.

台積電神曲！用 AI 創作《護國神山》

你看到的是一顆晶片 我看到的是一場文明的遠征 從最粗糙的砂 一路走到奈米世界的神準刀鋒 這不是生產 這是人類把秩序刻進混沌的過程 故事從一粒砂開始，被烈火點亮 高純度的矽，在坩堝裡熬成信仰 拉晶那一刻像把混亂慢慢拉長 一根單晶柱直衝天際，像科技的脊梁 接著切片、倒角、研磨、拋光登場 每片晶圓平得像深夜沒風的湖光 表面不能傷，缺陷不能藏 因為再小的偏差，後面都可能滿盤輸光 先氧化成牆，再把薄膜一層層鋪上 像在虛無裡替未來蓋出第一道城牆 這不是原料堆放，這是秩序在生長 把看不見的世界，雕成可運轉的形狀 無塵室像聖堂，連空氣都被馴養 不是求快，是把穩定做到近乎瘋狂 沒有人大吼，只有精度在發燙 每一奈米的後面，都有人把靈魂押上 白光下面沒有神話，只有一再校準的日常 真正的巔峰，不靠喧鬧，只靠每次都一樣 把偶然趕出現場，把紀律變成鋼 讓世界最小的尺度，也能承載最大的想像 從砂到神作，從光到脈搏 在看不見的深處裡，雕出時代輪廓 一刀都不能錯，一步都不能拖 把萬億次運算，先種進沉默 從圓到方格，從線到星河 一片晶圓背後，是無數日夜雕琢 你說這只是晶片，不，聽清楚了 這是工匠精神，把未來親手做活 重頭戲來了，微影像魔法上桌 光不是拿來照亮，是拿來切割山河 光阻先上，像給命運蓋上底色 對準、曝光、顯影，讓圖騰浮出輪廓 極紫外光一閃，整個世界都變窄 把宏大的電路，壓進奈米級舞台 兩千多道工序，最後還得七成站立 每一步的失誤，都得鎖進萬分之二禁區裡 沉積、蝕刻、離子佈植，一層層堆起奇蹟 CMP磨平高低，像把狂浪壓成平地 不是誰都敢，把極限當作日曆 台積電把工匠精神，刻進每一道工藝 晶體管開始列隊，密得像夜空繁星 金屬線彼此交會，像城市神經甦醒 電子奔跑其間，速度快到沒有聲音 而每一次通電，背後都是多少人守住精準命運 你手上的手機，只看見螢幕發亮 你問的每一句AI，只看見答案登場 但沒看到那條路，有多少人咬牙硬扛 把偏差壓到極小，把良率扛到極強 這不是運氣，這是千萬次驗證過的倔強 不是一次漂亮，是每一次都不能失常 真正厲害的，不是說自己有多鋒芒 是把最難的事，做到像呼吸一樣平常 前段完成之後，後段接力上場 量測、檢測像審判，把瑕疵一一曝光 該留下的留下，該退場的退場 能走到最後的，才扛得起世界重量 切割成晶粒，像從大地取出心臟 封裝不是包好而已，是讓力量重新組裝 散熱、傳輸、堆疊，每一步都要像樣 讓一顆晶片真的能進戰場，不只是好看發光 最後它進伺服器、進雲端、進車廂 進工廠、進醫療、進你我每天的日常 也許你只看到產品上市、跑分飛揚 但真正讓人敬佩的，是背後那群無名工匠 他們不把偉大掛嘴上，只把細節做到最強 把一微米再砍成千分，把誤差逼到退讓 從一粒砂走到世界中央，不靠幻想 靠的是把每一次重複，都做到無可取代的漂亮 從砂到神作，從光到脈搏 在奈米的深處裡，鍛造時代輪廓 每一步都比傳說，更沉默也更執著 把未來的引擎，先藏進沉默 從圓到宇宙，從微小到磅礴 一片晶圓背後，是整個文明在雕琢 你問誰值得敬佩，答案早就寫活 是那些把一生，磨進一顆晶片裡的人啊 當全世界都在驚嘆科技衝得多快 別忘了有人先把慢，做到極致的愛 在最小的尺度，做最大的安排 所以台積電讓人敬佩，不只是厲害 而是把每一個細節，都做到無可取代

AI 需求旺！台積電 Q3 毛利率預告下滑，短期的蹲低，是為了在 2030 年前的 AI 黃金期跳得更高

.

AI 需求旺！台積電 Q3 毛利率預告下滑，短期的蹲低，是為了在 2030 年前的 AI 黃金期跳得更高

台積電（TSMC）最新法說會剛交出 EPS 27.25 元、毛利率 67.7\% 的歷史級亮眼成績單，主引擎 HPC 更暴增 20\%。然而，財務長卻預告第三季毛利率將會承壓。在 AI 需求一路看旺到 2030 年的狂熱市況下，下步棋究竟是什麼？

從軟體架構的角度來看，台積電這不是獲利衰退，而是在進行一場大版本的「技術重構（Refactoring）與架構擴展」。

核心原因就在於 2 奈米（N2）先進製程的商業化爬坡。新製程上線初期，就像新系統剛上線進行壓力測試，必須先投入高昂的設備與材料折舊成本，且良率（Yield Rate）還在優化階段，這導致庫存天數從 80 天拉長到 87 天。雖然這會短期稀釋毛利率，但這筆「技術債」是為了築起更恐怖的長線護城河。一旦 2 奈米順利放量，台積電將徹底鎖定未來幾年 AI、HPC 與旗艦晶片的絕對定價權。短期的蹲低，是為了在 2030 年前的 AI 黃金期跳得更高。

AI booms! TSMC's Q3 margin dip is a brief pullback to leap higher by 2030.

TSMC just delivered a historic Q2 earnings report, crushing expectations with an EPS of NT$27.25, a 67.7\% gross margin, and a stellar 20\% growth in its HPC segment. However, management surprised the market by guiding for short-term gross margin pressure in Q3. With AI demand projected to boom through 2030, what is the silicon giant's hidden playbook?

From a software architecture perspective, this is not a sign of financial regression, but rather a massive "tech refactoring and infrastructural scaling" phase.

The core driver behind this temporary dip is the commercial ramp-up of the next-generation 2nm (N2) node. Launching a new node is akin to stress-testing a critical system upgrade; it demands heavy upfront capital expenditure and raw material costs, while yield rates are still being optimized. This has pushed inventory days up from 80 to 87 days. While this operational friction temporarily dilutes gross margins, it is a strategic investment to pay off "technical debt" and build an even more formidable long-term moat. Once 2nm production stabilizes and scales, TSMC will lock in absolute pricing power over future AI, HPC, and flagship smartphone silicon. This short-term dip is simply the giant bending its knees to jump higher into the 2030 AI golden decade.

台股全面喊漲！這波晶片漲價潮，誰是真贏家、誰被毛利反噬？

.

台股全面喊漲！這波晶片漲價潮，誰是真贏家、誰被毛利反噬？

2026年下半年的台股主旋律，無疑是由晶圓代工、被動元件到晶片材料掀起的「全面漲價潮」。從軟體架構的角度來看，這就像是上游雲端 API 強制漲價，下游 App 開發商迎來一場殘酷的生存 Debug。

在這場定價權的賽局中，上市櫃企業被劃分為兩大陣營：

成本轉嫁贏家（擁有護城河）： 以台積電（TSMC）為首的先進製程，以及吃到 AI 規格升級、市佔率極高的被動元件與 CCL（銅箔基板）大廠。因為市場對高階 AI 算力與光模組的需求是「剛性需求」，他們能輕易將成本轉嫁給客戶，毛利率反而逆勢擴張。

毛利反噬輸家（替代性高）： 缺乏產品差異化的中下游 IC 設計廠，或技術門檻較低的傳統電子元件廠。面對上游代工與材料的雙重夾擊，他們既不敢向終端品牌漲價，又必須吞下高昂成本，獲利結構將面臨斷崖式惡化。

投資人此時盲目追高很容易踩雷。看清誰能在這波漲價潮中轉嫁成本，才能像架構師一樣，精準挑選出系統中韌性最強的核心資產。

Taiwan Stocks Surge! The Chip Price Hike: Who Wins Big, and Who Suffers Margin Erosion?

The dominant theme for Taiwan stocks in the second half of 2026 is undoubtedly the "sweeping price hike" cascading from semiconductor foundries and passive components down to raw materials. From a software architecture perspective, this mimics a scenario where upstream cloud APIs hike their fees, forcing downstream App developers into a brutal operational debug.

In this battle for pricing power, listed companies are split into two distinct factions:

Pricing Power Winners (The Moat Holders): Led by TSMC’s advanced nodes, alongside top-tier passive component and CCL (Copper Clad Laminate) giants integrated into AI hardware upgrades. Because global demand for premium AI computing and optical modules is perfectly inelastic, these players can easily pass on costs, driving their gross margins even higher.

Margin Erosion Losers (The High-Substitutability Group): Mid-to-downstream IC design houses with commoditized products or legacy electronic component suppliers. Trapped between rising foundry rates and surging material costs, they lack the leverage to hike prices for end-user brands, leaving them to absorb the inflation and face a steep collapse in profitability.

Chasing the rally blindly now poses immense risks for retail investors. Identifying who can successfully pass on costs is the only way to pick the most resilient core assets in the ecosystem, much like a seasoned systems architect.

為什麼馬斯克的火星夢，需要「非同步電路設計」來續航？

.

為什麼馬斯克的火星夢，需要「非同步電路設計」來續航？

馬斯克（Elon Musk）的 SpaceX 火星計劃是人類史上最瘋狂的分布式系統工程。但在外太空極端輻射與超長效能需求的考驗下，傳統晶片架構正面臨物理極限。這時，晶片設計中冷門卻強大的「非同步電路設計（Asynchronous Circuit Design）」成了關鍵解方。

傳統晶片就像一隊跟著時脈訊號（Global Clock）齊步走的士兵，每次開關都在同步耗電。非同步電路則移除了這個中央時脈，改採「握手協定（Handshaking）」——有資料才運算，沒工作就完全靜止。這帶來了兩大火星級優勢：第一，極致省電，這對靠太陽能維生的火星探測設備是救命稻草；第二，高容錯與抗輻射。太空中高能粒子引發的瞬時干擾（Soft Errors）常會讓同步時脈訊號偏移、導致系統崩潰；非同步電路沒有統一時脈，對延遲和電壓波動極具彈性，能自動優雅降級（Graceful Degradation）而不死機。

從第一原理（First Principles）來看，要在火星生存，系統必須擺脫對中央集權（時脈）的依賴。非同步電路的去中心化架構，或許正是火箭晶片能在星際輻射中活下來的終極底層。

Why Elon Musk’s Mars Mission Needs Asynchronous Circuit Design

Elon Musk’s SpaceX Mars initiative is the most audacious distributed systems engineering project in human history. However, under the scrutiny of cosmic radiation and extreme power constraints, traditional chip architectures are hitting a physical wall. This is where a niche yet powerful hardware philosophy comes in: Asynchronous Circuit Design.

Traditional chips operate like a regiment of soldiers marching to a global clock signal, consuming power uniformly at every tick. Asynchronous circuits, by contrast, eliminate this central clock, relying instead on a localized "handshaking protocol"—computing only when data is present and remaining perfectly dormant otherwise. This paradigm shifts brings two Martian-scale advantages. First, extreme power efficiency, which is a lifeline for solar-dependent Mars exploration hardware. Second, high fault tolerance and radiation hardness. In deep space, high-energy particles cause soft errors that easily disrupt global clock lines, crashing synchronous systems. Asynchronous designs, being clockless, are inherently resilient to timing and voltage fluctuations, allowing the system to gracefully degrade rather than catastrophic failure.

From a First Principles perspective, surviving on Mars requires computing infrastructures to break free from centralized dependency. The decentralized architecture of clockless silicon may well be the ultimate foundational layer for rockets to survive interstellar radiation.

三皇之戰：TSMC、Intel、Samsung 誰才是 2 奈米世代的真贏家？

.

三皇之戰：TSMC、Intel、Samsung 誰才是 2 奈米世代的真贏家？

在當前由高效能運算（HPC）與 AI 晶片主導的賽局中，台積電、Intel、與三星的三方爭霸戰已經進入白熱化。這場晶圓代工大戰的勝負關鍵，不能只看行銷名稱，而是要從「製程穩定度」與「架構重構」的底層邏輯來解析。

從目前的技術推進來看，三巨頭正採取截然不同的戰略方向：

台積電（TSMC）：極致的漸進式 Debug。 台積電在 N3（3奈米）世代大獲全勝後，進入 N2（2奈米）世代才正式導入全環繞柵極（GAA）架構，並將背面供電技術（BSPDN）保留給之後的 A16 製程。這種「先求穩、再求快」的發布節奏，讓台積電在 2 奈米量產初期就交出高達 80% - 90% 的極致良率，持續贏得 Apple 與 NVIDIA 等核心客戶的信任。


三星（Samsung）：激進的先行者豪賭。 三星早在 3 奈米就搶先押寶 GAA 技術，試圖建立先行者優勢（First-mover Advantage），並計劃在 2027 年的 SF2Z 製程導入背面供電。然而，跨世代技術的併發（Concurrency）導致其製程複雜度大增，目前的 SF2 良率仍落後於台積電。


Intel：重構架構的背水一戰。 Intel 的 18A 為了扭轉劣勢，採取最激進的策略：同時應用 GAA（RibbonFET）與背面供電（PowerVia）。雖然這種「全整合」的藍圖成功吸引了部分客製化晶片訂單，但對代工廠而言，同時除錯兩項革命性硬體技術，折舊與產線爬坡的系統風險極高。

結論而言，製程戰爭的終點不是誰先發布漂亮的架構藍圖，而是誰能將產線良率優化到最高、成本控制到最低。在這場 2 奈米賽局中，台積電憑藉穩健的底層架構與純代工的生態系，依然穩坐這場全球晶片戰爭的最強主導者。

The Clash of Giants: TSMC, Intel, and Samsung—Who Wins the 2nm Era?

In the current era dominated by High-Performance Computing (HPC) and AI silicon, the sub-2nm foundry war among TSMC, Intel, and Samsung has reached a fever pitch. To decode the eventual winner, we must look beyond marketing nomenclatures and analyze the core engineering trade-offs between yield stability and radical architectural refactoring.

As the industry advances, the three giants are executing fundamentally divergent strategies:

TSMC: The Ultimate Iterative Debugging. Following its sweeping victory in the N3 generation, TSMC has adopted a calculated cadence for N2, introducing Gate-All-Around (GAA) nanosheets first, while delaying Back-Side Power Delivery Network (BSPDN) to the subsequent A16 node. This "stability-first" methodology has allowed TSMC to secure exceptional early 2nm yield rates estimated at $80\% - 90\%$, maintaining unshakeable trust among top-tier clients like Apple and NVIDIA.


Samsung: The Audacious First-Mover Gamble. Samsung boldly bet on GAA architecture early during its 3nm node and plans to deploy BSPDN on its enhanced SF2Z process. However, the high concurrency of manufacturing innovations significantly raised system complexity, leaving its SF2 yields trailing behind TSMC's execution curve.


Intel: A High-Stakes Architectural Refactor. Desperate to reclaim its crown, Intel's 18A node takes the most radical approach by simultaneously deploying both RibbonFET (GAA) and PowerVia (BSPDN). While this "all-in-one" blueprint has captured initial custom silicon pipelines, debugging two revolutionary hardware layers at once introduces massive capex depreciation and operational scaling risks.

Ultimately, the semiconductor race is not won by the most elegant marketing slideshow, but by the highest yield rate and the lowest total cost of ownership on the fab floor. In the 2nm landscape, TSMC's robust engineering discipline and trusted pure-play model continue to anchor it as the most irreplaceable foundational layer of the global tech ecosystem.

為什麼半導體機台不用無線網路？

.

為什麼半導體機台不用無線網路？

在追求最新科技的半導體晶圓廠（Fab）裡，生產線上的千億級機台卻幾乎一律使用傳統的實體網路線（Ethernet）。這看似復古的選擇，背後是軟體與通訊架構對「極致可靠度」的硬核要求。

核心原因可以總結為三個關鍵字：高併發（Concurrency）、低延遲（Latency）與抗干擾（Interference）。

半導體製程是極致的實體 Debug，機台每秒會產生海量的感測器數據、派工指令與即時的控制訊號。無線網路（如 Wi-Fi 或 5G）本質上是共享介質，容易受到 Fab 內複雜的晶圓傳送軌道（OHT）、金屬屏蔽與電磁波的反射干擾，進而引發封包遺失（Packet Loss）。在微秒（\mu s）必爭的生產線上，一次僅幾毫秒的網路瞬斷，就可能導致晶圓定位偏移、甚至是整批晶圓報廢。此外，實體線路的「資訊安全」屏障也是無線網路無法比擬的。

對工程師而言，架構的漂亮比不上維運的穩定。實體網路線雖然笨重，但它所帶來的零干擾、決定性（Deterministic）延遲與絕對的資安控制，才是支撐晶圓廠 24 小時高可靠度運作的最強大底層。

Why Don't Semiconductor Tools Use Wireless Networks?

Inside state-of-the-art semiconductor wafer fabs, multi-billion dollar tools still rely almost exclusively on traditional wired Ethernet cables. This seemingly outdated choice is driven by a hardcore architectural requirement for ultimate reliability.

The core reasons boil down to three key metrics: high concurrency, ultra-low latency, and interference immunity.

Semiconductor manufacturing is essentially the ultimate form of physical debugging. Tools generate massive streams of sensor data, dispatching commands, and real-time control signals every second. Wireless networks (like Wi-Fi or 5G) are inherently shared media. They are highly vulnerable to packet loss caused by electromagnetic interference from overhead hoist transport (OHT) systems, metal shielding, and signal reflections inside the fab. On a production line where microseconds (\mu s) matter, a network drop of just a few milliseconds can cause wafer misalignment or even scrap an entire batch. Furthermore, the air-gapped security provided by physical cabling is something wireless networks simply cannot match.

For engineers, a system's elegance always takes a backseat to operational stability. Wired networks might feel clunky, but their zero-interference nature, deterministic latency, and absolute security control remain the most robust infrastructure supporting 24/7 high-reliability fab operations.

TSMC vs. Terafab

. 

TSMC vs. Terafab

馬斯克（Elon Musk）聯手 Intel 打造的 Terafab 計劃，無疑是半導體產業近年最狂妄的巨賭。從軟體工程的「第一原理」來看，Terafab 追求的是極致的垂直整合，試圖將晶片設計、製造、先進封裝到系統維運全部收攏在「同一個屋簷下」，徹底砍掉供應鏈之間的溝通摩擦與利潤抽成。

然而，台積電（TSMC）的護城河不單是設備，而是硬體層面的極致 Debug 能力。先進製程的良率（Yield Rate）是靠數萬名頂尖工程師，在產線上日以繼夜累積的製程參數（Recipe）與微調經驗堆疊出來的。這就像一套經過數十年高併發流量驗證、毫無死角的底層微服務架構，絕非靠千億美元與幾顆明星晶片就能在一夕之間重構的。Terafab 雖然展示了未來智慧製造的終極架構，但台積電靠著純代工的信任資產與極致良率，依然是這場全球晶片賽局中最難被重構的核心底層。

TSMC vs. Terafab
Elon Musk’s Terafab initiative, partnered with Intel, is undoubtedly the most audacious gamble the semiconductor industry has seen in recent years. From the perspective of software engineering's "First Principles," Terafab pursues ultimate vertical integration, attempting to bring chip design, fabrication, advanced packaging, and system operations all "under one roof" to completely eliminate supply chain friction and margin stacking.

However, TSMC’s moat is not just about equipment; it is about the ultimate debugging capability at the hardware layer. Advanced node yield rate is forged by tens of thousands of top engineers over decades of day-and-night optimization of manufacturing parameters (recipes) on the fab floor. It is akin to a bulletproof, low-level microservices architecture battle-tested under massive concurrent traffic—something that cannot be refactored overnight simply by throwing hundreds of billions of dollars or a few high-profile chips at it. While Terafab envisions the ultimate blueprint for next-generation smart manufacturing, TSMC, backed by its trusted pure-play foundry model and peerless yield infrastructure, remains the most irreplaceable core layer in the global chip ecosystem.