把 NVIDIA 最近四季的淨利數字排成一列：$26.4B、$31.9B、$43.0B、$81.6B。不是成長，是加速——每一格跳幅都比上一格大。廣告業有個術語叫「飛輪效應」，意思是當品牌的曝光、轉換、口碑三者同時強化，成本反而遞減。現在 NVIDIA 的財報，就在演那個劇本。
毛利率為什麼是關鍵訊號
毛利率是公司在扣掉生產成本後，每一塊錢營收剩下多少。一般產業快速擴張時，為了衝量往往要壓價、多砸產能，毛利率隨之壓縮——這幾乎是製造業的鐵律。但 NVIDIA 這四季，毛利率從 72.4% 爬到 74.9%，量在漲，利潤率也在漲。這個反向走勢，才是真正不尋常的訊號。
能做到這件事，前提是定價權。當全球資料中心搶著排隊要同一顆 GPU，供應方就可以拒絕折扣。NVIDIA 在 AI 加速器市場的市占超過 82%，這個數字不是行銷說辭，是定價能力的硬底子。
79 位分析師的共識與少數聲音
覆蓋 NVDA 的 79 位賣方分析師中，74% 給出買入評級——強買 2 位、買入 58 位、持有 16 位、賣出 3 位。內部人（公司高管）雖有賣出紀錄，但屬於計畫性期權行使，和「大股東認為股票要跌了」是完全不同的動作。期權行使有預設時程，不能和臨時拋售劃等號。
少數派（3 位賣出評級）的邏輯也值得給空間：客戶集中、競爭者自研晶片、估值過高。這些風險都是真的，只是目前還沒有落地的財務後果。
估值分化，才是市場在說的話
同樣是大型科技股，五家公司的市淨率（PB）排出來是這樣：MSFT 7.0 倍、TSM 10.7 倍、AVGO 20.8 倍、NVDA 25.3 倍、AAPL 43.8 倍。從 7 倍到 43.8 倍，分布極度分散。
PB 是「市場願意用帳面淨資產的幾倍來買這家公司」的指標。倍數極端分化，代表市場在做的不是整體族群的情緒性追價，而是在這個族群內部精選勝者——每家公司被用完全不同的邏輯在定價。AAPL 的 43.8 倍和 MSFT 的 7.0 倍同時存在，說明市場不傻，針對每家公司的護城河深度給出不同溢價。
如果這是泡沫，應該是所有公司一起漲、估值一起膨脹、無差別追價。現在看到的分化，反而是理性定價的特徵。
廣告人看財報的角度
做廣告做久了，有個本能：看品牌的客戶是不是在幫它轉介紹。若是，這個品牌的後續獲客成本幾乎為零。NVIDIA 的 CUDA 生態系就在做同一件事——工程師在上面開發、公司買了 GPU 就深度綁定、切換成本極高，新客戶進來會跟著老客戶的工具鏈走。這套生態系讓競爭者自研晶片的難度，不只是硬體規格的差距，而是整個軟體護城河的高度。
ROE 0.742、ROA 0.385，這兩個數字在半導體產業裡屬於少見的高位。用最白話的方式說：每投入一塊錢資產，NVIDIA 能賺回來的比同業多很多，而且股東的資金被運用得相當有效率。
沒有答案的問題
四季淨利加速，毛利率逆勢拉升，分析師共識強烈，估值分化而非齊漲——這些訊號拼在一起，目前看不到泡沫的結構性特徵。
真正的問題不是「現在是不是泡沫」，而是「客戶集中度與競爭者自研晶片，要花多少個季度才會反映在財報上」。那個時間差，才是現在持倉最核心的風險變數。不是今天，但也不是永遠不來。
— 胡凱翊（Kevin）
本文為個人研究筆記，不構成任何投資建議。
AI Chip Margins Defy the Bubble Script
Line up NVIDIA’s net income for the last four quarters: $26.4B, $31.9B, $43.0B, $81.6B. The gap between each entry is larger than the one before. That’s not growth — that’s acceleration, and the shape of it matters more than any single number.
Why Gross Margin Is the Real Signal
Gross margin measures how much of each revenue dollar survives after production costs. In most industries, rapid expansion compresses margins — prices get cut to chase volume, capital pours into capacity. That’s textbook manufacturing logic. NVIDIA did the opposite: gross margin climbed from 72.4% to 74.9% across those same four quarters, while revenue surged. Both lines going up simultaneously is the anomaly worth staring at.
The only mechanism that allows this is pricing power. When global data centers are queuing for the same GPU, the supplier stops discounting. With over 82% market share in AI accelerators, NVIDIA’s position isn’t a marketing claim — it’s the hard floor beneath its ability to hold price.
What 79 Analysts Are Saying — and the Three Who Aren’t
Of the 79 sell-side analysts covering NVDA, 74% carry a buy rating: 2 strong buy, 58 buy, 16 hold, 3 sell. Insider selling has occurred, but through scheduled option exercises — a pre-planned liquidity mechanism, structurally different from a discretionary dump. The two shouldn’t be conflated.
The three sell-rated analysts aren’t wrong on risk: customer concentration, hyperscalers building proprietary chips, stretched valuation. Those risks are real. They just haven’t shown up in the financials yet. The gap between a credible risk and a realized one is where most of the market disagreement lives.
Valuation Dispersion Is the Story
Across five large-cap tech names, price-to-book (PB) ratios sit at: MSFT 7.0x, TSM 10.7x, AVGO 20.8x, NVDA 25.3x, AAPL 43.8x. A range from 7x to 43.8x within the same sector is unusually wide. That spread is information.
Bubbles tend to be indiscriminate — everything in the category inflates together, pricing becomes uniform, differentiation collapses. What’s visible here is the opposite: the market is assigning radically different multiples to companies within the same AI theme, which implies it’s still doing the work of distinguishing moat depth from narrative momentum. AAPL at 43.8x and MSFT at 7.0x coexisting in the same portfolio screen is not the fingerprint of a market that’s stopped thinking.
The Ecosystem Lock-In That Doesn’t Show Up on a Slide
NVIDIA’s CUDA software ecosystem functions like a high-switching-cost loyalty program — engineers build on it, enterprises commit to it, and new entrants inherit it because that’s where the talent and tooling already are. Competing on chip specs alone doesn’t crack that. The moat isn’t the silicon; it’s the stack built on top of it over years.
ROE of 0.742 and ROA of 0.385 are unusually high for the semiconductor sector. Plain translation: for every dollar of assets deployed, NVIDIA generates outsized returns compared to peers, and shareholder capital is being used efficiently. Those ratios don’t come from hype — they come from margin structure and asset velocity.
The Question That Doesn’t Have an Answer Yet
Four consecutive quarters of accelerating net income, margin expansion against the trend, strong analyst consensus, and valuation dispersion rather than uniform inflation — the structural fingerprints of a bubble aren’t present in this data set.
The real variable is timing: customer concentration risk and in-house chip development by major hyperscalers are credible threats, but translating into financial impact takes quarters, sometimes years. That lag — between when a risk becomes visible and when it hits earnings — is the actual exposure for anyone holding the stock today. Not imminent. But not theoretical either.
— 胡凱翊（Kevin）
This article is personal research notes only and does not constitute investment advice of any kind.
延伸閱讀
https://justfly.idv.tw/s/Mw1C68f