封面
市場調查報告書
商品編碼
2158077

超大規模市場追蹤報告(2026年第二季):資本支出突破7,000億美元;得益於Alphabet的業績表現,淨利潤率創下新高;自由現金流利潤率則降至新低。

Hyperscale Market Tracker, 2Q26: Capex Tops $700B, an Alphabet Gain Lifts Net Margin to a Record, and Free Cash Flow Margin Falls to a New Low

出版日期: | 出版商: MTN Consulting, LLC | 英文 | 訂單完成後即時交付

價格

2026年第二季概覽

2026年第二季度,超大規模業務營收達8,492億美元(按以季度為基礎,年增17.7%),年化營收達3.264兆美元(年增15.8%)。年化資本支出飆升至7,015億美元(年增75.3%),資本密集度達到營收的21.5%,遠高於電信業約16%的水準。有形固定資產淨成長率年增55.5%。淨利率達到創紀錄的26.1%,主要得益於Alphabet的一次性收益。自由現金流利潤率(更清楚反映現金產生能力的指標)再次下降至9.5%,為自2011年開始編制該資料庫以來的最低水準。 2026年第二季員工人數較去年同期成長6.2%。本報告將全面分析這些數據背後的因素、各公司的業績表現(包括業績不佳和表現不佳的情況)以及數據所揭示的未來市場趨勢。

閱讀這份報告的理由

MTN Consulting自2017年第四季(4Q17)起,透過詳盡的季度分析持續追蹤該市場,我們的資料庫可追溯至2011年第一季(1Q11)。我們在該市場擁有長期深入的研究經驗,並掌握獨特的工具和一致且超大規模資料中心業者,涵蓋全球領先的超大規模資料中心(以前稱為「網路規模資料中心」)的收入、資本支出(Capex)、研發費用(R&D)、盈利、就業和資產負債刮器指標。我們的研究涵蓋超過12家企業,包括亞馬遜、Alphabet、微軟、Meta、蘋果、甲骨文、阿里巴巴、騰訊,以及CoreWeave、Nebius、快手和小米等新晉參與企業。沒有其他分析公司能夠像我們一樣,對該市場進行如此深入的分析並擁有如此廣泛的數據網路。

本公司已將此市場認定為「泡沫」約兩年,本報告的分析也印證了這一觀點。 2026年第二季(2Q26)報告包含詳細的公司層級細分資料、區域資料、供應商層級的營收基準,以及2026年至2028年的三種資本支出展望情境。對於任何計劃根據超大規模企業的支出趨勢制定投資、採購或策略決策的人士而言,本報告都是理想之選。

  • 注記:
    • (1)本報告是 MTN Consulting 對超大規模市場的第 35 份季度評估報告,也是自 2017 年第四季開始的一系列報告的一部分。我們的數據和分析涵蓋了 62 個季度,從 2011 年第一季到 2026 年第二季度。
    • (2)本次調查中包含的公司包括幾家最近新增的公司:CoreWeave(2024 年新增)、快手、Nebius(Yandex 的衍生公司)和小米(2025 年第三季新增)。
    • (3)在 2025 年第四季版中,該行業的名稱從「Webscale」更改為更通用的「Hyperscale」。

營收:總營收達到 8,490 億美元(年成長 17.7%),其中 HPE、Meta 和 Yandex 在現有公司中實現了最大幅度的成長。

2026年第二季度,超大規模資料中心公司營收達8,492億美元,較去年同期成長17.7%。年化營收增至3.264兆美元(年增15.8%)。 CoreWeave(成長112%)和Nebius(成長454%)仍然是受訪公司中成長最快的兩家,儘管它們的絕對規模仍然相對小規模。在成熟企業中,HPE(成長33.7%,目前仍在進行Juniper Networks的整合)的成長最為強勁,超過了Meta(成長28.0%)和Yandex(成長26.0%)——這兩家公司很少出現在榜單前列。 Alphabet(年成長24.2%)、Oracle(成長20.6%,主要受惠於OpenAI相關的雲端合約)、亞馬遜(成長19.6%)、騰訊(成長17.9%)和微軟(成長17.8%)的營收均高於產業平均。百度、京東和阿里巴巴在上一季表現不佳,但本季均重回成長軌道,營收分別成長1.8%、3.2%和15.4%。本報告涵蓋的公司中,僅有富士通(-5.7%)和小米(-0.2%)本季營收出現下滑。

廣告收入仍是支撐該領域多家領先公司的重要支柱。 Meta雖然持續向硬體和人工智慧平台多元化發展,但其收入仍主要依賴廣告,是廣告依賴度最高的公司。 Alphabet的非廣告收入目前佔其總收入的25%以上。亞馬遜的廣告收入略低於10%。鑑於人們對2026年美國消費者支出的擔憂,特別是考慮到川普政府對伊朗採取的行動可能導致能源供應中斷,對廣告的依賴構成了真正的風險。

對於這個行業來說,一個更大的未解決的問題是,鑑於目前付費訂閱在早期採用者之外的普及率不高,人工智慧平台最終是否會依賴廣告。

資本支出(Capex):年度支出超過 7,000 億美元(年增 75%),甚至連產業本身的預測也不斷上調。

2026年第二季年度資本支出達7,015億美元,年增75.3%。資本密集度升至營收的21.5%,比電信業的約16%高出5個百分點以上。亞馬遜、Alphabet和Meta在本季均上調了2026年的資本支出預測。亞馬遜的預測值目前約為22億美元(先前為20億美元),Alphabet的預測值為19.5億美元至20.5億美元(先前為18億美元至19億美元),Meta的預測值為13億美元至14.5億美元(先前為12億美元至14.5億美元)。亞馬遜將上調的原因歸咎於零件價格上漲,而Alphabet則表示需要提前確保產能。其中一些說法屬實──甲骨文公司本季的合約累積訂單達到6,640億美元,光是一年就成長了2,000多億美元。然而,該行業總合累積訂單的很大一部分仍然反映在少數交易對象之間的周期性資金籌措上。 2026年8月,輝達同意為OpenAI在俄亥俄州新建的資料中心園區提供高達1,050億美元的租賃和電力付款擔保。預計輝達將成為該園區唯一的晶片供應商。

資本投資的擴張仍然受到類似軍備競賽的動態驅動。沒有一家公司願意顯得對人工智慧的投入不如競爭對手,而且投資規模越大,就越難退。由於已經投入了數千億美元,停止投資將導致資產減值,承認策略失敗,並給競爭對手留下超越的機會。因此,在真正的雄心壯志、害怕落後以及在經濟可行性受到嚴格審查之前搶佔市場地位的願望的驅動下,投資仍在繼續。

儘管電力本身正逐漸成為阻礙因素,但現在就斷言它已超越半導體成為行業最大瓶頸還為時過早。更大的問題在於時間。資料中心建置和半導體採購可以在幾個月內完成,但接取電網往往需要數年時間。目前已有資金投入以彌補這一差距。 Google、Intersect Power 和 TPG Rise Climate 已承諾投資 200 億美元,用於夥伴關係。 2026 年 6 月,KKR 也推出了「Helix 數位基礎設施」平台,這是一個由輝達、Vistra 等公司支援、規模超過 100 億美元的平台,專注於開發配備預置電力供應的超大規模資料中心。 9 月,三星集團旗下公司又向 Helix 投資了 10 億美元。保障能源供應正成為人工智慧基礎設施投資本身不可或缺的一部分。問題在於,這些私人電力供應系統能否比傳統的電網連接更快保障電力容量。

光纖仍然是主要限制因素。今年,Meta和亞馬遜與康寧公司簽署了數十億美元的光纖協議(Meta的協議金額可能高達60億美元)。 5月份,輝達向康寧投資5億美元,以擴大其光元件製造能力。晶片供應仍然令人擔憂,輝達執行長黃仁勳警告稱,供不應求可能持續數年。電力和網路連結也正成為限制因素,企業無法僅僅透過投入資金來解決。企業需要在資料中心投入運作之前就確保電力供應和網路接入。

客製化晶片仍是業界對抗輝達定價權的有效手段。亞馬遜執行長安迪傑西在7月的財報電話會議上表示,公司以Trainium和Graviton為首的晶片業務,目前年收入已超過250億美元,並以三位數的速度成長。他還補充說,公司本季的營收承諾幾乎翻了三倍。傑西也表示,如果亞馬遜像傳統半導體供應商一樣對外銷售晶片,年收入可達500億美元。亞馬遜估計,與依賴輝達相比,採用Trainium晶片將使其年度資本支出(CAPEX)減少數百億美元,並將利潤率提高數百個基點。谷歌的張量處理單元(TPU)專案和博通為Meta開發的晶片也處於類似的發展軌跡。隨著客製化晶片的成熟,它最終將從輝達手中奪取網路/IT資本支出的市場佔有率。然而,本季的數據尚未顯示出徵兆。網路/IT產業的資本投資成長了80.2%,超過了資本投資整體75.3%的成長率。

2026年第二季度,研發支出佔營收的13.4%,而資本支出佔比為21.5%。這一8.1個百分點的差距在今年持續擴大。在產業發展的大部分歷史時期,研發支出佔收入的比例都高於資本支出,這源自於其以軟體為中心的根基。然而,如今這種關係已經逆轉。該行業正在向硬體優先的成長模式轉型,而資本尚未回歸軟體或貨幣化領域。人均淨有形固定資產持續成長,年增55.5%,而員工人數僅成長了6.2%。

本季併購支出有所成長。以年計算,收購支出達到1,030億美元,比2026年第一季成長近30%,是去年同期水準的三倍多。然而,這仍然只佔總資本支出的一小部分,總資本支出按年計算高達7,015億美元。其根本邏輯仍然不變:傳統收購需要12個月以上的時間才能完成,但在這場快速發展的「軍備競賽」中,這樣的速度遠遠不夠。儘管傳統併購活性化,但資金仍然不成比例地流向基礎設施投資、有針對性的智慧財產權收購和人才收購。 Meta-Scale AI的交易(於2025年中期宣布,涉及超過140億美元的投資,獲得49%的股份)仍然是一個典範,它能夠立即獲得人才和數據標註知識產權,而無需像全面收購那樣面臨整合延遲和監管摩擦。

中國:華為加快「Assemble」藍圖,禁令依然無效。

華為已聲明其生產的昇騰晶片全部售罄。 10月初,輪值董事長舒勉表示,根據華為自身的市場追蹤數據,昇騰晶片在中國市場的銷售量可能已經超越輝達。彭博社於2026年9月10日報道稱,由於供不應求,華為在三個月內將其昇騰950DT加速器的建議零售價提高了約60%,至25萬元人民幣。受高頻寬記憶體(HBM)短缺導致產能受限的影響,華為預計今年僅能生產數十萬顆晶片。光是DeepSeek一家公司就計劃在其位於內蒙古的資料中心部署至少16萬顆晶片。在9月的華為全通大會上,華為宣布將下一代晶片昇騰960DT的發佈時間提前三個季度至2027年第一季,優先滿足中國客戶的需求。此外,華為已承認其現有晶片比輝達落後一到兩代,且目前獨立檢驗的基準測試結果仍然有限,難以將其與輝達硬體進行比較。因此,關於性能相當的說法主要來自於廠商和分析師。 DeepSeek正在加強與華為半導體生態系統的合作。 9月30日,該公司宣布已為華為昇騰AI晶片開發並開放原始碼了包括運算和通訊庫在內的程式設計工具。該計畫還包括一個以128顆昇騰950晶片為核心的超級節點,兩家公司將共同致力於最佳化運算和通訊效能。百度已單獨申請將其AI晶片業務崑崙鑫在香港上市,但估值和上市時間尚未確定。

根據史丹佛大學發布的「2026年人工智慧指數」,截至2026年3月,基於Arena排行榜,中美頂級人工智慧模型之間的效能差距為2.7%。同時,美國在私人人工智慧投資和頂級模型開發方面持續保持領先地位,而中國則在人工智慧相關論文、引用次數、專利數量以及工業機器人部署數量等領域領先。無論晶片層面的比較結果如何,出口限制並未阻止中國繼續建構其國內人工智慧生態系統。

在預測資本支出(Capex)時,這一點至關重要,體現在兩個方面。首先,如果中國模式能夠以​​相對較小的投資保持競爭力,這將削弱「美國超大規模資料中心業者只有透過巨額資本投資才能實現尖端人工智慧」的觀點。其次,中國的顯著技術突破可能會迫使美國重新評估其支出計劃,尤其是在宏觀經濟指標疲軟或人工智慧獲利能力令人失望的時期。

資本支出展望:我們已上調了所有三種情景,在基本案例下,我們預測 2026 年的資本支出約為 8,900 億美元。

2025 年的總資本支出約為 5,070 億美元,其中超大規模資料中心業者僅在 2026 年上半年就支出 4,143 億美元(2026 年第一第一季:1,833 億美元,2026 年第二季:2,310 億美元)。亞馬遜、Alphabet 和 Meta 自年初以來均已上調了其 2026 年的預測,分別約為 2,200 億美元、1,950 億至 2,050 億美元以及 1,300 億至 1,450 億美元。因此,我們也上調了所有三個情境的預測。每個情境的 2026 年預測值均有所提高,而 2027 年至 2028 年的預測軌跡與先前保持一致。在高預測情況下(2026 年:約 10,200 億美元,2027 年:約 10,520 億美元,2028 年:約 10,100 億美元),我們假設支出將在 2026 年下半年繼續加速成長,每季達到約 3,000 億美元。

基本案例(2026年:約8,900億美元,2027年:約8,830億美元,2028年:約8,520億美元)基於更保守的發展假設。具體而言,該情境假設輝達的市場佔有率將被客製化半導體蠶食,部分項目將被整合或推遲,能源和電力限制將減緩技術普及。最後一個假設已被證明是準確的。前文提到的資料中心專案停滯和電網建置延誤正是此情境預測的障礙。對於2026年,該情境預測2026年下半年的支出將與2026年第二季大致持平。

在低迷情境下(2026年約為8,500億美元,2027年約為7,010億美元,2028年約為6,430億美元),我們預計2026年下半年市場規模將逐步放緩至每季約2,180億美元,隨後從2027年開始出現市場調整。這可能由多種因素共同引發,包括宏觀經濟指標疲軟、人工智慧貨幣化進程放緩,以及中國人工智慧領域的突破性進展吸引了全球關注並重塑了全球預期。 OpenAI首次股票公開發行(IPO)的延遲或取消將是一個值得我們密切關注的早期訊號。 2026年6月8日,OpenAI宣布已私下提交了美國IPO的註冊聲明草案;6月26日,「紐約時報」報道稱,薩姆·奧特曼(Sam Altman)強烈敦促顧問將公司估值定為1兆美元。在9月接受「財星」雜誌採訪時,奧特曼表示「現在上市並不明智」,排除了2026年上市的可能性。 OpenAI目前正尋求私募融資。彭博社9月29日報道稱,該公司計劃以1.4兆美元的估值資金籌措至少300億美元。因此,此次延後上市並不代表OpenAI目前面臨資金籌措問題。

盈利:雖然由於 Alphabet 的一次性收益,淨利率達到了歷史新高,但基於現金流的盈利仍在持續下降。

淨利率與自由現金流(FCF)利潤率之間的差距正在擴大,部分原因是Alphabet的收益。 2026年第二季度,淨利率達到創紀錄的26.1%,這是我們自2011年資料庫建立以來的最高值。這一成長主要歸功於Alphabet股權相關資產帶來的一次性非現金收益。根據Alphabet 2026年第二季公佈財報,990億美元的稅前淨利為淨利潤貢獻了771億美元。另一方面,扣除資本支出(CAPEX)且不包含這些收益的自由現金流利潤率則呈現截然相反的趨勢。它再次下降至9.5%的歷史新低。由於自由現金流是經營現金流減去資本支出,這種不斷擴大的差距顯示人工智慧競爭的成本有多高。隨著超大規模資料中心業者支出的增加,其流動性進一步下降。儘管微軟、騰訊和Meta的淨利潤表現強勁,但它們的自由現金流利潤率卻在惡化。這是因為它們在圖形處理器(GPU)叢集上的支出超過了這些業務產生的現金流量。亞馬遜和甲骨文也呈現類似的趨勢,它們將流動性強的現金用於投資回報週期較長的基礎建設。

以人均計算,蘋果以81.6萬美元的人均自由現金流(2026年第二季年化率)領先,高於去年同期的77.3萬美元。 Meta緊追在後,為54.3萬美元,低於去年同期的61.9萬美元。 Meta的基礎設施投資開始超過其廣告業務產生的現金流。在主要公司中,微軟(30萬美元)和Alphabet(26.8萬美元)位居其後。就自由現金流利潤率而言,蘋果(29.3%)在受訪公司中表現最佳,其次是騰訊(23.7%,較去年同期的31.8%大幅下降)和SAP(22.9%)。同時,甲骨文(-35.2%)的「前期投資階段」比去年同期更加深入,低於去年同期的-20.7%。 CoreWeave(-179.9%)和Nebius(-434.2%)是早期專注於新雲端和GPU雲端服務的雲端服務供應商,由於積壓訂單而快速擴張,但目前仍面臨顯著下滑。阿里巴巴(-7.4%)和百度(-10.2%)雖然營收年增,但自由現金流卻轉為負值。營收復甦和現金盈利並不一定相關。

監管罰款對盈利的影響仍然很小。 2026年7月2日,歐盟法院駁回了谷歌針對與安卓系統相關的41億歐元反壟斷罰款的最終上訴。 7月23日,歐盟委員會根據「數位市場法」對谷歌處以約10億美元的罰款,原因是其操縱搜尋排名並限制了應用程式商店Play Store。超大規模資料中心業者仍然主要將這些罰款視為營運成本;也就是說,他們會挑戰法院裁決,承擔罰款,並透過公關活動來控制其影響。

就業:員工人數較去年同期成長 6.2%,與去年同期持平,仍低於長期趨勢。

2026年第二季(2Q26),員工人數年增6.2%,與上一季基本持平,但仍低於2019年至2025年7.9%的複合年成長率。由於亞馬遜、阿里巴巴和京東等公司的物流、履約和配送人員數量對其影響巨大,因此該行業的員工總數仍然是一個難以分析的指標。僅一家公司倉庫的招聘就可能對行業平均水平產生顯著影響,而人工智慧相關員工人數的趨勢則難以預測。如果我們特別關注以技術為中心的大型超大規模資料中心業者,則該趨勢仍在持續下降。

超大規模資料中心業者的經營模式基於巨大的規模經濟,投資集中在邊際生產成本接近零的領域。隨著人工智慧平台能夠分擔更多的工作負載,許多超大規模資料中心業者營運商的經營團隊將樂於大幅裁員,而這仍然是該產業明確的發展方向。過去兩到三年,每位員工的收入和每位員工的淨有形固定資產均大幅成長,而本季度的數據——尤其是淨有形固定資產成長55.5%而員工人數僅成長6.2%——表明這一趨勢沒有放緩的跡象。

區域收入:歐洲的成長率最高(年成長 20.4%),而美洲的表現優於產業平均。

2026年第二季度,美洲地區的營收達到3,962億美元(以單以季度為基礎,佔總營收的46.7%,年增19.1%)。這扭轉了該地區上一季成長乏力的局面,使其成為第二快成長速度的地區。歐洲地區的營收達到1,556億美元(佔總營收的18.3%,年成長20.4%),在本季和上一季均超過了亞太地區和中東及非洲地區,目前是成長速度最快的地區。亞太地區的營收為2,709億美元(佔總營收的31.9%,較去年同期成長14.3%),但成長速度最慢。儘管阿里巴巴、百度和京東在本季度均恢復了成長勢頭,但中國宏觀經濟的不利因素以及國內競爭的影響仍然對該地區構成沉重負擔。中東和非洲地區(MEA)市場規模達265億美元(成長3.1%,年增16.8%)。歐洲市場持續保持強勁成長勢頭,這主要得益於受監管行業的持續需求、人工智慧法律合規推動的雲端遷移以及政府主導的人工智慧舉措。同時,美洲市場的復甦表明,即使總部位於美國的超大規模資料中心業者正在向國際市場擴張,其國內市場仍有充足的空間來消化人工智慧驅動的新增需求。

目錄

  • 1. 報告要點
  • 2. 未來展望
  • 3.分析
  • 4.關鍵統計數據
  • 5.公司詳細分析
  • 6.公司間的比較
  • 7. 按地區分類
  • 8. 原始數據
  • 9.外匯
  • 10. 概述
Product Code: GNI-12102026-1

Snapshot view of 2Q26

In 2Q26, hyperscale revenues reached $849.2B (single-quarter, +17.7% YoY), with annualized revenues reaching $3.264T (+15.8% YoY). Annualized capex surged to $701.5B (+75.3% YoY), driving capital intensity to 21.5% of revenues, well above the telecom sector’s ~16% level. Net PP&E grew +55.5% YoY. Net profit margin hit a fresh record of 26.1%, boosted heavily by a one-time gain at Alphabet; FCF margin, a cleaner read on cash generation, fell again to 9.5%, its lowest since our database begins in 2011. Headcount grew 6.2% YoY in 2Q26. The full report analyzes what is driving each of these numbers, which companies are winning and losing, and what the data says about where this market goes next.

Why read this report

MTN Consulting has been tracking this market with quarterly deep dives since 4Q17, and our database begins in 1Q11. We have a long history of in-depth coverage of this market, and proprietary tools and consistent, comparable data across revenues, capex, R&D, profitability, employment, and balance sheet metrics for the world’s leading hyperscalers (previously called “webscalers”). We cover Amazon, Alphabet, Microsoft, Meta, Apple, Oracle, Alibaba, Tencent, and more than a dozen other players, including newer entrants like CoreWeave, Nebius, Kuaishou, and Xiaomi. No other analyst firm covers this market with this depth or our data coverage.

We have called this market a bubble for almost two years now, and the analysis reflects that view. The 2Q26 report includes detailed company-by-company breakouts, regional data, vendor revenue benchmarks, and three capex outlook scenarios for 2026-2028. If you are making investment, procurement, or strategic decisions that depend on where hyperscale spending goes from here, this report is built for you.

  • Notes:
    • (1) This is MTN Consulting’s 35th quarterly assessment of the hyperscale market, part of a series we launched in 4Q17; our data and analysis spans the 1Q11-2Q26 timeframe, i.e. 62 quarters.
    • (2) The companies in our study include several recent additions: CoreWeave (added in 2024), and Kuaishou, Nebius (Yandex spinoff), and Xiaomi (added in 3Q25).
    • (3) The 4Q25 edition adopts a new term for the sector, swapping “webscale” with the more common “hyperscale”.

Revenue: Topline tops $849B (+17.7% YoY), as HPE, Meta, and Yandex post the biggest gains among established players

Hyperscale revenues reached $849.2B in 2Q26, up 17.7% YoY, with annualized revenues rising to $3.264T (+15.8% YoY). CoreWeave (+112% YoY) and Nebius (+454%) remain the fastest-growing names in our coverage, though both remain small in absolute terms. Among established players, HPE (+33.7%, still absorbing Juniper Networks) posted the largest gain, ahead of Meta (+28.0%) and Yandex (+26.0%) – a name that rarely tops this list. Alphabet (+24.2%), Oracle (+20.6%, increasingly a beneficiary of OpenAI-linked cloud contracts), Amazon (+19.6%), Tencent (+17.9%), and Microsoft (+17.8%) all grew above the sector average. At the bottom of the table, last quarter’s laggards Baidu, JD.Com, and Alibaba all returned to growth, at 1.8%, 3.2%, and 15.4% respectively. The only two companies in our coverage with a revenue decline this quarter were Fujitsu (-5.7%) and Xiaomi (-0.2%).

Advertising still anchors several of the largest names here. Meta remains the most exposed, with ads still generating the overwhelming majority of revenue, even as it keeps trying to diversify into hardware and AI platforms. Alphabet’s non-ad share has climbed past 25% of total revenue. Amazon sits just under 10% of revenue from ads. That advertising exposure is a real risk given concerns about US consumer spending in 2026, concerns that have grown sharper following the energy disruption tied to the Trump administration’s actions in Iran.

The bigger open question for the sector: will AI platforms end up relying on advertising too, given how little traction paid subscriptions have found outside early adopters?

Capex: Annualized spend tops $700B (+75% YoY), and even the industry’s own guidance keeps getting revised upward

Annualized capex reached $701.5B in 2Q26, up 75.3% YoY, with capital intensity rising to 21.5% of revenues – now more than five points above the telecom sector’s ~16%. Amazon, Alphabet and Meta each raised their 2026 capex guidance this quarter: Amazon to ~$220B (from $200B), Alphabet to $195-205B (from $180-190B), and Meta to $130-145B (from $125-145B). Amazon cited higher component pricing and Alphabet faster delivery of capacity. Some of this is real – Oracle’s own contracted backlog reached $664B this quarter, up more than $200B in a single year. But a meaningful share of the sector’s combined backlog still reflects circular financing between the same small set of counterparties. In August 2026, Nvidia agreed to guarantee up to $105B of lease and power payments for OpenAI’s new data center campus in Ohio, where Nvidia will also be the sole chip supplier.

The capex escalation still carries the dynamics of an arms race: no company wants to appear less committed to AI than its rivals, and the scale of investment makes it harder to pull back. With hundreds of billions already committed, stopping would mean writing down assets, admitting the strategy has failed, and giving competitors room to advance. So the spending continues, driven by a mix of genuine ambition, fear of falling behind, and the hope of securing a market position before the economics come under scrutiny.

Power is emerging as a constraint in its own right, though it is too early to say it has overtaken chips as the industry’s biggest bottleneck. The bigger issue is timing: data centers can be built and chips procured within months, while grid connections often take years. Capital is already moving to address the gap. Google, Intersect Power and TPG Rise Climate have committed $20B to a partnership linking new solar, wind and storage capacity directly to data center projects. In June 2026, KKR also launched Helix Digital Infrastructure, a $10B+ platform backed by NVIDIA, Vistra and others, focused on developing hyperscale data centers with power secured upfront. Samsung group companies committed a further $1B to Helix in September. Securing energy supply is becoming part of the AI infrastructure investment itself. The question is whether these private power arrangements can deliver capacity faster than traditional grid connections.

Fiber remains a real constraint too: Meta and Amazon have both signed multi-billion-dollar fiber contracts with Corning this year (Meta’s agreement is worth up to $6B). NVIDIA invested $500M in Corning in May specifically to expand its optical manufacturing capacity. Chip supply remains a concern, with Nvidia CEO Jensen Huang warning that shortages could persist for years. Power and connectivity are also becoming constraints that companies cannot solve simply by spending more. They need to secure energy supplies and network access well before their data centers come online.

Custom silicon remains the industry’s hedge against NVIDIA pricing power. Amazon’s chips business, led by Trainium and Graviton, now has “an annual revenue run rate of over $25 billion” and is growing at triple-digit rates, CEO Andy Jassy said on the July earnings call. He added that revenue commitments nearly tripled in the quarter. Jassy has also said the business could generate $50B a year if Amazon sold the chips externally the way conventional chip vendors do. Amazon estimates Trainium saves it tens of billions of dollars in annual capex versus relying on Nvidia, plus several hundred basis points of margin. Google’s tensor processing unit (TPU) program and Meta’s Broadcom-developed silicon are on similar trajectories. As custom silicon matures, it should eventually take share from Nvidia in Network/IT capex. This quarter’s numbers show no sign of that yet: Network/IT capex grew 80.2%, faster than total capex at 75.3%.

R&D intensity was 13.4% of revenues in 2Q26, against capex intensity of 21.5%. The 8.1-point difference has grown every quarter this year. For most of this sector’s history, R&D spending as a share of revenue ran ahead of capex, a legacy of its software-centric roots. That relationship has now inverted: the sector has shifted to a hardware-first growth model, and capital is not yet flowing back toward software and monetization work. Net PP&E per employee keeps climbing, with net PP&E up 55.5% YoY against 6.2% headcount growth.

M&A spending rose this quarter: annualized acquisition spend reached $103.0B, up nearly 30% from 1Q26 and more than triple the year-ago level. That is still a fraction of the $701.5B annualized capex bill. The underlying logic has not changed: a traditional acquisition takes 12 months or more to close, too slow for an arms race moving this fast. Capital continues to flow disproportionately into infrastructure capex, targeted IP purchases, and acqui-hires alongside the pickup in traditional M&A. The Meta-Scale AI deal (more than $14B for a 49% stake, announced mid-2025) remains the template: immediate access to talent and data-labeling IP, without the integration lag or regulatory friction of a full buyout.

China: Huawei pulls its Ascend roadmap forward; the embargo still isn’t working

Huawei says it is selling every Ascend chip it can make, and rotating chairman Eric Xu said in early October that Ascend has likely overtaken Nvidia in China, based on Huawei’s own tracking. Bloomberg reported on 10 September 2026 that Huawei had raised the suggested price of its Ascend 950DT accelerator by around 60% over three months, to CNY 250,000, as demand outstripped supply. The company is expected to produce only a few hundred thousand units this year, with shortages of high-bandwidth memory (HBM) constraining output. DeepSeek alone plans to deploy at least 160,000 of them at a data center in Inner Mongolia. At Huawei Connect in September, Huawei brought its next chip, the Ascend 960DT, forward by three quarters to 1Q27, and said it is prioritizing Chinese customers. Huawei has also acknowledged that its current chips remain one to two generations behind Nvidia, while independently verified benchmarks against Nvidia hardware remain limited. Claims of performance parity therefore remain largely vendor- or analyst-sourced. DeepSeek is strengthening its ties with Huawei’s silicon ecosystem. On 30 September, it announced that it had developed and open-sourced programming tools for Huawei’s Ascend AI chips, including compute and communication libraries. The work also covers a supernode built around 128 Ascend 950 chips, with both companies optimizing computation and communication. Baidu has separately filed for a Hong Kong listing of its AI-chip unit, Kunlunxin, although the valuation and timing remain unsettled. Baidu has separately filed for a Hong Kong listing of its AI-chip unit, Kunlunxin, although the valuation and timing remain unsettled.

Stanford’s 2026 AI Index puts the gap between the top US and Chinese AI models at 2.7% as of March 2026, measured on the Arena leaderboard. At the same time, the US continues to lead in private AI investment and the production of top-tier models, while China leads in areas including AI publications, citations, patents and industrial robot installations. Whatever the chip-for-chip comparison, export controls have not prevented China from continuing to build out its domestic AI ecosystem.

This matters for capex forecasts in two ways. First, if Chinese models remain competitive at a fraction of the investment, it undercuts the argument that massive US hyperscaler capex is the only path to frontier AI. Second, a high-profile Chinese breakthrough could force a rapid reassessment of US spending plans, especially if it lands alongside weak macro data or disappointing AI monetization.

Capex Outlook: We raise all three scenarios, with 2026 at ~$890B in our base case

Capex closed 2025 at ~$507B, and hyperscalers spent $414.3B in the first half of 2026 alone ($183.3B in 1Q26 and $231.0B in 2Q26). Amazon, Alphabet and Meta have all raised their 2026 guidance since early in the year, to ~$220B, $195-205B and $130-145B. We have raised all three scenarios to match. Each case now starts from a higher 2026 level, and the 2027-28 path keeps the same shape as before. The high case (~$1,020B in 2026, ~$1,052B in 2027, ~$1,010B in 2028) assumes spending keeps accelerating through 2H26, to about $300B a quarter.

The base case (~$890B in 2026, ~$883B in 2027, ~$852B in 2028) assumed more conservative execution: Nvidia losing share to custom silicon, some projects consolidating or delaying, and energy and power constraints slowing deployment. That last assumption has held up: the stalled data center project and grid delays described above are the kind of friction this scenario anticipated. For 2026, it implies 2H26 spending roughly flat with 2Q26.

The low case (~$850B in 2026, declining to ~$701B in 2027 and ~$643B in 2028) assumes a market correction beginning in 2027, after a modest 2H26 slowdown to about $218B a quarter. The likely trigger is a combination of weak macro data, soft AI revenue monetization, and a high-profile Chinese AI breakthrough that resets global expectations. A delayed or canceled OpenAI initial public offering (IPO) is the early signal we watch. OpenAI said on 8 June 2026 that it had confidentially submitted a draft registration statement for a US IPO, and The New York Times reported on 26 June that Sam Altman was pressing advisers for a $1 trillion valuation. In a September interview with Fortune, Altman said that “right now would be an ill-advised moment to go public” and ruled out a 2026 listing. OpenAI is raising private money instead: Bloomberg reported on 29 September that it is targeting at least $30B at a $1.4 trillion valuation. So the delay does not yet point to a funding problem.

Profitability: Net margin reaches a record on a one-time Alphabet gain, while cash profitability keeps sliding

The divergence between net margin and FCF margin has widened, and an Alphabet gain explains part of it. Net profit margin reached a record 26.1% in 2Q26, the highest in our database, which goes back to 2011. A large share of that jump traces to a one-time, non-cash gain on equity securities at Alphabet. Alphabet’s 2Q26 earnings release shows the $99.0B pre-tax gain added $77.1B to net income. FCF margin nets out capex and does not include that gain, and it tells the opposite story: it fell again, to 9.5%, a new low. FCF is operating cash minus capex, so the widening divergence measures what the AI arms race costs. The harder hyperscalers spend, the more they draw down their own liquidity. Microsoft, Tencent and Meta report strong net income while their FCF margins erode, because graphics processing unit (GPU) cluster spending runs ahead of the cash these businesses generate. Amazon and Oracle follow the same pattern, trading liquid cash for infrastructure with long payoff horizons.

On a per-employee basis, Apple leads at $816K in FCF per employee (2Q26 annualized), up from $773K a year ago, followed by Meta at $543K, down from $619K. Meta’s infrastructure spending has started to outpace the cash its ad business generates. Microsoft ($300K) and Alphabet ($268K) follow among the larger players. By FCF margin: Apple (29.3%) leads our coverage, followed by Tencent (23.7%, down sharply from 31.8%) and SAP (22.9%). At the other end, Oracle (-35.2%) is even deeper in its build-ahead phase than a year ago, when it stood at -20.7%. CoreWeave (-179.9%) and Nebius (-434.2%) remain deeply negative as early-stage neoclouds, GPU-focused cloud providers, scale rapidly against contracted backlogs. Alibaba (-7.4%) and Baidu (-10.2%) are both more FCF-negative than a year ago, even as their revenue growth has turned positive. Top-line recovery and cash profitability do not move together.

Regulatory fines remain a minor drag on profitability. On 2 July 2026, the Court of Justice of the European Union dismissed Google’s final appeal against a €4.1B Android antitrust fine. On 23 July, the European Commission fined Google about $1B under the Digital Markets Act, over search ranking and Play Store restrictions. Hyperscalers continue to treat fines largely as a cost of doing business: they fight rulings in court, absorb the fines and use public relations to manage the fallout.

Employment: Headcount growth holds flat at +6.2% YoY, still trailing the long-run trend

Headcount grew 6.2% YoY in 2Q26, essentially flat with last quarter and still running below the 2019-25 CAGR of 7.9%. Total headcount remains a tricky metric for this sector, since the employee base is heavily influenced by logistics, fulfillment, and delivery staff at companies like Amazon, Alibaba, and JD.Com – a single company’s warehouse hiring can move the sector average more than any AI-related headcount trend. Among the more tech-centric hyperscalers, the direction continues to point downward.

The hyperscale business model is built around massive economies of scale, with investment concentrated in areas where the marginal cost of production can approach zero. Many hyperscaler executives would welcome a meaningfully smaller headcount if AI platforms can absorb more of the operational load – and that remains the clear direction of travel for the industry. Revenue per employee and net PP&E per employee have both climbed sharply over the last 2-3 years, and this quarter’s data – particularly the 55.5% jump in net PP&E against 6.2% headcount growth – shows no sign of that trend slowing.

Regional Revenues: Europe grows fastest (+20.4% YoY) and the Americas move above the sector average

Americas revenues reached $396.2B in 2Q26 (single-quarter, 46.7% of total, +19.1% YoY) – a reversal from last quarter, when the region grew below the sector average; it’s now the second-fastest-growing region. Europe reached $155.6B (18.3%, +20.4% YoY) and is now the fastest-growing region, ahead of Asia Pacific and MEA both this quarter and last. Asia Pacific contributed $270.9B (31.9%, +14.3% YoY), the slowest-growing region, still weighed down by Chinese macro headwinds and domestic competition, even as Alibaba, Baidu, and JD.Com all returned to growth this quarter. MEA reached $26.5B (3.1%, +16.8% YoY). Europe’s continued outperformance reflects sustained demand from regulated industries, AI Act compliance-driven cloud migration, and sovereign AI initiatives; the Americas’ rebound suggests US-based hyperscalers’ home market still has plenty of room to absorb fresh AI-driven demand even as they expand abroad.

Table of Contents

  • 1. Report highlights
  • 2. Outlook
  • 3. Analysis
  • 4. Key Stats
  • 5. Company Drilldown
  • 6. Company Benchmarking
  • 7. Regional Breakouts
  • 8. Raw Data
  • 9. Exchange Rates
  • 10. About

Figures & Charts

  • 1. Capex outlook by scenario ($B)
  • 2. Key Metrics: Growth rates, Annualized 2Q26/2Q25 vs. 2021-25
  • 3. Hyperscale Revenues: Single-quarter & annualized (US$M)
  • 4. Key Hyperscaler revenues: YoY % revenue growth in 2Q26
  • 5. Annualized profitability: HYPERSCALERS
  • 6. Free cash flow per employee, 2Q26 annualized (US$)
  • 7. Key Hyperscaler free cash flow margins: 2Q26 annualized
  • 8. Advertising revenues as % total (FY2025)
  • 9. Annualized capex and R&D spending: HYPERSCALERS (% revenues)
  • 10. Hyperscale capex by type, Annualized: 2Q15-2Q26 (US$M)
  • 11. Network & IT capex as share of revenues, 2Q26 annualized
  • 12. Key Hyperscaler R&D expenses, % revenues: 2Q26 annualized
  • 13. Acquisition spending vs. capex spending, annualized (US$M)
  • 14. Net PP&E per employee (US$’ 000) – 2Q26
  • 15. Ranking the Hyperscale Network Operators: Revenues; R&D; Capex; Network & IT capex – 2025 & 2Q26 (US$B)
  • 16. Annualized spending for key Hyperscalers since 2011 Capex: Network, IT and software
  • 17. Share of Hyperscale spending by company, 2Q26 and 2Q25 annualized (Capex: Network, IT and software)
  • 18. Energy consumption vs. Net PP&E for key Hyperscalers in 2025
  • 19. Hyperscale vs. Telco Market: Annualized Capex (US$B)
  • 20. Hyperscale vs. Telco Market: Annualized capital intensity
  • 21. USA: Hyperscale capex total ($M) and % of global market, 2011-25
  • 22. China’s Hyperscalers versus the big 4: Capex in 2025 ($B)
  • 23. Chinese Hyperscale capex on the rise again ($M)
  • 24. Hyperscale tech capex vs. DC-related chip revenues (annualized $M)
  • 25. Hyperscale tech capex vs. Data center-related systems revenues (annualized $M)
  • 26. Revenues: annual, single-quarter, and annualized (US$M)
  • 27. Profitability (Net Profit; Cash from operations; Free cash flow): annual, single-quarter, and annualized (US$M)
  • 28. Spending (R&D; M&A; Capex; Network & IT capex; Lease): annual, single-quarter, and annualized (US$M)
  • 29. Cash & Short-term Investments: annual and single-quarter (US$M)
  • 30. Debt (Total debt; Net debt): annual and single-quarter (US$M)
  • 31. Property, Plant & Equipment: annual and single-quarter (US$M)
  • 32. Key Ratios: Net margin; R&D/revenues; Capex/revenues; Network & IT capex/revenues; Free cash flow/revenues; Lease costs/revenues – annual and annualized (%)
  • 33. Total employees
  • 34. Revenue per employee, annualized (US$K)
  • 35. FCF per employee, annualized (US$K)
  • 36. Net PP&E per employee, annualized (US$K)
  • 37. Revenues & Spending (US$M)
  • 38. Revenues (US$M) & YoY revenue growth (%), single-quarter: by company
  • 39. Revenues, annualized (US$M): by company
  • 40. Annualized profitability margins: by company
  • 41. Annualized capex and capital intensity: by company
  • 42. Annualized capex and R&D spending as % of revenues: by company
  • 43. Share of hyperscale network & IT capex, Annualized: by company
  • 44. Total employees: by company
  • 45. Annualized per-employee metrics (US$000s): by company
  • 46. Net debt (debt minus cash & stock) (US$M): by company
  • 47. Top 10 Hyperscale employers in 2Q26: Global market
  • 48. Headcount changes in 2Q26 (YoY %): Global market
  • 49. Net PP&E: USA vs. RoW (by company)
  • 50. Net PP&E: total in $M and % global Hyperscale market (by company)
  • 51. Energy consumption, MWh and % Hyperscale total (by company)
  • 52. Share of Hyperscale energy consumption, net PP&E, and capex (by company)
  • 53. Energy intensity relative to Hyperscale average and select data center-focused CNNOs (by company)
  • 54. Energy intensity in Hyperscale sector, 2025: MWh consumed per $M in revenue
  • 55. Capex/revenues (annualized): Company vs. Hyperscale average
  • 56. Revenue per employee (US$000s) (annualized): Company vs. Hyperscale average
  • 57. 2018 vs. 2025: company benchmark by KPI (Revenues, R&D, Net profit, Cash from operations, Capex, Free cash flow, Cash & short-term investments, Net PP&E, Total debt)
  • 58. 2018 vs. 2025: company benchmark by key ratio (Capex/revenues; R&D/revenues; Net margin; FCF margin)
  • 59. Top 8 Hyperscaler share vs. Rest of the market: by KPI (Revenues, R&D, Net profit, Cash from operations, Capex, Free cash flow, Cash & short-term investments, Net PP&E, Total debt)
  • 60. Top 8 Hyperscalers benchmarking by Key ratio: Capex/revenues; R&D/revenues; Net margin; FCF margin)
  • 61. Total Hyperscaler Market Revenues, by region: Latest CY; Latest Quarter; Annual trend (2011-25); Single quarter (2Q17-2Q26)
  • 62. Hyperscaler Market: Revenues, single-quarter (YoY % change)
  • 63. Regional revenues by operator: Latest CY; Latest Quarter; Annual trend (2011-25); Single quarter (2Q17-2Q26)
  • 64. Top 10 operators by region: Latest CY; Latest Quarter