Variant-perception thesis · personal research變異觀點投資論點 · 個人研究
Applied Materials (AMAT)
The picks-and-shovels enabler of chip complexity. The reasoning is the model's; every number carries its source and date. Where the first draft ran on memory, live data rewrote the conclusion.晶片複雜度背後「賣鏟子」的設備商。推論由模型產生，但每個數字都標註來源與日期。第一版靠記憶寫成的地方，即時資料改寫了結論。
How this is built. A variant-perception framework — the discipline of ignoring consensus to find what the market may be misunderstanding, underestimating, or ignoring. It walks the same nine steps for any stock: business → bull → bear → the variant view → catalysts → management → moat → probability-weighted odds → what it takes to double. The rule that makes it trustworthy: no number is written from memory — every figure is grounded to a filing or live data with a date, and the share price is recomputed the moment this page is built.
headlineThe bull re-rating has largely already happened多方期待的「評價重估」其實大半已經發生
The variant bull case is "the market underestimates how long AI-driven chip complexity lifts equipment demand, so a re-rating is coming." But AMAT is at $589.04, +129.5% YTD, above its prior 52-week high (✓ IBKR · 06-16). Forward P/E here is ~57×, not the "mid-40s" the memory draft assumed. Trailing P/E is 68× GAAP / 62× non-GAAP (live price ÷ FY25 EPS ✓ 10-K).
The re-rating the thesis argues for is largely done. The real question is not "is the bull case right" — it's "is there edge left after a double, at a high-50s multiple." A polished page built on the stale number would have hidden that.
AMAT sells the steps that turn a blank silicon wafer into a chip. To judge the thesis you have to know what those steps are. Here is the wafer's journey, then a plain-language glossary.
AMAT 賣的是把一片空白矽晶圓變成晶片的那些製程步驟。要判斷這個論點，得先知道這些步驟是什麼。先看晶圓的旅程，再看白話術語表。
The whole bull case in one picture: AI chips need more layers, so the loop runs more times, so AMAT sells more tools. AMAT leads every step here except lithography (step 2 — that's ASML's EUV).一張圖看懂多方論點：AI 晶片需要更多層 → 循環跑更多次 → AMAT 賣出更多機台。圖中每一步幾乎都是 AMAT 主場，只有微影（第 2 步）是 ASML 的 EUV。
And here is what those steps physically do to the silicon — building one metal wire, in cross-section (the real copper "damascene" process):
This loop builds just one wiring layer. An AI chip stacks 10–15+ of them — so the loop runs 10–15+ times, and AMAT sells a tool for nearly every pass. More complexity = more passes = more AMAT.這套循環只做出一層線路。一顆 AI 晶片要疊 10–15 層以上 — 循環就跑 10–15 次以上，幾乎每一輪 AMAT 都賣得到機台。越複雜 = 跑越多次 = AMAT 賣越多。
Step 1 (deposit) isn't one thing — it's a ladder of tools for ever-harder 3D structures: PVD → CVD → ALD. The more vertical the chip, the more you need the precise (slower, pricier) end. This is AMAT's backyard.
第 1 步（沉積）不是單一技術，而是一排對應越來越難 3D 結構的工具：PVD → CVD → ALD。晶片越立體，越需要精準（但更慢、更貴）的那一端。這正是 AMAT 的主場。
PVDphysical · sputter物理氣相沉積
Blast a solid target into atoms that fly straight down. Fast, high throughput — great for metal layers, barriers, liners. Struggles to reach the bottom of deep 3D structures.把固態靶材打成原子，直線飛濺到晶圓上。快、吞吐量高 — 金屬層、阻障層（barrier）、襯墊層（liner）的主力。但難顧到深 3D 結構的洞底與側壁。
CVDchemical · grows化學氣相沉積
Gaseous precursors react on the surface — the film "grows." Better coverage and uniformity than PVD; the production workhorse. Hits its precision limit as structures get more 3D.氣態前驅物在表面反應，薄膜是「長出來」的。覆蓋性與均勻性優於 PVD，量產主力。結構越 3D，會碰到精度極限。
ALDatom by atom原子層沉積
Self-limiting — each cycle grows exactly one atomic layer, almost regardless of shape. Indispensable for FinFET, GAA, high-aspect-ratio. Slow and pricey, so used only where nothing else works. Key for GAA.自限制反應 — 每循環只長一層原子厚度，幾乎不受結構形狀影響。FinFET、GAA、高深寬比必備。慢又貴，只用在非用不可處。GAA 關鍵。
AMAT is the deposition leader — #1 in PVD and a top player in CVD and ALD. (EPI — growing a single crystal for the transistor channel — is a fourth AMAT strength.) Illustrations are conceptual.AMAT 是沉積龍頭 — PVD 市占第一，CVD 與 ALD 也是要角。（另有磊晶 EPI，長單晶當電晶體通道，也是 AMAT 強項。）圖為示意。
Whatever the method, deposition is judged on three things — and 3D chips make all three harder:
不論用哪種方法，沉積的好壞看三件事 — 而 3D 晶片讓這三件事都更難：
Every method chases the same three: even thickness everywhere (uniformity), full coverage down into 3D trenches (conformality), a smooth top (surface smoothness). The deeper and narrower the structure, the harder all three get — which is why the precise tools (ALD) keep gaining share. Recreated from a vocus.cc explainer.每種方法追求的都是同三件事：到處厚度一致（均勻性）、連 3D 深溝底部都完整包覆（覆蓋性）、表面平整（表面平整度）。結構越深越窄，這三件事越難 — 這正是精準工具（ALD）市占持續成長的原因。圖表重製自 vocus.cc 解說。
The process steps + 2 key terms製程步驟 ＋ 2 個關鍵詞
Deposition薄膜沉積（Deposition）
Lay an ultra-thin film onto the wafer. Four tools, by rising precision: PVD (sputter a solid, fast), CVD (gas grows a film, the workhorse), ALD (one atom-layer at a time, for 3D), EPI (grow a single crystal). See the diagrams above.
More AI = more complexity stacked per chip = more process steps AMAT sells into.AI 越多 = 每顆晶片堆疊的複雜度越高 = AMAT 能賣進去的製程步驟越多。
AMAT is the world's broadest supplier of materials-engineering equipment for making chips — the deposition, etch, CMP, implant and metrology tools from the decoder above. As chips move to GAA transistors, backside power and HBM stacking, each wafer needs more steps and higher-value tools. AMAT sells the steps.
AMAT 是全球產品線最廣的材料工程設備供應商 — 也就是上面解碼器裡的沉積、蝕刻、CMP、佈植、量測機台。當晶片走向 GAA 電晶體、晶背供電、HBM 堆疊，每片晶圓需要的步驟更多、機台單價更高。AMAT 賣的正是這些步驟。
How it makes money怎麼賺錢 FY2025 ✓ 10-K
Semiconductor Systems — $20.8B (73%). Capital tools to TSMC, Samsung, SK Hynix, Intel.
AGS — $6.4B (23%). High-margin recurring service on the installed base.
Rising chip complexity = more materials-engineering steps per wafer. AMAT is the diversified toll-taker on that complexity, not a bet on one node or customer.
GAA transistors, backside power and HBM/advanced packaging each add deposition, etch, CMP and metrology steps — AMAT's strongest areas. AI capex could stay high for years.
GAA 電晶體、晶背供電、HBM／先進封裝，每一項都增加沉積、蝕刻、CMP、量測步驟 — 都是 AMAT 最強的領域。AI 資本支出可能維持高檔好幾年。
Mix upgrade, not just a headwind不只是逆風，是品質升級
China fell from 37.2% of revenue in FY24 to 30.1% in FY25 ✓ 10-K. The bull read: mix shifts toward higher-ASP leading-edge tools for TSMC/Samsung/SK Hynix — better margins, steadier demand.
$6.4B of recurring, high-margin service grows with the installed base and smooths the WFE cycle — a layer the market underweights vs. lumpy systems sales.
Leadership in EPI for GAA and CMP for new materials; possible expansion into power semis and process-control software; CHIPS-Act onshoring as a 2026-27 tailwind. R&D $3.57B/yr funds the edge ✓ 10-K.
What the market may be missing:市場可能漏看的：the duration and breadth of AI-complexity intensity.AI 複雜度需求的持續時間與廣度。butsee the headline — much of this is now in the price.但見開頭結論 — 這大半已反映在股價裡。
Bear case — what could go wrong空方論點 — 哪裡可能錯
Valuation is the bear case now估值本身就是空方理由
At ~57× forward / 68× trailing after +129% YTD, any growth wobble = multiple compression. The stock already prices in durable, low-cyclicality growth. live · 06-16
If hyperscalers digest AI capacity, equipment orders can drop sharply — history says so. A "super-cycle" is itself an assumption.
若雲端大廠消化既有 AI 產能，設備訂單可能急跌 — 歷史有前例。「超級循環」本身就是一個假設。
China / geopolitics中國 / 地緣政治
~30% of revenue. AMAT guided a $600M FY2026 hit from expanded U.S. export curbs ✓ Reuters/TF; domestic substitution (NAURA) and Taiwan tail-risk add more.
Must hold for the thesis to fail:論點要被推翻，需同時成立：sustained high AI capex from a concentrated customer set, on-time node transitions, no new export shocks, continued pricing/mix gains. Break one and the multiple unwinds.集中客戶群的 AI 資本支出持續高檔、節點轉換準時、沒有新的出口管制衝擊、定價/組合持續改善。任一條斷掉，本益比就會回吐。
Variant perception變異觀點— the reason to read— 值得讀的地方
Consensus → variant → what the tape already reflects. The variant thesis was right enough that the market adopted it; the +129% run is the receipt.共識 → 變異觀點 → 股價已反映的。變異論點夠對，所以市場已採納；+129% 這波大漲就是收據。
Management & insiders經營層與內部人
Track record經營紀錄
CEO Gary Dickerson✓ public — long tenure, navigated multiple cycles, positioned AMAT for advanced nodes and AI. High-quality capital allocation: heavy R&D ($3.57B), buybacks, growing dividend.
執行長 Gary Dickerson✓ 公開 — 任期長、歷經多次循環、把 AMAT 卡位在先進節點與 AI。資本配置品質高：重研發（$3.57B）、買回庫藏股、配息成長。
Insider signal — net selling內部人訊號 — 淨賣出
Insiders own just 0.3% (~$1.1B) ✓ SWS. President Prabu Raja sold 50,000 shares on 2026-06-04 for $25,264,197 ($503.99–507.23, via living trust) ✓ SEC Form 4. Likely routine — but no buying near highs is a mild negative.
內部人僅持股 0.3%（約 $1.1B）✓ SWS。總裁 Prabu Raja 於 2026-06-04 賣出 5 萬股、$25,264,197（$503.99–507.23，透過生前信託）✓ SEC Form 4。多半是例行 — 但高檔無人買進，算輕微負面。
Probability-weighted outcomes機率加權情境
judgmentThese percentages are the model's subjective call, not a statistic. After +129% YTD a personal read would raise the bear odds — less room up, more room down.這三個百分比是模型的主觀判斷，不是統計數字。大漲 +129% 之後，若由我自己評估會把「空方」機率調高 — 因為上漲空間變小、下跌風險變大。
To double from $589 means a ~$1.2T market cap. A rising stock has two engines: (1) the company earning more, and (2) the market paying a higher P/E (re-rating). When the stale draft assumed a mid-40s P/E, engine 2 still had room — the easy half. At ~57× today it is largely spent, so a double now rests almost entirely on engine 1: real earnings growth. A materially harder bet than the memory draft implied.
