大立光（3008）深度研究：CPO 故事是真的，價格把 2028 先付了

A real moat, a real second curve — and a price that already spent it護城河是真的，第二條成長曲線也是真的 — 但這個價格，已經把兩者都預支光了

Key considerations主要考量

(1) The variable-aperture upgrade in the iPhone 18 Pro is real revenue, already in guidance. It is the biggest camera spec jump in a decade and lifts flagship lens ASP (average selling price) by roughly 50%, with Largan as one of the primary high-end suppliers (exact share unverified) — but this is a cyclical spec bonus that recurs every one to two upgrade cycles, not a reason to permanently re-rate the multiple.

這一段在做的事：把大立光真正的優勢與盲目的樂觀炒作劃清界線。公司的實力是真的，但拆開數據會發現 — 現在的股價已經把未來的利多全數「預支完畢」了。

1 · 可變光圈是真的，但這只是常態性的「換代紅利」

真的有賺頭 — iPhone 18 Pro 升級可變光圈，讓鏡頭平均單價（ASP）拉高約 50%，大立光也是核心供應商之一。
但這不值得暴漲 — 這種「手機相機升級」在過去十年裡，每隔 1–2 年就會發生一次。它是常態性的循環紅利，不是「股價可以永久重新估值」的理由。

2 · 2026 年看似驚人的成長，其實是「匯率魔術」

破解數字陷阱 — 2026 Q2 稅後純益看似暴衝 +352% ✓ ETtoday 2026-07-09・查證 #12，但本業的營業利益只微幅成長 +5.5% ✓ simpletechtrend 2026-07・查證 #13。同一季、同一家公司、同一份財報，這兩個數字卻天差地遠 — 差別全出在「業外」。
先解釋「業外」是什麼 — 損益表分兩層：本業是靠賣鏡頭賺的（營業利益），業外是本業以外的損益，主要有三塊：利息收入（大立光帳上 1,184 億現金，光利息就很可觀）、匯兌損益（外幣資產與應收帳款隨台幣升貶產生的帳面損益）、以及投資收益。業外不是賣東西賺來的，它跟公司做得好不好沒有直接關係。
真相是什麼？2025 Q2 台幣急升，大立光單季認列匯損 42.2 億，把當季 EPS 壓到只剩 7.73 元、13 年最低 ✓ 經濟日報／旺得富 2025-07・查證 #24。今年的 +352%，是拿今年跟那個被匯損砸出來的超低基期相比 — 分母太小，成長率自然嚇人。本業根本沒有暴賺，營業利益只成長 5.5% 就是證據。

3 · 市場把「AI CPO 的夢想」畫得太滿

現實只有一小張訂單 — 目前公開的 CPO 訂單只有一張最小規格的試水溫訂單，管理層說「最快」要 2027 年年中才能量產。
外資預估差了 10 倍 — 2028 年 CPO 營收貢獻，保守派估 1–2%、樂觀派（Citi）估 20%。
股價已經站上樂觀頂峰 — 現價對應 2028 年預估獲利高達 32.6 倍本益比；即便用最極端的 275 元 EPS 來算，依然 28.4 倍。代表股價已經把「CPO 必定大成功」的超高機率提前算進去了。

4 · 連董事長自己都說了實話

管理層警訊 — 林恩平在法說會坦言「AI 讓手機鏡頭愈不需要升級，是在吃掉我們的未來」 — 連經營者都承認手機鏡頭本業正走向成熟。
沒有第二隻腳 — 今天 95% 以上的營收依然靠手機鏡頭。CPO 如果沒有完美大爆發，現在的股價就失去支撐。
高層的真實動作 — 管理層買自家庫藏股的均價是 2,425 元，比現在的 7,810 元低 69%。連管理層自己都覺得，要 2,425 元才叫便宜。

TP:目標價： NT$5,850, 6-month horizon — the derivation is broken out below. 5,850 元，期間 6 個月 — 推導拆在下面四格。

The formula算式

2026/2027 EPS modelled at NT$195 / NT$225.
Target price = 2027E EPS 225 × 26× PER = NT$5,850.

本研究預估 2026／2027 年 EPS 為 195／225 元。
目標價＝ 2027E EPS 225 × 26 倍 PER ＝ 5,850 元。

Where 26× comes from26 倍是怎麼來的

Largan's own 2023–25 PER band is roughly 12–20× ◆ estimated概算. That band is a cyclical mid-point, not a ceiling — in the 2015–17 upgrade boom the stock traded well above 20× for long stretches. 26× is "top of the band plus about 30%": a judgment call, not a derivation.

大立光自己 2023–25 年的 PER 區間概算約 12–20 倍 ◆ 概算。這個帶是循環股的中樞，不是天花板 — 2015–17 規格升級大年，它曾長期在 20 倍以上交易。26 倍是「循環帶上緣再加約三成」：這是判斷，不是數學推導。

Why not the 35–42× the market pays那為什麼不給市場現在付的 35–42 倍

The premium is real: CPO optionality plus Largan's position as one of Apple's primary high-end lens suppliers. But 35–42× means booking an unverified 2028 bull case as today's cash flow. This analysis will not.

溢價是真的：CPO 的選擇權，加上大立光身為蘋果高階鏡頭主力供應商之一的地位。但 35–42 倍等於把還沒驗證的 2028 藍天，當成今天的現金流入帳。本研究不跟。

Cross-check交叉檢核

TP 5,850 implies a PBR of ≈4.7×, against the close's 6.29×.

目標價 5,850 對應 PBR 約 4.7 倍，收盤則是 6.29 倍。

Where this view could be wrong — separating two kinds of risk這個判斷可能錯在哪裡 — 先分清楚兩種完全不同的風險 ◆

Method premise, stated plainly. The target price is 2027E EPS × 26×, and it holds only if, over the next six months, the market moves its pricing benchmark back from "the 2028 blue sky" to "2027 earnings you can verify." If the market keeps pricing off 2028, time is not on this view's side.

這一段在做的事：最誠實的自我批判 — 主動列出這個判斷可能在什麼情況下被打臉，以及為什麼即便風險很高，本頁選擇「迴避」而不是「放空」。

1 · 時間不站在這個判斷這邊（市場還在作夢）

目標價（26 倍）建立在一個前提上：未來 6 個月內，市場會從熱炒 2028 年的遠大夢想，冷靜下來看 2027 年的真實獲利。萬一未來半年市場依然盲目相信 2028 年的 AI 夢，這個預測就會落空。

2 · 最誠實的弱點：就算獲利完全命中，股價也不一定會跌

估值（倍數）與獲利（EPS）是兩回事 — 就算 2027 年真的賺了 225 元、EPS 一毛沒估錯，只要市場給的本益比沒有壓縮、繼續維持在現在昂貴的 35 倍：225 × 35 ＝ 7,875 元，依然高於現在的收盤價。
核心風險 — 這份分析最大的弱點，在於估不出市場的情緒（本益比）何時會降下來。「迴避」的風險不在 EPS 估錯，在倍數不回歸。

3 · 有哪三件事會讓股價繼續居高不下？

(a) 拿到大訂單 — CPO 突然宣布拿到大規模量產訂單（已列入失效條件）。
(b) 整個生態系被重新估值 — 台積電 COUPE／NVIDIA Rubin 帶動所有 FAU 供應商一起飛（同業上詮就是正在發生的例子）。
(c) 被動資金硬挺 — ETF 與法人不管基本面，持續買進千金股。這一項無法用基本面對沖 — 承認它。

4 · 為什麼只選擇「迴避」，而不是「放空」？

市場發瘋的時間可能比你想的更長 — 過度高估的狀態，完全可以持續超過六個月。
兩者不是同一件事 — 迴避＝「我知道這裡有風險，我不參與這場遊戲」；放空＝「我賭它近期一定會暴跌」。市場若持續發瘋，放空會賠得很難看。
所以本頁的建議是 — 不要買，但也絕對不要跑去放空它。

Risk factors: a repeat of the 2025 Q2-style FX shock (NT$4.22bn quarterly loss precedent) · CPO timeline slipping into 2028 · Apple's dual-sourcing policy structurally capping ASP · Sunny Optical catching up in Android high-end.

風險因子：台幣急升重演2025Q2式匯損衝擊（單季42.2億前科） · CPO時程滑動至2028年 · 蘋果雙供應商政策結構性壓抑ASP · 舜宇光學(2382.HK)在安卓高階急起直追。

Sources for this section (consolidated here instead of inline): price & valuation = Yahoo股市/鉅亨網 (2026-09-02) + computed · Q2/H1 2026 actuals & monthly revenue = 鉅亨網 earnings coverage (2026-07/08) · variable-aperture & iPhone 18 = TechNews (2026-06) + 時報/豐雲學堂 (2026-08) · CPO/FAU disclosures & chairman quotes = 2026-07 earnings call transcript via simpletechtrend + 工商時報/經濟日報 (2026-07/08) · consensus = FactSet via 鉅亨網 (2026-08) · buyback = 中央社 (2026-02) · estimates, TP and view = this analysis ◆. Per-number ledger: see the Sources section.本節出處（統一註記於此，不逐句插標）：價格與估值＝Yahoo股市／鉅亨網（2026-09-02）＋本研究計算 · Q2/H1 2026實績與月營收＝鉅亨網財報報導（2026-07/08）· 可變光圈與iPhone 18＝TechNews（2026-06）＋時報／豐雲學堂（2026-08）· CPO/FAU揭露與董事長發言＝2026-07法說會轉述（simpletechtrend）＋工商時報／經濟日報（2026-07/08）· 共識＝FactSet經鉅亨網轉載（2026-08）· 庫藏股＝中央社（2026-02）· 預估、目標價與觀點＝本研究判斷◆。逐數字查證表見〈資料來源〉章。

Same facts, two readings — the variant this page argues同一組事實，兩種讀法 — 本頁的分歧

This is the hero of this page's argument, expanded in full under Variant perception inside the collapsed full case.本頁論點的核心圖，完整論證見收合區塊內的變異認知章節。

The full case — business & technology · moat · bull vs bear · variant · management · competition (click to expand)完整論證 — 事業與技術・護城河・多空對決・變異認知・經營層・競爭（點開展開）

The argument in three paragraphs論點全文 — 三段講完

1 · What Largan actually sells

Start with what Largan actually sells: precision, not design. A flagship phone's main camera stacks seven or eight plastic aspheric lens elements, each injection-molded and assembled to sub-micron tolerance. Largan's own reported edge is tolerance control around 0.3μm versus roughly 0.5–0.8μm for peers — a gap that shows up as a 5x-handle gross margin versus rivals' 2–3x handle. That precision culture is now being pointed at a new target: fiber array units (FAU) for co-packaged optics, the plumbing that couples light into and out of silicon-photonics chips inside AI data-center switches. The mechanism transfer — active-alignment assembly skill learned on phone lenses, applied to optical coupling — is real and is the entire basis for a lens company plausibly becoming an optics-interconnect supplier.

一句話總結：大立光靠手機鏡頭的「超高精密技術」跨足 AI 光通訊是事實，但市場現在把「未來的 AI 夢」捧得太高，忽略了它眼前的股價已經偏貴。

1 · 大立光真正厲害的是什麼？（技術本質）

大立光的強項不只是「會設計鏡頭」，而是製造精度高到離譜。

手藝有多強？現代旗艦手機鏡頭塞了 7–8 片塑膠鏡片，大立光能把組裝誤差壓在 0.3 微米，同業約 0.5–0.8 微米 — 精度差了一倍以上。這就是它毛利率能到 50%、同業只有 20–30% 的全部原因。
AI 新故事是什麼？AI 資料中心要把光訊號接進晶片（CPO／FAU），最考驗的正是「把光纖精準對齊」。大立光把組裝手機鏡頭的手藝拿去對齊光纖 — 這項技術轉移是真的，也是市場願意把它從「鏡頭廠」重新定價成「AI 光通訊概念股」的理由。

2 · 外星人級別的估值分歧（市場在畫大餅）

對於這個 AI 新故事到 2028 年能賺多少，賣方給出的推估差了整整 10 倍：

保守派 — AI 光通訊只占 2028 年營收的 1–2%（只是小點心）。
樂觀派（Citi）— 市占衝到 10%，占營收 20%（變成大樑柱）。
核心問題 — 不管你信哪一派，現在的股價都已經按最高標準把夢算進去了：現價對應 2028 年預估獲利是 28–32 倍本益比，連最看好的報告算出來的目標價，也只比現價高一點點。

3 · 真正在賺錢的引擎 vs. 被炒作的故事

大立光目前真正的業績支撐，根本不是 AI，而是舊有的手機業務：

眼前的飯碗（手機）— iPhone 18 Pro 升級可變光圈確實能讓它賺錢，但這只是它過去十年一直在吃的「手機升級週期」，不是什麼爆發性新業務。
未來的夢想（AI 光通訊）— 管理層自己都說「最快」要 2027 年年中才能量產，而且目前拿到的只是最小規格的訂單。

結論：一家靠著「老本行手機升級」加上「一個還沒實現的 AI 夢想」的公司，股價卻已經衝到 42 倍本益比 — 代表市場已經提前把還沒發生的美好未來，先在股價上收繳了。

Business & commercial model公司與商業模式

Finding: Largan's moat is manufacturing precision, not optical design. Optical design — how many lens elements, what curvature, what materials — is broadly known industry knowledge. What separates Largan is the ability to injection-mold plastic aspheric lenses and assemble seven-to-eight-element stacks with decenter and tilt tolerances controlled to roughly 0.3μm, versus an estimated 0.5–0.8μm for competitors ✓ 2026-07 earnings call, via simpletechtrend transcription2026-07法說會，simpletechtrend整理. That precision gap is the entire explanation for Largan's mid-40%s-to-50%s gross margin against peers running 20–30%.

核心摘要：大立光的競爭優勢不是「設計鏡頭」，而是極致的製造工藝。光學設計是公開祕密，但把微小鏡片完美堆疊且不偏心的硬實力，才是它能賺取高毛利的真正護城河。

白話拆解：大立光憑什麼賺大錢？

設計沒祕密，製造才是真功夫 — 鏡頭要用幾片、彎曲度多少，業界大家都知道（就像名廚的食譜是公開的）。但要真正做出來，靠的是把塑膠鏡片完美塑形與組裝的神級手藝。
誤差只有同業的一半 — 手機鏡頭裡塞了 7–8 片鏡片，大立光的組裝公差能壓在 0.3 微米，同業只能做到 0.5–0.8 微米 ✓ 2026-07法說會，simpletechtrend整理。這種微米級的差距，反映在照片上就是清晰與模糊的分別。
技術差距變現為毛利率 — 這項精度優勢直接寫在財報上：大立光毛利率 40–50%，同業通常只有 20–30%。

擬真渲染圖：七片超薄塑膠非球面鏡片以爆炸圖方式垂直懸浮堆疊，每片邊緣有黑色遮光框與精密安裝凸緣，右側放著一個霧黑鏡筒 — **This is the product.** Seven ultra-thin plastic aspheric elements, each a few millimeters across, shown exploded in the order they stack — plus the barrel they drop into, on the right. Every element carries a black light-blocking rim and precise mounting flanges. **Largan's entire moat lives in the gaps between these discs:** decenter and tilt held to about 0.3μm, against roughly 0.5–0.8μm for peers ✓ 2026-07 call via simpletechtrend. Stack seven parts each slightly imperfect and the errors compound — which is why yield, and therefore gross margin, separates a 5-handle from a 2–3-handle. **Provenance:** a photorealistic render generated for this page, not a photograph of a Largan part. It is drawn to match the published description of the component; treat it as an accurate diagram, not as evidence about Largan's specific product. Named sources for the component description are in the caption of the schematic above.這就是產品本身。七片超薄塑膠非球面鏡片，每片直徑不過幾毫米，依照實際堆疊順序拆開來看 — 右邊那個就是它們最後要裝進去的鏡筒。每一片都有黑色遮光框與精密安裝凸緣。**大立光的護城河，整個住在這幾片鏡片之間的縫隙裡：**偏心與傾斜控制在約 0.3μm，同業約 0.5–0.8μm ✓ 2026-07 法說，simpletechtrend 整理。七片各自略有瑕疵的零件疊起來，誤差會累積 — 這就是為什麼良率、進而毛利率，會把 5 字頭和 2–3 字頭分開。**圖片來源說明：**這是為本頁生成的擬真渲染圖，**不是大立光實際零件的照片**。它依照公開描述繪製，請當成「精確的圖解」看，不要當成大立光特定產品的證據。零件描述的具名出處見上方示意圖的圖說。

Value chain — where Largan sits, and who squeezes it價值鏈 — 大立光站在哪一格，又被誰夾住

Largan sits one step upstream of the module makers and two from the phone brand. Its pricing power comes from being un-substitutable inside a long qualification cycle; the ceiling on that power is Apple deliberately keeping a second source alive.大立光站在模組廠的上一格、離手機品牌兩格。它的議價力來自「在漫長的認證週期裡無法被替換」；而這個議價力的天花板，是蘋果刻意養著第二供應商。

Unit economics — one lens, and what it earns單位經濟 — 一顆鏡頭，能賺多少

One unit = one lens, not the full VCM module.

一個單位＝一顆鏡頭，不含 VCM 模組。

60–70%

July 2026 shipments at 10MP and up2026年7月出貨：1,000萬畫素以上占比 ✓ 鉅亨網 2026-08-05

+50%

High-end lens ASP lift from the iPhone 18 Pro variable aperture可變光圈讓高階鏡頭單價提高 ✓ TechNews 2026-06-02

49.41% / 48.3%

Gross margin: H1 2026 (✓) vs trailing 4Q (◆ derived) — two calibers, never conflated毛利率：H1 2026(✓) vs 近四季(◆推算) — 兩個口徑，絕不混用

Full July mix: 10MP+ 60–70% · 20MP+ 10–20% · 8MP 0–10% · sub-5MP 10–20%. Ranges, not exact figures — the company does not disclose finer splits. Both margin calibers sit below 50%: "the 50%+ era ended three quarters ago."

7 月完整結構：1,000萬畫素以上 60–70%・2,000萬畫素以上 10–20%・800萬畫素 0–10%・500萬畫素以下 10–20%。皆為區間、非精確數字 — 公司未揭露更細分項。兩個毛利率口徑都在 5 字頭之下：「5 字頭時代已在三季前結束」。

Technology layer — how the two engines actually work技術層 — 兩台引擎實際怎麼運作

Engine 1: the variable aperture — a mechanical iris shrunk into a phone lens引擎一：可變光圈 — 把相機的機械光圈縮進手機鏡頭

The problem today

A conventional phone lens has a fixed opening. The light it lets in never changes, so software has to compensate in dim or bright scenes.

What a variable aperture does

It adds several thin overlapping metal blades that physically rotate to widen or narrow the opening — exactly the iris in a full-size camera lens, shrunk to a few millimeters across and driven by a micro-actuator.

Low light — blades open wide, more light in.
Bright light — blades stop down, sharper and less blown out.

Why this favors Largan specifically

Fitting moving mechanical blades with sub-micron clearances into a phone-thin module is a precision-manufacturing problem before it is an optics problem — which is why the upgrade favors a lens maker with Largan's tolerance-control record over a generic module assembler.

一秒白話比喻：就像人的瞳孔。光線暗時瞳孔放大（多進光）、光線強時瞳孔縮小（避免過曝）。以前手機鏡頭的瞳孔是「死」的，現在要把它做成「活」的。
為什麼利好大立光？這本質上不是光學問題，而是極致的精密微型機械製造問題 — 要讓金屬葉片在幾毫米的空間裡精準轉動、間隙控制在微米等級才不會卡住或漏光。全天下沒幾家做得到，正好是大立光的獨家強項。

現在的問題

一般手機鏡頭的光圈是固定的。進光量永遠不變，所以暗處或亮處都得靠軟體去硬補。

可變光圈做了什麼

它在鏡頭裡加了幾片薄金屬葉片，能物理性地轉動、把開口變大或變小 — 原理跟大台相機的機械光圈一模一樣，只是縮小到幾毫米見方，用微型致動器驅動。

暗處 — 葉片開大，多進光。
亮處 — 葉片縮小，畫面更銳利、不過曝。

為什麼這件事特別有利於大立光

要在手機厚度的模組裡，塞進會動、而且間隙控制在次微米級的機械葉片，這先是一個精密製造問題，才是光學問題 — 這正是為什麼這個升級偏好大立光這種有公差控制紀錄的鏡頭廠，而不是一般模組組裝廠。

編輯示意圖：智慧型手機鏡頭內部的微型可變光圈機構，多片黑色金屬葉片以精密螺旋狀重疊排列，形成一個可調整大小的圓形開口 — AI-generated illustration of a variable-aperture mechanism inside a phone camera module.AI 生成示意圖：手機相機模組內部的可變光圈機構。

Engine 2: FAU — the connector between a fiber-optic cable and a chip引擎二：FAU — 光纖跟晶片之間的那個接頭

What CPO changes

Co-packaged optics moves the light-to-electricity conversion — the optical engine — from a pluggable module at the edge of a switch box onto the same package as the switch or accelerator chip. Shorter electrical path, less power burned just moving bits between chips.

So how does light get in and out?

Through the FAU (fiber array unit) — a precision part that holds a row of hair-thin optical fibers in exact alignment and couples their light into a silicon-photonics chip. This is the part Largan is selling.

What Largan's version is made of

A four-layer FA (fiber array) — fibers pinned in place by V-groove channels: tiny machined grooves that hold each fiber at a fixed position.
A dual-row, 80-channel PMLA (prism-integrated micro-lens array) — bends the light path 90 degrees, so the fiber bundle can come in from the side while the light still lands squarely on the chip surface below.

一秒白話比喻：AI 晶片專用的光纖插頭。以前 AI 資料中心的傳輸線是銅線，現在資料量太大、傳輸太熱，必須改用光纖。FAU 就是把幾十根比頭髮還細的光纖，精準插進 AI 晶片內部的那個「微型光纖轉接頭」。
大立光賣什麼？在基板上雕出微型 V 形溝槽，把幾十根頭髮細的光纖像排排站一樣卡死，再裝上微型鏡片讓光 90 度轉彎射進晶片。
為什麼利好大立光？要把幾十根光纖精準對齊到晶片上的極小一點，用的對位組裝技術，跟它做手機鏡頭堆疊的技術是完全通用的。

CPO 改變了什麼

共同封裝光學把「光轉電」的工作 —— 也就是光引擎 —— 從交換器機箱邊緣的可插拔模組，直接搬進跟交換器或加速器晶片同一個封裝裡。電訊號路徑變短，資料中心光是把資料在晶片之間搬來搬去所燒掉的電也跟著變少。

那光要怎麼進出？

靠 FAU（光纖陣列單元） —— 一個精密零件，把一整排髮絲般細的光纖固定在精確位置，再把光耦合進矽光子晶片。這就是大立光要賣的東西。

大立光的版本由什麼組成

四層堆疊的 FA（光纖陣列）—— 用 V-groove 把光纖釘住：機械刻出的微型 V 形溝槽，讓每一根光纖都待在固定位置。
雙排 80 通道的 PMLA（整合稜鏡的微透鏡陣列）—— 把光路轉 90 度，讓光纖束可以從側面進來，而光依然垂直打在下方的晶片表面上。

✓ 工商時報 2026-07-16 · 經濟日報 2026-08-24工商時報 2026-07-16 · 經濟日報 2026-08-24 Schematic, not to scale. Largan's disclosed route couples via grating (GC); Corning's competing Glass Bridge product couples via the fiber edge (EC). Management says the two routes are expected to coexist long-term, with GC currently the larger of the two — a management claim this analysis has not independently verified.示意圖，非實際比例。大立光揭露的路線走光柵耦合（GC）；康寧的競爭產品Glass Bridge走邊緣耦合（EC）。管理層說法是兩條路線長期會並存，目前GC市占較大 — 此為管理層說法，本研究未獨立驗證市占數字。

擬真渲染圖：一塊拋光玻璃塊上，一整排髮絲般細的光纖各自嵌在與光纖同向的 V 形溝槽中，玻璃塊貼合在深色矽光子晶片上方，晶片位於墨綠色基板上，右側有鑷子做尺寸對照 — **What the schematic above looks like as an object.** A polished glass block with a row of hair-thin fibers bonded into it, each seated in its own V-shaped groove — and note the grooves run *along* the fibers, which is what pins each one to a fixed lateral position. The block sits on a silicon photonics die; the tweezers give the scale. **The whole engineering problem is in that contact plane:** every fiber core has to land on its coupler within a fraction of a micron, across dozens of channels at once, and stay there through thermal cycling. That is the same alignment problem Largan already solves in phone lenses — the entire basis of the "capability transfers" argument. **Provenance:** a render generated for this page, not a photograph of a Largan part. **上面那張示意圖，做成實物大概是這樣。**

**看到什麼 —** 一塊拋光玻璃塊，一整排髮絲般細的光纖膠合在裡面，每根坐在自己的 V 形溝槽裡。**請注意溝槽是「順著」光纖走的** — 這才能把每一根釘在固定的橫向位置。玻璃塊貼在矽光子晶片上，右邊的鑷子是尺寸對照。

**難在哪裡 —** 整個工程難題就在那個接觸面：每一根光纖的纖芯，都要落在自己的耦合點上、誤差**次微米以內**，而且是幾十個通道同時做到，還要撐過反覆的熱脹冷縮。

**為什麼跟大立光有關 —** 這正是它在手機鏡頭上**已經解掉的同一個對位問題** — 也就是本頁在檢驗的那句「能力可以遷移」的全部根據。

**圖片來源說明：**為本頁生成的擬真渲染圖，**不是大立光實際零件的照片**。它依公開描述繪製，請當成「精確的圖解」看。

Demand transduction — three chains, only one of them structural需求從哪裡來 — 三條鏈，只有一條是結構性的

One line: the switch to optics in AI (CPO) is a long swell, but Largan holds only a very small ticket to it; the phone-lens upgrade is a passing ripple — and the market is valuing the ripple as if it were the swell.

一句話摘要：AI 換光纖（CPO）是長久的大浪，但大立光目前只拿到一張很小的入場券；手機鏡頭升級只是短暫的浪花 — 而市場正把浪花當成大浪在給估值。

Structural — the long swell結構性需求 — 長久的大浪

Demand exists because the industry's architecture changed. It does not go away when the cycle turns.這個需求存在，是因為產業的架構本身變了。景氣轉頭它也不會消失。

Example: AI compute explodes → copper interconnect overheats → data centers must move to optics.例子：AI 算力大爆發 → 舊的銅線傳輸會過熱 → 資料中心一定得換成光纖。

Cyclical — the passing ripple循環性需求 — 短暫的浪花

A one-off bonus from a spec upgrade. Earn it, then wait for the next upgrade to earn it again.規格升級帶來的一波性紅利。賺完就沒了，要等下一次升級。

Example: the iPhone camera upgrades every 1–2 years — what Largan has lived on for a decade.例子：iPhone 鏡頭每 1–2 年升級一次 — 大立光過去十年就是靠這個賺錢。

The core problem: the market is valuing the ripple as the swell. Of the three chains below, only the second is a swell — and on that chain Largan currently holds a single smallest-configuration order.核心問題：市場錯把「浪花」當「大浪」在給估值。下面三條鏈裡，只有第二條是大浪 — 而大立光在那條鏈上，目前只拿到一張最小規格的訂單。

The AI buildout's scale-up bandwidth demand is running into an electrical-interconnect power wall — moving bits between chips over copper burns power and heat that does not scale the way AI clusters need it to. CPO switches and accelerators are one response, and every CPO package needs an FAU to get light in and out, which is Largan's entry point via FA + PMLA. Read top to bottom, this chain is structural. Read bottom to right, the box it lands on today is a single named order at the smallest configuration — the gap between "the industry needs this" and "Largan has captured meaningful share of it" is exactly where this analysis's skepticism lives.

AI建設的scale-up頻寬需求，正撞上電互連的功耗牆 — 用銅線在晶片之間搬資料所燒的電與熱，跟不上AI叢集需要的規模。CPO交換器與加速器是其中一個解方，而每一個CPO封裝都需要FAU把光送進送出，這正是大立光透過FA+PMLA切入的入口。由左讀到右，這條鏈是結構性的。但讀到最右邊，它今天落在的那個方框，是一張具名的最小規格訂單 — 「產業需要這個」跟「大立光已經拿到這個裡有意義的份額」之間的落差，正是本研究保持懷疑的地方。

Block-level only. This is the standard way the industry draws the shift — pluggable optics at the faceplate versus optics moved onto the ASIC's own substrate. It shows where the parts sit, not the internal optical geometry of any one of them, which Largan has not published.

只到方塊層級。這是業界描述這個轉變的標準畫法 — 光模組掛在面板上，對比光學元件被搬到 ASIC 自己的基板上。它畫的是零件的相對位置，不是任何一顆零件的內部光學幾何（大立光並未公開，本頁不畫）。

Sourced: Largan's FAU comprises a four-layer FA (V-grooves hold each fiber) plus a dual-row 80-channel PMLA that turns the light path 90°, coupling into the chip by grating (GC) rather than the edge-coupling (EC) route Corning's Glass Bridge uses ✓ 工商時報 2026-07-16 · 經濟日報 2026-08-24 · 2026-07 法說. Management says GC and EC will coexist long-term with GC currently larger — a management claim, not independently verified ⚠.

有來源的部分：大立光的 FAU 由四層堆疊的 FA（V 形溝槽固定每根光纖）加上雙排 80 通道的 PMLA 組成，把光路轉 90 度，以光柵耦合（GC）進晶片，而非康寧 Glass Bridge 走的邊緣耦合（EC）✓ 工商時報 2026-07-16・經濟日報 2026-08-24・2026-07 法說。管理層說 GC 與 EC 長期並存、目前 GC 較大 — 此為管理層說法，未經獨立驗證 ⚠。

Moat, mechanism by mechanism — and named indicators to watch護城河機制化 — 逐一拆解，附具名觀察指標

Moat 1 — molding tolerance (protects the phone-lens business)護城河一 — 模造精度（保護鏡頭本業份額）

Mechanism: ~0.3μm decenter/tilt tolerance versus an estimated 0.5–0.8μm for peers, built on 30 years of mold-tool machining know-how and accumulated process parameters ✓ 2026-07法說會.

機制：約0.3μm的偏心／傾斜公差控制，vs同業估計0.5–0.8μm，建立在30年模仁加工know-how與累積製程參數庫上 ✓ 2026-07法說會。

Erosion tripwire: Sunny Optical(2382.HK) closing the yield gap in Android high-end — and a risk the chairman himself named: if AI makes phones need lens-spec upgrades less often, this precision edge has nowhere left to monetize, regardless of how wide it stays.

侵蝕觸發點：舜宇光學(2382.HK)在安卓高階的良率追趕；以及董事長自己點名的風險 — 若AI讓手機愈來愈不需要升級鏡頭規格，這個精密度優勢就算再寬，也沒有地方可以變現。

Moat 2 — self-built assembly equipment (CPO's entry ticket)護城河二 — 自建組裝設備（CPO的入場券）

Mechanism: no off-the-shelf equipment exists for CPO/FAU assembly at volume — Largan built its own automation, so the equipment itself is the barrier, not just the process know-how ✓ 2026-07法說會，simpletechtrend整理.

機制：CPO/FAU組裝目前市面上買不到現成量產設備 — 大立光靠自建自動化做到，設備本身就是門檻，不只是製程know-how ✓ 2026-07法說會，simpletechtrend整理。

Erosion tripwire: a third-party equipment vendor shipping a standardized FAU-assembly line — the moment the tooling becomes purchasable, this barrier democratizes.

侵蝕觸發點：設備商推出標準化的FAU組裝方案 — 一旦設備可以買到，這道門檻就民主化了。

Moat 3 — active-alignment skill transfer (protects "the legitimacy to enter FAU," not a guaranteed share of it)護城河三 — 主動對位技術遷移（保護「切入FAU的正當性」，不保證拿到份額）

Mechanism: the active-alignment technique used to focus-align phone lenses transfers directly to optical coupling alignment — this is the entire basis for "a lens maker doing FAU" making sense. Citi's read is that Largan can hold yield without needing high-spec V-grooves or ultra-precision fiber, which it frames as differentiation ⚠ reported, management/analyst claim報導轉述，管理層/分析師說法.

機制：手機鏡頭組裝的主動對焦/對位技術，直接遷移到光耦合對位上 — 這是「鏡頭廠做FAU」這件事合理的全部立論基礎。Citi的看法是大立光不依賴高規格V-groove與高精度光纖也能維持良率，並將此定位為差異化 ⚠ 報導轉述，管理層/分析師說法。

Erosion tripwire: this claim has not been independently verified against a rival's yield data — treat it as a hypothesis this analysis has not stress-tested, not a confirmed edge.

侵蝕觸發點：此說法尚未經對手良率數據獨立交叉驗證 — 應視為一個尚未被本研究壓力測試過的假說，不是已確認的優勢。

Moat 4 — route position (GC) (protects the GC camp's industry position, not Largan's share within it)護城河四 — 路線位置（GC）（保護GC陣營的產業位置，不保護大立光在GC陣營內的份額）

Mechanism: Corning runs edge coupling (EC) via its Glass Bridge; Largan runs grating coupling (GC). Management says the two routes coexist long-term and GC currently holds larger share ⚠ management claim, share % not independently verified管理層說法，市占數字未獨立驗證. The distinction that matters: CPO's structural growth is close to sell-side consensus — but at the FAU node specifically, Largan is a late entrant. Foxsemicon(3363.TT) is TSMC's sole COUPE-platform supplier, Tainbow Photonics(300394.SZ) is China's NVIDIA-ecosystem leader, and Point Optics(3163.TT) sits in the Spectrum-X ecosystem. A strong industry moat does not guarantee Largan's share of it.

機制：康寧走邊緣耦合（EC）路線，靠Glass Bridge產品；大立光走光柵耦合（GC）路線。管理層說法是兩條路線長期並存，目前GC市占較大 ⚠ 管理層說法，市占數字未獨立驗證。真正重要的區分：CPO產業結構性成長幾乎是賣方共識 — 但在FAU這個節點上，大立光是後進者。上詮(3363.TT)是台積電COUPE平台獨供、天孚通信(300394.SZ)是輝達生態的中國龍頭、波若威(3163.TT)在Spectrum-X生態系裡。產業護城河強，不代表大立光在其中的份額有保障。

Leading indicators — where to check, what the trigger is領先指標 — 去哪裡查、門檻是什麼

Indicator指標	Where to check去哪裡查	Threshold門檻
Largan's own monthly revenue + self-reported monthly profit大立光月營收＋自結單月純益	MOPS (公開資訊觀測站), ~10th of each month公開資訊觀測站，每月~10日	Aug revenue ≥NT$5.5bn (this analysis's own threshold) = seasonal ramp confirmed8月營收≥55億（本研究門檻）＝旺季軌道確認
TSMC(2330.TT) COUPE platform progress平台進度	TSMC technology symposium / earnings call台積電技術論壇／法說會	upstream-most CPO signal; Foxsemicon is its sole FAU supplierCPO生態最上游訊號；上詮為其FAU獨供
NVIDIA(NVDA.O) CPO switch shipments & Rubin/Rubin Ultra timeline交換器出貨與時程	GTC / earnings calls法說會	Largan supplying Rubin Ultra is currently market speculation — any official confirm/deny is a re-rating event大立光供貨Rubin Ultra目前是市場臆測，任何官方確認/否認都是重定價事件
Foxsemicon(3363.TT) / Point Optics(3163.TT) monthly revenue月營收	MOPS	FAU-peer volume ramp, corroborating or contradicting the chainFAU同業放量旁證
Tainbow Photonics(300394.SZ) quarterly report季報	SZSE filings深交所公告	China's FAU-leader CPO revenue magnitude, as a scale reference中國FAU龍頭的CPO收入量級對照
Genius Electronic Optical(3406.TT) monthly revenue月營收	MOPS	Apple high-end lens dual-source share shift蘋果高階鏡頭雙供份額變化
USD/TWD	Central bank / Bank of Taiwan posting央行／台銀牌告	2025 Q2's NT$4.22bn FX-loss precedent — a sharp TWD rally hits EPS directly2025Q2單季匯損42.2億前科 — 台幣急升＝EPS直接侵蝕

What actually turns into volume — three chains什麼會放量 — 三條需求鏈

1. Phone-spec upgrade cycle — happening now1. 手機規格升級鏈（現在進行式）

AI-camera demand → Apple's biggest camera upgrade in a decade (variable aperture, iPhone 18 Pro, launching 2026-09-10) → high-end lens ASP +~50% → Largan as one of the primary suppliers (share unverified) ✓ 豐雲學堂 2026-08 → revenue grows through mix, not unit volume. Level: ASP/mix. Nature: cyclical (spec bonus recurs every 1–2 years). Deceleration risk: variable aperture confined to Pro Max only, yield issues, Apple reallocating share to a second source.

AI手機影像需求 → 蘋果十年最大相機升級（可變光圈，iPhone 18 Pro，2026/9/10發表）→ 高階鏡頭ASP+約50% → 大立光為高階主力供應商之一（份額未證實）✓ 豐雲學堂 2026-08 → 營收靠mix而非量成長。層級：ASP/mix。屬性：循環性（規格紅利1–2年一波）。減速條件：可變光圈僅Pro Max獨占、良率不順、蘋果分單。

2. CPO optical-interconnect chain — the second curve from 20272. CPO光通訊鏈（2027起的第二曲線）

AI data-center scale-up bandwidth demand → electrical-interconnect power/density wall → CPO switches/accelerators → FAU demand → Largan enters via FA + dual-row 80-channel PMLA, extending toward a full FAU module ✓ 工商時報 2026-07-16／經濟日報 2026-08-24. Level: new TAM. Nature: structural but not yet realized — a mass-production-tier customer's formal spec document received, samples shipped July, pilot line launching September, mass production "at the earliest" mid-2027. Deceleration risk: customer qualification fails, timeline slips to 2028, Foxsemicon/Tainbow lock in the position first.

AI資料中心scale-up頻寬需求 → 電互連功耗/密度撞牆 → 光引擎進封裝（CPO）→ 每顆CPO交換器/加速器需要FAU把光耦合進出矽光子晶片 → 大立光以FA元件+雙排80通道PMLA切入，並延伸到FAU模組一條龍 ✓ 工商時報 2026-07-16／經濟日報 2026-08-24。層級：新TAM。屬性：結構性但未兌現 — 已收到量產型客戶正式規格文件、7月送樣、試產線9月啟動、最快2027年中放量。減速條件：客戶驗證不過、時程滑到2028、上詮/天孚通信先卡位。

3. New-device optionality3. 新裝置選擇權

A foldable iPhone (expected to debut 2026-09-10) adds lens positions; humanoid-robot vision-lens orders are already in hand ✓ 今周刊 2025-01; AR glasses remain a further-out option. Level: unit volume. Nature: optionality, all currently small in magnitude. Chairman Lin En-ping told the earnings call foldable-phone lenses are "not comparatively difficult" — i.e. for Largan, foldables are a volume add-on, not an ASP step-up.

摺疊iPhone（預期2026/9/10亮相）增加鏡頭位；人型機器人視覺鏡頭訂單已到手 ✓ 今周刊 2025-01；AR眼鏡是更遠的選擇權。層級：量。屬性：選擇權，目前量級都小。林恩平在法說會親口說摺疊機鏡頭「困難度沒有相對比較高」— 即摺疊機對大立光是量的加項、不是ASP跳升項。

Bull case多方論點

The spec-upgrade cycle is real and already in hand規格升級週期真實且在手上

Variable aperture is "the biggest camera upgrade in a decade"; high-end lens ASP +~50%; Largan is one of the primary suppliers (share unverified); Q4 "will be busier than usual years" per the 2026-06-09 shareholder meeting ✓. All of this is already inside guidance.

可變光圈是「十年最大相機升級」、高階鏡頭單價+50%、大立光為高階主力供應商之一（份額未證實）；Q4「將比往年忙」（股東會6/9）✓。已在guidance內。

CPO is moving faster than anyone expectedCPO進度快於所有人預期

April "under evaluation" → July "spec document + samples shipped" → September pilot line → mid-2027 mass production; five separate land purchases totaling ~NT$6.3bn in 2026 is capital expenditure voting with its feet; self-built equipment + 0.3μm tolerance + active-alignment transfer is genuine differentiation. Still an unproven option, but the milestones are landing densely.

4月「評估中」→7月「規格書＋送樣」→9月試產線→2027年中放量；2026年五度購地共約63億元＝用資本開支投票；自建設備＋0.3μm精度＋AA技術遷移是真差異化。屬未證實的選擇權，但里程碑在密集兌現。

Profitability profile, ex-FX獲利體質

Net profit margin can return to 40%+ without FX noise (43.6% in 2024); operating margin 37%+; net cash + book value NT$1,240/share provides a downside cushion (relatively speaking — with PBR at 6.29x, that cushion is thin).

無匯損干擾下NPM可回40%+（2024年43.6%）；營益率37%+；淨現金+每股淨值1,240元提供下檔緩衝（相對而言 — PBR6.29倍下緩衝有限）。

Capital allocation turning shareholder-friendly, plus unpriced optionality資本配置轉向股東，選擇權未計價

Semi-annual dividends plus treasury-share cancellation, and robotics/foldable/AR optionality not yet in any model.

資本配置轉向股東（半年配＋庫藏股註銷）＋機器人/摺疊/AR選擇權未計價。

Bear case空方論點

Structured as fear → company claim → independent evidence → this analysis's judgment.

結構：市場恐懼 → 公司說法 → 獨立證據 → 本研究判斷。

1. The accounting illusion of growth1. 成長的會計幻象

Q2 2026 after-tax profit +352% YoY versus operating profit +5.5% YoY is the headline mismatch; July revenue also fell -10.66% YoY ✓. Full argument (the three H1 growth rates side by side, and why the gap matters) is argued once under Variant perception — not repeated here. This analysis's judgment: the core business is a mid-single-digit-growth mature franchise, not a re-accelerating one.

Q2 2026稅後+352% YoY對比營業利益僅+5.5% YoY，是頭條數字的落差；7月營收也年減10.66% ✓。完整論證（H1三個成長率並列，以及落差為何重要）僅在變異認知一次講完，此處不重複。本研究判斷：本業是中個位數成長的成熟事業，不是重加速。

2. Structural margin drift downward2. 毛利率結構性下移

Fear: gross margin has fallen below 50% for three consecutive quarters: 4Q25 47.75% → 1Q26 49.4% → 2Q26 49.41% ✓. Company claim: yield is the key variable and it is gradually improving. Independent evidence: H1 gross margin fell 4.77pp YoY, and this analysis's own trailing-4Q derivation lands at ≈48.3% ✓◆ derived推算. This analysis's judgment: 2026–27 gross margin lands at 50–51%, not back above 2024's 52.4%.

市場恐懼：毛利率連續三季跌破50%：4Q25 47.75% → 1Q26 49.4% → 2Q26 49.41% ✓。公司說法：良率是關鍵，逐步改善。獨立證據：H1毛利率年減4.77pp，本研究自行推算的近四季毛利率約48.3% ✓◆ 推算。本研究判斷：2026–27毛利率50–51%，回不到2024年52.4%之上。

3. CPO magnitude and timeline3. CPO量級與時程

The first order is a single-row FA (the smallest configuration); mass production is "at the earliest" mid-2027; the traditional sell-side camp estimates CPO at just 1–2% of 2028 revenue. Even a meaningful 2027 win adds only a low single-digit percentage to EPS that year — yet the price has already discounted Citi's 2028 bull case (20% of revenue).

首張訂單是單排FA（最小規格）；量產「最快」2027年中；傳統賣方估2028僅占營收1–2%。以2027年貢獻推估：即使拿下可觀份額，對EPS的增量在2027年是個位數百分比。價格卻已按Citi 2028藍天（20%營收占比）折現。

4. Competition and customer structure4. 競爭與客戶結構

Foxsemicon/Tainbow Photonics have already staked out the FAU node; Apple's dual-sourcing policy permanently caps ASP; the CEO himself admits AI is eroding lens-upgrade demand.

FAU節點上詮(3363.TT)/天孚通信(300394.SZ)已先卡位；蘋果雙供政策長期壓ASP；CEO自己承認AI侵蝕鏡頭升級需求。

5. FX is an invisible lever5. 匯率是隱形槓桿

A single quarter (Q2 2025) FX loss of NT$4.22bn ate 80% of that quarter's profit, dropping EPS to NT$7.73, a 13-year low ✓. A sharp TWD rally would force a full recalc of the 2026H2/2027 model.

2025Q2一季匯損42.2億（吃掉單季獲利八成、EPS剩7.73，13年低點）✓。台幣若再現急升，2026H2/2027模型全部重算。

6. Valuation is the primary risk6. 評價風險是主風險

41.5x trailing versus the stock's own three-year (2023–25) cyclical mid-point of roughly 12–20x ◆ estimated概算 — a mid-point, not a ceiling: in the 2015–17 spec-upgrade boom Largan traded well above 20x (same caliber as the TP box); the shareholder base (margin traders, day traders, passive ETF buyers) built up during a 21-day doubling amplifies any drawdown if the narrative cools.

41.5倍trailing vs自身近三年（2023–25）循環中樞約12–20倍 ◆ 概算 — 是中樞不是天花板：2015–17規格升級大年曾長期高於20倍（與TP區塊同一口徑）；21天翻倍後的籌碼結構（融資/當沖/被動ETF追買）任何敘事降溫都放大回撤。

Variant perception (the hero, shown in the summary above)變異認知（本頁核心，圖見上方摘要）

“

AI making phone camera specs less necessary to upgrade is eating our future.AI 讓手機鏡頭規格「愈不需要升級」，等於是在吃掉我們的未來。

Lin En-ping, Chairman of Largan — July 2026 earnings call, as reported ✓ simpletechtrend. Why this quote anchors the page: the structural risk to the core business is not an outside analyst's thesis. It is the person running the company, saying it out loud, on the record, about the 95%+ of revenue that is not CPO.大立光董事長林恩平 — 2026 年 7 月法說會，媒體轉述 ✓ simpletechtrend。為什麼這句話撐得起整頁：本業的結構性風險不是外部分析師的推論，是經營這家公司的人，自己在法說會上公開講出來的 — 而且講的是那 95% 以上、跟 CPO 無關的營收。

Stance — this page argues with the tape, not with the sell-side. The FactSet target-price median across 19 analysts (NT$5,150) sits 34% below the close, and even the verified highest foreign-broker target (Goldman Sachs, NT$6,231) is 20% below it. This analysis stands on the same side as the sell-side. That settles where the burden of proof lies: this page does not have to show the analysts are wrong — only what the tape is paying for.姿態：跟盤面吵，不跟賣方吵。19 位分析師的 FactSet 目標價中位數 5,150 元比收盤低 34%，連已驗證的最高外資目標價（高盛 6,231）也低 20%。本研究與賣方站在同一邊。這決定了本頁的舉證責任：不需要證明賣方錯了，只需要說明盤面在為什麼付錢。

Market narrative (since 2026-08): "the camera king becomes an AI optical-interconnect stock" — Largan re-rated from a phone-cycle name to a CPO-infrastructure supplier, sell-side target-price medians trailing behind the stock price, with the latest reports racing to raise (Nomura NT$4,310→6,000→reportedly a 5-digit call).

市場敘事（2026-08之後）：「鏡頭王變AI光通訊股」— 大立光從手機循環股重定價為CPO基礎設施供應商，FactSet目標價中位數落後股價，最新報告競相上調（野村4,310→6,000→據報導五位數）。

What this analysis agrees with (the opponent's strongest card, stated first): the manufacturing-precision moat is real and transferable — 0.3μm tolerance, self-built equipment, active-alignment transfer to optical coupling; this is not a hollow story-multiple. The milestones (spec document → samples → pilot line → land purchases) are landing densely, and a management team with a track record of blunt, honest disclosure gives "mass production mid-2027" real weight. The direction is right.

本研究同意的部分（先講對手最強的牌）：製造精度護城河真實且可遷移 — 0.3μm公差、自建設備、AA對位遷移到光耦合，這不是本夢比空殼；里程碑（規格書→送樣→試產線→購地）在密集兌現；管理層誠實紀錄讓「2027年中放量」這句話有分量。方向是對的。

Where this analysis diverges — three facts the market has skimmed past:

本研究的分歧 — 三個被市場略過的事實：

Divergence 1分歧 1

The growth is an accounting echo成長是會計的回音

+11% / +1% / +44%

H1 2026 revenue / gross profit / after-tax. Three numbers, three different stories — the market quotes the third.H1 2026 的營收／營業毛利／稅後。三個數字，三個故事 — 市場只引用第三個。

Divergence 2分歧 2

CPO barely moves 2027 EPSCPO 幾乎不影響 2027 EPS

NT$3 / 225

CPO is worth about NT$3 of the NT$225 estimate — 1.4%. The price is paid as if it were the whole story.CPO 在 225 元裡只值約 3 元 — 1.4%。但價格是按「它就是全部」在付。

Divergence 3分歧 3

The price needs a year that never happened現價要求一個沒發生過的年份

+32% vs +21.7%

Revenue growth 2027→2028 implied by today's price, against the best year of 2023–25.今天的價格隱含的 2027→2028 營收成長，對照 2023–25 最好的一年。

Most of 2026's "high growth" is an FX-loss base effect — this is the full shape of that argument, in one place: in H1 2026, three growth rates tell three different stories — revenue +11%, operating (gross) profit +1%, after-tax profit +44% ✓. The gap between them is gross-margin compression (three straight quarters below 50%); underneath that sits the 2025 Q2 FX-loss base (NT$4.22bn). Zoom into Q2 alone and the same shape repeats: after-tax profit +352% YoY against operating profit of only +5.5% YoY ✓. The market tells this story on after-tax YoY; this analysis tells it on gross and operating profit.
CPO's revenue magnitude and timeline are being compressed by an optimistic narrative: single-row FA first order → mass production "at the earliest" mid-2027 → traditional sell-side estimates 1–2% of 2028 revenue vs Citi's 20% — a ten-times spread, and the price sits on the optimistic side of it. The close is 32.6x the 2028 consensus-median EPS, and still 28.4x against the consensus-high EPS (NT$275) — even if the 2028 bull case lands in full, three years of return has already been prepaid by today's buyer.
The person selling the telescope told the truth about it: Lin En-ping said "AI making lens specs less necessary to upgrade" = "eating our future." The phone core business (95%+ of today's revenue) is structurally maturing; CPO has to succeed, or the current price has no second leg.

2026年的「高成長」大半是匯損基期效應 — 這個論證的完整形狀，一次講完：H1 2026 三個成長率各說各話 — 營收 +11%、營業毛利僅 +1%、稅後 +44% ✓。三者的落差，中間是毛利率壓縮（連三季跌破50%），下面墊著2025 Q2匯損基期（42.2億）。只看Q2單季，同樣的形狀又出現一次：稅後純益 YoY +352%，對比營業利益僅 YoY +5.5% ✓。市場用稅後YoY說故事，本研究用毛利與營益說話。
CPO收入的量級與時程被壓縮敘事：單排FA首單→2027年中「最快」量產→傳統賣方2028占比1–2% vs Citi藍天20%，十倍分歧，而價格站在樂觀那一端。現價是2028共識中位EPS的32.6倍，對照共識最高EPS（275元）仍是28.4倍 — 即便2028藍天全部兌現，三年報酬也已被今天的買家預支。
賣掉望遠鏡的人自己說了實話：林恩平「AI讓鏡頭規格愈不需要升級=吃掉我們的未來」。手機本業（今天95%+的營收）處於結構性成熟，CPO必須成功，否則現價沒有第二隻腳。

One-line variant:一句話變異認知： this is a "right story, prepaid price" — this analysis does not argue against the CPO narrative; it argues against a price that has already compressed 2028's discount rate to zero. 這是一個「對的故事、透支的價格」— 本研究不與CPO敘事作對，與「把2028藍天貼現率壓到零」的價格作對。

Management & capital allocation經營層與資本配置

Chairman Lin En-ping — second-generation family leadership — is known in the local press for "saying it straight" at earnings calls (media nickname: "the most honest chairman"). He discloses bad news himself: the 2025 Q2 FX loss of NT$4.22bn, and AI eroding lens-upgrade demand. That candor makes his guidance more credible, but it also means the guidance carries no padding.

林恩平（董事長，創辦家族第二代經營者）以「法說會有一說一」聞名（媒體稱「最老實董座」）。負面訊息自己講：2025Q2匯損42.2億、AI吃掉鏡頭升級需求。這種管理層的guidance可信度高，但也意味guidance沒有灌水緩衝。

Capital allocation is improving資本配置在改善

30 consecutive years of dividends; switched to semi-annual payouts from 2026 (H1 paid NT$41.5, payout ratio 50.39%) ✓. In 2025-12 the board approved its second-ever treasury-share programme (statutory ceiling NT$179.66bn — the legal maximum, not the actual size); by 2026-02-02 it had bought back 2,670 board-lots at an average of NT$2,424.82, NT$6.47bn in total, for cancellation ✓ 中央社. Management bought its own stock at NT$2,425; seven months later the market pays NT$7,810 — management's entry is 69% below where the market is paying today.

一句話摘要：大立光開始注重股東權益（改半年配息、實施庫藏股），而且管理層買自家股票買在超級便宜的底部；但因為市值已經膨脹到破兆，手上的千億現金，防禦力其實相當有限。

1 · 資本配置在變好（對股東變大方了）

改為半年配息 — 以前一年才發一次，現在一年發兩次（2026 上半年就發了 41.5 元，配發率 50.39%），股東更快拿到現金。連 30 年配息未中斷。
低點買回庫藏股 — 2025-12 董事會核准史上第二次庫藏股（法定上限 179.66 億元，是法定可動用上限、非實際規模）；2026-02-02 完成買回 2,670 張、均價 2,424.82 元、共 6.47 億元，將辦理註銷 ✓ 中央社。註銷後總股數變少，每股價值提高。
重點亮點 — 管理層買在均價 2,425 元，對照 7 個月後的市價 7,810 元，買點比現在低 69%。代表公司高層很精明，只在股價被嚴重低估時才動用公司的錢。

2 · 董監持股與籌碼

董監持股約 19%（2026-05）✓，大股東跟公司綁在一起；未見近期內部人大額申讓報導（未逐筆查證申報 ⚠）。

3 · 口袋極深，但「防守保護力」不夠大

現金多到嚇人 — FY2025 年底 1,159 億元（每股約 885 元）✓ 富邦資產負債表年表；2026 Q2 末 1,184.82 億元，等於每股口袋裡塞了約 905 元純現金 ✓ 合併財報現金流量表期末現金餘額118,482,131千元。每股淨值 1,240 元、庫藏股法定上限 179.66 億元皆為佐證。
但殘酷的現實是 — 現在市值高達 1.02 兆。1,185 億放在 1.02 兆面前，只占 11.6%（1,184.82÷10,231 億）。公司很有錢、不會倒，但股價太貴，萬一遭遇利空暴跌，這點現金比例擋不住跌勢。「有現金」不等於「有下檔保護」。

Insider ownership is roughly 19% (board + supervisors, 2026-05) ✓; no recent large-block insider-transaction reporting was found (this page did not check individual filings line by line ⚠). Cash & equivalents were NT$115.9bn at FY2025-end (≈NT$885/share) ✓ 富邦 balance-sheet timeline富邦資產負債表年表, and NT$118.482bn at Q2'26-end (≈NT$905.75/share) ✓ consolidated cash-flow statement, closing cash balance合併財報現金流量表期末現金餘額. Book value NT$1,240/share and the treasury-programme ceiling of NT$179.66bn corroborate distribution capacity. But scale it against the market cap: at PBR 6.29x, the cash cushion is only ≈11.6% of the ≈NT$1.02tn market cap (118.48 ÷ 1,023.1) — "having cash" is not the same as "having downside protection."

Competitive position競爭位置

Comparison比較項目	Largan 3008.TT	Genius Electronic Optical 3406.TT	Sunny Optical 2382.HK	Foxsemicon 3363.TT / Tainbow Photonics 300394.SZ上詮 3363.TT / 天孚通信 300394.SZ
Main battlefield主戰場	Apple's primary high-end lens supplier蘋果高階鏡頭主供	Apple's second source蘋果第二供應商	Android leader, automotive lenses安卓龍頭、車載鏡頭	CPO FAU (Foxsemicon: sole for TSMC's COUPE; Tainbow: China's NVIDIA-ecosystem leader)CPO FAU（上詮：台積電COUPE獨供；天孚：輝達生態中國龍頭）
iPhone 18 variable apertureiPhone 18可變光圈	One of the primary high-end suppliers (share unverified); "near-sole source" is media inference — a counter-rumor ("declined extra orders") also circulates, neither direction confirmed高階主力供應商之一（份額未證實）；「近獨供」為媒體推斷，另有「婉拒加單」等相反傳聞，兩個方向都未證實 ⚠ unverified, both directions兩方向皆未證實	Investing NT$1.86bn to expand capacity to compete for it斥資18.6億擴產搶進 ✓ CMoney 2026-08	Rumored to take 40–50% of module assembly傳聞分食模組端40–50% ⚠ unconfirmed rumor未經證實	n/a不適用
CPO positionCPO位置	Late entrant, end-to-end (FA→FAU), mass-production-tier customer spec in hand後進者、一條龍（FA→FAU）、量產型客戶規格書在手	none無	no disclosed positioning無公開布局	First mover, already inside the ecosystem先行者、已在生態系內

Geographic concentration: production is centered in Taichung (five separate land purchases in 2026 totaling ~NT$6.3bn to expand capacity, mostly in Nantun/Xitun districts) ✓ 工商時報／NOWnews 2026-08. Customer concentration: Apple is a high share of revenue (the company has not disclosed the exact ratio ⚠).

地理集中：生產基地集中台中（2026年五度購地共約63億元擴產，多在南屯/西屯）✓ 工商時報／NOWnews 2026-08。客戶集中：蘋果占比高（公司未揭露精確比例 ⚠）。

Comparable valuations — same day's close對照組估值 — 同日收盤

Company公司	Close (2026-09-02)收盤(2026-09-02)	Trailing-4Q PER近四季PER	Market cap市值	Role角色
大立光 3008.TT	7,810	41.5×	NT$1.02tn1.02兆	Apple's primary high-end lens supplier + CPO late-comer蘋果高階鏡頭主供＋CPO後進者
玉晶光 3406.TT	1,050 (+4.48%)	27.2–28.4×	~NT$118bn約1,184億 ⚠ single source, not cross-checked單一來源未二次核對	Apple's second source. PER caliber differs by source (Yahoo 27.19 / 鉅亨 28.41) — the trailing-EPS input differs; both sites agree on the price蘋果第二供應商。本益比有口徑差（Yahoo 27.19／鉅亨 28.41，差在近四季 EPS 取數）；兩站股價一致 ✓ Yahoo+鉅亨 2026-09-02
Sunny Optical 2382.HK舜宇光學 2382.HK	HK$71.50	~14×約14×	HK$77.4B	Android lens leader (EPS 5.10)安卓鏡頭龍頭(EPS 5.10) ✓ stockanalysis 2026-09-02
上詮 3363.TWO	780 (−0.76%) mkt cap NT$88.6bn市值886億	n/a無	NT$88.6bn886億	Sole FAU supplier for TSMC's COUPE — trailing-4Q EPS is −1.22, so no PER exists台積電COUPE平台FAU獨供 — 近四季EPS −1.22元，算不出本益比 ✓ 鉅亨 2026-09-02
TFC 300394.SZ天孚通信 300394.SZ	¥250.00 (−0.70%)	≈118×	¥272.7bn ≈ NT$1.29tn¥2,727億 ≈ NT$1.29兆	China's NVIDIA-ecosystem FAU leader — 1.26× Largan's market cap輝達生態中國FAU龍頭 — 市值是大立光的1.26倍 ✓ investing.com 2026-09-02 · FX CNY/TWD 4.7307·匯率CNY/TWD 4.7307

Two readings — the bull's strongest card goes first兩個讀法 — 先講多方最強的牌

1. The bull's strongest card, and it holds. Foxsemicon is the sole FAU supplier on TSMC's COUPE platform — the most orthodox beneficiary at the CPO end of this chain — and it carries a trailing-four-quarter loss of NT$1.22 per share, which is why no PER can be computed for it at all. Pricing at the CPO end has detached from earnings entirely. Against that kind of pure story stock, Largan at least has NT$188 of trailing EPS underneath its price. The argument that "Largan is the cheapest way to own CPO exposure" is a real argument, and this analysis concedes it.

一句話摘要：大立光確實是 CPO 概念股裡最有獲利支撐的一檔，但別把「比瘋狂的泡泡股便宜」當成「真的便宜」；它 95% 的收入依然靠手機鏡頭，而 41.5 倍本益比已經嚴重偏貴。

1 · 先承認多方最強的牌：大立光確實是 CPO 概念股中最有底氣的

對手全在講故事 — 上詮這種純 CPO 概念股，近四季每股還虧損 1.22 元，根本算不出本益比，股價完全在炒夢想。
大立光有真金白銀 — 近四季實打實賺了 188 元 EPS。「大立光是 CPO 籃子裡最安全、最便宜的選擇」這個說法完全成立。

2 · 打破迷思：不是 CPO 市場太小，而是市場給的估值太不理性

對手估值誇張 — 天孚通信（輝達供應鏈的 FAU 龍頭）市值 1.29 兆，比大立光還大，本益比衝到 118 倍。
所以 — 拿 118 倍或虧損的股票來跟大立光比，並不能證明大立光便宜；它只證明了市場對「CPO 概念」的定價完全失控、全憑想像力。

3 · 用「鏡頭本業」這把尺來量，大立光已經太貴

大立光 41.5 倍，對比玉晶光 27–28 倍高出約 50%，對比舜宇光學 14 倍更是貴了約 3 倍。
多出來的這筆溢價，買的完全是「AI CPO 的未來選擇權」，不是它的鏡頭本業。

4 · 本研究的最終回應

「相對便宜」不等於「真的便宜」 — 跟 118 倍的泡沫比起來便宜，只代表你買這檔的槓桿比較低，不代表它沒有跌落的風險。
萬一重力回歸（夢醒時分）— 現在大家都用「夢想」給 CPO 估值；但大立光今天有 95% 以上的營收依然來自手機鏡頭。一旦市場退燒、重新用「獲利」打分數，股價會被拉回鏡頭廠該有的水準（14–27 倍）。
結論 — 大立光無疑是這個領域品質最好的一家公司，但 41.5 倍的股價，已經遠遠超越它該有的合理價值。

4. This analysis's answer. "Cheap relative to 118×" is not cheap — it is the same trade with less leverage. "Cheap relative to the pure story stocks" only holds while the CPO theme keeps being priced as a theme. The moment pricing reverts from story to earnings, the lens ruler (27× / 14×) is the direction gravity points — and 95%+ of Largan's revenue today is still lenses. This analysis does not dispute that Largan is the best-quality name in this basket. It disputes that 41.5× is a fair price for that quality.

4. 本研究的回應。「相對 118 倍便宜」不是便宜，是同一筆交易的低槓桿版本。「相對純故事股便宜」只在 CPO 主題持續被當成主題定價時成立。一旦定價機制從故事回到盈餘，鏡頭這把尺（27 倍／14 倍）才是重力的方向 — 而大立光今天 95% 以上的營收仍然在鏡頭。本研究不否認大立光是這個籃子裡品質最好的一檔，只否認 41.5 倍是這個品質的合理價。

Pricing the option itself — what the market is paying for Largan's CPO市場為大立光的 CPO 付了多少錢？把它算出來 ◆ derived, inputs verified推算，輸入已✓

If Largan's lens business deserves Yuchang's multiple (27.5× on trailing EPS of 188.02), the lens business is worth about NT$5,171 a share. Subtract that from the NT$7,810 close and the market is paying NT$2,639 a share — about NT$345.8bn — for the CPO option alone. Valuing the lens business off a different peer does not make the number look better.

一句話摘要：拆解大立光 7,810 元的股價會發現：有 5,171 元是靠手機鏡頭撐起來的，剩下的 2,639 元（相當於 3,458 億市值）完全是市場買「AI CPO 夢想」疊上去的溢價。

第一步 · 先算「實打實的手機鏡頭本業」值多少錢？

近四季實賺 188.02 元（EPS）。借用同業玉晶光的合理本益比 27.5 倍：
188.02 × 27.5 ＝ 5,171 元／股。

第二步 · 抓出「AI 夢想」被偷塞了多少溢價？

7,810 元（現價）− 5,171 元（本業價值）＝ 2,639 元／股。
乘上總股數 ⇒ 市場單單為了這個 AI 夢想，就多給了 3,458 億元的市值。

第三步 · 換一家同業來比，只會更難看

下表是三種比法的結果 — 不論拿哪一家來比，這筆「夢想溢價」都下不來。

Which multiple the lens business is valued at鏡頭本業比照誰的倍數	Lens value (NT$/sh)鏡頭價值(元/股)	Implied CPO option隱含CPO選擇權	× Foxsemicon's entire market cap (NT$88.6bn)＝上詮整家市值(886億)的
Yuchang Optics, 27.5×比照玉晶光 27.5 倍	5,171	2,639 = NT$345.8bn	3.9×
Largan's own 2023–25 cyclical mid-point, 16×比照自己 2023–25 的循環中樞 16 倍	3,008	4,802 = NT$629.0bn	7.1×
Sunny Optical, 14×比照舜宇 14 倍	2,632	5,178 = NT$678.3bn	7.7×

Whichever peer the lens business is valued against, what the market pays for Largan's CPO option is 3.9 to 7.7 times the entire market capitalization of the sole FAU supplier on TSMC's COUPE platform — for a business where Largan is a late-comer, whose first disclosed order is the smallest configuration (a single-row fiber array), and which is worth about NT$3 of 2027 EPS on this analysis's own bridge. That is the word "overdrawn" expressed as an amount of money.

不論拿哪一家同業來替鏡頭本業定價，市場付給大立光「CPO 選擇權」的金額，都是台積電 COUPE 平台獨供商整家公司市值的 3.9 到 7.7 倍 — 而大立光在這個節點上是後進者，首張揭露的訂單是最小規格的單排光纖陣列，依本研究自己的橋只值約 3 元的 2027 年 EPS。這就是「透支」兩個字換算成金額的樣子。

Numbers, valuation & scenarios數字、估值與情境

NT$100M except per-share (NT$)億元，每股例外(元)	2023A	2024A	2025A	2026E ◆	2027E ◆
Revenue營業收入	488.42	594.58	611.48	637	695
Revenue YoY營收YoY	—	+21.7%	+2.8%	+4.2%	+9.1%
Gross margin毛利率	48.53%	52.41%	50.37%	50.1%	51.0%
Net income稅後純益	179.08	259.15	212.82	256	295
EPS (NT$)EPS(元)	134.18	194.17	159.46	195	225
This analysis vs consensus (EPS)本研究 vs 共識(EPS)	—	—	—	+4.0%	+7.8%

Share count basis: ~131M shares (post 2026-02 treasury-share cancellation). 2026E basis: H1 actual EPS 82.35 + H2E 113 (Q3E 55 / Q4E 58) = 195.35, rounded to 195. The +20.3% net-income YoY for 2026E is mainly the 2025 FX-loss base effect, not core acceleration (operating profit estimated YoY +6–8%). Consensus: FactSet, 19 analysts via 鉅亨網 (2026-08): 2026E EPS 187.55 / 2027E EPS 208.75 / 2028E EPS 239.29 (high 275.07); TP median NT$5,150. This analysis's EPS runs above consensus (peak season + variable-aperture mix); the target price is above the consensus median but 25.1% below the close — the disagreement is not about how optimistic the fundamentals are, it's about what multiple to pay.股本基礎：約1.31億股（2026/2庫藏股註銷後）。2026E基礎：H1實際EPS82.35+H2估113（Q3E55／Q4E58）＝195.35，取整為195。純益YoY+20.3%的主因是2025匯損基期，非本業加速（營業利益估YoY約+6–8%）。共識：FactSet，19位分析師經鉅亨網轉載（2026-08）：2026E EPS 187.55／2027E EPS 208.75／2028E EPS 239.29（最高275.07）；TP中位數5,150元。本研究EPS高於共識（旺季+可變光圈mix），目標價高於共識中位數、但低於現價25.1% — 分歧不在基本面樂觀度，在「付多少倍」。

TP derivation目標價推導 ◆

2027E EPS NT$225 × 26x PER = NT$5,850 (see the full derivation and multiple-justification argument in the summary's TP box above — not re-argued here). Cross-check: TP 5,850 implies PBR ≈4.7x versus the close's 6.29x.

2027E EPS 225元 × 26倍PER ＝ 5,850元（完整推導與倍數論證見上方摘要TP區塊，此處不重複論證）。交叉檢核：目標價5,850對應PBR約4.7倍（收盤6.29倍）。

◆ Anchor reconciliation: the probability-weighted scenario value is NT$5,797.5 ≈ 5,800 — within 1% of the target price. The canonical anchor stays the base-case formula (225×26=5,850); the weighted figure agrees closely enough that there is no need to argue for one over the other.

◆ 錨定調節：機率加權情境值為5,797.5≈5,800元 — 與目標價差距在1%以內。正式錨點仍以基準情境公式（225×26=5,850）為準；加權值與其高度一致，無需擇一辯護。

Breaking 2027E down — what NT$225 of EPS is actually made of2027E 拆解 — 225 元的 EPS 是由哪些東西湊出來的 ◆

The company does not disclose CPO segment revenue. The table below is this analysis's assumption set, laid open so NT$225 can be audited — not a disclosed figure.

公司不揭露 CPO 分部營收。下表是本研究的假設組，攤開來讓 225 元可以被稽核 — 不是揭露數字。

Where the revenue comes from (NT$100M)營收來自哪裡(億元)	2026E	2027E	Δ	Assumption假設依據
Phone lens core (incl. variable-aperture mix)手機鏡頭本業(含可變光圈mix)	≈637	≈681	+44	Units flat, ASP/mix +7% (variable aperture goes full-year + coverage widens) — not unit growth出貨量持平、ASP/mix +7%（可變光圈進入全年＋覆蓋擴大）— 非量成長
CPO / optical interconnect (FA→FAU)CPO／光通訊(FA→FAU)	≈0	≈14	+14	Mass production mid-2027, half a year of ramp only; 2.0% of revenue = midpoint of the base case's 1–3%2027年中量產、只認半年爬坡；占營收2.0%＝基準情境1–3%的中點
Total合計	637	695	+58

Unit math: turning "1–3% of revenue" into EPS單位數學：把「占營收 1–3%」換算成 EPS ◆ derived, inputs verified推算，輸入已✓

On core-business economics (operating margin 37%, tax 20%, 131M shares; non-operating income excluded, because interest on the NT$118.5bn cash pile and FX do not scale with CPO revenue): roughly every NT$440M of revenue ≈ NT$1 of EPS. Three consequences follow, and they are arithmetic, not opinion.

換算基準：大立光每多賣多少錢，EPS 才會增加 1 元？
扣掉營業成本與稅（營益率 37%、稅率 20%、股數 1.31 億股；業外不計 — 因為 1,184 億現金的利息與匯兌不會隨 CPO 營收成長）：大約每多賣 4.4 億元，EPS 才多 1 元。這是純粹的數學，不是意見。

2027 CPO revenue of NT$1.4bn ≈ NT$3 of EPS — 1.4% of the NT$225 estimate. CPO barely moves 2027 EPS at all. This is exactly where this analysis parts company with the price.
For CPO to contribute 10% of EPS (NT$22.5) in a single year, it needs about NT$10bn of revenue = 14% of 2027E total — between the traditional camp's 1–2% and Citi's 20%, and that is a 2028 story.
Working backwards from the price: 7,810 ÷ 28.4× ⇒ 2028 EPS of 275. Going from 225 to 275 is +NT$50, needing roughly NT$22bn of incremental revenue, +32% YoY. The best year in 2023–25 was 2024's +21.7%.

算式結果一 · 2027 年 CPO 根本賺不到什麼錢

預估 CPO 只能貢獻 14 億元營收 → 換算下來 EPS 只增加約 3 元。相較於整年預估賺 225 元，這 3 元只占 1.4%，對獲利幾乎沒有影響。這就是本研究跟市場樂觀派最大的分歧點。

算式結果二 · 要讓 CPO 顯著貢獻獲利，需要多大的營收？

如果要讓 CPO 貢獻 10% 的 EPS（增加 22.5 元），CPO 營收必須衝到 100 億元，相當於占全公司總營收的 14%。但保守派認為到 2028 年 CPO 也才占 1–2% — 要達到 14% 幾乎是天方夜譚。

算式結果三 · 從現價 7,810 元反推，市場的要求有多誇張？

現價 7,810 元要合理（給 28.4 倍本益比），2028 年 EPS 必須飆到 275 元。要從 2027 年的 225 元衝到 275 元（+50 元），營收必須額外爆發 220 億元、年增 +32%。但大立光過去三年表現最好的一年，營收成長也才 +21.7% — 現在的股價，等於要求它交出歷史上從未發生過的成長。

Gross-margin check — opening up 50.1% and 51.0%毛利率驗算 — 攤開 50.1% 與 51.0% ◆ derived, inputs verified推算，輸入已✓

2026E at 50.1%: H1 actual gross profit NT$14.432bn on revenue NT$29.209bn (49.41%) → full-year 637 × 50.1% = gross profit NT$31.91bn ⇒ H2 implied gross margin 50.68%, a 1.3pp recovery from H1. That assumes new-model yield ramps on schedule. If H2 merely holds H1's 49.41%, the full year is 49.4% and 2026E EPS falls from 195 to about 192.

這一段是在做財務模型的壓力測試 — 拆解預測的毛利率到底合不合理，以及如果沒達標，EPS 會少掉多少。

1 · 2026 年預估 50.1%：下半年良率不能掉

上半年實際 — 毛利率 49.41%。
下半年挑戰 — 要讓全年達到 50.1%，下半年必須拉到 50.68%（比上半年高 1.3 個百分點）。完全取決於 iPhone 新機鏡頭的良率能不能順利提升。
萬一沒達標 — 下半年只維持上半年的 49.41%，全年 EPS 就從 195 元降到約 192 元。

2 · 2027 年預估 51.0%：這個假設給得太過樂觀

模型內建的超樂觀假設 — 要讓 2027 年毛利率到 51.0%，等於要求「新增的那段營收」毛利率高達 61%。這需要「可變光圈」和「CPO」兩大新業務都非常賺錢才辦得到 — 這是整個模型裡最樂觀的單一假設。
回歸現實 — 如果新增業務的毛利率只跟公司平均差不多（51%），2027 年全年毛利率就降到 50.2%，EPS 從 225 元掉到約 222 元。

3 · 怎麼驗證？T8 設定的「10 月法說會毛利率是否 ≥52%」就是在驗這一格 — 檢查這個超高毛利率的預測到底有沒有吹牛。

Stress tests — one variable at a time壓力測試 — 每次只動一個變數 ◆

The T7 bear case (NT$3,330) stacks three bad things together. These do not — each moves one variable, so the reader can see which one actually matters.

T7 的空方情境（3,330）是三件壞事疊加。以下三個測試不是 — 每個只動一個變數，讓讀者看清楚哪一個才真的有份量。

Test 1 — CPO slips to 2028測試 1 — CPO 滑到 2028

2027 CPO revenue goes from NT$1.4bn to zero → EPS −3 → 222; ×26× = NT$5,772, just 1.3% below the TP. A CPO delay on its own barely touches this analysis's target price — because this analysis never put CPO into 2027 EPS to begin with. What it does hit is the multiple: a delay removes the narrative holding up 40×. If the multiple compresses to 20× at the same time, the answer is NT$4,440. A CPO delay is a multiple event, not an earnings event — the single most important sentence in this box.

2027 年 CPO 貢獻自 14 億歸零 → EPS −3 元 → 222；×26 倍＝ 5,772 元，只比 TP 低 1.3%。單獨的 CPO 延誤幾乎不影響本研究的目標價 — 因為本研究本來就沒把 CPO 放進 2027 年的 EPS。真正被打到的是倍數：延誤會拿掉支撐 40 倍的敘事。若倍數同時壓到 20 倍＝ 4,440 元。CPO 延誤是倍數事件，不是盈餘事件 — 這是本框裡最重要的一句。

Test 2 — the TWD spike repeats測試 2 — 台幣急升重演

2025 Q2's single-quarter FX loss of NT$4.22bn ate 80% of that quarter's profit and left EPS at 7.73 ✓. One repeat of the same magnitude → 2027 after-tax −NT$4.2bn ≈ EPS −32 → 193; ×26× = NT$5,018. Watch the signal shape: an FX loss hits after-tax profit, not gross margin — the market will comfort itself with "the core business is fine," and that comfort does not hold, because a meaningful share of the NT$225 estimate rests on non-operating income. Proxy to watch: the slope of USDTWD's monthly line, and the central bank's language after each board meeting.

2025 Q2 單季匯損 42.2 億、吃掉單季獲利八成、EPS 只剩 7.73 ✓。同量級再現一次 → 2027 稅後 −42 億 ≈ EPS −32 元 → 193；×26 倍＝ 5,018 元。注意訊號的形態：匯損打的是稅後，不是毛利 — 市場會用「本業沒事」安慰自己，但這個安慰不成立，因為 225 元裡有相當比重靠業外。觀察 proxy：USDTWD 月線斜率、央行理監事會後的措辭。

Test 3 — variable aperture stays Pro Max only測試 3 — 可變光圈只上 Pro Max

The ASP/mix gain halves from +7% to +3.5% → 2027 core revenue 681→659, total 673 (−NT$2.2bn); at 61% incremental gross margin (the same caliber as the gross-margin check above — this slice of revenue is the high-margin mix gain by construction), 20% tax and 131M shares → after-tax −NT$1.07bn ≈ EPS −8.2 → about 217; ×26× = NT$5,642. This is the only one of the three that resolves on 9/10 itself, and the only one that cuts both ways — coverage better than expected triggers a review of upgrading 2027E.

ASP/mix 增益自 +7% 砍半到 +3.5% → 2027 本業營收 681→659、合計 673（−22 億）；以增量毛利率 61%（與上方毛利率驗算同一口徑 — 這段營收本來就是高毛利的 mix 增益）、稅率 20%、1.31 億股推算 → 稅後 −10.7 億 ≈ EPS −8.2 元 → 約 217 元；×26 倍＝ 5,642 元。這是三個測試裡唯一在 9/10 當天就會揭曉的，也是唯一雙向的 — 覆蓋範圍優於預期即觸發 2027E 上修檢討。

How to read all three together三個測試的共同讀法

Each single-variable test lands the target price at NT$5,018 / NT$5,642 / NT$5,772 (NT$4,440 if the CPO delay also compresses the multiple) — a NT$4,400–5,800 band — all of them still far below the NT$7,810 close. Which is to say: this analysis's "avoid" conclusion does not depend on any one bad thing happening. That is where the negative asymmetry comes from — the good news is already in the price, and the bad news is not.

三個單變數測試各自把目標價打到 5,018／5,642／5,772（CPO 延誤若同時壓縮倍數則到 4,440），區間 4,400–5,800 — 全部仍遠低於收盤 7,810。也就是說：本研究的「迴避」結論不依賴任何一個壞消息發生。這就是負不對稱的來源 — 好消息已經在價內，壞消息還沒有。

Overdraft monitor — scored monthly alongside T8透支訊號監測表 — 與 T8 逐月同時打分 ◆

What to watch觀察	Healthy健康	Overdrawn透支	Now現況
Monthly revenue vs the stock月營收 vs 股價	Monthly revenue YoY turns positive and accelerates, stock moves with it月營收YoY轉正且加速、股價同向	Stock keeps making highs while monthly revenue YoY is still negative股價續創高但月營收YoY仍負	OVERDRAWN透支 July −10.66% against +96% in 21 days7月−10.66%，同期股價21天+96% ✓
Where gross margin comes from毛利率的來源	50%+ and the company attributes it to yield50%+且公司歸因良率改善	Margin recovers on mix or FX rather than yield毛利率回升但靠mix／匯率而非良率	watch觀察 3 straight quarters <50%; October call decides連三季<50%，10月法說驗
CPO milestone languageCPO里程碑用詞	"on schedule," "samples shipped," "line installed"「如期」「已送樣」「已架設」	Shifts to "under evaluation," "timeline adjusted"改口為「評估中」「時程調整」	healthy健康 end-Sept pilot line pending9月底試產線待驗
Shareholder base籌碼結構	Foreign investors buying back, local funds still buying外資回補、投信續買	Foreign investors keep selling, retail and margin buyers take the other side外資續賣、由散戶／融資接手	OVERDRAWN透支 34.15%→27.30%; net buyers on only 2 of 10 August sessions34.15%→27.30%；8月十個交易日僅2天買超 ✓
Sell-side models vs the price賣方模型 vs 價格	A broker's target catches up to the price with an auditable model behind it有券商TP追上現價並附可審模型	Median target stays far below the price; upgrades follow the price, not a modelTP中位仍遠低於現價、上調只跟價不跟模型	watch觀察 median 5,150 vs close 7,810; high 6,231中位5,150 vs 收盤7,810；最高6,231

The rule, mechanized rather than felt: any two cells sitting in the "Overdrawn" column counts as a tripwire, even in a month when revenue looks good. Two are red right now → the tripwire is already triggered. This is the mechanical reason for the "avoid" view — it is not a mood. The rule runs in reverse too: two cells turning back to "Healthy" means this view is due for re-examination.規則，機制化而非憑感覺：任何兩格落在「透支」欄＝視同絆線，即使當月營收好看。目前已有兩格轉紅 → 絆線已經觸發。這就是本研究給「迴避」的機制化理由，不是主觀情緒。反向也成立：兩格同時轉回「健康」＝本觀點必須重審。

Top: the three scenarios by probability weight. Bottom: the same three scenario prices on a shared linear axis, with the current close marked — the close sits closer to the bull scenario than to the base case.上：三個情境按機率權重呈現。下：同樣三個情境價位畫在同一條線性數值軸上，並標出現價 — 現價離多方情境比離基準情境更近。

Scenario情境	Prob.機率 ◆	Key assumptions關鍵假設	2027E EPS	PER	Scenario price情境價
Bear空方	25%	iPhone 18 sales soft / variable aperture confined to Pro Max only; TWD rally repeats; CPO slips to 2028iPhone 18銷售平淡/可變光圈僅Pro Max；台幣急升重演；CPO滑至2028	185	18×	3,330
Base基準	50%	Spec-upgrade bonus lands, GM 50–51%; CPO small-volume mid-2027 (1–3% of revenue); FX stable新機規格紅利兌現、毛利率50–51%；CPO 2027年中小量產（占營收1–3%）；匯率平穩	225	26×	5,850 =TP
Bull多方	25%	Variable aperture extends to the full 2027 lineup; CPO ahead of schedule with a disclosed order magnitude (Rubin Ultra-tier customer confirmed); 2028 visibility supports a story multiple可變光圈2027下放全系列；CPO提前且訂單量級公告（Rubin Ultra級客戶確認）；2028能見度支撐本夢比	255	32×	8,160

Asymmetry check不對稱性檢查 ◆

From the close (NT$7,810): to the base TP, -25.1%; to the bear scenario, -57.4%; to the bull scenario, +4.5%. This analysis's own bull scenario (8,160) is only 4.5% above today's close — the price you'd pay today already sits almost inside this analysis's bull case. For the stock to double from here (to ~15,620), 2028 EPS would need to reach roughly NT$450 at 35x — CPO contributing beyond even Citi's bull case, with the market permanently awarding an optical-interconnect multiple. That is beyond both management's own cautious wording ("at the earliest mid-2027") and every published sell-side model.

現價7,810元 → 基準TP：-25.1%；→ 空方情境：-57.4%；→ 多方情境：+4.5%。本研究的多方情境（8,160）只比現價高4.5% — 你今天付的價格幾乎已經是本研究的bull case。股價要從現價翻倍（約15,620）：需要2028 EPS達約450元×35倍 — 即CPO貢獻超越Citi藍天、且市場永久給光通訊倍數。超出管理層自己的說法（「最快2027年中放量」的謹慎措辭）與所有已公布的賣方模型。

Sensitivity grid (T7) — 2027E EPS × PER敏感度矩陣（T7）— 2027E EPS × PER

2027E EPS \ PER	18× (band floor)18×（歷史帶下緣）	20× (band ceiling)20×（歷史帶上緣）	26× (TP multiple)26×（TP倍數）	32× (story multiple)32×（本夢比）
185 (bear)(空方)	3,330	3,700	4,810	5,920
225 (base)(基準)	4,050	4,500	5,850 = TP	7,200
255 (bull)(多方)	4,590	5,100	6,630	*8,160*

Solid highlight = the TP cell (225×26=5,850). Dashed outline = the cell the close most closely implies: at NT$7,810, close ÷ 32× ≈ EPS 244 (between base 225 and bull 255) and close ÷ 255 ≈ 30.6× (between the TP multiple 26× and the story multiple 32×) — the current price sits just inside the bull-EPS × dream-multiple cell (8,160), only 4.3% below it. The market is simultaneously paying for the bull case on EPS and the bull case on the multiple.實心標色＝TP所在格（225×26=5,850）。虛線框＝現價最接近的格：現價7,810元÷32×≈EPS244（介於基準225與多方255之間），7,810÷255≈30.6×（介於TP倍數26×與本夢比32×之間）— 現價落在「多方EPS×本夢比」（8,160）那一格附近，只比它低4.3%。市場同時押了EPS的bull case和倍數的bull case。

Same company, one 12-month span: valuation moved from ~11× to ~43×. The close sits near the top of that range, essentially where the 52-week high was set.同一家公司，12個月內評價從約11倍走到約43倍。收盤價緊貼區間頂端，幾乎就是52週高點所在的評價水準。

Catalysts (T8)催化劑（T8）

Timing時間	Event事件	Threshold → reaction門檻 → 反應 ◆	Status狀態
2026-09-10 (Taipei time)(台北時間)	Apple fall event: iPhone 18 Pro + foldable iPhone蘋果秋季發表會：iPhone 18 Pro＋摺疊iPhone	Variable aperture across the full Pro lineup = matches expectation (already priced in); explicit rollout to the 2027 standard lineup → review upgrading 2027E; confined to Pro Max only → mix contribution shrinks, review downgrading 2027E可變光圈覆蓋全Pro系列＝符合預期（已定價）；明示下放2027標準版→2027E上修檢討；僅Pro Max獨占→mix貢獻縮水、2027E下修	scheduled已排定
~2026-09-10	August monthly revenue8月營收公告	≥NT$5.5bn (+13% MoM — this analysis's own threshold, extrapolated from July's NT$4.877bn, not company guidance) → "August is even better" confirmed, seasonal track intact; <NT$5.0bn → momentum missed, review downgrading 3Q26E≥55億（月增+13% — 本研究門檻，由7月48.77億＋月增13%推算，非公司數字）→「8月更好」兌現、旺季軌道確認；<50億→動能跳票、3Q26E下修	scheduled已排定
~2026-09-30	CPO pilot line completes (company-guided)CPO試產線完成架設（公司預告）	On schedule = neutral (already priced in); delayed → narrative damaged如期＝中性（已定價）；延誤→敘事受創	company-guided公司預告
~2026-10 (date TBD)(日期未定)	Q3 earnings + callQ3財報＋法說	GM ≥52% and EPS ≥55 → yield issues resolved, review upgrading; GM <48% → new-model yield drag confirmed, base case in doubt毛利率≥52%且EPS≥55→良率過關、上修檢討；毛利率<48%→新機良率拖累、基準情境存疑	scheduled已排定
2026 Q4	CPO pilot-line yield / customer qualification progressCPO試產良率/客戶驗證進度	Passes → 2027 mass-production probability raised; fails → timeline slips, the bull-case valuation loses its anchor通過→2027量產機率上調；未過→時程滑動、藍天派估值失錨	speculative推測
2027 mid-year年中	CPO mass production (company's own "at the earliest" language)CPO量產（公司說法「最快」）	Actual mass production + disclosed order magnitude → the second curve begins to be real, rebuild the model實際量產＋公告訂單量級→第二曲線開始兌現，重建模型	speculative推測

T8 scoring rulesT8 打分規則 ◆

Scored every month alongside the overdraft monitor in the numbers section.

與〈數字〉章的透支訊號監測表逐月同時打分。

Any T8 "review upgrading" threshold being met means rebuild the model, not add automatically — a threshold is a re-examination trigger, not a buy signal.
Two cells in the overdraft monitor turning red is a tripwire that overrides T8 scoring, however good that month's catalysts look.
An invalidation condition firing means this view is void and gets rewritten — not patched.
9/10 is the only day when two thresholds resolve at once (the Apple event and the August revenue print). Score them separately; do not let them merge into one impression.

這一段講的是監控與評估這個觀點的「紀律 SOP」。為了避免自己主觀硬撐，事前先把「檢驗門檻」與「作廢紅線」通通寫死。

1 · 打分門檻是「重審信號」，不是「無腦加碼」

每月針對 T8 的觀察項目打分。就算某項超乎預期、觸發了「上修檢討」門檻，也只會重新計算模型，不會直接自動買進 — 門檻是提醒你「該重新評估了」，不是進場訊號。

2 · 只要透支訊號轉紅，利多再好都沒用

如果透支訊號監測表裡任兩格亮紅燈，這是最高等級的絆線警訊，優先權高於一切。當月消息再利多，也必須保持戒備。

3 · 錯了就砍，不修修補補

一旦事前設定的「失效條件」被觸發，代表核心邏輯已經崩塌 — 直接把這份報告作廢重寫，不硬找藉口、不做部分微調來自欺欺人。

4 · 關鍵時間點（9/10）要客觀冷靜

當天同時公布蘋果發表會與大立光 8 月營收，是雙重試金石。打分時必須把兩件事分開獨立評估，不能混在一起憑「整體感覺」給分。

Playbook操作框架

Playbook (not investment advice)操作框架（非投資建議）

If not holding未持有	◆ Do not chase. Wait for a real entry zone, or wait for a formal CPO order-magnitude announcement before rebuilding the model.不追價。等進場區，或等CPO訂單量級正式公告後重建模型再評。
If holding若已持有	◆ Default to scaling out in tranches (harvest the windfall down to a core position). If the 09-10 event beats expectations on variable-aperture coverage and October's call shows gross margin ≥52%, the core position can continue to be held and watched.預設分批減碼（將暴利了結至核心倉位）；若9/10發表會後可變光圈覆蓋優於預期且10月法說毛利率≥52%，核心倉可續抱觀察。
Entry zone (one ruler)進場區（一把尺）	◆ First tranche is priced only. NT$4,300–5,000 (≈2027E 19–22×, at or below 15% under the base TP): entering that band buys the first tranche, with no evidence gate. Second tranche is evidence only. A formal CPO mass-production order amount is disclosed, or the 3Q26 call shows gross margin ≥52% — one of the two is enough to add; if both go the wrong way, do not add. The tripwire overrides every tranche: two cells red on the overdraft monitor, or an invalidation condition firing → stop all tranches regardless of price.第一批只看價格。4,300–5,000 元（≈2027E 19–22 倍、基準 TP −15% 以下）：進入這個區間即建第一筆，不設證據關卡。第二批只認證據。CPO 量產訂單金額正式公告，或 3Q26 法說毛利率 ≥52% — 一項達標即加、兩項皆反不加。絆線凌駕所有批次：透支監測表任兩格轉紅、或失效條件觸發 → 不論價格，所有批次先停。
Why not more aggressive (a short)為何不更激進（放空）	◆ A four-digit-priced stock with only ~131M shares outstanding, daily traded value in the NT$10bn range, CPO news flow still in an acceleration phase, dual-sided event risk from the 09-10 launch, and no low-cost TW borrow environment — negative asymmetry is a reason not to buy, not a reason to short.千金股在外流通僅1.31億股、日成交值百億級、CPO新聞流仍在加速期、9/10發表會事件風險雙向、台股無低成本借券環境 — 負不對稱是「不買」的理由，不構成「放空」的期望值。
Invalidation失效條件	◆ (1) A formal CPO mass-production order (named customer/amount) is disclosed with annualized contribution >NT$5bn → the 2028 valuation center shifts up, this view is due for review. (2) 3Q26 EPS ≥65 and gross margin ≥53% → this analysis has underestimated the spec-upgrade windfall, review upgrading.(1) CPO量產訂單客戶/金額正式公告且年化貢獻>50億元→2028中樞上移，本觀點重審；(2) 3Q26 EPS≥65且毛利率≥53%→本研究低估規格紅利，上修。

Jargon decoder名詞解碼

14 terms used across this page.

本文用到的14個術語。

14 terms — expand when you hit one you don't know14個術語 — 遇到看不懂的名詞再點開

CPO Co-Packaged Optics，共同封裝光學: Moving the optical engine (light-to-electricity conversion) from a pluggable module at the switch's edge directly onto the same package as the switch/accelerator chip. Why it matters: this is Largan's claimed second curve — the entire bull case rests on Largan capturing a meaningful slice of this new TAM. 把負責收發光訊號的光引擎，從交換器邊緣的可插拔模組，直接搬進跟交換器/加速器晶片同一個封裝裡。為何重要：這是大立光聲稱要切入的第二曲線 — 整個多方論點的立論基礎，就是大立光能在這塊新TAM裡拿到有意義的份額。
FAU Fiber Array Unit，光纖陣列單元: The precision part that holds a row of optical fibers in exact alignment and couples their light into/out of a silicon-photonics chip — the "connector" for light. Why it matters: this is the specific product Largan is entering CPO through, and it is the single number that determines whether CPO ever shows up in revenue. 把一整排光纖固定好、對準晶片上光路入口的小零件，是光訊號進出晶片的「接頭」。為何重要：大立光切入CPO的具體產品就是FAU，直接決定它能不能真正吃到CPO的營收。
FA Fiber Array，光纖陣列: The part of an FAU that pins each fiber in place using V-groove channels. Why it matters: Largan's first disclosed CPO order is a single-row FA — currently the only known production starting point. FAU裡負責固定光纖位置的部分，用V型溝槽把每一根光纖卡在精準的位置上。為何重要：大立光首張CPO訂單就是單排FA，是目前唯一已知的量產起點。
MLA/PMLA Micro-Lens Array / Prism-integrated MLA，微透鏡陣列／整合稜鏡微透鏡陣列: A row of lenses smaller than a grain of rice, paired with a prism that bends the light path 90 degrees so a fiber bundle can approach from the side while landing squarely on the chip below. Why it matters: Largan's differentiation claim is a dual-row, 80-channel PMLA that integrates this optical path into one part. 一排比米粒還小的迷你透鏡，配合稜鏡把光路轉個90度彎，讓光纖跟晶片能上下對接而不是硬要對齊同一平面。為何重要：大立光的差異化賣點是雙排80通道PMLA，把光路整合進一個零件裡。
GC vs ECGC 光柵耦合 vs EC 邊緣耦合 Grating Coupling vs Edge Coupling: GC sends light straight down into the chip through a grating structure etched on its surface; EC sends light in sideways, docking against the chip's edge like a plug. Why it matters: Largan uses GC, Corning's competing Glass Bridge uses EC — management says the two routes coexist long-term, with market share moving between them. GC是讓光從晶片正上方垂直照射進去、靠表面刻的柵欄狀結構把光導入；EC是讓光從晶片側邊水平打進去，像插頭一樣對接。為何重要：大立光走GC路線，康寧Glass Bridge走EC路線，管理層說兩條路線會長期並存，市占互有消長。
Silicon photonics矽光子 Silicon Photonics: Using the same silicon material chips are made from to build optical circuits, so light signals can be "printed" onto a chip and processed alongside electronic ones. Why it matters: the silicon-photonics chip is exactly what an FAU has to connect to — it's the computing substrate the whole CPO story is built on. 用做晶片的矽材料來做光學迴路，讓光訊號可以跟電子訊號一樣被「印」在晶片上處理。為何重要：矽光子晶片就是FAU要對接的目標，是整個CPO故事的核心運算載體。
COUPE: TSMC's own CPO packaging platform, integrating the optical engine into the chip-packaging process. Why it matters: Foxsemicon(3363.TT) is COUPE's sole FAU supplier — Largan's most direct competitor on this exact node. 台積電自己的CPO封裝平台，把光引擎整合進晶片封裝製程。為何重要：上詮是COUPE平台的FAU獨家供應商，是大立光在CPO賽道上最直接的競爭對手。
Variable aperture可變光圈 Variable Aperture: The lens opening can widen or narrow automatically, like a full-size camera's iris — wide open in low light, stopped down in bright light. Why it matters: the iPhone 18 Pro's variable aperture raises high-end lens ASP by ~50% — this is the most direct revenue engine Largan has in hand right now. 手機鏡頭裡的光圈大小可以像相機一樣自動調整，暗處開大光圈多進光、亮處縮小光圈避免過曝。為何重要：iPhone 18 Pro導入可變光圈，讓高階鏡頭單價提高約50%，是大立光眼前最直接的營收動能。
Plastic aspheric lens塑膠非球面鏡片 Plastic Aspheric Lens: A plastic lens with a specially curved (not spherical) surface, correcting image distortion with fewer elements than a spherical design would need. Why it matters: this is Largan's core business — stacking 7–8 of these to sub-micron tolerance is the company's entire manufacturing identity. 不是球面（表面弧度處處相同）而是特殊曲面設計的塑膠鏡片，能用更少片數修正影像變形。為何重要：這是大立光的本業命脈，公司核心能力就是把7–8片這種鏡片疊在一起、公差控制在次微米級。
7P/8P: A lens stacks 7 or 8 plastic elements — generally, more elements signals a higher-end spec. Why it matters: element count is a quick proxy for lens tier, and a clue to ASP. 一顆鏡頭裡疊了7片或8片塑膠鏡片的意思，片數越多通常代表規格越高階。為何重要：片數是鏡頭規格的簡易指標，也是ASP（平均售價）高低的線索。
Active alignment (AA)AA 主動對位 Active Alignment: While assembling a lens or optical part, image quality is measured in real time and the part is micro-adjusted until it hits the optimal position, rather than assembled to a fixed blueprint position. Why it matters: Largan's claim to legitimacy in CPO is transferring this exact skill from phone-lens assembly to optical-fiber coupling. 組裝鏡頭或光學元件時，一邊即時量測成像品質、一邊微調零件位置，直到對到最佳點才固定，而不是照圖面死板組裝。為何重要：大立光把手機鏡頭的AA技術遷移到FAU的光耦合對位，是它切入CPO領域的技術正當性來源。
V-groove V型溝槽: A row of V-shaped micro-channels machined into a substrate, holding each fiber at a fixed, precisely spaced position. Why it matters: this is the core mechanical structure of an FA — if the precision isn't there, the light simply doesn't line up with the chip. 在基板上刻出一排V字形的細溝，讓每一根光纖能卡進固定位置、彼此間距精準一致。為何重要：是FA（光纖陣列）定位光纖的核心機械結構，精度不夠光就對不準晶片。
PER Price-to-Earnings Ratio，本益比: Stock price divided by earnings per share — how many years of today's profit you're paying up front for one share. Why it matters: Largan's trailing PER has jumped above 40x, well outside its own 3-year band of roughly 12–20x — the core of this page's valuation argument. 股價除以每股盈餘，代表你用今天的價格，要花幾年的獲利才能「回本」。為何重要：大立光近四季PER已跳到40倍以上，遠高於過去三年12–20倍的區間，是本文評價論點的核心。
Cyclical stock循環股 Cyclical Stock: A stock whose profits swing up and down with an industry spec-upgrade or capacity cycle, unlike a steady consumer-staple name. Why it matters: Largan's earnings ride the phone-lens spec cycle; CPO would be a genuine second curve if it lands, but it has not yet changed the stock's underlying cyclical nature. 獲利會隨產業規格升級週期忽高忽低的股票，不像民生消費股那樣穩定。為何重要：大立光的獲利跟著手機鏡頭規格週期起伏，CPO若成功會是額外的「第二曲線」，但目前尚未改變它循環股的本質。

Data provenance — every number, sourced資料來源 — 每個數字都有出處

54 claims: 47 ✓ verified · 5 ⚠ flagged · 2 ◆ this analysis's derivation — expand the ledger61項查證：55✓已查證 · 3⚠未核實 · 3◆本研究推導 — 點開展開逐項表

Claim主張	Verdict判定	Value · source · as-of數值 · 來源 · 日期
Price & valuation (masthead)價格與估值（頁首）
1. Close NT$7,810 (+3.03% from NT$7,580), settled 13:30 TPE — ✗→✓ corrected from an earlier NT$7,850 last-intraday-quote estimate used mid-build收盤7,810元（較前收7,580 +3.03%），台北時間13:30已收盤 — ✗→✓ 修正自建置期間曾使用的7,850盤中估計值	✓ settled close已收盤	Yahoo股市 2026-09-02 13:30
2. 52-week range 2,020–7,99552週區間2,020–7,995	✓	鉅亨網 2026-09-02
3. Market cap ≈NT$1.03tn市值約1.03兆	✓	鉅亨網 2026-09-02
4. Paid-in capital NT$1.307bn / ~131M shares outstanding股本13.07億元／流通約1.31億股	✓	鉅亨網＋中央社 2026-02-02
5. PBR 6.29× / book value NT$1,240.07/sharePBR 6.29倍／每股淨值1,240.07元	✗→✓ was 6.35x — a ratio left over from the build-time price, not recomputed when the close was corrected; 7,810 ÷ 1,240.07 = 6.298 and 鉅亨 shows 6.29 the same day前版 6.35 倍是建置期價格對應的舊比值，收盤更正時未跟著重算；7,810÷1,240.07＝6.298，鉅亨同日亦顯示 6.29	鉅亨網 2026-09-02＋本研究計算鉅亨網 + this analysis
6. Trailing 4Q EPS NT$188.02近四季EPS188.02	✓	財報狗 2026-09
7. Dividend yield ≈1.02% (80 ÷ 7,810)殖利率約1.02%（80÷7,810）	✗→✓ 鉅亨 shows 1.23%, but 1.23% × 7,810 = NT$96, which does not reconcile with the FY25 dividend of NT$80 stated on this page; its numerator and caliber are undisclosed, so the auditable 80 ÷ close is used instead鉅亨顯示 1.23%，但 1.23%×7,810＝96 元，與本頁自列的 FY25 股利 80 元對不起來；分子與口徑未揭露，故改用可稽核的 80÷收盤	本研究計算（股利80元 ✓ 財報狗）this analysis (dividend ✓ 財報狗)
8. Foreign ownership: 34.15% (2025-05) → 27.30% (2026-08-28); foreign investors were net sellers through August; local funds (投信) bought 5 straight sessions from 8/26 (they were still net sellers of 107 board-lots on 8/25)外資持股：2025-05為34.15% → 27.30%（2026-08-28）；8月外資淨賣；投信自8/26起連5買（8/25尚賣超107張）	✓ Yahoo chip page and ifa.ai agree at 27.30%; the reviewer's 9/1 read of 26.64% is consistent with a further ~670 board-lots of foreign selling on 8/29–9/1 (≈0.5pp)Yahoo籌碼頁與ifa.ai同為27.30%；審閱記9/1為26.64%，與8/29–9/1外資續賣約670張（≈0.5pp）方向一致 ✗→✓ "17 straight sessions" was wrong — corrected to 5原「投信連17買」為錯誤，更正為連5買	Yahoo股市籌碼頁＋ifa.ai 2026-09-02
9. Insider (board+supervisor) ownership ≈19.02% (2026-05-17)董監持股約19.02%（2026-05-17）	✓	ifa.ai
Financial actuals財務實績
10. Q2 2026 revenue NT$13.665bn (QoQ -12.1% / YoY +17.1%)Q2 2026營收136.65億（QoQ-12.1%／YoY+17.1%）	✓	鉅亨網 2026-07-09
11. Q2 2026 EPS 35.72 / after-tax profit NT$4.67bnQ2 2026 EPS35.72／稅後46.7億	✓	鉅亨網／時報 2026-07-09
12. Q2 2026 after-tax profit +352% YoYQ2 2026稅後YoY+352%	✓	ETtoday 2026-07-09
13. Q2 2026 operating profit only +5.5% YoYQ2 2026營業利益YoY僅+5.5%	✓	simpletechtrend 2026-07
14. H1 2026 revenue NT$29.209bn / after-tax profit NT$10.793bn (+44.38%) / EPS 82.35H1 2026營收292.09億／稅後107.93億（+44.38%）／EPS82.35	✓	鉅亨網 2026-07-09
15. H1 2026 gross margin 49.41% (YoY -4.77pp); Q2 YoY -4.22ppH1 2026毛利率49.41%（YoY-4.77pp）；Q2 YoY-4.22pp	✓	鉅亨網 2026-07-09
16. July revenue NT$4.877bn (MoM +31.4% / YoY -10.66%)7月營收48.77億（MoM+31.4%／YoY-10.66%）	✓	鉅亨網 2026-08-05
17. Jan–Jul cumulative revenue NT$34.086bn (+7.49%)1–7月營收340.86億（+7.49%）	✓	鉅亨網 2026-08-05
18. July self-reported net profit NT$1.826bn / EPS 13.94; Jan–Jul EPS 96.297月自結純益18.26億／EPS13.94；1–7月EPS96.29	✓	鉅亨網 2026-08
19. July shipment mix: 10MP+ 60–70%7月出貨結構10MP+占60–70%	✓	鉅亨網 2026-08-05
20. FY2025: revenue NT$61.148bn / GM 50.37% / after-tax NT$21.282bn (-17.9%) / EPS 159.46FY2025：營收611.48億／毛利率50.37%／稅後212.82億（-17.9%）／EPS159.46	✓	鉅亨網法說 2026-01-08
21. FY2024: revenue NT$59.458bn / GM 52.41% / after-tax NT$25.915bn / EPS 194.17FY2024：營收594.58億／毛利率52.41%／稅後259.15億／EPS194.17	✓	鉅亨網／ETtoday 2025-01-09
22. FY2023: revenue NT$48.842bn / GM 48.53% / after-tax NT$17.908bn / EPS 134.18FY2023：營收488.42億／毛利率48.53%／稅後179.08億／EPS134.18	✓	鉅亨網 2025-01
23. 4Q25 EPS 50.4; 3Q25 EPS ≈53 (derived)4Q25 EPS50.4；3Q25 EPS≈53（推導）	✓◆ derived推導	鉅亨網法說 2026-01-08 + this analysis's derivation鉅亨網法說 2026-01-08＋本研究推導
24. Q2 2025 FX loss NT$4.22bn / EPS 7.73 (13-year low)Q2 2025匯損42.2億／EPS7.73（13年低）	✓	經濟日報／旺得富 2025-07
25. Q1 2025 EPS 48.28 / gross margin 54.63%Q1 2025 EPS48.28／毛利率54.63%	✓	Yahoo／時報 2025-04
Dividends & treasury shares股利與庫藏股
26. FY2025 dividend NT$80 total; 2026 H1 semi-annual payout NT$41.5 (payout ratio 50.39%, ex-dividend 08-13); 30 consecutive years of dividends2025年度股利合計80元；2026H1半年配41.5（配發率50.39%，8/13除息）；連30年配息	✓	財報狗＋鉅亨網 2026-07-27
27. Treasury shares: 2,670 board-lots, avg. NT$2,424.82, NT$647M total, to be cancelled, statutory ceiling NT$17.966bn (a market record for the ceiling)庫藏股：2,670張、均價2,424.82、6.47億、將註銷、上限179.66億創紀錄	✓	中央社 2026-02-02＋工商時報 2025-12-27
iPhone 18 variable apertureiPhone 18可變光圈
28. Variable aperture, decade's biggest camera upgrade, lens cost +50%iPhone18 Pro可變光圈、十年最大升級、鏡頭成本+50%	✓	TechNews 2026-06-02
29. Largan is one of the primary high-end suppliers for the iPhone 18 variable aperture (share unverified)大立光為iPhone 18可變光圈高階主力供應商之一（份額未證實）	✓ "one of the primary suppliers" — reported「主力供應商之一」報導轉述 ⚠ "near-sole source" = rumor, media inference, never confirmed by the company or Apple (a counter-rumor also circulates — see #54)「近獨供」＝傳聞：媒體推斷，未經公司或蘋果證實（另有反向傳聞，見#54）	時報／豐雲學堂 2026-08
30. Sunny Optical rumored to take 40–50% of module assembly舜宇分食可變光圈模組40–50%	⚠ unconfirmed supply-chain rumor供應鏈傳聞，未經證實	supply-chain trade press 2026供應鏈媒體 2026
31. Genius Electronic Optical investing NT$1.86bn to expand capacity, targeting variable aperture玉晶光18.6億擴產劍指可變光圈	✓	CMoney 2026-08
32. Apple event Taipei time 09-10; foldable iPhone possibly debuting蘋果發表會台北時間9/10；摺疊iPhone可望亮相	✓	瘋先生／eprice 2026-08-27
CPO / FAU
33. CPO: mass-production-tier customer spec document + samples shipped July + pilot line by end-Q3 + mass production "at the earliest" mid-2027CPO：量產型客戶規格文件＋7月送樣＋Q3底試產＋最快2027年中放量	✓	鉅亨網／ETtoday法說 2026-07-09
34. CPO pilot line launching September, 0.3μm precision (2026-06-09 shareholder meeting)CPO試產線9月啟動、精度0.3μm（股東會6/9）	✓	豐雲學堂 2026-06
35. First order is a single-row FA首單為單排FA	✓ reported轉述	simpletechtrend 2026-07
36. FAU = FA + MLA; Largan's 4-layer stacked FA + dual-row 80-channel PMLA; extending end-to-end to the FAU moduleFAU=FA＋MLA；大立光四層堆疊FA＋雙排80通道PMLA；一條龍到FAU模組	✓	工商時報 2026-07-16／經濟日報 2026-08-24
37. Supplying NVIDIA's Rubin Ultra = market speculation (not officially confirmed)供貨輝達Rubin Ultra＝市場臆測（非官宣）	✓ confirmed AS speculation確認為臆測	經濟日報 2026-08-24
38. GC vs EC routes; Corning's Glass Bridge uses EC; GC currently has larger market share (management framing)GC vs EC路線、康寧Glass Bridge走EC、GC市占目前較大（管理層說法）	✓ management claim (share % not independently verified)管理層說法（市占未獨立驗證）	simpletechtrend轉述法說
39. Tolerance 0.3μm vs peers' 0.5–0.8μm; self-built equipment公差0.3μm vs同業0.5–0.8μm；自建設備	✓ reported from earnings call轉述法說	simpletechtrend 2026-07
40. 2026: five land purchases totaling ~NT$6.3bn to expand capacity2026年五度購地共約63億擴產	✓	工商時報／NOWnews 2026-08-26 起
41. Chairman Lin: AI reducing lens-upgrade need = "eating our future"; foldable lenses "not comparatively difficult"林恩平：AI減少鏡頭升級=「吃掉我們的未來」；摺疊鏡頭難度不高	✓ reported from earnings call轉述法說	simpletechtrend 2026-07
42. Humanoid-robot vision-lens orders already in hand (2025-01 earnings call)機器人視覺鏡頭訂單到手（2025-01法說）	✓ reported報導轉述	今周刊 2025-01-09
Consensus, sell-side targets & price action共識、賣方目標價與股價動能
43. FactSet consensus: 19 analysts; 2026E EPS 187.55 / 2027E EPS 208.75 / 2028E EPS 239.29 (high 275.07); TP median NT$5,150FactSet共識：19人；2026E187.55／2027E208.75／2028E239.29（高275.07）；TP中位5,150	✓	鉅亨網 via LINE 2026-08
44. Goldman Sachs NT$3,423→6,231 (7/1, verified highest); Citi NT$3,060→5,325; four major foreign brokers' highest NT$6,125 (07-13); Nomura NT$4,310→6,000; a US broker NT$3,535→5,325高盛3,423→6,231（7/1，已驗證最高）；Citi3,060→5,325；四大外資最高6,125（7/13）；野村4,310→6,000；美系3,535→5,325	✓	工商時報 2026-07-02／TechNews 2026-07-01／udn／經濟日報
45. Nomura's latest "5-digit" target price野村最新「五位數」目標價	⚠ reported, original figure not obtained報導提及，原始數字未取得	search transcription 2026-09搜尋轉述 2026-09
46. Stock doubled in 21 trading days (NT$3,670 on 07-30 → NT$7,195 on 08-27/28, +96%); limit-up to NT$6,290 on 08-2621天股價翻倍（7/30 3,670→8/27-28 7,195，+96%）；8/26漲停6,290天價	✓	今周刊 2026-08-26／工商時報 2026-08-26
47. Traditional sell-side estimates CPO at 1–2% of 2028 revenue; Citi's bull case models 10% FAU share → 20% of revenue賣方傳統派估CPO 2028占營收1–2%；Citi藍天估2028市占10%→占營收20%	✓ reported, two camps stated side by side轉述，兩派並列	simpletechtrend／經濟日報 2026-08-24
48. A well-known local commentator ("the NT$100M professor"): CPO could push EPS above NT$300; "NT$6,000 is Largan performing at its potential"億元教授：CPO放量EPS可逾300、6,000元＝發揮實力	✓ opinion屬意見	今周刊 2026-08-26
2026-09-02 external review round2026-09-02 外部審閱修訂
49. 4Q25 gross margin 47.75%; 1Q26 gross margin 49.4%, EPS 46.63 (directly confirmed)4Q25毛利率47.75%；1Q26毛利率49.4%、EPS46.63（直接證實）	✓	Yahoo／豐雲學堂 2026-04法說
50. Trailing-4Q gross margin ≈48.3% (= (FY25 gross profit 308.0 − H1'25 142.9 + H1'26 144.32) ÷ (611.48 − 262.5 + 292.09)); H1'26 gross profit 144.32, +1% YoY近四季毛利率≈48.3%（=（FY25毛利308.0−H1'25 142.9＋H1'26 144.32）÷（611.48−262.5＋292.09））；H1'26營業毛利144.32億，年增僅1%	✓◆ inputs verified, ratio derived輸入均查證，比率為本研究推算	時報 2026-07-09＋本研究推算
51. Goldman Sachs initiated/reiterated Buy 7/1, TP NT$3,423→6,231 (market's highest at the time); 2026-28 net-profit estimates raised 6%/11%/11%高盛7/1評等買進、TP 3,423→6,231（當時市場最高）；2026-28淨利上調6%/11%/11%	✓	工商時報 2026-07-02／TechNews 2026-07-01
52. Cash & equivalents NT$115.942bn at FY2025-end (≈NT$885/share)現金及約當現金2025年底115.942億元（每股約885元）	✓	富邦資產負債表年表
53. Cash & equivalents NT$118,482,131k = NT$118.482bn at Q2 2026-end (≈NT$905.75/share)2026 Q2末現金及約當現金118,482,131千元＝1,184.82億元（每股約905.75元）	✓ closing cash balance on the consolidated cash-flow statement; note the summary pages at Yahoo / 富邦 / goodinfo do not carry this line, so no secondary source was cross-checked this round合併財報現金流量表期末現金餘額；註：Yahoo／富邦／goodinfo摘要頁皆不列此欄位，本輪未取得二次源複核 ✗→✓ ZH cell previously read 118.4億 — a 10x unit error, corrected原中文欄誤植「118.4億元」（差10倍），已更正	合併財報現金流量表 2026Q2＋ifa.ai（每股現金）consolidated cash-flow statement 2026Q2 + ifa.ai
54. A counter-rumor ("Largan declined extra orders") circulates alongside the near-sole-source rumor「大立光婉拒加單」等反向供應鏈傳聞	⚠ rumor, neither direction is a load-bearing argument傳聞，兩個方向都不作為論證支柱	外部審閱轉述 2026-09-02
55. Yuchang Optics 3406.TT close NT$1,050, trailing PER 27.19x玉晶光 3406.TT 收盤 1,050 元、近四季本益比 27.19 倍	✓	Yahoo股市 2026-09-02 14:30
56. Sunny Optical 2382.HK close HK$71.50, PER 14.09x, EPS 5.10舜宇光學 2382.HK 收盤 HK$71.50、本益比 14.09 倍、EPS 5.10	✓	stockanalysis 2026-09-02
57. Foxsemicon 3363.TWO close NT$780 (−6.00 / −0.76%, prev close 786), market cap NT$88.64bn; trailing-4Q EPS −1.22 (3Q25 +0.20 / 4Q25 −0.22 / 1Q26 −0.98 / 2Q26 −0.22), hence no PER上詮 3363.TWO 收盤 780 元（−6.00／−0.76%，昨收 786）、市值 886.40 億；近四季 EPS −1.22 元（2025Q3 +0.20／4Q25 −0.22／1Q26 −0.98／2Q26 −0.22），故無本益比	✗→✓ first draft printed NT$7,806 — Yahoo's page concatenated the price "780" with the change "6.00", and that page's own figures were self-contradictory (786 + 6.00 ≠ 7,806). Cross-verified at 鉅亨 as 780.初稿誤植 7,806 — Yahoo 頁面把股價「780」與漲跌「6.00」串接，且該頁自身數字即自相矛盾（昨收 786＋6.00 ≠ 7,806）。經鉅亨交叉驗證為 780。 ✓ quarterly EPS reported; the four-quarter sum is this analysis's calculation四季EPS為公告數，加總為本研究計算	鉅亨 2026-09-02（價格）＋Yahoo股市（季度EPS）鉅亨 2026-09-02 (price) + Yahoo (quarterly EPS)
58. TFC 300394.SZ ¥250.00 (−1.77 / −0.70%, prev close 251.77); PE ≈118×; market cap ¥272.7bn ≈ NT$1.29tn (1.26× Largan's)天孚通信 300394.SZ ¥250.00（−1.77／−0.70%，昨收 251.77）；本益比約118倍；市值 ¥2,727億 ≈ NT$1.29兆（＝大立光的1.26倍）	✗→✓ first draft used a stale 2026-08-17 quote of ¥275.69. The 250.00 / −1.77 / −0.70% read and investing.com's same-day 250.70 / −1.07 / −0.43% both work off the same prev close of 251.77, so they agree; PE and market cap are investing.com's, rescaled to 250.00初稿誤用2026-08-17舊報價¥275.69。250.00／−1.77／−0.70% 與 investing.com 同日 250.70／−1.07／−0.43% 同以昨收251.77起算，兩者一致；PE與市值取自investing.com並按250.00折算	investing.com 2026-09-02
59. FX rate CNY/TWD = 4.7307 (used for TFC's market-cap conversion)匯率 CNY/TWD ＝ 4.7307（用於天孚市值換算）	✓	Google Finance 2026-09-02
60. Foxsemicon shares outstanding 113,640,567; 780 × 113.64M = NT$88.6bn market cap, matching 鉅亨's display上詮在外流通股數 113,640,567 股；780 × 1.1364億股＝市值886億元，與鉅亨顯示一致	✓ share count reported; the product is this analysis's reconciliation股數為公告數，乘積為本研究對帳	Yahoo股市 profile 2026-09-02＋鉅亨Yahoo profile 2026-09-02 + 鉅亨
61. Implied value of Largan's CPO option NT$345.8bn (lens business priced at Yuchang's 27.5×); NT$629.0bn / NT$678.3bn at 16× / 14× anchors大立光 CPO 選擇權隱含價值 3,458 億元（鏡頭本業以玉晶光27.5×定價）；換錨16×／14× 則為 6,290／6,783 億	◆ this analysis's derivation; inputs (EPS 188.02, close 7,810, 131M shares) all verified本研究推導；輸入（EPS 188.02、收盤7,810、股數1.31億股）皆已✓	本研究計算this analysis

Verdict count: 55 VERIFIED / 3 UNCERTAIN (#30, #45, #54) / 3 derived calculations (#23, #50, #61). Two items were upgraded from UNCERTAIN to VERIFIED this round (#8 foreign-ownership level, #53 Q2 cash) and three were corrected in place (#1 close price, #8 the 投信 buying streak, #53 a ZH unit error). UNCERTAIN items are consistently flagged ⚠ or presented with rumor/reported-speech framing on the page — never used as an argument's load-bearing support.判定統計：VERIFIED 55／UNCERTAIN 3（#30、#45、#54）／派生計算3（#23、#50、#61）。本輪2項自UNCERTAIN升級為VERIFIED（#8外資水準值、#53 Q2現金），3項原位更正（#1收盤價、#8投信買超天數、#53中文欄位單位）。UNCERTAIN項在頁面一律標⚠或以「傳聞/報導」語氣呈現，不作為論證支柱。

Update log更新紀錄

Append-only, and reserved for material events — earnings, guidance changes, a catalyst resolving, or a full /thesis-update refresh. Drafting and layout work is not recorded here.附加式紀錄，且只記重大事件 — 財報、財測變動、催化劑揭曉，或一次完整的 /thesis-update 更新。撰稿與排版過程不列入。

◆ initial初次評估 2026-09-02 — Initial coverage 2026-09-02 — 初次報告

✗→ 2026-09-02 Entry preserved as written; the close price in it was later corrected. This entry used NT$7,850 (a last-quote-before-close estimate, taken while the settled print was not yet retrievable) and an implied return of -25.5%. The verified settled close is NT$7,810 and the implied return is -25.1% — see the entry above and claim #1 in the Sources ledger.條目照原樣保留；其中收盤價事後經修正。本條目使用的7,850元是建置期間（尚未能取得結算價時）的最後報價估計，隱含報酬-25.5%。已驗證結算收盤為7,810元，隱含報酬-25.1% — 詳上方條目與資料來源查證表#1。

First build of this page. Research view: Avoid. Target price NT$5,850 (6-month horizon, 2027E EPS 225 × 26x PER), implied return -25.5% vs the NT$7,850 close. No prior model exists to revise against. Close price note: repeated live pulls during this build returned the same last-available intraday quote (13:03 TPE, 25 minutes before the 13:30 bell); the settled official close could not be retrieved — see the ⚠ note under the masthead.

本頁首次建置。研究觀點：迴避。目標價 5,850元（6個月，2027E EPS 225×26倍PER），對收盤7,850元隱含報酬-25.5%。無前一版模型可供對照。收盤價說明：本次建置多次即時拉取皆回傳同一筆最後可得盤中報價（台北時間13:03，距13:30收盤25分鐘）；未能取得交易所正式結算收盤價 — 詳頁首下方⚠說明。

Date日期	Action動作	View觀點	TP	Close收盤
2026-09-02	Initial coverage初次報告	Avoid迴避	5,850	7,850 ✗→

Append-only ledger. Entries and rows are never deleted, only corrected in place (✗→✓) or superseded by a new row.附加式(append-only)紀錄。條目與列永不刪除，僅原位修正（✗→✓）或以新列取代。

Research view band研究觀點區間	6-month absolute return6個月絕對報酬
Buy買進	> +15%
Range-bound區間	-10% ~ +15%
Avoid迴避	< -10%
Watch觀察	—