先看一組反差。雲端軟體股的本益比動輒 40 到 80 倍，而 NTAP 29.1 倍、HPE 29.4 倍、CSCO 28.95 倍，這批賣儲存與網路設備的公司，PE 中位數整整齊齊落在 28 到 30 倍。同一個 AI 題材，一邊賣夢想，一邊賣機櫃，市場給的價碼差了一倍以上。
更有意思的是股價位置。這週篩了一輪美股網路傳輸與混合雲儲存標的，超過 75% 站在 200 日均線（過去 200 個交易日的平均收盤價，常被當成長期趨勢的分界線）之上。NTAP 距年線 +68%，HPE +94%，ANET +35%。電視上沒有聲音喊它們是 AI 概念股，漲幅卻不輸。
算力買回家，資料還在門外
先看錢從哪來。Google、Amazon、Microsoft、Meta 在 2026 年 4 月 29 日公布的財報顯示，四家公司單季 AI 資料中心資本支出合計首度超過 1,300 億美元。GPU 搶購的規模就是這樣。
問題接在後面。晶片到貨之後，得有人把資料餵進去，得有地方把訓練出來的東西存好。這就像廣告公司砸重金買了一台頂規攝影機，結果發現硬碟不夠、傳輸線太細，片子拍得出來，卻導不出來。企業 IT 部門這一年多半在補的，就是這張硬碟與傳輸線的帳單。
數字長什麼樣
HPE 最新一季營收 122.1 億美元，毛利 49.0 億美元。NetApp 營收 20.2 億美元，毛利率 70.3%。毛利率 70% 的意思是：每賣 100 塊的東西，扣掉直接成本還剩 70 塊，這個水準通常是軟體公司的專利，現在由賣儲存的拿下了。混合雲交付（資料同時放在自家機房與公有雲）賺的是長期合約與維護費，客戶一旦把資料搬進來，就很難搬走。
PE 收斂在 28 到 30 倍，加上高自由現金流，這批股票的故事很無聊：訂單、毛利、現金。無聊的好處是，股價不太需要靠敘事撐著。
巨頭自己也在算帳
算力那一端也出現了鬆動的訊號。Alphabet 第二季錄得史上首次負自由現金流，燒掉 59 億美元。Meta 投入約 1,450 億美元資本支出建設 AI 基礎設施後，開始籌組事業部門，把超額算力以雲端服務賣給外部客戶，消息一出股價單日漲 9%。買得最兇的買家，已經在想怎麼把用不完的算力變現。
當市場焦點從「擁有多少 GPU」轉到「怎麼有效利用算力」，被點名的就是資料進得去、存得下、跑得動這條管線。儲存與網路公司賣的正好是這一段。
該留意什麼
這個劇本有個前提：企業 AI 預算持續從實驗轉向落地。若巨頭的資本支出降溫，這批股票也不會置身事外。年線以上 +94% 的 HPE，追高的風險也比 +35% 的 ANET 大得多。下一份該盯的是下一季財報裡，儲存與網路設備的訂單積壓數字有沒有繼續長。
— 胡凱翊（Kevin）
本文為個人研究筆記，不構成任何投資建議。
The GPU Rush Sent the Bill to Storage
Start with a gap. Cloud software trades at 40 to 80 times earnings. NTAP sits at 29.1x, HPE at 29.4x, CSCO at 28.95x, and the median for this storage-and-networking group lands between 28 and 30x. Same AI theme, one side sells the dream, the other sells the racks, and the market prices them at roughly half the multiple.
The chart positions are just as odd. In this week’s screen of US networking and hybrid-cloud storage names, more than 75% sit above their 200-day moving average, the average closing price over the past 200 trading days, a common line between long-term uptrend and downtrend. NTAP is +68% from its yearly line, HPE +94%, ANET +35%. Nobody is shouting “AI play” about them on TV, and they’re rising anyway.
The chips arrived. The data didn’t.
Start with where the money comes from. Google, Amazon, Microsoft and Meta reported on April 29, 2026 that their combined quarterly AI data center capex topped $130 billion for the first time. That is the scale of the GPU rush.
The trouble comes after delivery. Someone has to feed data into those chips and somewhere has to store what comes out. It’s like an ad agency buying a top-spec camera and then discovering the hard drives are too small and the cables too thin. The footage exists. It just can’t be exported. A lot of corporate IT spending this past year has gone to paying that bill for drives and cables.
What the numbers look like
HPE’s latest quarter: $12.21 billion in revenue, $4.90 billion in gross profit. NetApp: $2.02 billion in revenue at a 70.3% gross margin. A 70% margin means that out of every $100 sold, $70 remains after direct costs. That used to be software territory, and now a storage vendor is posting it. Hybrid cloud delivery, with data living both in a company’s own facility and in public clouds, earns long contracts and maintenance fees. Once a customer’s data moves in, it rarely moves out.
A P/E held at 28 to 30x, plus strong free cash flow, makes for a boring story: orders, margins, cash. The upside of boring is that the stock price needs less narrative to hold it up.
The big spenders are doing the math too
The compute side is showing strain. Alphabet posted its first-ever negative free cash flow in Q2, burning $5.9 billion. Meta, after roughly $145 billion in capex on AI infrastructure, is forming a unit to sell excess compute to outside customers, and its stock jumped 9% in one day. The biggest buyers are already figuring out how to monetize what they can’t use.
Once the question shifts from “how many GPUs are owned” to “how efficiently they are used,” the spotlight lands on the pipeline: can data get in, can it be stored, can it keep running. Storage and networking companies sell that stretch.
What to watch
The script has a precondition: enterprise AI budgets keep moving from experiments to deployment. If the giants’ capex cools, this group won’t be immune. And HPE at +94% above its yearly line carries more chase risk than ANET at +35%. The next thing to check is whether order backlogs for storage and networking gear keep growing in the next quarterly reports.
— 胡凱翊（Kevin）
This article is a personal research note and does not constitute investment advice.
延伸閱讀
https://justfly.idv.tw/s/c3QtvSk