版權說明:本文檔由用戶提供并上傳,收益歸屬內容提供方,若內容存在侵權,請進行舉報或認領
文檔簡介
AIeatstheworld
BenedictEvans
November2025
BenedictEvans––November20252
Thenextplatformshift
Mainframes
Smartphones
GenerativeAI
Every10-15years,aplatformshiftreshapestechnology
PCs
Web
BenedictEvans––November20253
Whathappensinaplatformshift?
Whoisaffected,andhowmuch?
Insidetech
Newgatekeepers,
newvaluecapture
Newandbigger
markers
Outsidetech
Isthisanewtool,newrevenue,oranexistentialthreat?
TheNewThing!
Alltechinnovation,
investmentand
companycreation
switchesover
BenedictEvans––November2025
4
Source:Apple,Google,Gartner
Dominanceiswonandlost
MicrosoftdominatedthePCera,butwhenthecentreofgravityshiftedtosmartphonesitbecameirrelevant
MicrosoftOSshareofglobalcomputerunitsales
100%
75%
50%
25%
0%
19751980198519901995200020052010201520202025
Source:Gartner
5
BenedictEvans––November2025
Earlyleadersoftendisappear
The‘first’wasnotthewinnerinPCs,browsers,search,socialorsmartphones
GlobalPCunitsalesshare
IBM-compatibleAllothers
Apple
100%
75%
50%
25%
0%
197519801985199019952000
BenedictEvans––November20256
Howwillthenewthingwork?Wedon’tknow
Foreverynewplatform,weforgethowmanyideasfailedandhowuncleareverythingwas
Internet
AOL,Yahoo,
Pointcast,Flash,plugins,portals…
Sun?Netscape?
Mobileinternet
GenerativeAI?
i-mode,J2ME,WAP,
Browsers?Agentic?
Voice?MCP?
Wearables?GEO?
mesh,DVB-H,
keyboards…
Nokia?RIM?
Noise,hype,anti-hype
Whenthingsareexciting,peoplegetexcited
BenedictEvans––November20257
Thisoftenbringsbubbles
Peopledrawstraightlinesonlogscalecharts
Theyforgetthatexponentialgrowthisgenerallyanexponentialcurve
Andalwayssay“thisisdifferent”
Thetroubleis,they’regenerallyright-everybubbleisdifferent!Butitcanstillbeabubble
BenedictEvans––November20258
Source:RosenfeldHCMST
9
BenedictEvans––November2025
Butwhenthedustsettles,theworldhaschanged
TheinternethasgonefromtheNewThingtoabasicpartofdailylife
60%
50%
40%
30%
20%
10%
0%
USheterosexualcoupleswhometonline,byyearofmeeting
199520002005201020152020
*Source:ProductivBenedictEvans––November202510
Newplatformsmeannewtools(andnewrevenue)
SaaSmeansthetypicallargeenterpriseintheUSAnowuses4-500apps
Mainframes
Oneapp
SaaS
4-500apps*
LLMs
Whatcanwe
automate
next?
PCs
Dozensof
apps
Onewaythisplatformshiftisdifferent,though
ForPCs,theweborsmartphones,weknewthephysicallimitsofwhatcouldhappennextyear
WithLLMs,wedon’tknowhowmuchbetterthiscouldget
BenedictEvans––November202511
“TheracetoAGIisafoot”SergeyBrin
“AGIneedsmultiple
furtherbreakthroughs”DemisHassabis
BenedictEvans––November202512
Anotherplatformshift,ormore?
?
Mainframes
Smartphones
GenerativeAI
Weknowthiswillgetbetter,butwedon’tknowhowmuch
PCs
Web
BenedictEvans––November202513
Sohowwillthenewthingturnout?
Ifthisis‘only’asbigasmobileortheinternet,thatseemslikeenough
Howwillthis
work?
Whereisthe
distribution,
valuecapture,
andvalue
destruction?
Howwillthisbeuseful?
BenedictEvans––November202514
Insidetech
“Theriskofunder-investingissignificantlygreaterthantheriskofover-investing”
SundarPichai,Q22024
BenedictEvans––November202515
“Theveryworstcasewouldbethatwehavejustpre-
builtforacoupleofyears”
MarkZuckerberg,Q32025
BenedictEvans––November202516
ThreeyearsofFOMOinBigTech
“Thisisahugenewmarket,ahugethreattoallourexistingbusinesses,andwecan’tmissit”
“Thisisahugeopportunityand
threat
Wecan’tmissit”
Howmuch
better?
Howmuch
capex?
“Itkeeps
gettingbetter
withmore
capex”
FOMOdrivesacapexsurge
~$400bnin2025forthebigfouralone(forcomparison,globaltelecomsis~$300bn)
Capex($bn)
CoreweaveOracle
Meta
AlphabetAWS*
Microsoft
500
400
300
200
100
0
2010201120122013201420152016201720182019202020212022202320242025e
Source:Companies,companyguidance.IncludescapitalleasesBenedictEvans––November202517
*AmazondoesnotbreakoutAWScapexdirectly
Source:Companies,companyguidance.IncludescapitalleasesBenedictEvans––November202518
*AmazondoesnotbreakoutAWScapexbutreportsitas‘themajority’
Planned2025growthalmostdoubled-in2025
~$350bnin2025forthebigfouralone(forcomparison,globaltelecomsis~$300bn)
Capex($bn)
CoreweaveOracle
Meta
AWS*
AlphabetMicrosoft
500
400
300
200
100
0
2010201120122013201420152016201720182019202020212022202320242025eJanestfor2025
Anewinvestmentcycle
USdatacentreconstructionovertakingoffices
USconstructionvalue(2025$bn,seasonallyadjustedannualrate)
Offices
Retail
Warehouses*Datacentres**
125
100
75
50
25
0
Jan1995Jan2000Jan2005Jan2010Jan2015Jan2020Jan2025
Source:USCensus.*Excludesmini-storage**ExcludescomputeBenedictEvans––November202519
NB:dataafterJuly2025delayedbyUSgovernmentshutdown
Source:Companies
20
BenedictEvans––November2025
Nvidiacan’tkeepup
TryingtobuildanewSunMicrosystems(thoughChinaandhyperscalers’ownchipsarecomingupbehind)
Quarterlyrevenue($bn)
Intel
Nvidia
60
50
40
30
20
10
0
200020052010201520202025
Source:Companies
21
BenedictEvans––November2025
Nvidiacan’tkeepup(neithercanTSMC)
TSMCunwilling/unabletoexpandcapacityfastenoughtomeetNvidia’sbook
Quarterlyrevenue($bn)
Intel
NvidiaTSMC
60
50
40
30
20
10
0
200020052010201520202025
BenedictEvans––November2025
22
Source:SchneiderElectricindustrysurvey
USpowerbacklogsbecomingamajorissue
USpowerdemandgrowthis~2%,andAImightadd1%that’shardtobuildfast(thisisnotanissueinChina)
Mainconstraintstodatacentreconstruction,USA(February2025)
SomewhatsignificantVerysignificant
Extremelysignificant
100%
75%
50%
25%
0%
LandPublic
Accesstonaturalgas
UtilitypowerAccesstochipsFibrePermittingWater
opposition
availability
BenedictEvans––November202523
“It’sbeenalmostimpossibletobuildcapacityfastenoughsinceChatGPTlaunched”
KevinScott,MicrosoftCTO
BenedictEvans––November202524
“WenowexpecttheFY26growthratetobehigherthanFY25”-Microsoft
“Capexdollargrowthwillbenotablylargerin2026”-Meta
“Weexpectasignificantincreasein2026”-Alphabet
Source:Companies
25
BenedictEvans––November2025
Datacentrecapacitytriples?For$3tr?$5tr?More?
Someverylargenumbers(althoughsomeofthe‘bragawatts’maybemoreperformativethanreal)
Globaldatacentrecapacityestimates,H12025(GW)
Current300
Est.of2030plans
200
100
0
Omdia/IEABloombergBNEFGoldmanSachsMorganStanleyBainBCGMcKinsey
Source:ITU,IEA,ONSBenedictEvans––November202526
“Threetrilliondollars!”
AnnualisedAIcapexaspirationsareasimilarmagnitudetomatureglobalcapital-intensiveindustries
Globaltelecoms~$300bn
USprivate
sector
~$20.1tr
GenAI
$500-750bn
annual?
Oil&gas‘upstream’~$540bn
Source:Companies,consensusestimates.Calendaryear.BenedictEvans––November202527
Thehyperscalerscanaffordit…
BigTechcashflowhassurgedsincethePandemic,andmostofthatgrowthisgoingonAIcapex
AnnualCapexandFreeCashFlow,2010to2025e($bn)
CashcapexFCF
200
150
100
50
0
-50
AlphabetAmazonMetaMicrosoft
Source:Companies,consensusestimates.Calendaryear.BenedictEvans––November202528
Thehyperscalerscanaffordit…uptoapoint
Capitalleasesarenotnew,butthey’vegotalotbigger
AnnualCapexandFreeCashFlow,2010to2025e($bn)
Leases
CashcapexFCF
200
150
100
50
0
-50
AlphabetAmazonMetaMicrosoft
Uptoapoint
Hyperscalersaddleasesanddebt,whilesomeanalystssuggestOracle’scloudcapexmightbe>100%ofrevenue
BenedictEvans––November202529
OpenAIjoinstheclub
Announcedcommitmentsfor30GW+ofcapacityat$1.4tr
Aspirationfor1GW/weekofnewconstructionat$20bn/GW=~$1trannually…
Source:OpenAI,October282025BenedictEvans––November202530
Equivalentto2/3oftotalcurrentglobalbase,everyyear
“Circularrevenue”
Withoutitsowncashflows,OpenAIpartnerswithNvidia,Oracle,Softbank,petrodollars…
OpenAIisbuyingNvidiachipswithNvidia’scashflow…
Whichcomesfromthehyperscalers…
andusingNvidia’scashtoturnAMDintoanNvidiacompetitor,andpayBroadcomtomakeitsownchips…
BenedictEvans––November202531
*Source:NvidiaBenedictEvans––November202532
Rationalactors?
Whatwouldyoudoifyourcompanywassittingonabubble?
Nvidiahas$72bnofTTMFCF*andTSMCcan’tkeepupwithdemand
Useyourexcesscashtobuydemand,FOMOandplatformlock-in
OpenAIhasmindshare
andexpensivestock,butnoplatform,infraormoat
Swapyourpaperforhardassetsandmarketposition
Oracleisacash-generativelegacybusinesslosingsharetocloudandnowtoAI
Gearupandburnyourwayintothenewthing?
BenedictEvans––November202533
Yes,butwherehasallthismoneygotus?
Afterthreeyears,lotsmorescienceandengineering,butnorealclarityontheshapeofthemarket
Noapparent
moats
Noclarityon
productorvalue
capture
Modelsstill
improving
Farmoremodels,
China,OSS
Lotsofnew
acronyms
Source:ArtificialAnalysis
34
BenedictEvans––November2025
Farmoremodels
Everyweek-newmodels,new(problematic,gamed,saturated)benchmarks,newacronyms
Modelsbyaggregatebenchmarkscore
OpenAI75
Allother
50
25
0
Nov2022Nov2023Nov2024Nov2025
BenedictEvans––November2025
35
Source:LMArena,ArtificialAnalysis
Modelsconvergeandleaderschangeweekly
Dozensof(saturated)benchmarkstochoosefrom,butonthemostgeneral,theleadersareveryclose
Bestscoringmodelrelativetoleader,October2025
Anthropic
OpenAI
Other:China
Other:Western
NBtruncatedaxis
100%
95%
90%
LMArenaTextMMLU-Pro
Source:ReutersInstitute
36
BenedictEvans––November2025
Techversusbrandversusdistribution?
Themodelsmaybeclosetocommodities(especiallyforgeneraluse),butmarketpositionisnot,sofar
WeeklyactiveusersofgenerativeAItoolsasshareofthepopulation,June2025
ChatGPT
GoogleGemini
MetaAI
MicrosoftCopilotGrok
Claude
30%
20%
10%
0%
USAUKFranceDenmarkJapan
Source:OpenAI.NB:roundnumbersreportedatscheduledeventsBenedictEvans––November202537
“Everyoneisalreadyusingthis!”
800mweeklyusers,butapparentlyonly5%arepaying-andwhyannounceWAUandnotDAU?
ChatGPTglobalweeklyactiveusers(m)
1,000
750
500
250
0
Dec2022Jun2023Dec2023Jun2024Dec2024Jun2025Dec2025
Stillmoreexperimentationthandailyuse
Sofar,manymorepeopleusechatbotsoccasionallythanmakethempartoftheirdailylives
MonthlyWeeklyDaily
60%
50%
40%
30%
20%
10%
0%
ConsumergenerativeAIchatbotuse:USA
202320242025
ConsumergenerativeAIchatbotuse:UK
202320242025
Source:Deloitte.SurveysconductedinJune.BenedictEvans––November202538
Definedas“Useofapurpose-builtGenAItool,likeChatGPT,MicrosoftCopilot,GoogleGemini,etc”
Source:ConsumersurveysconductedasofdateshownBenedictEvans––November202539
Mostdatashowsthesamepicture
Surveysareearly,scatteredandinconsistent,butanengagementgapseemsclear
HowmanypeopleusegenerativeAIchatbotsintheUSA?
WeeklyactiveDailyactive
50%
40%
30%
20%
10%
0%
Pew,
Aug2024
CCIA,
Nov2024
BlicketalRPS,Nov2024
Bain,
Dec2024
Hartleyetal,Dec2024
YouGov,Mar2025
Hartleyetal,Jun2025
ReutersInst.,Jun2025
Deloitte,Jun2025
Bain,
Sep2025
BenedictEvans––November202540
Isthisjustearly?Oraharderproblem?
WhydomostusersofChatGPTonlyuseitalittlebit?
Howmanyuse
casesarean
obviouseasy
fit?
Foreveryone
else,doyou
needtowrap
thisintooling
andproduct?
Whohasa
flexiblejob…
Andconsciously
looksforwaysto
optimise?
BenedictEvans––November202541
“Peopledon’tknowwhattheywantuntilyoushowittothem”
“You’vegottostartwiththeexperienceandworkbackwardstothetechnology”
SteveJobs
BenedictEvans––November202542
Sowheredoesamodellabcompete?
Whereisthevaluecaptureforaresearch-heavy,capital-intensivecommodity?
GoupthestackWinonnetworkeffects&product? (See:software)
Ifthemodelsarecommodities?
Anddon’thavenetworkeffects?
Godownthe
stack
Winonscale?(see:aircraft,AWS,chips)
BenedictEvans––November2025
43
Source:Microsoft.Includescapitalleases
Microsoft’sshiftawayfromnetworkeffects?
Fromcompetingonnetworkeffectstocompetingonaccesstocapital?
Microsoftcapex/sales
50%
40%
30%
20%
10%
0%
Mar1995Mar2000Mar2005Mar2010Mar2015Mar2020Mar2025
BenedictEvans––November202544
ForOpenAI,“yes!’toeverything
Everything,everywhere,yesterday(onotherpeople’sbalancesheets),beforethemarketslipsaway
Infradeals
withOracle,
Nvidia,Intel,
Broadcom,
AMD…
Appplatform,socialvideo,webbrowser
RobotsJonyIveBiotech
Ecommerce
integrations,
ads,vertical
datasets
BenedictEvans––November202545
“Therearetwowaystomakemoney.Youcanbundle,oryoucanunbundle”
JimBarksdale
OpenAIbundlesandunbundlesusecases
What’stherightexperience?Therightdistribution?Andwhyisn’titjusttheChatGPTapp?
BenedictEvans––November202546
Source:YCombinator
47
BenedictEvans––November2025
Startupsexisttounbundleusecases
ThecomingwaveofAIstartupstryingtounbundleGoogle,Excel,emailandOracle…andChatGPT
YCombinatorstartupsbyfield
OtherAI
500
400
300
200
100
0
W2015W2016W2017W2018W2019W2020W2021W2022W2023W2024F2024S2025
BenedictEvans––November202548
Whereisthevaluecapture?
Ifmodelsarenear-commodities,andwedon’tknowtherightproduct,wherewillthevaluebe?
Bestmodel?Mostcapital?
Distribution
>M?
Product?
UX?
Building
‘normal’
software
companies?
Proprietaryverticaldata?
BenedictEvans––November202549
Outsidetech
BenedictEvans––November202550
“What’sourAIstrategy?”Well,what’sthepattern?
Howdowealwaysdeploynewtechnologies?
Innovate
Newproducts,bundlingandunbundling
Absorb
Disrupt
Redefinethequestion
Automateobvious
use-cases
Makeitafeature
BenedictEvans––November202551
Sofar,mostsuccessfuluse-casesare‘absorb’
WhereisiteasyandobvioustousegenerativeAI?
Coding
Customersupport
Automation
Marketing
AIcodingasthenewAWS
“Vibecoding”asthenewabstractionlayer,afterAWS,libraries,operatingsystems…
Anewstepchangereductioninsoftwarecreationcosts
BenedictEvans––November202552
Source:Accenture
53
BenedictEvans––November2025
Howdoyouknowwhattoautomate?
Stepone:askyoursystemsintegrator
Accenturereportednewquarterly‘generativeAI’contracts($m)
2,000
1,500
1,000
500
0
Feb2023Aug2023Feb2024Aug2024Feb2025Aug2025
Source:Palantir
54
BenedictEvans––November2025
Howdoyouknowwhattoautomate?
Steptwo:buysomeSaaSfromDrEvil
Palantirquarterlyrevenuebysegment($m)
Enterprise
Government
1,250
1,000
750
500
250
0
Mar2020Mar2021Mar2022Mar2023Mar2024Mar2025
Source:McKinsey
55
BenedictEvans––November2025
Pilotscomefirstanddeploymenttakestime
“Agentic!’is2025’sbuzzword,butdeploymenttakeslonger
AI‘agent’usebybusinessfunction,whereGenerativeAIisalreadyused.June2025
ExperimentPilot
Deploying
Deployed
30%
20%
10%
0%
ITKnowledgeMarketingServiceProductSoftwareHRSupplyManufacturing
managementdevelopmentchain
BenedictEvans––November202556
Noteverythingworks?Welcometotech
“WhydidourAIpilotfail?”That’saCTOquestion,notanAIquestion
Security,
privacy,IPR,
errorrates,
legal
Findingthe
rightsolution
fortheright
people
Thesame
issuesas
deployingany
newtech
Data
integration&legacysystems
BenedictEvans––November2025
57
Source:MorganStanleyCIOSurvey
Thefuturecantaketime
AquarterofCIOshavelaunchedsomething-but40%don’tplananythinguntilatleast2026
CIOexpectedtimingforfirstLLMprojectsinproduction,September2025
Alreadydeployedatleastoneproject
H22025
2026
Laterornoplan
0%25%50%75%100%
BenedictEvans––November2025
58
Source:GoldmanSachsCIOSurvey
Butthefuturealwaystakestime
Cloudisoldandboring-butstillonly30%ofworkflows
Enterpriseworkloadsinpubliccloud
Today
Expectedin3years
60%
50%
40%
30%
20%
10%
0%
2013201420152016201720182019202020212022202320242025
Source:FMIBenedictEvans––November202559
Sometimes‘automation’aloneisabigdeal
UPCs,barcodesanddatabasesletretailersmanage5xmoreSKUs
AverageSKUspersupermarket,USA
60,000
Grocerybarcodeslaunchedin1974
50,000
40,000
30,000
20,000
10,000
0
19501960197019801990200020102020
BenedictEvans––November202560
But…
“We’veallseenlotsofAIpresentations now,andwe’vedeployedabunchofstuff.Isthatit?What’snext?”
F100RetailerCMO,summer2025
BenedictEvans––November202561
Whatnext?
Whatcomesafterautomatingtheobvious,easythings?
Absorb
Innovate?
Newproducts,bundlingandunbundling
Disrupt?
Redefinethe
question
Automateobvious
use-cases
(20yearsmore
deploymentto
come!)
BenedictEvans––November202562
Wheremightwelookforchange?
WhatcanLLMautomationunbundle?Whatthingsdidwenotrealisewerebundles?
Internet
unbundlingcreated
newaggregators
HowdoLLMsdo
thatbetter?
Onlinedistribution
unbundledphysical
assets
WhatdoLLMsunbundle?
BenedictEvans––November202563
“AIgivesyouinfiniteinterns”
BenedictEvans––November202564
Howdoweuseautomationthatmakes‘mistakes’?
Wehavenoindicationthaterrorrateswillgoaway,sowhere’sthehumanintheloop?
Do‘errors’matter?
Ishumanverificationefficient?
Howmuch
doweneed
towrapthe
LLMin
software?
Canyou
automate
verification?
BenedictEvans––November202565
Whatdo‘automatedinterns’change?
TheJevonsparadox-appliedpriceelasticity
Doyoudothe
sameworkwith
fewerpeople?
Ormorework
withthesame
people?
Whatbecomes
possiblewhen
youdon’tneed
millionsofpeople
todothat?
Wasemploying
lotsofpeople
yourmoat?
Source:Landes
66
BenedictEvans––November2025
300minterns?Jevonsparadoxatwork
SteamenginesgaveBritaintheequivalentlabourof(veryroughly)5xitstotalpopulationby1900
UKpopulationandsteamenginelabourunitequivalent(m)
Population
Steamengineequivalent
300
250
200
150
100
50
0
184018701896
BenedictEvans––November202567
Where’sthehumanintheloop?
Allrecommendationsystemstodayworkbydriving,capturingandanalysinguseractivity
GenAI?
Pre-internet
Humaneditors,forbothretail&media
Physicalassetsas
moat
Internet
CananLLMdo
Allofusare
thisbetter?
mechanicalturks
CananLLMdoit
feedingalgorithms
withoutneeding
Networkeffectsas
auserbase?
moat
Valuetocapture!
Brandsspendatrilliondollarsayeartotalktoconsumers-plusrent,shipping,marketing,returns…
Globaladrevenue,2024($bn)
GogoleNetwork
YouTube
Amazon
Allotherex-China
China
Meta
GoogleSearch
01002003004005006007008009001,0001,100
Source:Companies,WPPMediaBenedictEvans––November202568
NB:‘China’representsspendinginChina,notglobalspendingbyChinesecompanies
“OurnewAIrecommendationmodeldrove5%moread
conversionsonInstagramand3%onFacebook”
Meta,Q22025
BenedictEvans––November202569
“AdvertisersthatactivateAIMaxinSearchcampaigns
typicallysee14%
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,請聯系上傳者。文件的所有權益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網頁內容里面會有圖紙預覽,若沒有圖紙預覽就沒有圖紙。
- 4. 未經權益所有人同意不得將文件中的內容挪作商業(yè)或盈利用途。
- 5. 人人文庫網僅提供信息存儲空間,僅對用戶上傳內容的表現方式做保護處理,對用戶上傳分享的文檔內容本身不做任何修改或編輯,并不能對任何下載內容負責。
- 6. 下載文件中如有侵權或不適當內容,請與我們聯系,我們立即糾正。
- 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔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。
最新文檔
- 2025年東北林業(yè)大學輔導員招聘考試真題匯編附答案
- 2025年天津輕工職業(yè)技術學院輔導員考試參考題庫附答案
- 2025年衢州職業(yè)技術學院輔導員招聘考試真題匯編附答案
- 2025廣西桂林市灌陽縣公安局警務輔助人員招聘6人備考題庫及答案1套
- 2025福建泉州德化縣消防救援大隊招聘政府專職消防員4人備考題庫新版
- 證券市場分析報告編制規(guī)范
- 2025年旅游產品開發(fā)與市場推廣指南
- 電商平臺客服部工作規(guī)程規(guī)章制度守則
- 《EJT 999-1996核燃料后處理廠自控儀表工程設計規(guī)定》專題研究報告深度
- 2024年炎黃職業(yè)技術學院輔導員考試參考題庫附答案
- 《底層邏輯》劉潤
- 甲狀腺手術甲狀旁腺保護
- 幼兒園《企鵝遇險記》原繪本故事
- 多波多分量地震勘探規(guī)范
- (高清版)TDT 1057-2020 國土調查數據庫標準
- 曼娜回憶錄的小說全文
- 管道工培訓課件
- 2024版未來食品加工技術趨勢:智能化與自動化培訓課件
- 無人機測繪操控員培訓計劃及大綱
- 父親給孩子的一封信高中生(五篇)
- 動角問題專項訓練(30道)
評論
0/150
提交評論