不是已經充分兌現的結果。

接下來仍要觀察時程、良率、量產貢獻與成本改善是否真的落地，否則路線圖本身還不能直接等同於長期競爭力。

對投資人來說，接下來比爭論題材更有用的，是持續追蹤幾個可操作的驗證點。

第一，看傳統DRAM價格韌性，尤其是DDR4、DDR5在補庫存告一段落後，合約價能否仍維持高檔，而不是很快反轉。

第二，看終端需求到底是一次性的提前拉貨，還是有更持久的拉貨基礎。如果客戶庫存回補後需求明顯轉弱，這輪價格紅利的持續性就要打折。

第三，看南亞科的新製程與高階產品推進是否如預期。重點不只是宣布方向，而是良率、成本與量產貢獻有沒有逐季改善。

第四，看市場現在給高獲利的估值，是否已經內含過於樂觀的常態化假設。一旦投資人開始下修毛利率與獲利中樞，評價就可能先於基本面回落。

結論

我認為，南亞科現在最值得理解的，不是一句籠統的「AI受惠股」，而是一個更接近真實的命題：AI確實正在改變記憶體產業，但它帶給南亞科的主要好處，不是直接的HBM紅利，而是全球產能重新分配後，傳統DRAM供給被壓縮所帶來的價格與利潤彈性。

這個邏輯比題材更紮實，也更值得被拆開驗證。

結論很簡單：南亞科的投資邏輯，不應被簡化成AI概念股或HBM受惠股。更關鍵的是，AI帶動HBM與伺服器記憶體需求，促使大廠轉移產能與資本，進而壓縮傳統DRAM供給，提升南亞科在DDR4、DDR5等產品上的定價能力。

但這輪高獲利是否能延續，仍取決於供給緊張能維持多久、製程升級進度是否跟上，以及公司能否把景氣紅利轉成長期競爭力。AI需求可能是結構性的，南亞科當前的超高獲利卻未必是。把這兩件事拆開看，才比較不容易在最熱的時候，看錯真正的價值。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

Subscribe now

如果你把京元電子 (2449.TW) 只看成一檔沾上 AI 題材的封測股，很可能會在最關鍵的地方看錯。

這家公司近兩年的吸引力，不只是 AI 伺服器出貨增加、供應鏈輪到誰受惠，而是高價值 AI 晶片正在把測試這個過去偏末端、偏執行的環節，推向更核心的位置。對投資人來說，真正值得回答的問題不是「京元電子有沒有吃到 AI」，而是它能不能把眼前的產能稀缺與工程能力，變成更長期、較難被取代的技術壁壘。

這個差別之所以重要，是因為兩種敘事對估值的含義完全不同。如果只是景氣循環中的短期受惠，那麼營收創高、產能吃緊、法說樂觀，都未必足以支撐長期高估值；但如果測試的價值密度真的因 AI 而上升，京元電子的角色就不再只是提供機台與產能的測試廠，而有機會升級為高階晶片驗證流程中的工程夥伴。

也因此，後面真正該追的，不只是營收，而是五個更有穿透力的指標：月營收年增、毛利率韌性、資本支出對應的訂單承諾、非 GPU 業務占比，以及自由現金流與投資報酬率。

為什麼 AI 會讓測試從配角變成關鍵環節

京元電子長期提供半導體後段測試服務，包括晶圓針測、成品測試、預燒與系統級測試。過去這不是最容易引發市場想像的環節，因為投資人更常把注意力放在設計、製造或先進封裝；但 AI 晶片改變了排序。

當 GPU、ASIC 或高效能 CPU 的單價、功耗與系統複雜度同步提高，測試失敗的代價就不再只是單顆晶片報廢，而可能擴大成模組良率、伺服器穩定性，甚至客戶部署時程的風險。

這就是京元電子投資價值的核心命題：晶片愈昂貴、功耗愈高、系統愈複雜，測試的重要性就愈高。對低價、標準化程度高的晶片，市場更重視效率與成本；但對先進 AI 晶片，可靠度本身就是價值的一部分。

也因此，客戶買的不再只是「有沒有空機台」，而是一套能把高價值晶片穩定送進終端系統的驗證能力。從這個角度看，京元電子受益的不是泛 AI 概念，而是測試價值鏈地位的實質提升；如果它近年投入的高功耗預燒、液冷散熱與系統級測試能力能持續深化，角色就可能從測試產能供應商，升級為與客戶共同定義測試流程的工程夥伴。

至少到目前為止，已公告的營運數字對這個方向提供了初步支持。若依公司 2025 年財報與 2026 年第一季、2026 年 6 月營收公告來看，2025 年營收 349.33 億元、年增 30%，毛利率由 34.79% 升至 35.85%；2026 年第一季營收約 101.92 億元、毛利率 39.7%，到 2026 年 6 月單月營收 36.22 億元、年增 28.64%，上半年累計營收 213.34 億元、年增 36.08%。

這些結果至少說明，高階測試組合、稼動率或定價能力正在改善；但若直接把它們解讀成「AI 測試需求已被完全證實」，仍然太快，因為產品組合、景氣循環與非 AI 業務也可能同時影響營收與毛利率。

另一個不能省略的修正，是獲利品質。依公司揭露，2025 年 EPS 9.01 元包含出售中國子公司京隆科技的利益，因此不能直接視為完全可重複的正常化獲利。這不是小細節，而是估值判斷的基本功。

換句話說，目前可以較有把握地說，京元電子的營運數字對高階測試需求提供了初步支持；但還不能因為帳面獲利改善，就斷言它已進入毫無爭議的高品質長期成長軌道。

真正的賭注，是資本支出能否變成護城河

真正讓市場分歧迅速擴大的，不是營收，而是資本支出。

依公司對外規劃，2025 年資本支出上修至 370 億元，2026 年原先規劃 393.72 億元，之後再提高至 500 億元，規模甚至超過 2025 年全年營收。新增投資的重點放在成品測試、預燒與系統級測試設備，同時也承租苗栗頭份與桃園楊梅廠區，新增約 3.5 萬坪無塵室空間，使設備可使用空間增加約 50%。

這不是一般程度的擴產，而是一次相當激進的資本配置押注。

多頭會認為，這正是最有說服力的訊號。高功耗測試不是需求明朗後再補機台就能追上的生意，它牽涉廠房、電力、散熱、設備認證與客戶驗證週期，準備時間長，切入門檻也高。若管理層願意在短期內投入數百億元，合理的解讀是公司已看見一定程度的中期需求能見度，而不只是被股價氣氛推著走。

再加上 AI 基礎設施投資仍在擴大，需求來源也不只 GPU，還包含自研 ASIC 與客製化加速器，這確實讓京元電子的成長故事不再侷限於單一產品。

更值得注意的是二階效應。市場習慣用晶片出貨量去推估供應鏈，但 AI 測試的價值不一定與出貨量同步。若每顆晶片的測試時間更長、預燒要求更高、系統整合驗證更複雜，那麼每顆晶片對應的測試價值，就可能成長得比晶片數量更快。

這也是京元電子多頭論點最強的地方：它未來的成長不必完全依賴整體半導體景氣總量，而可能來自單位服務內容與單位測試時間的提升。

風險排序：先看稼動率與固定成本，再看集中與競爭

但空頭最強的反駁同樣成立，而且最該先看的，不是抽象的「景氣風險」，而是產能利用率能否撐住新增固定成本。

五百億元不是免費的選擇權，它會實際轉化為折舊、人力、電力、維護與財務壓力。只要需求遞延、客戶拉貨節奏改變，或新產能爬坡不如預期，毛利率就可能迅速反轉。依財報數字，2025 年底負債占資產比率已升至 50.15%；至於流動比率與速動比率下滑，若沒有更直接的管理層說法，較保守的表述應是：這可能與設備款支付壓力有關。

也因此，後續最重要的驗證，不是營收創高本身，而是折舊逐步進來後毛利率是否仍具韌性，以及自由現金流何時回正。

客戶與應用集中，是第二個會放大主風險的變數。京元電子客戶數量看起來很多，但獲利增量未必分散；現階段真正推升市場情緒的，很可能仍是少數 AI GPU 與 ASIC 專案。如果最大客戶調整委外比例、導入第二供應商，或產品世代交替期間出現空窗，短期波動可能比表面上的客戶覆蓋率更高。

所以投資人真正該追的，不只是服務多少大客戶，而是前五大客戶營收占比、AI 營收比重，以及 GPU 以外的業務是否正形成第二條成長曲線。

第三個風險才是競爭，而且它和前兩者互相連動。京元電子在純測試領域確實具備規模與工程經驗，但市場並非沒有替代者。大型封測集團可以把封裝與測試整合銷售，其他專業測試廠也可能擴充高階產能，一些 IDM 與大型晶片公司則保留內部測試能力。

如果整個產業因 AI 而同步擴產，今天的產能稀缺，兩三年後未必不會變成供給壓力。京元電子要證明的，不能只是「我先買到設備」，而是客戶為什麼在設備普及後仍需要它；真正的護城河，應該建立在認證週期、良率資料與客製化測試方案的黏著度上。

估值現在在交易什麼

把這些因素合起來看，問題就回到估值。

若以截至 2026 年 7 月的市場預估與券商報告區間來看，市場對 2026 年 EPS 的估計大致落在 9.5 至 10.2 元，目標價區間約 338 元至 400 元。這種評價方式，已經不是在看一家傳統封測循環股，而是在提前反映 AI 測試需求延續、毛利率維持高檔，以及新增產能順利爬坡的未來。

換句話說，現在的核心問題不是公司會不會成長，而是未來幾年的成長，能不能超過股價已內含的高期待。

也因此，京元電子現在不是「有成長就夠了」。哪怕未來幾季營收仍然亮眼，只要毛利率不再提升、資本支出回收變慢，或管理層對需求的說法開始轉弱，股價都可能承受壓力。反過來說，如果非 GPU 客戶開始放量、ASIC 測試占比提升、訂單能見度拉長，市場也可能進一步上修對它的獲利與評價。

高估值能否站穩，取決於成長是不是高品質、可延續，而且能跨客戶驗證。

最後回到五個驗證點

因此，後續最值得追蹤的，不是單一目標價，而是前面提過的五個指標。

第一，月營收年增率能否在高基期下維持，證明需求不是提前拉貨。第二，新增折舊進來後，毛利率是否仍具韌性。第三，500 億元資本支出對應的，究竟是具承諾的訂單，還是管理層對景氣高點的提前押注。第四，GPU 以外的 ASIC、CPU、網通或其他高階測試需求，能否逐步提高占比。第五，自由現金流與投資報酬率何時改善，因為只有回到現金回收，擴產才真正創造股東價值。

總結來看，京元電子的故事值得認真研究，但不適合用簡化版的 AI 概念股框架處理。它受惠的是一項合理而重要的產業變化：當晶片價值與系統風險同步上升，測試就不再只是成本中心，而可能變成價值鏈中更關鍵的驗證平台。

2025 年至 2026 年上半年的營收與毛利率改善，的確替這個方向提供了一些初步證據；但若談到獲利，仍必須記得 2025 年的 EPS 含一次性處分利益，不能直接當作完全正常化的長期水準。

所以，京元電子值不值得投資，答案不在於你相不相信 AI，而在於你相不相信它能把短期產能稀缺轉化為長期技術壁壘。真正的多頭情境，是它從測試產能供應商升級為 AI 晶片不可替代的驗證平台；真正的空頭情境，則是市場把一輪高度集中的 AI 資本支出週期，錯誤定價成永久成長。

未來幾季，答案會慢慢出現在稼動率、毛利率、自由現金流與客戶多元化之中，而不只是在營收新高或目標價上修裡。

Subscribe now

免費訂閱，不錯過後續分析與更多有價值的內容。

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

Disclaimer: For informational purposes only. Not financial advice. Invest at your own risk.

Subscribe now

NVIDIA (NVDA) is no longer a debate about whether it matters. It sits near the center of the AI economy, supplying the compute, networking, systems, and software that many leading models and AI services depend on. The investment question now is narrower and harder: can future demand, margins, execution, and customer returns stay strong enough to justify expectations that are already extremely high?

That shift has made the company unusually easy to admire and unusually hard to value. The basic quality story is not controversial. NVIDIA has strong products, a deep ecosystem, impressive execution, and financial performance that has been extraordinary by any large-cap standard. The harder investment question is not whether the business is excellent. It is whether the next several years can be good enough to exceed what investors already seem to expect.

That is the frame worth keeping in mind. If you are evaluating NVIDIA today, the core debate is not collapse versus dominance. It is whether the company’s current level of growth, profitability, and strategic control can remain durable as supply constraints ease, alternatives improve, and customers become more focused on returns.

NVIDIA’s moat is broader than the chip

It is still common to discuss NVIDIA as if its edge begins and ends with faster processors. That understates what the company has built.

NVIDIA designs accelerated computing platforms. It relies on manufacturing partners for fabrication, while concentrating its own efforts on architecture, systems, networking, and software. That division matters because it helps explain why the company can look somewhat asset-light relative to the scale of value it captures, even as it remains deeply exposed to the realities of advanced packaging, memory supply, and leading-edge foundry capacity.

The technical foundation is well understood. GPUs are well suited to highly parallel workloads, which made them useful first for graphics and later for scientific computing and machine learning. But NVIDIA’s strategic position in AI comes from more than the hardware itself. CUDA, software libraries, developer tools, model optimization, interconnects, rack design, and deployment maturity all matter. In practice, customers are not buying an isolated chip. They are buying a system that has to work at scale, stay utilized, and get into production quickly.

That is why benchmark debates often miss the economic point. A competing processor can look attractive on paper and still be the wrong choice in a real environment if it creates more software friction, longer deployment times, weaker utilization, or higher operating complexity. NVIDIA’s moat is not simply performance. It is the combination of performance, ecosystem, workflow, and speed.

That distinction matters for investors because it helps explain why margins have been so strong. NVIDIA is not being paid like a vendor selling interchangeable components. It is being paid more like a platform owner whose product sits at the most valuable bottleneck in a rapidly expanding market.

The business is extraordinary. The stock still has a harder job.

There is no need to soften the point: NVIDIA’s recent financial performance has been exceptional. In the quarter ended April 26, 2026, Data Center accounted for the overwhelming majority of revenue, non-GAAP gross margin remained around 75%, and cash generation stayed enormous. For a company often described as a semiconductor supplier, those economics are rare.

But strong businesses do not automatically produce easy investments. Once a company becomes the obvious winner in the market’s favorite theme, the threshold for surprise rises sharply.

That is where the discussion often gets muddled. Over the next two to three years, the key issue is not whether NVIDIA can keep posting strong numbers in a given quarter. It is whether growth can remain high, margins can stay unusually elevated, hyperscaler spending can hold up, major product transitions can stay smooth, and the valuation can keep supporting all of that at once. The bull case is not hidden. It is broadly understood, widely modeled, and heavily discussed.

So the next phase likely depends less on proving that NVIDIA is important and more on proving that the opportunity is broader, longer, and more defensible than even bullish investors assume. If the market already expects smooth product transitions, continued hyperscaler spending, durable high margins, and manageable competitive pressure, then simply delivering strong results may not be enough to drive the same kind of stock response that investors have grown used to.

That does not make the stock unattractive. It does mean the debate has matured. Business quality is no longer the question doing most of the work.

The next demand wave may be real, but it needs to monetize

The next demand wave should be judged through one filter above all others: customer return on AI infrastructure. Inference, broader enterprise adoption, sovereign AI projects, and physical-world applications such as robotics could all expand NVIDIA’s market. But they only strengthen the long-term thesis if they produce durable revenue, cost savings, or strategic advantage for the buyers funding the compute.

That broader demand picture matters because it could reduce NVIDIA’s dependence on a relatively small set of giant buyers. If future growth comes only from a handful of hyperscalers, investor confidence will remain closely tied to the capital allocation decisions of a few companies. If the customer base broadens materially, the revenue base becomes more resilient and the long-term thesis improves.

There is also a reasonable argument that efficiency gains do not automatically hurt NVIDIA. Lower cost per token can expand usage. If inference becomes cheaper, developers may place AI into more products, run more frequent workloads, and allow agents to handle more tasks. In that world, efficiency is not the enemy of demand. It is one of the mechanisms that expands it.

So the key downstream question is not simply whether more workloads appear. It is whether those workloads support attractive returns on the infrastructure customers are buying.

That is where the real long-term scrutiny belongs. NVIDIA’s largest customers are spending enormous sums on AI infrastructure. Some of that spending is clearly strategic. Some is tied to real revenue opportunities. Some is probably defensive, because no large platform company wants to discover too late that it underinvested in a foundational shift. Those motives can coexist, but they are not economically equivalent.

If customers are building profitable AI services, improving monetization, and earning strong returns on deployed infrastructure, then NVIDIA’s current economics may prove more durable than skeptics think. If instead much of the spending is driven by competitive fear ahead of proven monetization, then the industry could eventually face a more normal capital cycle: overbuilding, lower returns, more pricing pressure, and sharper procurement discipline.

That risk may show up outside NVIDIA before it shows up inside NVIDIA. Investors should watch cloud rental economics, utilization, commentary from customers on AI returns, and the health of smaller AI infrastructure operators. Orders can remain strong for a while even if end economics are getting weaker. Lead times, supply commitments, and prearranged capacity can delay the signal.

Competition is real, but the economics matter more than the headlines

NVIDIA’s leadership should not be confused with permanent insulation from competition. The company’s success creates a strong incentive for customers to reduce dependence on it.

AMD remains the most visible merchant competitor, but the more structural issue is custom silicon. Large cloud and platform companies have the scale, technical talent, and economic motivation to design specialized chips for selected workloads. Partners that help hyperscalers build these systems add to that threat. This is not a theoretical risk. It is a rational response to NVIDIA’s strength and pricing power.

But the existence of alternatives does not resolve the investment debate either. Custom silicon does not need to eliminate NVIDIA’s position to matter, and NVIDIA does not need to maintain every point of share to keep growing. The more relevant question is where alternatives become good enough economically across real production workloads.

That means looking beyond peak benchmark claims. What matters is total production economics: software cost, deployment speed, utilization, power efficiency, networking performance, and how well a given solution handles a mix of training and inference needs. A hyperscaler running a narrow, high-volume inference workload inside its own software stack has a clear reason to favor custom silicon if the workload is stable and utilization is high. NVIDIA is harder to displace where customers need fast iteration, broad model support, and a platform that can move from training to varied inference workloads without rebuilding the stack.

This is why moderate share loss does not automatically break the thesis. If the overall AI compute market grows fast enough, NVIDIA can still expand even while some customers route stable inference workloads to internal chips. The company’s future does not require monopoly. But its current valuation likely assumes that any erosion in share happens gradually and is offset by broader demand and deeper system-level revenue.

Product cadence is now part of the moat

One underappreciated change in NVIDIA’s business is that it increasingly has to execute like a systems company, not just a chip company.

Annual platform transitions help keep the company ahead technologically and make it harder for competitors to catch a static target. But they also raise the operating challenge. Each new cycle depends on coordination across processors, high-bandwidth memory, packaging, networking, power, and cooling. A problem in one layer can affect the whole system.

That is why the transition from Blackwell to Rubin matters as more than a normal product launch. If NVIDIA can keep moving quickly without disrupting customer plans, margins, or supply reliability, that reinforces the idea that its moat includes execution itself. If transitions become bumpier, the market will start asking whether annual cadence adds fragility as well as strength.

Margins will be one of the clearest signals here. Some pressure during a new ramp can be normal. Persistent pressure because of weaker pricing power, mix changes, or competitive concessions would tell a different story. Investors should care less about whether margins move in isolation and more about why they move.

China is not just foregone revenue

China also deserves a more strategic reading than the usual simple framing of lost sales.

Restrictions on advanced AI exports limit NVIDIA’s access to an important market and constrain one potential source of revenue. But the deeper issue is what happens when a major market is forced to build around the absence of the incumbent leader. Policy support, guaranteed domestic demand, and enough time can help local alternatives improve. Even if those alternatives start behind, insulation can give them room to become credible.

So China matters in two ways at once. It is foregone opportunity in the present, and it is a possible incubator for future competition. Investors do not need to predict that outcome with confidence to take it seriously.

A great company can still be a demanding investment

The cleanest way to describe NVIDIA today is that it remains an exceptional business facing exceptional expectations. That makes the right conclusion less a simple verdict than a set of signals to watch.

The bullish case is straightforward. AI infrastructure demand is still expanding. NVIDIA’s full-stack platform is difficult to replicate. Inference, enterprise adoption, sovereign AI, and physical-world applications could all widen the market. The company may also capture more value per deployment through systems, networking, CPUs, and software.

The cautious case is straightforward too. Customers will eventually demand proof of returns. Custom silicon will keep improving. Product transitions will become more operationally complex. China will remain strategically complicated. And once scarcity fades, the durability of current margins will face a more serious test.

That leaves investors with a more subtle conclusion than either extreme allows. NVIDIA does not need to stop being great for the stock to become harder. It only needs reality to land closer to already-high expectations than far above them. At the same time, skepticism that ignores the company’s genuine platform advantages risks underestimating how durable those economics could be.

The investment thesis, then, is not about choosing between perfection and collapse. It is about judging how long today’s rare combination of growth, margins, ecosystem power, and execution can persist as the rest of the market adapts.

The signals that matter most are fairly clear: evidence that customers are earning real returns on AI infrastructure, margin behavior through major product transitions, and the pace at which custom silicon starts taking meaningful share in stable workloads.

That uncertainty does not invalidate the thesis. It defines it.

Subscribe now

Subscribe for free and never miss future analysis and other valuable insights.

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”

Disclaimer: For informational purposes only. Not financial advice. Invest at your own risk.

Subscribe now

Unity (U) is no longer just a cleanup story. That framing is now too narrow. In its first-quarter 2026 results, Unity reported 49% year-over-year growth in Strategic Grow, alongside improving margins and free cash flow. That is enough evidence to move the debate forward. The more useful investor question now is whether Unity can become a more focused Create-plus-Grow platform, with its advertising engine doing most of the heavy lifting, and whether that growth can hold up once the easy recovery comparisons fade.

That shift matters because it changes what investors should watch. The old argument was about stabilization after strategic overreach, advertising disruption, repeated restructuring, and a serious loss of trust with developers. The new argument is about durability. Unity has shown enough improvement to earn renewed attention. It has not yet shown enough to deserve blind confidence.

The cleanest way to think about the business now is through its two connected segments. Create is the software developers use to build games and other interactive media applications. Grow covers the monetization and user-acquisition side of the business. The investment case increasingly turns on whether those two businesses can reinforce each other in a way competitors will struggle to match.

That is why Vector now sits at the center of the thesis. If Vector can sustain strong growth and convert Unity’s position inside games into better ad targeting and better advertiser outcomes, the company starts to look less like a repaired game engine and more like a differentiated software-plus-advertising platform. If not, the recent improvement may prove real but less valuable than the market hopes.

A better business picture is emerging

For many investors, Unity still carries the residue of its prior narrative. That is understandable. This was once one of the most admired software listings of the pandemic era, backed by a broad real-time 3D vision that extended far beyond gaming. But that ambition was eventually undermined by slowing growth, an acquisition-heavy strategy, advertising-platform problems, internal complexity, and the runtime-fee policy that alienated many developers.

The company now looks more disciplined. Create remains strategically important because it keeps Unity embedded in the development and operation of interactive content. That position gives Unity distribution and workflow relevance even when competition is intense. Unity says its tools are used to create roughly 70% of mobile games, though that company statistic should be read as a reach measure, not proof of equal strength across genres or platforms. Unreal remains stronger in many high-end and visually demanding projects, while Godot and other alternatives appeal to developers seeking open-source or lower-cost options.

Still, Create matters for more than direct revenue. Subscriptions, enterprise contracts, and related services are only part of the story. The deeper value is that Unity sits inside the software stack where games are built, updated, and observed. That makes the segment strategically more important than a simple revenue comparison might suggest.

Grow, meanwhile, is where the recent momentum has become more visible. This business helps publishers acquire users and monetize players through advertising, and Vector is Unity’s machine-learning system for predicting which users are likely to install, engage, and spend. In theory, this is where Unity’s broader strategy has always been heading: help create the game, see what happens once players are inside it, and use that information to improve the economics of growth and monetization.

For years, that logic was more attractive than the execution. Apple’s privacy changes hit mobile advertising hard. Unity’s own systems had weaknesses. The company also became too sprawling, with reported revenue that looked larger than the strategic coherence of the business underneath it. That disconnect is now narrowing.

The turnaround is measurable, but not complete

Unity’s recent operating results suggest this is no longer just a cost-cutting story. In Unity’s reporting, Strategic revenue is meant to isolate the businesses management is still building around by excluding operations being exited, wound down, or treated as non-core. On that basis, growth, especially in Grow, has been strong enough to suggest product-level improvement rather than expense control alone. Strategic Grow has been the standout. Strategic Create has also returned to healthier growth, even if investors still need to separate genuine adoption from the effects of pricing.

That distinction matters because portfolio simplification can make the headline numbers look weaker while improving the underlying business. Unity has been exiting or shrinking lower-priority operations, including the legacy ironSource network and the Supersonic publishing business. Those businesses added revenue, but they also inflated the topline, muddied the operating picture, and pulled attention away from the core Create-plus-Grow model. As they leave the mix, reported totals can understate the health of what remains.

This is why total reported revenue is a weaker scorecard during the transition. Strategic Create, Strategic Grow, Vector performance, margins, and cash flow say more about the business Unity is trying to keep than a mechanical comparison to older revenue bases that included operations it is now exiting.

The margin story is improving as well. Adjusted profitability has moved in the right direction, and free cash flow has strengthened. That deserves recognition. But it is equally important not to confuse better adjusted numbers with fully repaired earnings quality. Unity still reports substantial stock-based compensation and shows a meaningful gap between adjusted performance and GAAP results. A company can be genuinely improving and still not be financially clean.

That is where some investor discipline is required. The turnaround is real in the sense that the business is more focused, margins are improving, and the most strategic parts of the company are growing faster. The turnaround is not complete in the sense that GAAP profitability, earnings quality, and the long-term durability of the growth engine still need to be demonstrated rather than assumed.

Vector is now doing most of the explanatory work

The stock increasingly trades on one variable: how long Vector can keep outperforming expectations.

That is a much higher bar than the one Unity faced when investors were simply asking whether the advertising business could recover at all. Now the question is whether Vector can stay strong once the rebound gets harder. That helps explain the volatility. Reported improvement is no longer enough by itself, because the market is already assigning value to sustained Vector growth, margin expansion, and at least some future benefit from Unity’s data position.

The practical takeaway is that Unity is now an expectations story, not a simple recovery trade. A modest slowdown can matter a lot if investors are underwriting continued outperformance, while modest headline revenue can still work if Strategic Grow, guidance, and margins keep supporting the core thesis.

The most interesting part of the bull case is not just that Vector is growing. It is why it might keep growing.

Unity’s argument is that its position inside games gives it access to behavioral information that traditional ad platforms often do not see in the same way. An install is not the same as long-term engagement. A player who abandons a game after ten minutes is not equal to one who returns repeatedly, advances, and eventually spends money. If Unity can observe richer in-game behavior and translate that into better model performance, it could improve advertiser returns in a way that supports higher spend and better monetization.

This is the strategic logic behind runtime data. Management has said that more than 90% of eligible developers have opted in and that Unity was integrating runtime engine behavioral data into Vector’s AI model in the second quarter of 2026. If that company account translates into better advertiser outcomes, Unity could build a reinforcing system in which more Unity-based content produces more behavioral data, better data improves Vector, stronger ad performance attracts more budgets, and better monetization makes the ecosystem more valuable to developers.

That is the strongest version of the thesis. It is also the part that deserves the most skepticism.

High opt-in is not the same as a proven moat. Access to more data does not automatically create durable advantage. The data has to be distinctive, clean, predictive, and hard for rivals to match through other means. Unity also has to show that it can turn those signals into commercially meaningful outcomes at scale, not just into an attractive narrative.

So far, the market has more reason to view runtime data as promising optionality than as established evidence. The commercial proof still needs to deepen. Investors should want harder evidence on advertiser returns, retention, campaign expansion, and whether runtime signals produce measurable gains that competitors such as AppLovin (APP), Moloco, Meta, and Google are less equipped to replicate.

Until then, the runtime-data story is a legitimate source of upside, but not yet something that should be fully capitalized as though the moat has already been built.

AI expands the upside and the uncertainty

AI matters here because it can either widen demand for Unity’s stack or reduce the value of the engine layer, which is why it remains hard to value cleanly.

The bullish version is straightforward. If generative tools make asset creation, prototyping, and scripting cheaper, but teams still rely on Unity for deployment, physics, monetization, analytics, and cross-platform operation, then AI could increase project volume without displacing Unity from the workflow.

The bearish version is just as clear. If AI-native tools start producing more complete interactive products and shift value toward model, cloud, and distribution platforms, Unity’s role could narrow. For investors, the near-term test is practical: is AI causing more projects to start and finish inside Unity, or making Unity easier to route around?

Create looks healthier than it did. Grow looks meaningfully stronger. The portfolio is more coherent. Adjusted profitability and cash generation are improving. All of that is real, and it means Unity has crossed an important threshold. It is no longer simply promising a second act. It is beginning to perform one.

But the recovery is still concentrated. Vector is doing much of the explanatory work. The competitive advantage around runtime data is still more argued than proved. Create has improved, but investors still need to know how much of that improvement reflects real adoption rather than pricing or temporary stabilization. AI could widen Unity’s opportunity, or gradually compress its relevance.

That leaves the investment case in a disciplined middle ground rather than a clean buy. Unity deserves more serious attention than it did when the story was mostly about damage control, but the stock now depends on a narrower and more demanding set of forward assumptions. If Vector sustains high growth, if runtime data starts showing clear commercial value, and if AI reinforces rather than weakens Unity’s position, the upside could be meaningful. If those pieces remain only partially supported while expectations keep climbing, the risk rises even if the reported numbers continue to improve.

For investors, the most useful framework is simple. Watch whether Vector is taking durable share or just enjoying a strong improvement cycle. Watch whether runtime data leads to measurable advertiser outcomes rather than management promises. And watch whether AI makes Unity more central to production or easier to route around.

Unity’s second act is underway. The uncertainty has moved from survival to durability. For now, the evidence supports a watch-and-verify stance more than a full-conviction buy.

Subscribe now

Subscribe for free and never miss future analysis and other valuable insights.

Buy me a coffee

“If you find my research helpful, you’re always welcome to support my work with a coffee.”