2026+AI+繁榮通關(guān)指南:泡沫何時(shí)現(xiàn)形+A+Brief+Guide+to+Boom+or+Bust_第1頁
2026+AI+繁榮通關(guān)指南:泡沫何時(shí)現(xiàn)形+A+Brief+Guide+to+Boom+or+Bust_第2頁
2026+AI+繁榮通關(guān)指南:泡沫何時(shí)現(xiàn)形+A+Brief+Guide+to+Boom+or+Bust_第3頁
2026+AI+繁榮通關(guān)指南:泡沫何時(shí)現(xiàn)形+A+Brief+Guide+to+Boom+or+Bust_第4頁
2026+AI+繁榮通關(guān)指南:泡沫何時(shí)現(xiàn)形+A+Brief+Guide+to+Boom+or+Bust_第5頁
已閱讀5頁,還剩51頁未讀, 繼續(xù)免費(fèi)閱讀

下載本文檔

版權(quán)說明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請進(jìn)行舉報(bào)或認(rèn)領(lǐng)

文檔簡介

AdrianCox|WouldtherealAIbubblepleasestandup?|December,2025

DeutscheBank

Research

WouldtherealAIbubblepleasestandup?

Abriefguidetoboomorbust

December2025

AdrianCox|StefanAbrudan

DeutscheBankResearchInstitute

IMPORTANTRESEARCHDISCLOSURESLOCATEDINAPPENDIX1.

1AdrianCox|WouldtherealAIbubblepleasestandup?|December,20251

IsthereanAIbubble?Oraretheremorethanone?

IsthereanAIbubble?Withvaluationssurging

andaneconomyridingonAIcapitalexpenditurejustthreeyearssincethelaunchofChatGPT,

thisisthetopquestiononthemindofinvestors–particularlythegenerationscarredbythedot-combubbleofthelate1990s.

Ifthisisabubble,itisstillinitsearlystages.

Retreatingnowrisksleavingsignificantgainsonthetable.Itwasmorethanthreeyearsafter

then-USFederalReserveChairmanAlan

Greenspanwarnedof“irrationalexuberance”inDecember1996thatthedot-combubbleburst.

Infact,thedot-combubblewasreallytwo

bubbles:onewithlightlycapitalised,never-

profitableIPOsandtheotherwithheavily

indebtedtelecomscompaniesthatlaiddownfibreopticcablesthatremaineddarkforyears.

Thereisalsomorethanoneboom(orbubble)thistime.Thechargeisledbywell-established

bigtechcompanieswithmultiplerevenue

streams,whoarepayingfortheirinvestmentindatacentresmostlyoutoffreecashflowand

fromwhichtheyaregeneratingimmediate

returnsfromenterprisecustomers.The

unprofitablecompaniesatthecuttingedgeof

modelandapplicationdevelopmentarestill

private,withspendingcommitmentsthatmayormaynotactuallybefulfilleddependingonhowtheirbusinessmodelsevolve.

Partoftheissuecomesdowntonaming.JustastheclichégoesthatInuitpeoplehave100wordsforsnow,thereisnotjustonekindofbubble.

Thisreportaimstoseparateoutatleastthree

differentkindsofbubble:valuations,investmentandtechnology.Weexplainwhywethinkthat

reportsofabubbleareexaggerated(fornow)aswellaswhatcouldgowrong.

WouldtherealAIbubblepleasestandup?

1.Valuations

2.Investment

3.Technology

2AdrianCox|WouldtherealAIbubblepleasestandup?|December,20252

WouldtherealAIbubblepleasestandup?

Redflags

1.Elevatedvaluations

2.Historicinvestment

3.FlawedtechnologyGreenflags

1.Valuations:rangebound,earnings-led,publicvsprivate

2.Investment:intrend,realreturns,cashflowfunded

3.Technology:stillscaling,demand-led,fallingcostsWhythismatters

Whatcouldgowrong

1.Circularfinancevaluations

2.Debt-fuelledinvestment

3.Technologicalhurdles

4.Socialorpoliticalbacklash

5.Supplybottlenecks

Source:DeutscheBankResearch;imagebyChatGPT

3AdrianCox|WouldtherealAIbubblepleasestandup?|December,20253

Redflags:1.Valuations:levelsarenearingthehistoricpeaksofthedot-comera

TheShillerCyclicallyAdjustedPrice/Earningsratioishaspassed40,nottoofarbelowitspeakofalmost44in2000

ShillerCAPEratiofortheS&P500

50

CyclicallyAdjustedPrice/Earnings(CAPE)ratio

45

40

35

30

25

20

15

10

5

0

187218811890189919081917192619351944195319621971198019891998200720162025

Source:ShillerCyclicallyAdjustedPrice/Earnings(CAPE)ratio,DeutscheBankResearch

4AdrianCox|WouldtherealAIbubblepleasestandup?|December,20254

Redflags:2.Investment:AIdatacentrecapexmayhithistoric$4trilliontotalby2030

Capex,ledbyhyperscalers,isforecasttoexceed10xinflation-adjustedcostofApolloprogrammewithnoguaranteedreturn

$bn

700Hyperscalercapexprojectedtoreach$500bnin202670%

60060%

50050%

40040%

30030%

20020%

10010%

00%

20152016201720182019202020212022202320242025e2026e2027e2028e2029e2030e

AmazonGoogleMetaMicrosoftOracleYoYgrowth

Source:BloombergFinanceLP,DeutscheBankResearch

5AdrianCox|WouldtherealAIbubblepleasestandup?|December,20255

Redflags:3.Technology:generativeAIisflawedandmayhitawall

Thetechis(still)pronetoerrorandhallucinations;canbehardtoapplyatscale;andfacesbottlenecksinfurtherscaling

AI’srapidscalingmaysoonhitphysicalbarrierssuchaslimitsonhowfastdata

canbemovedbetweenchips,someresearcherssay

Source:EgeErdil,andDavidSchneider-Joseph.‘Datamovementlimitstofrontiermodeltraining,.ArXiv[cs.DC],2024.arXiv.

/abs/2411.01137

6AdrianCox|WouldtherealAIbubblepleasestandup?|December,20256

Greenflags:1.1Valuations:equitiesarestillatthelowendofthepost-2022trend

Historicallyelevatedvaluationsareduetostructuralandcyclicalfactorsandsupportedbyarobustdemand-supplybalance

Log

3500

S&P500trendchannelsinceOct2022

Jul-22

Oct-22

Jan-23

Apr-23

Jul-23

Oct-23

Jan-24

Apr-24

Jul-24

Oct-24

Jan-25

Apr-25

Jul-25

Oct-25

Jan-26

Trendgrowthrate(Oct2022toMar2025)=22.7%(annualised)Range=+/-4.6%

Log

3500

Source:BloombergFinanceLP,DeutscheBankResearch

AdrianCox|WouldtherealAIbubblepleasestandup?|December,20257

Greenflags:1.2Valuations:techstockrallyhasbeendrivenbyearningsgrowth

The60%valuationpremiumfortechhasbeenjustifiedby20%+earningsgrowthdifferential;nowearningsarebroadening

Bigtechvaluationsarenotnearlyasstretchedasinthetechbubble

Mega-capgrowth*andtechstocksnext-12-months(NTM)P/ErelativetorestofS&P500

2.8

2.6

2.4

2.2

2.0

1.8

1.6

1.4

1.2

1.0

0.8

1990

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

Recession

2.8

2.6

2.4

2.2

2.0

1.8

1.6

1.4

1.2

1.0

0.8

*MSFT,AAPL,AMZN,GOOGL,GOOG,FB,V,MA,NVDA,NFLX,ADBE,TSLA(sinceDec2020)Source:DeutscheBankResearch

8AdrianCox|WouldtherealAIbubblepleasestandup?|December,20258

Greenflags:1.3Valuations:highestvaluationsarenon-profitableprivatecompanies

ThepricetosalesratioofOpenAIis38andAnthropicis44,vsNvidiaat22,Microsoftat12,Googleat9.9andAmazonat3.5

$5tr

$4tr

$3tr

$2tr

$1tr

$0tr

Valuationsofnotablebigtechcompaniesinthepublicandprivatemarkets

Marketvaluation(LHS)Revenue(RHS)

NvidiaGoogleMicrosoftAmazonMetaTeslaOpenAIxAIAnthropicPerplexity

$700bn

$600bn

$500bn

$400bn

$300bn

$200bn

$100bn

$0bn

Source:Sacra,BloombergFinanceLP,DeutscheBankResearch;OpenAIisbasedonitsreportedforecastrevenueof$13bnfor2025,ratherthanitsestimateof$20bnannualrecurringrevenueatyear-end

9AdrianCox|WouldtherealAIbubblepleasestandup?|December,20259

Greenflags:2.1Investment:capexgrowthisstillinlinewithstrongtrendfrom2013

Atlessthan40%ofEBITDA,techcapexisstillwellbelowlevelsoflate1990sandisnowjustatlevelsofrestofS&P500

Globaltechnologycapex(quarterly,$bn,inlogs)

5.65

5.25

4.85

4.45

4.05

3.65

3.25

5.65

5.25

4.85

4.45

4.05

3.65

3.25

Q42026consensus

Trendsince2013:12.3%,annualrate

Width:+/-19%

2012

2013

2014

2015

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

2027

*MCG&Tech+othermembersofS&P1200GlobalTechindex+FTSEChinaTechindex

Source:BloombergFinanceLP,DeutscheBankResearch

10AdrianCox|WouldtherealAIbubblepleasestandup?|December,202510

Greenflags:2.2Investment:bigtechreturnshavegrownsincestartofAIcycle

HyperscalersaregeneratingnewreturnsfromAIviacustomerclouddemand,AI-poweredtoolsandcostsavingsoncoding

Aggregatereturnoninvestedcapital(%)

MCG&Tech——RestoftheS&P500*

21

19

17

15

13

11

9

7

5

1995

1997

1999

2001

2003

2005

2007

2009

2011

2013

2015

2017

2019

2021

2023

2025

*ExcludesFinancials,RealEstateandUtilities

Source:BloombergFinanceLP,DeutscheBankResearch

21

19

17

15

13

11

9

7

5

11AdrianCox|WouldtherealAIbubblepleasestandup?|December,202511

Greenflags:2.3Investment:hyperscalersarefundingcapexmainlyviafreecashflow

With,egGooglegenerating$48bninoperatingcashflowinQ3,hyperscalerscanaffordtoinvestwithoutexternalcapital

Hyperscalersmostlyhaveacapitalexpenditure/operatingcashflowratioofwellunder1

1.6

1.4

1.2

1.0

0.8

0.6

0.4

0.2

0.0

Amazon

Google

Meta

Microsoft

Oracle

20152016201720182019202020212022202320242025

Source:BloombergFinanceLP,DeutscheBank

12AdrianCox|WouldtherealAIbubblepleasestandup?|December,202512

Greenflags:3.1Technology:scalingisstilldeliveringimprovementsincapabilities

LaunchofGoogle’sGemini3inNovembershowedthatAIhasnotyethitawall,withsignificantmultimodaladvances

50%

40%

30%

20%

10%

0%

Gemini3exceedsallpreviousmodelsinHumanity'sLastExam,whichtestsreasoningandknowledge

Gemini3DeepGemini3ProGemini2.5ProGPT-5ProGPT-5.1ClaudeSonnet

Think4.5

Gemini3sawthebiggestimprovementinvisual

reasoning,whereitscoredthreetimeshigherthan

GPT-5ProinARC-AGI-2

50%

40%

30%

20%

10%

0%

Gemini3DeepGemini3ProGemini2.5ProGPT-5ProGPT-5.1ClaudeSonnet

Think4.5

Source:Google,DeutscheBank

13AdrianCox|WouldtherealAIbubblepleasestandup?|December,202513

Greenflags:3.2Technology:AIrolloutisjustgettingstartedanddemandissurging

GooglesaidinOctoberitisprocessing1.3quadrilliontokensamonth,upfrom9.7trillioninApril2024,asAIworkloadsrise

10%

9%

8%

7%

6%

5%

4%

3%

Evennow,fewerthan10%ofUSbusinessesareusingAI,BusinessTrendsandOutlookSurvey

shows

202320242025

Pricesforelectroniccomponentsandaccessoriesare

surgingamidboomindemand

6%

5%

4%

3%

2%

1%

0%

-1%

-2%

-3%

-4%

20152016201720182019202020212022202320242025

Source:USCensusBureau,BureauofLaborStatistics,DeutscheBankResearch

14AdrianCox|WouldtherealAIbubblepleasestandup?|December,202514

Greenflags:3.3Technology:demandisalsofuelledbytumblingcostsandnewuses

AIfollowsJevon’sParadox,wheregreaterefficiencyandlowercostsboostconsumption,meaningnochipsarelyingidle

CostinUSDpermilliontokens

Costofcheapestlargelanguagemodel(LLM)witha

100

10

1

0.1

0.01

GPT-3.5($20)

Llama27b

Llama

GPT-3.5-turbo

Llama38b

3.18b

minimumscoreof42inMMLUbenchmarkhas

decreasedbyafactorof1,000

GPT-3($60)

Llama

3.23b

100%

75%

50%

25%

0%

TheshareofreasoningtokensusedbycustomersofOpenRouternowexceedsnon-reasoningtokens

Nov24Jan25Mar25May25Jul25Sep25Nov25

ReasoningNon-reasoning

Source:GuidoAppenzeller/together.ai,OpenRouter,DeutscheBankResearch

15AdrianCox|WouldtherealAIbubblepleasestandup?|December,202515

Whydoesthismatter?InvestmentinAI-relatedsectorsiscriticaltoGDPgrowth

USwouldbeclosetorecessionthisyearifitweren’tfortech-relatedspending,asotherspendinghasflatlinedpost-Covid

2.5

USunderlyingrealGDPgrowthattributedtotech

spending(%qoq,saar)

2

1.5

1

0.5

0

-0.5

Realfixedinvestmentssplitbetweenitemsrelatedto

140

AIanditemsnon-related

AI-sensitiveitemsItemsexcludingAI

130

Index;Q42019=100

120

110

100

90

80

201920202021202220232024

Source:USBureauofEconomicAnalysis,BureauofEconomicAnalysis,DeutscheBankResearch;underlyingGDPmeasuredasrealfinalsalestoprivatedomesticpurchasers;techspendingincludesthesoftwareandITequipmentcomponentsoffixedinvestmentanddomesticsales

16AdrianCox|WouldtherealAIbubblepleasestandup?|December,202516

Whatcouldgowrong?1.Valuations:circularfinancingmayleadtoopaquevaluations

Recentcomplexagreements,suchasOpenAI’s$1.4trnincomputecommitmentsovereightyears,maycarrysystemicrisks

SemiconductorCompanies

PrivateAICo's/GovernmentEntitiesHyperscaleCloudCompanies

OCIbuysupto~800kNVDAGPUs

OCItobuy~50kMI450GPUs

NVDAinvests~$100binOAI;OAIbuys~6GWofsystems(DBe$100binsales)

OAIpurchases~$300b

(~4.5GW)worthofcomputefromOCI(startingin2027)

NVDAinvests~$5binINTC;INTCgivesup~4%stake

OAIbuys~6GWofAMDGPUs(DBe$100bin

sales)andgetswarrantstobuy~160mshares

USGgets~15%ofAMD'schipsalestoChina

OAIspends$38b

onAWS,gaining

accesstohundredsofthousandsof

NVDAGPUs

MSFThasinvested$13binOAIsince2019;Stakeisnowestimatedtobe$130b

USGov

USGgets~15%ofNVDA'schipsalestoChina

USGgets~10%equityinINTC;INTCgets~$9bcash

OAIgetscustomASICs;AVGOgets~10GWworthofproductsales(DBe~$100binsales)

NVDA&MSFTinvest$15b(NVDA~$10b)inAnthropic,who,in-return,commitsto

buying~$30bofcomputingcapacity(equiv.to~1GWofNVDABlackwell/RubinGPUs)

HUMAINwillbuy~0.2GWofQCOM'sAI200and

AI250rackstoadvanceSaudiArabia'sDCinfra.

Source:DeutscheBankResearch,companypressreleases.Note:Schematicaboveisupdatedonanongoingbasisfollowingofficialcompanyannouncements;Thedealsoutlinedabovearenotexhaustive;NotethatthereareothermajorAI-centricdealsinvolvingsoftware/cloudcompanies;FormoredetailsreachouttotheDBSemiconductorsTeam

17AdrianCox|WouldtherealAIbubblepleasestandup?|December,202517

Whatcouldgowrong?2.Investment:costscouldspiral,forcingcompaniesintodebt

Eventhecash-richhyperscalershavebeguntoissuefarmoredebt,withdownstreamcompaniesalsosettoincreaseissuance

5

4

3

2

1

0

-1

-2

-3

-4

NetDebt/EBIDTA

——AmazonGoogleMetaMicrosoftOracle

20152016201720182019202020212022202320242025

35

30

25

20

15

10

5

0

GrossUSDinvestmentgradeissuanceperyear($bn)

n202220232024n2025

MSFTGOOGLMETAAMZNORCL

Source:BloombergFinanceLP,DeutscheBankResearch

18AdrianCox|WouldtherealAIbubblepleasestandup?|December,202518

ArtificialAnalysisIntelligencescore

Whatcouldgowrong?3.Technology:practicalhurdles,prohibitivecostsorslowgains

Asscalingshowsdiminishingreturns,developersmaynotkeepfindingotherwaystodelivervaluethatworks“inthefield”

70

60

50

40

30

20

10

0

Everyincreaseinintelligencerequiresanexponentialincreaseincost

Grok4

DeepSeekR1

Grok3

GPT-4.5

DeepSeekV3

ClaudeSonnet3.5Llama4

Gemini1Ultra

Llama2

GPT-3

Trainingcomputecost(USD,Logscale)

1,000,00010,000,000100,000,0001,000,000,000

100%

80%

60%

40%

20%

0%

Artificialgeneralintelligence(AGI)iskeyto

justifyinginvestments,yetitsimpliedtiminghasbeendelayed,despiterisingdatacentrespending

2022202320242025

ProbabilityofdevelopingAGIin5yearsbasedondatacenterspending

ProbabilityofdevelopingAGIin5yearsbasedonscientificliteratureandmodeldevelopment

Source:ArtificialAnalysis,EpochAI,DeutscheBankResearch

19AdrianCox|WouldtherealAIbubblepleasestandup?|December,202519

Whatcouldgowrong?4.Social,politicalbacklashonfearsofcontrol,privacy,jobs

RisingscepticismaboutAIcouldleadtocustomerboycotts,employeeresistanceandrestrictiveregulation

25%

20%

15%

10%

5%

0%

PercentageofpeoplewhoareveryconcernedthatAIwilltaketheirjobsoverthecomingyears

12months2years5years12months2years5yearsUK+EUUS

MorepeopleindevelopedmarketsresistgreaterAIuse,whilethoseinBrazilandChinaembraceit

IrejectAIIembraceAI

US

UK

Germany

China

Brazil

-50%-40%-30%-20%-10%0%10%20%30%40%50%60%

Source:dbDataInsightssurveyof10,000respondentsinUS,UK,France,Germany,ItalyandSpain,Sept-Nov2025;2025

EdelmanTrustBarometerFlashPoll:TrustandAIataCrossroads

;DeutscheBankResearch

20AdrianCox|WouldtherealAIbubblepleasestandup?|December,202520

CentsperKWh

Whatcouldgowrong?5.Practicalorgeopoliticalsupplybottlenecks,egenergy,chips

Thebiggesthurdletoadoptionandmonetisationissupply,particularlyelectricity,wherecapacitywilltakeyearstobuild

TWh

1000

750

500

250

Demandforelectricityisprojectedtobefourtimes

largerin2030thanin2020

HistoricalProjected

2024

ChinaEurope

2030

2020

UnitedStates

Asiaexcl.ChinaRestofworld

18

UShouseholdsarealreadypayingrecordelectricitypricesthisyearasdemandincreases

17

16

15

14

13

12

11

10

9

8

Source:IEA,DeutscheBankResearch

AdrianCox|WouldtherealAIbubblepleasestandup?|December,202521

VisittheDeutscheBankResearchInstitutesiteformore

DeutscheBankResearchInstitute:Technology

Happybirthday,ChatGPT?OpenAIfacesthreethreatsasAltmanissues"codered"

(Dec3,2025)

AIsemiconductorsupplychainoverview

(Nov28,2025)

AI101:Economy:FivewaysAIisdrivinggrowth

(Nov18,2025)

EuropeanspendingonChatGPThasstalledsinceMay:dbDataInsights

(Oct14,2025)

AI101:Technology:hype-freeguideforusers

(Oct2,2025)

“AIbubble”bubblebursts

(Sept30,2025)

TheSummerAITurnedUgly:Part2

(Sept4,2025)

TheSummerAITurnedUgly:Part1

(Sept2,2025)

TheauthorswouldliketothanktheirDeutscheBankResearchcolleaguesforinspirationandcharts,notablyfromthefollowingreports(onlyavailabletoclients)

CapitalGoods:2026sectoroutlook

(GaeldeBrayetal,Dec7,2025)

IG&HYStrategy:2026Outlook:Fire&Ice:RoundII

(SteveCaprioetal,Nov25,2025)

AssetAllocation:2026USEquitiesOutlook:StayingConstructive

(BinkyChadha,ParagThatte,Nov25,2025)

22AdrianCox|WouldtherealAIbubblepleasestandup?|December,202522

Authors

AdrianCoxisaManagingDirectorandThematicStrategistatDeutscheBankResearchinLondon,focusingontheimplicationsofAIforinvestors,enterprisesandsociety.HeisalsoManagingDirectorofthe

DeutscheBankResearchInstitute

.HejoinedDeutsche

Bankin2009afterworkingasanaward-winningjournalistandeditoratBloombergNewsandtheFinancialTimesinLondon,BrusselsandNewYork.HeisagraduateofCambridgeUniversityandhasanMBAfromCityUniversityinLondon.

StefanAbrudanisaresearchanalystontheThematicResearchteaminLondon,

focusingonartificialintelligenceanddigitalassets.HeholdsaBScinEconomicsfromErasmusUniversityRotterdamandpreviouslyworkedinsalesandtradingasan

emergingmarketssalesperson.

23AdrianCox|WouldtherealAIbubblepleasestandup?|December,2025

溫馨提示

  • 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
  • 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
  • 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁內(nèi)容里面會(huì)有圖紙預(yù)覽,若沒有圖紙預(yù)覽就沒有圖紙。
  • 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
  • 5. 人人文庫網(wǎng)僅提供信息存儲(chǔ)空間,僅對用戶上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理,對用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對任何下載內(nèi)容負(fù)責(zé)。
  • 6. 下載文件中如有侵權(quán)或不適當(dāng)內(nèi)容,請與我們聯(lián)系,我們立即糾正。
  • 7. 本站不保證下載資源的準(zhǔn)確性、安全性和完整性, 同時(shí)也不承擔(dān)用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。

評論

0/150

提交評論