CARLYLEOctober2023

GlobalInsights

BRAVENEWWORLD

AIanditsDownstreamImplications

2

CARLYLE

EXECUTIVESUMMARY

→TheadventofArtificialIntelligencemayrepresentawatershedinhumanhistory,withthe

potentialtotransformdailylivestoanextentthatmaybedifficulttoappreciatefullyatthis

momentintime.ButasunprecedentedasthetechnologicalshockfromGenerativeAImayprovetobe,thecapitalmarketresponsetoitalreadyfollowsfamiliarpatterns.

→Ratherthansimplyseparaterealityfromhype,successfulinvestorsmustbeabletomapthat

realityontocompanyfundamentals.Thisrewardssecond-and-thirdorderthinking,asthemostsalientfeatureofthetechnologicalrevolution–escalatingrevenuegrowthatcompaniesattheepicenterofthetechnologicalquake–mayultimatelyprovetobeasmallfractionofthetotaleconomicvalueitdelivers.

Aswiththeadventofelectrification–aturningpointtowhichthedevelopmentofAIsystems

hasbeencompared–themainriskforinvestorstodaymaybeviewingtheAIrevolutiontoo

narrowly.Theproductivitygainsfrominvestmentinsoftwaredevelopmentandlifesciences,

contentgeneration,andCRMsystemsalreadysuggestthattheassetsbestpositionedtobenefitfromAImayhavenotyetlandedonthebroadermarket’sradar.

CARLYLE

3

Itisdifficulttooverstatethetransformationpotentialof

ArtificialIntelligence(AI).Wemaysoonliveinaworldwherecomputersystemscangeneratenewscientificknowledgeandperformvirtuallyanyhumantask.Asunprecedentedasthetechnologicalshockmayprovetobe,thecapitalmarketresponsetoitalreadyfollowsfamiliarpatterns.

Whenafoundationaltechnologyentersthepublics

consciousness,investorsnaturallyfocusonthetechnology

itselfandcompaniesthoughttobeoperatingatitsfrontier.GenerativeAIhasbeennoexception.Assetpricesquickly

reachlevelsdifficulttorationalizeusingconventional

financialmetrics;“value”comestobeassociatedwith

subjectiveimpressionsofthetechnologyspotential,barrierstoentry,andultimatescalability.

Debatesregardingthevaluationofnascenttechnologyoftendegradeontwoaxes.Enthusiasts,typicallyfromthetechsectoritself,recastinvestorskepticismasignorance;

anunwillingnesstodeployaggressivelyintothespace

revealsalackoftechnicalunderstanding.Detractors,

fortheirpart,oftendismissnovelvaluationmethodsand

optimistic“totaladdressablemarket”forecasts(Figure1)

astell-talesignsofahypecampaigndesignedtoseparatecredulousinvestorsfromtheircapital.Portfolioscanbe

deridedasuninformedorna?ve,dependingonperspective.

Suchdiscussionselideacrucialpoint.Whiledismissing

AIstransformationalpotentialcouldprovetobeavery

expensivemistake,returnsultimatelydependonhownew

technologygetsadoptedandmonetized.Andthisprocess

canoccuroverlonghorizonsandmanifestonincome

statementssomedistanceawayfromtheinitialshock.As

withtheadventofelectrificationaturningpointtowhichthedevelopmentofAIsystemshasbeencomparedthe

mainriskforinvestorstodaymaybeviewingtheAIrevolutiontoonarrowlyandfailingtoperceiveallofthedownstream

opportunities(andrisks)itcreates.

Figure1.

AIMarketSizeExpectations($Billions)

$1,600$1,500$1,400$1,300$1,200$1,100$1,000$900

$800

$700

$600

$500

$400

$300

$200

$100

$0

20l820l92020202l202220232024202520262027202820292030

GlobeNewswire(June2022)GrandViewResearch(June2020)IDC(February2021)Tractica(March2020)

Figure1.Source:IDC,Tractica,GrandViewResearch,Statista,GlobeNewswire,JefferiesEquityResearch.

CARLYLE

4

GROUNDBREAKINGCAPABILITIES

softwarescapacitytoidentifypatternsindataandanticipate

&ADOPTIONRATES

sequencesfasterandmorepreciselythanhumans.GenerativeAIrepresentsthenextstepinthisevolution,withsoftware

Investorinterestin“artificialintelligence”hasspikedover

nowabletosynthesizedataandcurateresponsesbeyond

thepastyearthankstothereleaseofGenerativeAItools

thosedirectlyintendedbytheprogrammer(Figure3,p.5).

capableofproducingcontentandanalysesofunprecedented

Andthereisstillampleopportunitytoreinventthelanguage

sophisticationandbreadthinresponsetonaturallanguage

toolsthathelpengineersdevelopnewgenerationsof

prompts.MostnotablehasbeenOpenAIsreleaseofChatGPT,whichreached100millionusersinjusttwomonths,asmall

softwareevenmoreefficiently.1

fractionofthetimeittookFacebookandothersocialmedia

OnenotablesubsetofGenerativeAIislargelanguage

platformstoachievesimilarscale(Figure2).Thesemodels

models(LLMs).Impressiveasthisclassofdeep-learning

canreasonprobabilistically,havebeentrainedonvirtually

algorithmis,itrepresentsbutonesteponalongerroad

theentireinternetcorpus,andcanbedirectedtoprocess

to“ArtificialGeneralIntelligence”autonomouscomputer

thatinformationthroughconventionaltextthatonemight

systemsthatcanlearntoperformvirtuallyanytaskof

otherwiseputintoanemail(notarcanecode).

scientificoreconomicvalue.WhilemanyAIresearcherswouldarguethatwereonthecuspofthisworld-historicalturning

GenerativeAIalreadyrepresentsanhistorictechnological

point,otherscontendthatAGImaybedecadesawayifits

leap,atleastasmeaningfulasinternet-basedsearchengines

everachievedatall.Muchofthedisagreementcenterson

displacementofreferencelibraries.Butwhereasthat

arcaneCartesianquestionsofself-awarenessandmysteries

revolutionliberatedinformationfromthephysicalconstraints

surroundingthebiochemistryofhumanconsciousnessand

oftheanalogworld,AIliberatesinformationflowsfromhuman

cognition.2Themorepracticalandeconomicallyrelevantthe

intermediation.MachineLearningalgorithmsdemonstrated

definition,theclosertoAGIwemaybe.

Figure2.

Timeto100MillionUsers

THETIMEITTOOKFORSELECTEDONLINESERVICESTOREACHl00MILLIONUSERS

2

Months

9

2.5

Months

3.5

Years

4

4.5

Years

Years

Years

5

Years

8

l0

Years

ll

Years

Years

2008l9992008200620042008200920l020l62022

Figure2.Source:VisualCapitalist,February2023.

1."DeveloperTools2.0,”SequoiaCapital,March2023.

2.C.f.Landgrebe,J.andB.Smith.WhyMachinesWillNeverRuletheWorld.Routledge,2022.

CARLYLE

5

Figure3.

NextStepinEvolutionofSoftware

SOFTWARE1.0

SOFTWARE2.0

SOFTWARE3.0

Statistically-based

MachineProgram

Optimizer/Compiler

Programmer-readablecode

Natural-language

likeinstructions

MachineProgram

Statistically-based

MachineProgram

AIagent

Data

Programmer-readablecode

NNarchitecture

MachineProgram

Interpreter/Compiler

Programmer-readablecode

Interpreter/Compiler

MachineProgram

Programmer-readablecode

Data+NNarchitecture

IMMEDIATEAPPLICATIONS

preciselytailored,bothintermsofthecontentofadvertisingcampaignsandthetargetingofaudiencesmostlikely

Allmajortechnologyandsoftwarevendorsarecurrently

toactonthem.AIwillrevolutionizecustomerrelations

embeddingGenerativeAIintotheirstack.Desktop

management(CRM)acrossindustries,generatingupselling

applications(email,word-processing,etc.),e-commerce,

proposalsinrealtimebasedontextfromtheconversation

internetsearch,socialmedia,andcontentconsumption

cross-referencedwithinternalcustomerdata,external

willallintegrateAIfunctionality.Sucheffortsremainina

markettrends,andotherrelevantinformation.Chatbotsmay

betastagewithlimitedvisibilityintomonetization.Butthe

soonaccountforthebulkofconsumer-facinginteractions

userexperienceislikelytoimproveimmeasurablyacross

intravel,finance,ande-commerceandeventuallyguide

eachofthesedimensions,withsignificantscopeforlaborproductivitygainsfromacceleratedinformationgathering

customers’entireshoppingexperience.

andideaandtextgeneration(Figure4,p.6).

Theapplicationsformediaandeducationareobvious.

GenerativeAIapplicationscanproducenewmusic,fictional

Moreconsequentialmaybetheevolutionofbusiness

narratives,poetry,visualartwork,anddigitalimagery.The

modelsandcorporatestrategy.Managementteamscould

recentScreenActorsGuild(SAG)andWritersGuildofAmerica

increasinglyrelyonAItoformulatemarketingstrategies

(WGA)strikeshavebeenfomented,inpart,byconcernsabout

andpricingdecisionsanddiligencepotentialacquisition

AI’sdisplacementpotential.AI-generatedcontentraisesnovel

targets.Digitalmarketingislikelytobecomeevenmore

copyrightissuessinceexistingworksareaccessedtoproduce

Figure3.Source:ItamarFriedman,Software3.0—theeraofintelligentsoftwaredevelopment,May2022.

6

CARLYLE

Figure4.

AIUseCases

CONTENTSOFTWAREIMAGE

NEWPRODUCTSALES&Q&A

GENERATIONDEVELOPMENTGENERATION

DEVELOPMENTMARKETINGINTERFACES

?Content

?SEO

?Primaryresearch

?Synthesis

?Alertgeneration

?Supportticketingsystems

?Languagetranslation

?Createwebsitedrafts

?Automaticcodegeneration

?CoPilots

?Regenerativecode

?Testscriptgeneration

?Bugfixes

?Customgeneratedphotos

?Imagetouchup

?Bannercreation

?Medicalimaging

?Productdetailpage

imagegeneration

?“Tryiton”AR

?Interactivedataproducts

?Conversational

interface&querying

?Whitelabeled1st

partytrainedmodels

?UXdesign

?Translationfromdesigntocode

?Contentcreation

?Leadgeneration

?Salesforecasting

?Personalizedads

?Orgspecificsalescollateral

?Customersupport

?Conversionrateoptimization

?A/Btesting

?R&Dideageneration

?Identityverification

?Ordertaking

?AdvancedChatbots

?Disasterplanning&recovery

?Strategy

development

?Competitorresearch

“substantiallysimilar”outputs.3Technologically,thehorseis

Whileguidancefromexperiencedengineersisfundamental

outofthebarn;thequestioniswhetherownersofexisting

toenableLLMstowritecode,LLMscreatesignificant

copyrightswillbetheonlyoneslegallysanctionedtoemploy

efficienciesbyfillingincodinggapsinsimplifiedprompts.

AItoassistintheformulation,production,andmarketingof

Eventualgainsfromsuchautomationmaybeespecially

cinematic,televisual,andaudioworks.

pronouncedamongvideogamemakersoperatingattheintersectionofAI-generatedcontentandsoftware.

ChatGPTeasilypassedtheUniformBarExaminationtaken

byU.S.lawschoolgraduatesandwouldearnarespectable

CompanieswillincreasinglyrelyonGenerativeAItocleanexisting

3.4gradepointaverage(ona4-pointscale)ifenrolledas

dataandproduceprototypedesignsandaccelerateproduct

afreshmanatHarvardCollege.4GenerativeAIsprowess

development.Lifesciencescompanies,forinstance,alreadyuse

writingessaysandtakingtestsraisethornyissuesaboutthe

AItogeneratesequencesofaminoacidsandDNAnucleotides

futureofeducationalintegrity,butalsoopenthedoorto

toshortenthedrugdesignphasefrommonthstoweeks.Existing

anewgenerationofdigitaltutors,autodidacts,andmore

developmentprogramsrequireresearcherstosortthrough

flexibleeducationalarrangements.

millionsofpotentialchemicalreactionstosynthesizeatargetmolecule.AImodelstrainedonexistingchemicalreactions

Hugeproductivitygainsarealreadyevidentinsoftware

datahavealreadyyieldeda15%reductionindevelopment

development,whereGenerativeAIhashalvedthetime

costs.6Weshouldexpecttoseecomparableproductivitygains

necessarytowriteandtestnewcode(Figure5,p.7).LLMs

whereverR&Ddependsontime-consuming,iterativeprocesses

canpredictthenextlinesofcodebasedonthecode

basedoncomplexinteractionsbetweenvariablesorinputs.

alreadywrittenandgeneratenewcodeinresponseto

tailoredpromptsfromsoftwareengineerswhoareskilledin

ManufacturerscannotonlyuseGenerativeAItodesignnew

naturallanguagedescribingsoftwarestructures.AsLLMs

products,butalsooptimizesupplychainsandautomate

becomefamiliarwiththefunctionalityandstructureof

shippingandproductionprocesses.Theautomotiveindustry

programminglanguages,promptscanbecomelessprecise,

hasbeenespeciallyaggressiveinitsadoptionofAIand

allowingneophytestocodelikeseasonedprofessionals.5

antecedentalgorithmictechnologiestotheseends.

Figure4.Source:CarlyleAnalysis,2023.

3.ABAJournal,March2023.“ChatGPTgoestoHarvard,”Substack,July2023.

4.“BeyondTheHype:HowGenerativeAIIsTransformingSoftwareDevelopment,”TowardsDataScience,May2023.

5.G2Retroasatwo-stepgraphgenerativemodelsforretrosynthesisprediction,CommunicationsChemistry,May2023.

6.G2Retroasatwo-stepgraphgenerativemodelsforretrosynthesisprediction,CommunicationsChemistry,May2023.

CARLYLE

7

Figure5.

AcceleratedSoftwareDevelopment

20-30

35-45

45-50

TASKCOMPLETlONTlMEUSlNGGENERATlVEAl,%

<l0

100

80

60

40

20

0

CodedocumentationCodegenerationCoderefactoringHigh-compIexitytasks

WithoutgenerativeAIWithgenerativeAI

RISKS&JOBLOSS

onthespeedwithwhichcompaniesadoptAIcapabilitiestocutcostsandincreasescalability.Competitivepressurethis

TheJanusfaceofnewtechnologyisobsolescence.Itisestimated

greatnaturallyopensthedoortocharlatanism.Companies

thatGenerativeAIapplicationscouldeventuallyautomate

willmarketthemselvesopportunisticallyand,occasionally,

60%to70%ofemployeeworkloads,7andthisnaturallyarouses

deceptively.Mentionsof“AI”oncorporateearningscalls

fearofjobloss.Itisimportanttonotethatthisestimaterefers

hasrisenexponentially(Figure8,p.9),andthemore“AI”

toemployeetasksnottheemployeesthemselves.Formost

isinvokedbycompetitors,themoresusceptiblelaggard

occupations,wesubscribetotheviewthatAIwon’ttakeyour

managementteamsbecometoimprudentbudgetingand

job;someoneusingAIwill.Thiswillresultindynamicadjustmentsinlabordemandacrossoccupationsandactivitiesratherthan

fairy-talesolutions.

jobloss(Figure6,p.8).Workloadautomationshouldincrease

Wemustalsobemindfulofthe“hallucinationproblem”with

throughputvolumes,naturallyincreasingproductivitylevels

LLMs,ortheirtendencytogeneratefactuallyincorrecttext

(outputperhourofwork);andbyfreeingmanagers’finitetime

thatmayseemsemanticallyorsyntacticallyplausiblebased

andattentionandspeedingmorejunioremployees’progression

onthecorpusofdataonwhichithasbeentrained.These

upthelearningcurve,AIalsocouldfacilitateasustainedincrease

statisticalmodelspredictthenextwordbasedonmassive

inproductivitygrowthratesashumancapitalgetsdeployed

volumesofdataandpastcontext.Theyarebuiltforfluency

morecreatively(Figure7,p.9).

ratherthanreason,whichmeanshumanverificationoftheiroutputswillstillberequiredinmanycases,andtheiruse

Obsolescencemaybeofgreaterconcernforbusinesses

inmissioncriticalapplicationslikeaeronauticsordefense

andbusinessmodels,ascompetitionincreasinglydepends

couldlayveryfarinthefuture.

Figure5.Source:McKinsey,2023.

7.“EconomicPotentialofGenerativeAI,”McKinsey&Co.June2023.

CARLYLE

8

Figure6.

DynamicAdjustmentinLaborDemand

Midpointautomationadoptionby2030,%

EmpIoyment,absoIute

EstimatedIabordemandchangeandgenerativeAlautomationacceIerationbyoccupation,US,2022-30

Change

inIabor

demand,%

HeaIth

professionaIs

STEM

professionaIs

HeaIthaides,

technicians,

andweIIness

lncreasingIabordemandandmodestchangeof

workactivities

BuiIdersManage

Transportationservices

MechanicaI

instaIIation

andrepair

rs

nity

Creativesand

artsmanagement

Businessand

IegaIprofessionaIs

Educationandworkforce

training

5

Food

services

l520

Customerservice

andsaIes

O代ce

support

35

30

25

20

l5

l0

5

0

-5

-l0

-l5

-20

l5-

25

25-

35

5Ml0M

40

35-

lncreasingIabordemandandhighchangeof

workactivities

Property

maintenance

Commu

services

AgricuIture

●

l0

Productionwork

DecreasingIabordemandandmodestchangeof

workactivities

lncreaseinautomationadoptiondrivenbygenerativeAlacceIeration,percentagepoints

Figure6.Source:

/mgi/our-research/generative-ai-and-the-future-of-work-in-america

.

CARLYLE

9

Figure7.

Economy-WidePositiveProductivityShock

ProductivityRelativeto2022Baseline

2023

2024

2025

2026

2027

2028

2029

2030

203l

2032

2033

2034

2035

2036

2037

2038

2039

2040

204l

2042

BaselineOne-TimeGenerativeAIShock——PersistentAcceleration

215%

195%

175%

155%

135%

115%

95%

75%

Figure8.

MentionsofAIonCompanyEarningsCalls

90

80

65

56

5352

42

40

30

2l

20

10

0

0

NvidiaAlphabetMetaMicrosoftSalesforceAMDAmazon

Q12022Q12023

70

60

50

83

38

l2

l2

0

8

7

Figure7.Source:CarlyleAnalysis,BrookingsInstitution,2023.

Figure8.Note:Includesmentionsof“AI”inanalyst/journalistquestions.Source:Companydata,Statista,GoldmanSachsGlobalInvestmentResearch.

CARLYLE

BARRIERSTOENTRY

Atthisstage,mostofthemarketdiscoursehasfocusedon

thosecompaniesdirectlyresponsibleforthedevelopmentofLLMs.And,giventheenormouscostsinvolved,thishasbeenandislikelytocontinuetobedominatedbymassive,cash-

richincumbents.Developingastate-of-the-artGenerativeAImodelrequiresmassivecomputationalresources,

specializedhardwarelikeGraphicsProcessingUnits(GPUs)andTensorProcessingUnits(TPUs),andvastdatasetsthatmustbecollected,stored,andcurated.Asingletraining

runforamodelcomparabletoChatGPTrequiresmillionsofdollars.8Ratherthancompetewithbetterfundedandmoresophisticatedincumbents,enterprisesseekingtointegrateAIintotheirproductsandservicesaremorelikelytopartnerwiththem.Thishasledtoaboominthemarketvaluesof

industry-leadinghardware,software,anddatacloud

platforms(Figure9)includinga$700billionincreasein

NvidiasmarketcapitalizationsinceChatGPTsreleaseandcreatessignificantheadwindsfornewentrantsandsmall

companiesacrossmuchofthevaluechain.

Thishasnotstoppedcapitalfromflowingtonewerand

youngercompanies,however.Overthepastyear,virtuallyanyassetwithknown“AIupside”hasbecomeveryrichly

valued,especiallyonarelativebasis(Figure10,p.11).Whileallindustrieshavebeenaffectedbythedeclineinventureandgrowthcapitaloverthepastyear,AIcompanieshavecapturedalargershareofthatfunding,especiallythose

focusedonnovelapproachestoAGI.IntheU.S.,AIsshareoffundingroundsreached23%inQ2-2023,morethantriplingoverthepast10yearsandnowthehighestamongall

industryverticals(Figure11,p11).Intermsofinvestedcapital,AIssharehasincreasedevenmoreoverthepastyear

thanks,inlargepart,toMicrosofts$10billioninvestmentinOpenAIandStripes$6.3billionSeriesIround.9

Figure9.

MegaCapAICompanies’ShareofTotalReturns

ShareofS&P500Return

AppIe

Microsoft

AIphabet

Amazon

TesIa

Meta

NVlDlA

S&P500

ContributiontoS&P500Returnin

PercentagePoints

AppIe

Microsoft

AIphabet

Amazon

TesIa

Meta

NVlDlA

S&P500

BREAKDOWNOFRETURNBYCOMPANY

20%

18%

16%

14%

12%

10%

8%

6%

4%

2%

0%

5.6%

2.2%

l.3%

l.3%

l.4%

l.l%2.3%

Top7Stocks,

PPContribution

l2.6%

2.9%

SHAREOFRETURNBYCOMPANY

100%

30.8%

l2.l%7.2%

7.2%7.9%

6.3%l2.7%

l5.7%

90%

80%

70%

60%

50%

40%

30%

20%

10%

0%

Top7Stocks,TotalShare

69.2%

Figure9.Source:CarlyleAnalysisofBloombergData,July21,2023.

8.“CanYouBuildLargeLanguageModelsLikeChatGPTAtHalfCost?”UniteAI,May2023.

9.GlobalPrivateMarketsQuarterlyQ2-2023,CarlyleGlobalInvestmentSolutions,July2023.

l0

CARLYLE

11

Figure10.

RiseinAIAttention&Valuations

AI-RELATEDSTOCKSDROVEVIRTUALLYALLTHE

SGAINEWSFLOWINDICATORCONTINUETOSURGE

RETURNSOFTHES&P500THISYEAR

4200

4200

4100

4000

3900

3800

3700

3600

4500

4000

3500

3000

2500

2000

1500

1000

500

0

4500

4000

4100

3500

4000

3000

2500

3900

2000

3800

1500

1000

3700

500

3600

0

Jan/l5

Aug/l5

Mar/l6

Oct/l6

May/l7

Dec/l7

Jul/l8

Feb/l9

Sept/l9

Apr/20

Nov/20

Jun/2l

Jan/22

Aug/22

Mar/23

Jan/23

Feb/23

Mar/23

Apr/23

May/23

S&P500

S&Pex-AlBoomstocks

Figure11.

AI’sIncreasingShareofVCFunding

6,000

NumberofVC&GrowthCapital

5,000

FundingRounds

4,000

3,000

2,000

1,000

0

30%

AIFundingRoundsin%ofTotalRounds

25%

20%

15%

10%

5%

20l4Hl

20l4H2

20l5HI

20l5H2

20l6Hl

20l6H2

20l7Hl

20l7H2

20l8Hl

20l8H2

20l9Hl

20l9H2

2020Hl

2020H2

202lHl

202lH2

2022Hl

2022H2

20l3Hl

0%

Non-AIAIAIShare

Figure10."SGAINewsflowIndicatorContinuetoSurge"Source:Factiva,SGCrossAssetResearch/EquityStrategy.Dataasof08/05/2023.

"AI-RelatedStocksDroveVirtuallyAlltheReturnsoftheS&P500ThisYear"Source:Datastream,SGCrossAsset/Research/EquityStrategy.Dataasof11/05/2023.

Figure11.Source:CarlyleGlobalInvestmentSolutions,GlobalPrivateMarketsQuarterly,Q3-2023.

CARLYLE

LESSONSFROMELECTRIFICATION

Onewondersifbyfocusingnarrowlyontheassetsclosesttotheepicenterofthistechnologicalquake,investors

mayberepeatingthemistakesofthepast.GenerativeAIhasbeenanalogizedtotheadventofelectricity,andthiscomparisonmaybeaptforreasonsthatextend

wellbeyonditstechnologicalsignificance.Though

discoveredinthe1880s,electriccurrentonlybeganto

transformsocietyinthe1920swhenmasselectrification

wasmadepossiblebyhigh-pressuresteampowerplantsandcentralizedgeneration,distribution,andsystem

management.Injustafewyears,electriccompanies

revenuesgrewbymorethan3.4x(~35%CAGR)during

aperiodofconsumerpricedeflation.Thevaluations

assignedtothosefundamentalsdoubledduringthistime(Figure12,p.13),asinvestorsaggressivelybidupthemarketvaluesofcompaniesoperatingatthefrontierofthis

technologicalrevolution.

Asitturnedout,farmoreeconomicvaluewasbeingcreatedbythecompaniesbuyingthatpower.Electrificationallowedmanufacturerstousealargenumberofcomplexmachinessimultaneously,whichmademassproductionprocesses

possibleandsharplyreducedthecostofproducing

consumerdurableslikerefrigerators,washingmachines,andradios(Figure13,p.13).Andsincetheseproductshadtobepluggedintooperate,masselectrificationnotonlydrovedownmanufacturersproductioncosts,butalso

stimulateddemandfortheirproducts.

Inthetenyearsfromthestartofthesustainedboomin

electricitygeneration,durablegoodsmanufacturers

generateda200%totalreturn,onaverage,inthedepthsoftheGreatDepression(!),whichwasmorethan2xthe

averagetotalreturntoelectriccompaniesoverthesameperiod(Figure14,p.14).Nosanepersoncouldcontend

thatmasselectrificationwasmere“hype,”aseventual

marketdemandforelectricitymetorexceededthemostoptimisticforecasts.Butthedisplacementofkerosene-firedilluminationwasbutthetipoftheiceberg,asthe

vastmajorityoftheeconomicvalueaccruedtothe

downstreamusersofthenewtechnologyratherthanthecompaniesresponsibleforitsintroduction.

Thesamedynamicsarelikelyatplaytodaywith

GenerativeAI.Specializedsemiconductorsalesmay

indeedgothroughtheroof,justasdemandforthemostadvancedboilersroseexponentiallyduringtheperiod

ofmasselectrification.Astep-functionincreaseinthe

volumeofdatagenerated,stored,andanalyzedby

companieswillalmostsurelybenefitcloudplatformsjustasacomparablejumpintheregionaltransmissionof

electriccurrentbenefitedelectricutilities.Futuregrowthintheutilitysectorwillrequiresignificantinvestmentin

GenerativeAItosupportpowergriddevelopment.Andcompaniesattheforefrontofthedesignofadvanced

AIsystemstodaywilllikelybeasinfluentialtoeconomic

developmentasthoseresponsiblefordevelopingthe

latestiterationofhigh-pressuresteamturbinesthen.

Butthebulkoftheeconomicvaluemay,onceagain,be

createdbythecompaniesmostadeptatcapitalizingonthesetrendsbyslashingproductioncostsanddevelopingthenewproductsandservicesmadepossiblebythese

newtechnologies.Thisislikelytobeespeciallytrueinsoftware,pharmaceuticals,andothersectorswhereGenerativeAIcanreducetheenormoussumsspentdevelopingintangibleassetsthatcanbeinfinitely

reproducedatnearlyzeromarginalcost.

"Butthebulkoftheeconomic

valuemay,onceagain,becreatedbythecompaniesmostadeptatcapitalizingonthesetrendsby

slashingproductioncostsand

developingthenewproductsandservicesmadepossiblebythesenewtechnologies."

l2

CARLYLE

13

Figure12.

RiseinValuationRatios,1925-29

Retail

Oil&Gas

DurableGoods

Railroad/Other

ConsumerProd

Telecom

Industrials

Tech

HealthCare

Utilities

150.0%

100.0%

50.0%

30%

0.0%

-50.0%

95%

Figure13.

Two-YearDeclineinProductionCostsbyItem

PeakTwoYearPriceDecline,1926-1936

0%

-10%

-20%

-30%

-40%

-50%

-60%

-70%

-80%

CofeeMaker

Electric

BlanketRadioFanCookerWasherToasterRefrigeratorFlatironRange

-17%

-28%

-34%

-41%

-44%

-50%

-52%

-54%

-58%

-69%

Figure12.Source:CarlyleAnalysis;CRSPDatabase,December2021.

Figure13.Source:RonaldC.Tobey,1997,“TechnologyasFreedom:TheNewDealandtheElectricalModernizationoftheAmericanHome.”

CARLYLE

Figure14.

TotalStockMarketReturnsbySector

4.5x

4.0x

CumulativeMOIC

3.5x

3.0x

2.5x

2.0x

1.5x

1.0x

0.5x

0.0x

3.02x

1.96x

1.37x

Jul-26

Oct-26

Jan-27

Apr-27

Jul-27

Oct-27

Jan-28

Ap