高通 Qualcomm (NASDAQ: QCOM)

The full argument論點全文

Qualcomm in one sentence: it used to make its money selling phone chips. That business is shrinking now, but as long as anyone in the world sells a phone that gets online, Qualcomm still collects a patent fee on it. At the same time it is trying to sell chips into cars and cloud data centers. The stock price sits stuck in the middle: the old engine hasn't finished winding down, and the new engine's revenue hasn't shown up yet.

用一句話說高通：它以前靠賣手機晶片賺錢。現在手機這門生意在縮，但只要世界上還有人賣能上網的手機，高通幾乎都能抽一筆專利費。同時它在試著把晶片賣進汽車和雲端機房。股價現在卡在中間：舊引擎還沒掉完，新引擎的營收還沒出現。

Many people think of Qualcomm as simply "the phone-chip company." That is only half true.

很多人以為高通「就是做手機晶片的」。這個說法其實只對一半。

Most of Qualcomm's revenue, about 87%, comes from selling chips (a unit called QCT). The other 13% or so comes from patent fees (a unit called QTL, which works like a toll booth). The fee business is smaller but far more profitable: out of every $100 of revenue, chips keep about $30 before tax, while patent fees keep about $70.

高通的營收主要來自賣晶片，約占八成七（這個部門叫 QCT）。剩下約一成三來自收專利費（這個部門叫 QTL，等於一座收費站）。專利費的錢雖然少，賺得卻比較多：每 100 元營收，晶片約賺 30 元，專利費約賺 70 元。

For the past year the market has worried about two things. The first is Apple: after six years of work, Apple built its own modem chip, and most new iPhones in 2026 use it instead of Qualcomm's. The second is memory: memory chip prices spiked, and phone brands cut specs and shipments to control cost.

過去一年，市場一直在為兩件事擔心。第一件是蘋果：蘋果花了六年自己做出數據機，2026 年新 iPhone 大多改用自家晶片。第二件是記憶體：記憶體價格大漲，手機品牌為了控制成本，砍了規格和出貨量。

Both are real, and both already show up in the earnings reports and guidance. In the third quarter of fiscal 2026, handset-chip revenue fell 20% year over year.

這兩件事都是真的，也都已經寫在財報和財測裡。2026 財年第三季，手機晶片營收年減 20%。

But two things get talked about less. First, even when Apple doesn't use Qualcomm's chip, it still has to pay Qualcomm's patent fee — and it just renewed that license in September. Second, Qualcomm landed its first large cloud customer in data centers, Amazon, and the contract is written into a filing with the SEC.

但有兩件事比較少人講。第一，蘋果不用高通的晶片，還是要付高通專利費，而且 9 月剛續約。第二，高通在資料中心拿到了第一張大型雲端客戶的合約，對象是亞馬遜，合約寫進了向證管會申報的文件。

So this report's judgment is: Qualcomm is not a sinking phone-chip company, and it is not yet an AI-chip company either. It is a company whose profit is near a trough while it swaps engines. Today's price roughly reflects that state, which is why the view here is range-bound. The real answer arrives with the December-quarter earnings report.

所以這份報告的判斷是：高通不是一家正在沉沒的手機晶片公司，也還不是一家 AI 晶片公司。它是一家獲利處在谷底、正在換引擎的公司。現在的股價大致反映了這個狀態，所以觀點是區間。答案要等 12 月那一季的財報揭曉。

The company and its business model公司與商業模式

Qualcomm makes money two ways: selling chips, and collecting patent fees — and the fee business runs more than twice the margin of the chip business.

高通靠兩門生意賺錢：賣晶片，和收專利費，後者利潤率是前者的兩倍多。

• Chip division (QCT, Qualcomm CDMA Technologies): FY2025 revenue $38.37B, up 16% year over year, 87% of total revenue. It splits into three parts: handsets $27.79B, automotive $3.96B, IoT (internet of things — connected devices outside phones and PCs) $6.62B. Pre-tax margin 30%.
• 晶片部門（QCT，Qualcomm CDMA Technologies）：2025 財年營收 383.7 億美元，年增 16%，占總營收 87%。裡面分三塊：手機 277.9 億、車用 39.6 億、物聯網 66.2 億。稅前利潤率 30%。
• Licensing division (QTL, Qualcomm Technology Licensing): FY2025 revenue $5.58B, flat year over year, only 13% of total revenue. Pre-tax margin 72%.
• 授權部門（QTL，Qualcomm Technology Licensing）：2025 財年營收 55.8 億美元，和前一年持平，只占總營收 13%。稅前利潤率 72%。
• FY2026 by quarter: handset-chip growth ran +3%, -13%, -20% across the first three quarters — getting worse. Automotive ran +15%, +38%, +61% — getting better. IoT ran about +9% each quarter. Licensing stayed roughly flat.
• 2026 財年的變化：手機晶片三季年增分別是 +3%、-13%、-20%，越來越差。車用是 +15%、+38%、+61%，越來越好。物聯網每季約 +9%。授權部門大致持平。
• In other words: Qualcomm's largest business is shrinking, while its smallest ones are growing. The company sits in the middle of a handoff.
• 換言之，高通最大的生意在縮，最小的幾塊生意在長，公司正處在交棒的中間。

Qualcomm designs its chips but owns no chip factories. It hands production to outside contract chipmakers, a setup called fabless. ✓ 10-K · FY25

高通只設計晶片，自己不蓋工廠。晶片交給外面的晶圓代工廠（專門替別人生產晶片的工廠）生產，這種模式叫無晶圓廠（fabless）。 ✓ 10-K · FY25

The two new flagship chips launched in September 2026 are made by TSMC on its 2-nanometer process, TSMC's most advanced. ✓ ETtoday · TechTimes · 09-23

2026 年 9 月發表的兩顆新旗艦晶片，由台積電用最先進的 2 奈米製程生產。 ✓ ETtoday · TechTimes · 09-23

Why it matters: the most advanced wafers cost the most. In July the finance chief said costs were rising across wafers, packaging, testing and memory, and that Qualcomm is raising its own chip prices to pass them on. So the handset-chip profit margin now depends on those price increases sticking. ✓ call transcript · 07-29

這跟高通的利潤有關：製程愈先進，晶圓愈貴。財務長在 7 月說，晶圓、封裝（把晶片包成可以焊上主機板的樣子）、測試和記憶體的成本全面上漲，高通正在把這些成本反映到自己的晶片售價上。所以手機晶片的利潤率，要看漲價能不能守住。 ✓ 逐字稿 · 07-29

The handset-chip weakness has two causes. First, Apple switched to its own chip. Second, memory prices: the memory cost inside a phone jumped, so brands either cut corners on the processor or simply built fewer phones. The finance chief said the 2026 handset market would run about ten-some percent below 2025.

手機晶片變差的原因有兩個。第一是蘋果改用自家晶片。第二是記憶體漲價：一支手機的記憶體成本大漲，品牌只好在處理器上省錢，或乾脆少做幾支。財務長說 2026 年手機市場會比 2025 年少十幾個百分點。

This point moves opposite to memory stocks generally: the more expensive memory gets, the more memory makers earn, and the harder Qualcomm's handset chips are to sell.

這一點和記憶體股是反向連動的：記憶體越貴，記憶體廠越賺，高通的手機晶片越難賣。

The technology layer — what Qualcomm actually sells技術層：高通到底賣什麼

Qualcomm's core technology has three layers: one integrated phone chip, one modem that is genuinely hard to build, and one library of communications patents.

高通的核心技術有三層：一顆整合的手機晶片、一套很難做的數據機、一整座通訊專利庫。

1. The phone chip (System on a Chip, SoC)一、手機晶片（系統單晶片，SoC，System on a Chip）

One phone chip does four jobs at once: it thinks, it draws the screen, it runs AI, and it makes the call. Engineers call this a "system on a chip" (SoC — several once-separate chips squeezed onto one piece of silicon, like fitting a whole small computer onto a postage stamp). Qualcomm's Snapdragon packs in: a CPU for general computing (Qualcomm's own Oryon core), a GPU for graphics (Adreno), an NPU for AI computing (Hexagon), and the modem and radio system that connects to the cellular network. The newest generation, launched in September 2026, comes in two versions: the 8 Elite Gen 6 and a higher-performance 8 Elite Extreme Gen 6. Both are made by TSMC on its 2-nanometer process, and their main cores reach 5GHz for the first time. So the flagship chips keep improving on schedule. The handset revenue problem is Apple leaving and memory costs squeezing Android phones, not a weak product. ✓ Qualcomm · 09-22

一顆手機晶片同時做四件事：想事情、畫畫面、跑 AI、打電話。工程師把這種晶片叫「系統單晶片」（把好幾個本來各自獨立的晶片，縮在同一片矽上，像把一整台小電腦塞進一張郵票）。高通的 Snapdragon 裡面有：負責一般運算的 CPU（高通自己設計的 Oryon 核心），負責畫面的 GPU（Adreno），負責 AI 運算的 NPU（Hexagon），以及負責連上行動網路的數據機和射頻系統（收發無線電訊號的零件）。最新一代在 2026 年 9 月發表，分成兩顆：8 Elite Gen 6，以及效能更高的 8 Elite Extreme Gen 6。兩顆都由台積電用 2 奈米製程生產，主核心時脈首次達到 5GHz。旗艦晶片仍按時更新。手機營收下滑，原因是蘋果退出和記憶體漲價壓縮安卓，不是產品跟不上。 ✓ Qualcomm · 09-22

Think of a phone as a company. The CPU is the general manager. The GPU is the art department. The NPU is the newly formed AI team. The modem is the switchboard connecting to the outside world. Qualcomm packs the entire company into a chip the size of a fingernail.

比喻：如果手機是一間公司，CPU 是總經理，GPU 是美術部，NPU 是新成立的 AI 部門，數據機是對外的電話總機。高通把整間公司塞進一顆指甲大小的晶片。

Without the picture: the chip has four areas. They handle general computing, graphics, AI, and the phone network. The phone-network area is the modem.

不看圖也能懂：這顆晶片分成四區，分別負責一般運算、畫面、AI 和連上電信網路。連網那一區就是數據機。

Early Snapdragon (QSD8250), HTC Desire phone board早期 Snapdragon（QSD8250），HTC Desire 手機主機板

Snapdragon 410 (MSM8916), a processor with a built-in modemSnapdragon 410（MSM8916），處理器和數據機做在同一顆

2026: Snapdragon 8 Elite Gen 6 and Extreme Gen 6 (official badge; both share it)2026：Snapdragon 8 Elite Gen 6 與 Extreme 版（兩顆共用這個官方標誌）

2. Why the modem is hard to build二、數據機為什麼難做

The hard part about a modem (the part of the chip that connects a phone to the cell network) isn't calculating fast — it's connecting to every phone carrier's network on earth. Apple took six years to build its first one, and it still lacks one fast frequency band. Underneath that: the modem turns a phone's data into radio signals and sends them to a cell tower, and it has to work across 2G through 5G and hundreds of frequency-band combinations worldwide, passing certification with every carrier, one at a time.

數據機（讓手機連上電信網路的那一塊）真正難的地方不是算得快，而是全世界每一家電信公司的網路都要連得上。蘋果花了六年才做出第一顆，而且還缺一種高速頻段。底層的技術是：數據機負責把手機的資料轉成無線電訊號，送到基地台，它要能連上 2G 到 5G 幾代網路、全世界上百種頻段組合，還要一家一家通過電信商的認證。

Apple is the clearest example. In 2019, Apple paid $1 billion for Intel's phone-modem business, taking on about 2,200 engineers. It then took almost six years before its first in-house modem, the C1, shipped in 2025. The C1 and C1X still don't support 5G mmWave (millimeter wave, a very fast but short-range high-frequency band) even today.

蘋果是最好的例子。蘋果在 2019 年花 10 億美元買下 Intel 的手機數據機部門，接收約 2,200 名工程師。之後又花了將近六年，2025 年第一顆自家數據機 C1 才上市。C1 和 C1X 至今仍不支援 5G 毫米波（mmWave，一種速度很快、但傳不遠的高頻段）。

In other words, a modem is a technology you can only build up slowly, through time and certification. Apple managed it, but at a large cost; most phone brands cannot.

換言之，數據機是一個要靠時間和認證慢慢堆出來的技術。蘋果做得到，但付了很大的代價；大部分手機品牌做不到。

3. How the patent toll booth collects money (QTL)三、專利收費站怎麼收錢（QTL）

Think of Qualcomm as a highway toll booth: it used to both sell the car and collect the toll. Apple now builds its own car, but it still drives on this road, so the toll still gets paid — up to about $13 per phone, no matter whose chip is inside.

高通像一條高速公路的收費站：以前它同時賣車又收過路費。蘋果現在自己造車，但還是要走這條路，過路費照付，一支手機最多約 13 美元，不管裡面裝的是誰的晶片。

Here's the mechanism behind that toll. Any phone that makes calls or connects to the internet has to comply with international standards like 3G, 4G and 5G. Those standards are built on patented technology called a Standard Essential Patent (SEP — a patent you cannot avoid using if you want a standard-compliant phone).

收費的機制是這樣的：手機只要能打電話、上網，就必須遵守 3G、4G、5G 這些國際通訊標準。這些標準裡寫進了很多專利技術，叫做「標準必要專利」（SEP，要做得出能打電話的手機就躲不掉的專利）。

Whoever holds these patents must license them to every phone brand under the FRAND principle (Fair, Reasonable and Non-Discriminatory — the patent has to be licensed at a fair price to every phone brand, not just sold to insiders). Qualcomm's published 5G rate: for a license covering standard-essential patents only, 2.275% of wholesale price for a single-mode phone, 3.25% for a multi-mode phone, capped at a $400 wholesale price for the calculation.

持有這些專利的公司，依照 FRAND 原則（專利必須用公平價格授權給每家手機品牌，不能只賣給自己人）授權給所有手機品牌。高通公開的 5G 費率是：只授權標準必要專利，單模手機收批發價的 2.275%，多模手機收 3.25%；計費的批發價最高只算到 400 美元。

Without the picture: every phone pays Qualcomm a patent fee, whoever made its chip. Only phones with a Qualcomm chip also pay Qualcomm for the chip itself.

不看圖也能懂：不管手機用誰的晶片，都要付高通專利費。只有用高通晶片的手機，才會再付高通一筆晶片錢。

4. Data center — what Qualcomm is building四、資料中心：高通在做什麼

AI answering a question after it has learned — a step called inference — burns electricity every single time, and it happens far more often than "training," the earlier step where the AI first learns. So cloud operators care most about the electricity cost of each answer. Qualcomm wants to sell power-efficient cards for that job to cloud data centers; it already has a contract, but shipments are just getting started.

AI 學會之後回答問題的這個階段叫「推論」，每回答一次都要耗電，而且次數遠多於「訓練」（AI 學會之前的那個階段）。所以雲端業者最在意每回答一次要花多少電費。高通想賣比較省電的卡給雲端機房做這件事；合約已經有了，出貨才剛要開始。

Data-center revenue is not zero today, but it is small. After the Alphawave purchase, its connection chips count as data-center revenue: in the first nine months of FY2026 that added about $182M compared with a year earlier ($88M in the third quarter alone). Custom-chip revenue has not started yet. ✓ 10-Q · FQ3 FY26

資料中心營收現在不是零，只是很小。併入 Alphawave 之後，它的連線晶片算進資料中心營收：2026 財年前九個月比去年多約 1.8 億美元，其中第三季多 8,800 萬美元。客製晶片的營收還沒開始。 ✓ 10-Q · FQ3 FY26

Why Alphawave matters: it is a UK company that makes the technology chips use to pass data to each other at very high speed, and it also designs custom chips for clients. Qualcomm bought it for an enterprise value of about $2.4 billion, and the deal closed on December 18, 2025, about a quarter ahead of plan. The logic is simple. In an AI data center, thousands of chips must constantly send data to one another. If that link is slow, fast chips sit idle. The high-speed optical links in the Amazon contract use this kind of technology. ✓ Qualcomm · 2025-12-18

為什麼要買 Alphawave：它是一家英國公司，專做讓晶片之間高速傳資料的技術，也替客戶設計客製晶片。高通用約 24 億美元的企業價值（連同對方負債一起算的收購總價）買下它，2025 年 12 月 18 日完成，比原訂時間早約一季。道理很簡單：AI 機房裡成千上萬顆晶片要不停互傳資料，傳得慢，算得再快的晶片也只能等。亞馬遜合約裡的高速光連接，用的就是這類技術。 ✓ Qualcomm · 2025-12-18

Qualcomm's strategy is to cut into the inference market using the power-efficiency skills it built for phone chips:

高通的策略是用它做手機晶片練出來的省電能力，切進推論市場：

• AI200 (2026) and AI250 (2027): inference accelerator cards. The AI200 carries 768GB of LPDDR memory per card — the low-power memory type used in phones, cheaper than the high-bandwidth memory (HBM) usually used for AI training.
• AI200（2026 年）與 AI250（2027 年）：推論加速卡。AI200 每張卡配 768GB 的 LPDDR 記憶體，這是手機用的低功耗記憶體，比 AI 訓練常用的高頻寬記憶體（HBM）便宜。
• Custom chips (ASIC — a chip built for one cloud customer and one job): customers include ByteDance (reported by Bloomberg, May 2026) and Amazon AWS (SEC filing, September 2026).
• 客製晶片（ASIC，為單一雲端客戶量身做的晶片）：客戶包括字節跳動（2026 年 5 月 Bloomberg 報導）與亞馬遜 AWS（2026 年 9 月向證管會申報）。
• Optical interconnect: Qualcomm completed the acquisition of Alphawave in December 2025, enterprise value about $2.4 billion, gaining chip-to-chip high-speed-data technology (SerDes — a circuit that turns wide parallel data into a fast serial signal). The Amazon contract covers 1.6-terabit-per-second optical connections.
• 光通訊：2025 年 12 月完成收購 Alphawave，企業價值約 24 億美元，取得晶片之間高速傳資料的技術（SerDes，把大量平行資料轉成高速串列訊號的電路）。亞馬遜合約涵蓋每秒 1.6 兆位元的光連接。
• Server CPU: Dragonfly C1000, with Meta as a customer, mass production in the second half of 2028.
• 伺服器 CPU：Dragonfly C1000，Meta 是客戶，2028 年下半年量產。
• Software: in June 2026 Qualcomm acquired Modular for about $3.9 billion, letting AI software run across chips from different vendors.
• 軟體：2026 年 6 月以約 39 億美元收購 Modular，讓 AI 程式可以在不同廠牌的晶片上跑。

5. Automotive — from design wins to revenue五、車用：從設計訂單到營收

Cars are increasingly "computers on wheels" — the industry calls this the software-defined vehicle. The chips a car needs now stretch from the dashboard and navigation all the way to driver-assistance systems. Qualcomm bundles all of this as the Snapdragon Digital Chassis.

汽車越來越像「裝了輪子的電腦」，這叫軟體定義汽車。一台車需要的晶片從儀表板、導航，一路擴到駕駛輔助。高通把這整套叫 Snapdragon Digital Chassis。

Auto chips have one quirk: an automaker picks its chip years before the car actually ships — usually a two-to-four-year gap. So the "design-win pipeline" (the total value of chip orders an automaker has already committed to, for future cars) is a preview of future revenue, not money booked this year. Qualcomm's pipeline grew from $30 billion in September 2022 to $65 billion in June 2026 — orders already won, not yet all shipped.

車用晶片有一個特性：車廠先選定晶片，兩到四年後新車上市才開始出貨。所以「設計訂單」（design-win pipeline，車廠已選定、未來會出貨的訂單總額）是未來營收的預告，不是今年進帳的錢。高通的設計訂單從 2022 年 9 月的 300 億美元，到 2026 年 6 月的 650 億美元，也就是已經接下的訂單，還沒全部出貨。

Moats (advantages rivals struggle to take) and early-warning signals護城河（別人很難搶走的優勢）與提早示警的訊號

Qualcomm's deepest moat is the patent toll booth. Its shallowest is data center, where it is still the new arrival.

高通最深的護城河是專利收費站；最淺的是資料中心，那裡它還是新來的。

Leading indicators — public, checkable by any reader領先指標（每一個都是讀者自己查得到的公開資料）

What will grow什麼會放量

Over the next two years, automotive and data center will grow; Apple's share will keep shrinking; handsets depend on when memory prices cool off.

接下來兩年，會長大的是車用和資料中心；會縮的是蘋果；手機要看記憶體什麼時候降溫。

1. Automotive (structural growth): cars need more computing, so the chip value per car rises, and the $65 billion design-win pipeline converts into revenue over the next three to five years. Evidence: third-quarter revenue up 61% year over year; fourth-quarter guidance implies about +60%.
2. 車用（結構性成長）：汽車需要的運算變多，每台車用到的高通晶片價值上升，650 億美元的設計訂單在三到五年內陸續變成營收。證據：第三季年增 61%，第四季財測年增約 60%。
3. Data center (a brand-new market, but the timing is uncertain): AI is moving from training to inference, so cloud operators want power-efficient custom chips. As Amazon's purchase orders arrive, the warrant unlocks in batches. Management's target is $5 billion in FY2027 and over $15 billion in FY2029. One caution: this kind of custom-chip business runs a far lower gross margin than Qualcomm's core business — the finance chief said it will drag the chip division's margin down by 1.5 to 2 percentage points.
4. 資料中心（全新市場，但時點不確定）：AI 從訓練走向推論，雲端業者想降低電費，所以找省電的客製晶片。亞馬遜的採購單一到，認股權證就解鎖一批。管理層目標是 2027 財年 50 億美元、2029 財年超過 150 億美元。要注意：這類客製晶片的毛利率（賣掉之後扣掉成本還剩多少）遠低於高通本業，財務長說會把晶片部門毛利率拉低 1.5 到 2 個百分點。
5. IoT and PCs (steady growth): industrial, networking and robotics products grow at double-digit rates, but tablets are hurt by memory costs. About +9% a quarter; fourth-quarter guidance is flat.
6. 物聯網與個人電腦（穩定成長）：工業、網通、機器人產品雙位數成長，但平板受記憶體影響。每季約 +9%，第四季財測持平。
7. Handsets (a reversal — a trough candidate): Apple-related revenue falls below $2 billion in FY2027; Android revenue is compressed 20% by memory costs. The finance chief called this more-than-$1.50 EPS hit "a potential tailwind once the market normalizes." So FY2027 may be the trough year for handset revenue.
8. 手機（反向，谷底候選）：蘋果產品營收 2027 財年低於 20 億美元；安卓營收被記憶體漲價壓低 20%。財務長把這超過 1.5 美元的每股盈餘衝擊，稱為「市場正常化後的潛在順風」。所以 2027 財年可能是手機營收的谷底年。

Bull case多方論點

Bear case空方論點

Where this view differs from the market我們跟市場看法不一樣的地方

The sell-side story is "a phone-chip maker that lost Apple, plus an unproven AI story." This analysis thinks that framing mixes two different things together.

賣方說「失去蘋果的手機晶片商，加一個還沒證實的 AI 故事」；本研究認為這個說法把兩件不同的事混在一起了。

Without the picture: the left side lists four things the market believes; the right side gives our answer to each. The same points are written out in full just below.

不看圖也能懂：左邊是市場的四個看法，右邊是我們逐條的回答。同樣的內容，下面有完整文字版。

Where this analysis breaks from consensus本研究的分歧

First, the multiple problem is smaller than it looks — but it is not nothing. Qualcomm trades near 18 times today, more expensive than any of the past three years. The finance chief said pricier memory cost more than $1.50 per share of this year's profit. Even adding all of it back, today's price is only about 16 times normal profit — the top of the historical range. In other words, the market is paying a little extra for the data-center business, but far less than the headline 18x suggests.

第一，倍數的問題比看起來小，但不是沒有。高通今天約 18 倍本益比，比過去三年都貴。財務長說，記憶體漲價讓今年每股盈餘少了超過 1.5 美元。就算全部加回，現價也約是正常獲利的 16 倍，落在歷史區間頂端。換言之，市場替資料中心付了一點溢價，但遠少於表面的 18 倍所暗示的程度。

Second, "Apple risk" is really two separate risks. The chip supply contract only covers iPhones launched between 2024 and 2026 — that part is confirmed to disappear. The patent license is a separate contract, and it was just renewed. The market often talks about these two as one thing.

第二，「蘋果風險」其實是兩個風險。晶片供應合約只涵蓋 2024 到 2026 年發表的 iPhone，這部分確定會消失。專利授權是另一份合約，剛剛續約。市場常把兩件事混在一起講。

Third, this analysis does not claim the data-center target will definitely be hit. The $15 billion FY2029 figure and the $5 billion FY2027 figure are both management targets, not guidance. This analysis's own base case counts only 60% of that.

第三，本研究不主張資料中心目標一定會達成。2029 年 150 億美元、2027 年 50 億美元都是管理層的目標，不是財測。本研究的基準情境只算了六成。

So the conclusion: the sell-side consensus is broadly right, and this analysis's own profit estimate is almost identical to it. The real difference is timing — if data-center revenue really shows up in the December quarter, the market will start believing the 2027 target. If it doesn't, today's multiple won't hold.

所以結論是：賣方的共識大致正確，本研究的獲利估計和賣方幾乎一樣。真正的差別是時間點：12 月那一季如果真的出現資料中心營收，市場會開始相信 2027 年的目標；如果沒有，現在的倍數就撐不住。

A recurring-fear ledger恐懼年表

Qualcomm has been the subject of many market fears over the past decade. Three of the first four times, the patent toll booth held up — Huawei is the only licensing customer that was genuinely lost. The fifth fear is now happening — to the chip division.

高通過去十年被市場擔心過很多次。前四次裡有三次，專利收費站都撐住了；華為是唯一真的流失的授權客戶。第五次正在晶片部門身上發生。

Management, and how the company spends its cash經營層，以及公司的錢怎麼花

Management has been selling shares, but the amounts are small and most sales follow a pre-scheduled plan. The company's cash is shifting from buybacks toward the data-center business.

管理層在賣股，但金額都小，多數是事先排好的計畫性賣出；公司的錢正從買回股票轉向資料中心。

• CEO Cristiano Amon; CFO and COO Akash Palkhiwala.
• 執行長 Cristiano Amon，財務長兼營運長 Akash Palkhiwala。
• Institutional ownership runs 74-78%, insider ownership 1-2.6% — data providers don't agree on the exact figure. ⚠ data providers disagree各家資料不一致
• 機構持股 74 到 78%，內部人持股 1 到 2.6%，各家資料不一致。⚠ 各家資料不一致
• Every insider transaction in 2026 was a sale, and every one was small. All three of Amon's sales followed a pre-scheduled trading plan set up in December 2025 (a Rule 10b5-1 plan — a pre-scheduled plan set in advance to avoid any appearance of trading on inside information). Palkhiwala's May and September sales also followed a pre-scheduled plan; the filing for the July sale did not note one. No purchases at all. Pre-scheduled sales say little either way; they are not a sign that management has turned negative.
• 2026 年內部人交易全部是賣出，而且金額都小。Amon 的三筆都依 2025 年 12 月設定的預先交易計畫（10b5-1 計畫，事先約定時間與條件自動賣出，用來避免內線交易疑慮）。Palkhiwala 在 5 月與 9 月的兩筆也依預先交易計畫，7 月那一筆的申報資料沒有註明。沒有任何買入。事先排好的賣出，能透露的資訊很有限，不代表管理層不看好。
• Capital allocation: FY2025 bought back 56 million shares for $8.79B, returning $12.6B to shareholders in total. In the first nine months of FY2026, buybacks totaled $6.8B — but only $1.4B in the third quarter alone, a clear slowdown. The cash has gone toward acquisitions instead: about $2.4B for Alphawave, about $3.9B for Modular, and Autotalks (which triggered the China investigation).
• 資本配置：2025 財年買回 5,600 萬股、87.9 億美元，總共回饋股東 126 億。2026 財年前九個月買回 68 億，但第三季只剩 14 億，明顯放慢。錢流向收購：Alphawave 約 24 億、Modular 約 39 億、Autotalks（引發中國調查）。
• Judgment: buybacks are slowing and cash is shrinking. That fits the "swapping engines" story, but it also means the balance-sheet cushion is thinner than before if the data-center bet fails.
• 判斷：買回在放慢，現金在變少。這符合「換引擎」的說法，但也代表資料中心若失敗，資產負債表的緩衝比過去薄。

Competitive position競爭格局

In phones, Qualcomm ranks second by volume and stays strong at the high end; in data center, it is the challenger.

在手機，高通以量計排第二、以高階計仍強；在資料中心，它是挑戰者。

• Phone processors (Counterpoint, Q2 2026, by shipment volume): MediaTek 31%, Qualcomm 23%, UNISOC 13%, Samsung 9%.
• 手機處理器（Counterpoint，2026 年第二季，以出貨量計）：聯發科 31%、高通 23%、紫光展銳 13%、三星 9%。
• Phones above $600: Apple holds 65% (first half of 2026). No reliable public data exists for Qualcomm's own share of the high-end Android segment.
• 600 美元以上的高階手機：蘋果占 65%（2026 年上半年）。高通在高階安卓的份額，找不到可靠的公開數據。
• 5G patents: LexisNexis's January 2026 report ranks Qualcomm in the top tier, alongside Huawei, Samsung and Ericsson.
• 5G 專利：LexisNexis 2026 年 1 月報告把高通列在最前段，與華為、三星、Ericsson 同組。
• Data center: the general-purpose accelerator-card rivals are Nvidia and AMD; the custom-chip rivals are Broadcom and Marvell.
• 資料中心：通用加速卡的對手是 Nvidia、AMD；客製晶片的對手是 Broadcom、Marvell。
• Valuation comparison (stockanalysis, 09-28): Broadcom (AVGO) trades near 20.4x forward earnings; Arm (ARM) near 130x; MediaTek (2454.TW) near 51x (a second source says 43x — the two sources don't agree). Qualcomm is still cheaper than its AI custom-chip peers, but it is no longer trading at a "phone stock" discount.
• 估值對照（stockanalysis，2026-09-28）：博通（AVGO）以明年獲利計約 20.4 倍；Arm（ARM）約 130 倍；聯發科（2454.TW）約 51 倍（另一個來源算出 43 倍，兩邊算法不同）。高通仍比 AI 客製晶片同業便宜，但已經不是「手機股折價」。