華泰證券：DeepSeek全球AI影響應用（47頁）.pdf

《華泰證券：DeepSeek全球AI影響應用（47頁）.pdf》由會員分享，可在線閱讀，更多相關《華泰證券：DeepSeek全球AI影響應用（47頁）.pdf（47頁珍藏版）》請在三個皮匠報告上搜索。

1、2025!1DeepSeek!#!#AI$%&()$%&()#$%#$%SAC NO.S0570519080006|SFC NO.BQZ938&(&(SAC NO.S05701220801381!#$!#$)*+,-.)*+,-./0)*+,-.123456789:;?ABCD1)*+,-.EFGH AI IJ2%KLCD1MNOPQRS3TUV6WX/89YZV6WX89YZV6WX/0 AI 8989YZV6WX1?H_:S3;ab2%VcCD/defghijkl*mnQodefghijkl*mnQo11VpqIrstVpqIrst/uvefwuvefwxxyzwyzwAI ARR|11C

2、apEx2025-01-30ghefghef1DeepSeek!#$%&!#$%&3!#$%&()DeepSeek%&()*+,-./012,34/01%&()*+,-./012,34/01DeepSeek!#$%&()#*+,!#$%&()#*+,-DeepSeek./01./01iOS./23456./234567789:;?A89:;?ABBCDEF()#CDEF()#GGHI%HI%BB=!#J*()=!#J*()GG!#$%DeepSeek()*Bloomberg*+,-.4!#$%&()DeepSeek vs Meta5/06789:;5/06789:;Llama#KLMNOPQR

3、#KLMNOPQRSSTTLlama License 1.0UUVVWXYZP/_ab#-cd6efF/gh7iT-Llama()jklmnopqrstpqrstMetauvwx2yz!n|uvwx2yz!n|7/VQ23/Llama#yz-/7“|/ghVT”“|/SMAUU 7-Meta fFPQ-Q|7VT”%-Llama%API-Q Meta PQ-jk/7DeepSeek-R1/MITPQRPQR yz()Vyz()V Q/Q/_ab#_ab#-/;|7Q/#-|7ZP/#yz-T(bC7/Q:#yz_y7ZP/-oh#-yz(b7/Q#_yzhab7PQ/7MIT PQRZP#|/

4、7 Q#/;|?%)*=?;()#6;()#6-#$2#$2-$-M./nM./n77DeepSeek;Token!-ZP/!#$%&d(#7Anthropic)*+z,-./012SDario AmodeiU?%-y345-Q6107_y78GPT-39$%o:;-$2?$%;!=#?$%-:;#$%5:;#$%577DeepSeek-R1jklmQjklmQBB#?#?GG77/DeepSeek-R1ABScuratedU80CD-EFGQwenLlamaC()#7H”Il-JK?LMNOPJK?LMNOP#$%#$%7:;?#-Q./SFT-/RLSRHSRLQ!T#UU7&#()VDe

5、epSeek-R1-ZeroDeepSeek-R1;QwenLlamaADeepSeek-R1?VWX#S1.5B7B8B14B32B70BU7!#$%&(!#$%&(1!#$%&()*+,-./0,-!2!1+,-2#3456789.:;?ABC89DEF*GHIJ!3!K1L,-.M6,-!NO:;PQR;?P#U?P#U-;:P#EFyz;:P#EFyzRL77Y;Y;DeepSeek-R1ZP#rZP#rRLvv77rQwen-32B-Base/_STEMyz!nRL#-#10K3-?#Uyz?#U-;r;rQwen2.5 32B./Ob./Ob771UUO!#?$P#QcdH”O!#

6、?$P#QcdH”-efg!nefg!nRLP#P#-Qh!Qh!-4545Qi-?UQi-?U772UUXY?jkltmdnXY?jkltmdn-copqrcopqr-QsYO!=#!nObQsYO!=#!nOb77=#Ot=#Ot-uP#(DdvuP#(Ddv77wxM:;wxM:;RL77ST*UVRL*UVTherefore,we can draw two conclusions:First,distilling more powerful models into smaller ones yields excellent results,whereas smaller models

7、relying on the large-scale RL mentioned in this paper require enormous computational power and may not even achieve the performance of distillation.Second,while distillation strategies are both economical and effective,advancing beyond the boundaries of intelligence may still require more powerful b

8、ase models and larger-scale reinforcement learning.!#$%DeepSeek()*Bloomberg*+,-.2()*+()*+DeepSeek!#,-!#,-9!#$%&()DeepSeek%CD)*EF%CD)*EFAIGHGHInfraModelSoftware!#$%/&()*bloomberg*+,-.10!#$%&()IJKLMNOPQRIJKLMNOPQRDeepSeek!#$%/&()*bloomberg*+,-.11!#$%&()IJKLMNOPQRIJKLMNOPQRDeepSeek!#$%/&()*bloomberg*+,

9、-.;yz|;yz|BBGG=“|&Qr:F&-,ef=F/e,ef=F/e77AzureAWSDeepSeek-R1#7712MSFT%DeepSeek-R1STSTUVWXYZUVWXYZDeepSeek R1#Foundry GitHub#“S#“SModel CatalogUU b”b”-X$SSRed TeamingUU&S&SContent SafetyUU&S&SSecurity ScanningUU7 Phi P#;#SP#;#SSLMsUU 2323 2000 CC-F-i SSBayerUUPaige.AISSRockwell AutomationUUSSSiemensUU

10、CS 30#-z|o./9z|o./9 7-=15%M Copilot+PC-!s-!PC M Copilot+PC7F-R1?#-(/?#-(/Windows!GPU 66 yz!7yz!7:OpenAISS sY|7T-tr Azure Lu!-45r SSbookingsUUL-M M 56S56Sfirst trancheU7U7;d;d vSvSROFRU-U-6Ssi7 6Ssi7MY-$8M$-rd|Sjk|Sjk 67cA-OGM-r 6!XS6!XSfungible fleetUU-mr#S#StrainingUU$%S$%SinferenceUU-S&a”Sgeo-dist

11、ributedUU7rr&d-Q&d-Q DeepSeek-i-i OpenAISS u!-Tu!-T GPT#$7-r#$7-r$%S$%Sinference optimizationUU L!-:;L!-:;$AI bb 5757r AI*-66 MVM$#SM$#Sfrontier modelU-U-T”x$%$-8dT”x$%$-8d-%D67-!#-$%$2$%$2-AZ$%/$%/7 Mr%!X%!X z7&F-M-6U6U-%(o)(o)*2 7U+7U+-Q 10 7+7+7-jkM,-+.!XS,-+.!XSupgrade the fleetUU_!/S_!/Smoderni

12、ze the fleetUU012012SSage the fleetUU73-45r ISISmonetizationUU 66 789z7:USUSfungibilityUUZ;|S;|Sscale more long-term businessUU7!#$%/&()*bloomberg*+,-.13Meta%Metas panicWW True or FalseWXYdeepseek v3=ZKLlama 429_abcdefg*h=i$%jk500lm5000l#8n*op_qrstuv0wxyz|deepsek=_w*$dmk“”N1u*7Ah*dw?=N1u*$%it*O#deep

13、seek v3*7A=wk3-%Z*itDeepseek r1ed=WrZAh$%uv*L-=Y2_=%d!#$%/&()*bloomberg*+,-.14Meta%/0H5%/0H5QVVPQaPQa77=?A%B=?A%B-=:;()Lz|xiCDMT#E=:;()Lz|xiCDMT#EII-=()LTr$F=()LTr$FGG|H|H77AVVd7;()-?%:IdJiMr(!SOpen ComputeULuKw-YM567-Y(-xLddM!v7-aW7Ycz|/u-%xNO$”7:xyzP-!.QRSWyzP6-*$n7?%r.M-T Llama U$%/-T-VW.x-x API(

14、b,XdQY:W-/W$-$6erZewx2/_y7-?%jk-dn-M-sY:,7?%d-?%Wrb/7?%;_6e?%;_6e-CD:aCD:aDeepSeek-?%Il?%Il-rbaKwrbaKw6M&d6()P6M&d6()P77?%A=cvd-M M6=P7-?e:DK-45&r/p-?%-T”Idfg(h-_6eio9j-k M.LmLl7Qm;DeepSeek no7?%u6epKw-sr7d6eyd6ey-45r45rq9-aWMMCD:a7d!*dy-*db”-d6ey-x(zAbL7s-r8:=1qCLtufgdOvQi%w7.(b6cxv7td;r/!=1qd/;#-T

15、=J$%#$%#Sr$%LM?y LlHIba7!#$%/&()*bloomberg*+,-.15SAP%DeepSeek is good newsSAP CFO Dominik Asam said that Chinese startup DeepSeeks AI model is actually“good news”for the company.SAP)*z,)*z,Dominik AsamI-=|I-AI#:IMBdG7#:IMBdG7CEO Klein:DeepSeekI7I-I7I-)/:)/:AI#y#yAI-$89:-$89:SAP AI77B:SB:SSAPUUAIkIMd

16、7kIMd7A6(8I-)rA6(8I-)rAI=1q!;#S=1q!;#SLLMUULyzS7LyzS79 oo$%6:7G9 oo$%6:7GUnUnDeepSeekM?UM?USAP-KleinI-X-l=7I-X-l=7SAP)*z,)*z,Dominik AsamI-I-T”-$QUCo-T”-$QUCo-SAPQ/#-“|:“|SQ/#-“|:“|SB to BUU22AI.QBnoN7.QBnoN7BA d-Q$MdA7GBA d-Q$MdA7GAsamI-I-AIQ-%Q-%SAP!K)7!K)7SAP-T”BLX-Tm-T”BLX-Tm L-T”#(bLQRl:yz-RG“-

17、L-T”#(bLQRl:yz-RG“-SAPQ/#7Q/#7!#$%/&()*bloomberg*+,-.16NOW%XY)*=?_5abc%XY)*=?_5abc1.AI=#$2:=#$2:ServiceNow”SSWilliam McDermottUUAI=#ST LLMsUrSCommoditizationU-x$2457:ServiceNow;-M6!dVVTVT AI$2$2-ServiceNow-Qd;ServiceNow AI BL7ServiceNowzCDrzCDr AI#-rxrx AIXXSSOrchestrationUU|SSOperationalizationUUVA

18、I Ma563-ServiceNow Z“|AI H%Q+z|z7TVBno+“|6C72.ServiceNow:DeepSeekServiceNow:AI/,-AI#&UQzU#&UQzUSSBB-!6!6GGUU7;B AI”SResponsibleAIU-m AI BLS“|Sn&P7ServiceNow r DeepSeek-Amit x(-crm&UQUoM?/73.ServiceNow AI6”6”SSAmit ZaveryUUServiceNow,5L!#;#SLLMsU-c AI.X$7ServiceNow/ZPX$AI*-m3 AI BL7AI+HSMHSM ServiceN

19、owzCDzCD-/%:/%:77DeepSeek M ServiceNow-AI B6-drxr./-sQ-yzX$7!#$%/&()*bloomberg*+,-.3./012345!6789:;=./012345!6789:;=182025de%de%AIa+Nfg2h)ga+Nfg2h)g202320242025|nn？!#$%/&()*bloomberg*+,-.2025de%de%AIa+Nfg2h)ga+Nfg2h)g)*+,)*+,-*+,-*+,././0123450123456789:;6789:;89?89?ABCABC.DE.DEMicrosoftGoogleAmazon

20、OracleFGHIJCFGHIJCMetaKLMNKLMNKLOKLOMongoDBKLPQKLPQSnowflakeKLMNKLMNConfluentKLMNKLMNElasticAIHIHI2BRQHIRQHISAPPalantirServiceNowSalesforceC3.AIWorkday2BSTHISTHIAppLovinShopifyDocusignGitlabAdobe2CHIHIRedditDuolingo!#$%/&()*bloomberg*+,-.20!#$%&()24Q124Q224Q324Q4AIUVWXUVWX20%#GYZ*AI I_aDbc#def%g RIS

21、E h GROW ijklPmn AI opqEWXPrD 30%s AI I#GY_tuvwxytuz|#aP#89 AI#mEWXPDEWXPD50%ss AI I#GYI#GYJoule NVIDIA)ftuvk Joulehh Joulew AI;tu#RISEq“f”fIah4 E fww 30%tugD Joule#)If-hH-hH_z Joule 500 AftuDg6 80%#9h+GYqJoule AI fy fy#l#lqtu 1,300 AfVIa 80%I#VqgDkgDk 30,000 aIaI SAP Business AIqqtuwgtuwg 2025-AIff

22、DD 30,000 ZtZtu#u#AI#Iq#Iqss 2025 wMNIawMNIa 30%qqJoule8989Q2a(:Mahindra and Mahindra h./DKKa Joule#ry)AI IfDrDr 3!#!#f$%;&(qAIII23Q4)fSAP#E01*+P,30#AI-.f2024wD 100 01/g01Pz,60 gen AI IfDg6 100-.tu62gtu#345P6;100 AI I#qg78Hh4ft9s6;=AI 4:#;5tug 2024 a 130 Gen AI Ifktu#deqSAPKL=SAPRQP#KLs?uD#D=#KLANDB

23、CDNDEFNDGKLNDHIf$JKq,SAP LMNOLMNO_LMNO%KP#9lLMf Gen AI Q;SAP RS#TKLUVhWXtuwTKLhYTKLSAP hY SAP KLf&IZHIqk4=ftuw AI nqwgwg 2!13#Business Unleashed VPf_qVPf_q+a+a6D27000aISAP#SAP Business AI01fk60aaIBTP#GenAI)bIAI-.f6/g 80 IqBTP6Dkk 90+acd+acdI/gefK_)gshi#jklmtu/gkMj2nolV4pqnefI BTP#gen AI Prqsdtu#-vwK

24、wxqgen AI hub 4(tu#+acd#o+acd#opypy-zyz;|ftu#a#opy;q2025de%de%AIa+Nfg2h)ga+Nfg2h)g!#$%/&()*bloomberg*+,-.212025de%de%AIa+Nfg2h)ga+Nfg2h)g“|“|AIAjRHl&“|RAjRHl&“|R|6*/;e k6-iT zz .Q.Q*-xH”7cT-xv-_6&EI(dyzj(dyzj7T-rb$-3$“|Copilot%6P&-&“|7 MY xv7|QMA.(-c jn.c“|.$%7Q/3$-T$!R7-P“|A Copilot Chat(-Y3./97“|Q

25、 Copilot-x$%6 P 6&722)/)/11$-1,000 c“|c“|SS8,400ccUU77Agentic AI/M;/M;Assist-Agent-+/So:/So:77Agentic AI/;P NowAssist“7 6LQmRQn AI-(J7 M&U-r%“|H#AI X*7kz|V!_Ikz|V!_I-6xv!r 6xv!r 2025s,-s,-77Q=*r IT*8-1,250-x7 M%fgTdBnGSDepartment of Government EfficiencyU!7T”ServiceNow r=$-n-AC$-CP7Agentic AIVZI%eVZ

26、I%e77!AI AgentITSM AI Agent#$#$AI Agent%&%&AI Agent(AI Agent2024.112025)*)*+,+,-.-./0/01212Agent33Agent45452025de%de%AIa+Nfg2h)ga+Nfg2h)g!#$%/&()*bloomberg*+,-.23!#$%&()*ijk%abc5lmn#)*ijk%abc5lmn#67678989AI:;$&#“|y“|y)#)#()G()G()d()d)&)&ddHH66&vgvg|HHHH2024lo%p0)*qr?5stuvlo%p0)*qr?5stuv:$%EF:$%EF-tt

27、&./tbX./tbSU6aX72UU&V&VSnowflake246$Snowflake Cortex-%/EFrxXyz AI./(b73UU./tV./tVPalantirOntologySalesforceData cloudServiceNowRaptorDBo%“|“|./-AI./=7llAI./(b77產品成熟度產品成熟度2024.06推出推出Snowflake Cortex2023.05推出推出Microsoft Fabric2022推出推出DataCloud2024.09推出基于推出基于DataCloud的的AgentForce2024.09推出推出Raptor DB201

28、6年推出年推出Ontology目前目前AIP產品基于成熟的產品基于成熟的Ontology模塊模塊!#$%/&()*bloomberg*+,-.43ServiceNowr66Xr66X-nSnS AI_%_%SSAI AgentsUUSSWorkflowsUUHHSSWorkflow DataFabricUURaptorDBX$X$SSIntegration HubUU77 i2i2-!“|.sr!“|.sr*.:-X$A.:-X$A*-rZrZewx2ewx2-45!y45!y-MN_yMN_y77kr!Lkr!L-TT)*,)*,SSChief Data Officer,CDOUU )*)*

29、AI,SSChief AI Officer,CAIOUU-ServiceNowsYMsYM CIOzhXzhXSSControl PlaneUU-M“|M“|AIb3#Llb3#Ll77Workflow Data FabricVV3;%&ServiceNow b$%7T-$Workflow Data Fabric y63Q89STAMU7(-_%AI,-N;+89:Workflow Data Fabric RaptorDB Pro-&NTT C/7%y63NO/-ServiceNow”Oracle Google Cloud&X$-y63+Data Fabric BL$7RaptorDB Pro

30、U(U(VVr RaptorDB Pro)ow89*7 2024 9 b”-$+J 5cS(:cS(:SSACVUU 100CC 7 x=r;2024 5,&/RaptorDB Pro-U+7(.+-3.5 7-(-30%-AM$.72024lo%p0)*qr?5stuvlo%p0)*qr?5stuv!#$%/&()*bloomberg*+,-.432024lo%p0)*qr?5stuvlo%p0)*qr?5stuv“|ra-“|ra-50%IT77Ryz!-P“|ss8a/0711U2or934|7%-r9wx2-AI _%sb7 SAP;6-B AI*A7SAP6HH6HH-SSSAPSA

31、P-3,93,977_%77Joule$%e$%e AI_%.,_%.,SSsuper orchestratorUU-+z6+z6-y!y!77rr213BBBusiness Unleashed”!-=b 6=b 677!#$%/&()*bloomberg*+,-.2025lo%stwg_5)*xyz5|lo%stwg_5)*xyz5|eEF./teEF./tGG1UU|AgentJ./J./77!-:;#$|72UU?/./?/./77#$2-:;Jt|AM+73UUdd77-EFtBl;r?012!A#B9CD?012!A#B9CD2025de%&de%&AIaa2B2CNnNnp3p33

32、0!#$%/&()*Wind*+,-.31!#$%&()*%)*%AIaWstcaWstc67678989AI:;$&#“|y“|y)#)#()G()G()d()d)&)&ddHH66&vgvg|HHHH32DeepSeek!#$#$%&%&()*()*25E+,-+,-25E PE25E././25E PS012314231415856789:14;?DeepSeekABCDEFDEF46GHIE78EJKLM?NDeepSeek Pre-ACOE0.85544.8210PQRSPQRS79T4UVWXY01ZDeepSeekHU_abcdefghijkhijk973lmDeepSeek9:

33、nohpqrst26.5937184.675uvwxuvwx780DeepSeekyz|fhpAI28.63271122.951cdEE40NT4UVOab4ii38NDeepseekMZ“”0.162329.084ii31NDeepseekMZ4Z 56I0.33954.497164NDeepSeek-OY2.566412.0514!#$%/&()*Wind*+,-.AgentaMNkaMNkAgent+,+,././25E25E PEj+aj+an|AgentfXjg 50-200 x2.5264tAA80+afgM9W|Agentfd+f=r2.73.1425“”“”127+af2-+f

34、 10-.Agentf+a.nAgent1.6924jOKjKj73RQfjUMaQChatFilemAIKjAgent2.2033ff1602ifind AI jff AI(RQfaAgent17.2093I“”I“”339YonGPT 2.02100jHIf|MEP-93.26104Agent+,+,././25E%;%;25E PS.662024,PDF AI fb=4.99 G/!f49.99 G/8.398kimi693,AI 01f2412!y 3.59 f6z90.188KLPQAA5024Sophon LLMOpsnjRQf29AgentHI#KL8.066/GG2024z|C

35、U&dOz|CU-D#=D#=AgentJLlEyJLlEy776M-X#hEFEF-|QGl;X|n+7_IV/H_IV/HI=I=%JKJKM-LJz|./M_-|QGl;X|n+7_IVNO_IVNOSSOOUUPQpPQpSS|UUSSRRUUSTSTSS)UU(UV(UVSSWWUUKK././$9$9AIj+a8j+a8I“”I“”9-eERPEmHm-maRm,YonGPT 2.0f2100jHI|MEP-P09vv9EERP93459E9AIfvEq)9*AIRQfnjRQCosmicStudioAIHIRQCosmic*j“”“”9fm.OA9nHjj.hjMNKLKLHID

36、K“”DAI#)APM01Server7.0#S)jRQDataview-v!?9abIT345-aL2.0RQ(jaKLjOjO#$%$/#&#IT345,#9S,njx(fR)KHI01fx(Kx(#x(.F ff*+-ha#vHithinkGPTnjf,-opj-01iFinDv9AIjAIfHIAI+KjKjERPjM9./“901eRQRQ#*f2jUMaQChatFilemAIKjIjAIHI-.OpenAIv.v.9h.2024E3fWPS AI 2.0fAI aAIAIKLAId_DAI#HI45f2/T/C/Kaf2#f,67/KimiChatbotqMiniMaxjOjO!

37、#$%/&()*+,-.“2%()*MNE%()*MNENN././$9$9AIj+a8j+a8jKjH9f6PQ9tfERP01#282023 8!+AIGC RQ H-COPILOT”f24H12r 2500 xG%f S/4HANA#AIGC j345f9:;|tVtVAAjj=M99Wv?r1.45G#nj-.HI+f fg./Gwa0ABfoCL0Dnj!njMNnjHIMNhKLMNYVVAgent-!#$%/&()*+,-.“3%st&jl%st&jl:3+:3+XX=%Y;EF=%Y;EF-!tXZS!tXZS-Xd5$%./(bSS(b6(b6UU77=V_r95,C7#$#

38、$cdcdcd)cd)-iiTDHYcd=JKYcd=JK11cdTDSYcd-Ycd-cd)cdEJKmSophon LLMOpsY.JKY.JKab)Y-YewxwxcdhKcdhKcdcdcdhKcdhKcdE)6-JK-JKS+AIZg8-WWSage MLDBAIcd)Sage MLDBAIcd)Sage LLMOpsgV Agent)scdEN!#$%/&()*+,-.5CapEx012EFGHIJKLM012EFGHIJKLM38(h)l(h)l(-77DeepSeek$-dgwx2#nn”-1UDeepSeek-R1ag?jkltmdnXY?jkltmdn-copcopqrqr

39、-QsYO!=#!nObQsYO!=#!nOb772U!teYr:n2r:n2-,CapEx-=O#U=O#U73UPtrdg2/RL#-#v-d;$%a-d;Transformer+Scaling Law,-$c,-$c-#$%n,-+76U=VWd=VWdyV7Therefore,we can draw two conclusions:First,distilling more powerful models into smaller ones yields excellent results,whereas smaller models relying on the large-scal

40、e RL mentioned in this paper require enormous computational power and may not even achieve the performance of distillation.Efwn#jp#TFfGHIJFy#KL#jjMN%Bn#dAfOPQ#y!#DqSecond,while distillation strategies are both economical and effective,advancing beyond the boundaries of intelligence may still require

41、 more powerful base models and larger-scale reinforcement learning.fR$!STUVDfW%Xj#YZN%n#jhn#KLqfR$!STUVDfW%Xj#YZN%n#jhn AWS GPU:rPOde:rPOde77JJ H200Y7Y7:fe:fe-%GPUgh(!gh(!77:p:p77 i i H100g:ib6!HIg:ib6!HI77SemiAnalysis40(h)l(h)l(!neYM=!#E!7OpenAIsYM(j&sYM(j&RL+CoTEtEt77Claude 3.5 Sonnet#$kC#$kC-T”MT

42、”MAnthropic$-g8AlmnXg8AlmnX-AopqnXAopqnX77U.rwx2-rdg26LM_Ys_#-6L=s_E-%tY7T”uMEDEC:A Benchmark for Medical Error Detection and Correction in Clinical Notesvo1-previeww300B-go1.Mx.-DeepSeekM671B-Y;o1w2x7M?Ilo1iddGMEDEC:A Benchmark for Medical Error Detection and Correction in Clinical Notes=2024.12.26

43、We BelievePre-trainingPost-training(RL)Inference(CoT)Most LLMsPre-trainingPost-training(RLHF)Inference.Once!GPT-4o0120123456OpenAI=MEDEC:A Benchmark for Medical Error Detection and Correction in Clinical Notes=2024.12.26IJ(M%IJ(M%CapEx OP3 OP3!#$%/&()*Wind*+,-.202220232024E2025E2026E2027EMSFT284 412

44、 739 889 980 1,173 45%79%20%10%20%GOOGL315 323 516 597 657 703 2%60%16%10%7%AMZN636 527 777 898 951 984-17%47%16%6%4%META320 281 392 621 663 681-12%39%58%7%3%ORCL67 69 146 153 163 167 4%110%5%7%2%Stargate1,000 1,300 1,400 202220232024E2025E2026E2027EMSFT18%26%29%21%21%23%GOOGL19%20%20%14%14%14%AMZN3

45、9%33%30%22%20%19%META20%17%15%15%14%13%ORCL4%4%6%4%3%3%Stargate24%28%27%IJ(M%IJ(M%CapEx OP3 OP3!#$%/&()*Wind*+,-.02004006008001,0001,2001,4001,600202220232024E2025E2026E2027EGMSFTGOOGLAMZNMETAORCLStargateMSFTVVFY2025!yz!yz800-FY2026!,-!,-N6z|N6772025!a-lz|!a-lz|77|sM!N;6|sM!N;6-rri!i!77085085!#$%/&(

46、)*Wind*+,-.?=?ABABNVIDIADeepSeek-R1 NIM C6)DEFGH NVIDIA HGX H200 IJKLMNOPQ 3,872 Htoken01-30AMDAMDRSDeepSeek-V3TUVInstinct MI300XGPUKDFAMD RyzenAIWRadeonKXYDeepSeek R1Z4_301-25QUALCOMMCabFScdNPUefg DeepSeek-R1 hijkl Copilot+mnopqrQualcomm Snapdragon X s701-29tuvtuvwxy3v6)z/SiliconCloud$|%oK|tuvv6)gD

47、eepSeek-V35DeepSeek-R1,)d5&!U:i#,)d5&!7744-#$%&#$%&#$%()*+,-./0123456789#$%:;1*?1ABCDEFGHIJ=KLMNOP,-.QR1STUV=./W1BCXY=ZY_abc&defbc&def+,-ghi*jklmnopqSrst*uvw41*123aJ(hiv+,-.BCHd-b1,-a(+,-.1*1(x,?1=a+no FINRA 1v(P#$%l FINRA 1#$%/Sl FINRA#$%1,ahi,-2Hn(+,-.123HUVR(=.*1=a23HUVMKAx,1STx,1H|Fd(J:x,UVax,#(

48、M!“”+,-(+,-x,#$1%b&a:(=“”+,-.)1b*+,(hid-./01a21#3*x,4=#.*x,156=78-9N:a;(+,-s.”1?1(A/BG/I(BG1?/I+_1-Cahi-9D*-C1_vFEI(#$s.1-1FG(FG1HIvJKLM.-1_ahidJ.1NO(+,-.1*?1*/+_hBCdUVb1ij/=XY23ahilklBCdUVa,-.lYW?1m|ahi1,n!opHqopFdPcx,|oprs+,-s1BC=UVb1x,#$ax,hid/=PJLM+,-2321tJR4uvax,wxk+,-x,=Pc#1%by(alg61STefwz+,-o

49、a+,-MHJ6/0=unUaPH1=(DMHH“”+,-1?“hi7=6/0=un1=a+,-+no.laK+no56U(=AFHH/H”=#c(Nk=o#)2+noa+noBW?”H1(JU1NO“”(MJ“”rs1/0BC(F5”H6“”/1(”hi*.(:+,-W?lB1”Ha+no11Ra.l+,-s“”1;H|;d;+no1;H|;d;a45stst+,-ghi*jklmno,Jghi jplmnoa*ahi jplmno*d|uvun(hit jlmno1,no(+hi*?1al1ef(whi jplmno1https:/.hk/stock_disclosurePcwCg tt-

50、unefunefattJt+,-ghi*/#ahi*1_h(hi*ptlmno(1934*XY/p15a-6Fdt*XYv(:+,-1ahi*ptlmnoqno1#$%SltunpFINRA#$%1,(hi*ptlmno1(lFINRA#$%;1noHn f6.XY*1mahi*ptlmnohit jlmno1,no(+hi*jklmno1,noaXYhi*ptlmnol,-MO+,-.*W?XY1e(Bhi*ptlmnoW?XYatt-unefunef#$%+deM.+,-.Qd1;1*?1H=a#$%de+,-.Qd1;1*?MN|R4a+efs.Qd1eFINRAm#$%1a#$%(hi

51、*1TR(nox,Z?|1ahi*jklmnoHPno/=Pqno,d/=vF=A!2adhi*.no1*/S(jdp_hi*.no1*/S(jdp_ahi*jklmnoHPno/=Pqno,d/=Pn!Hvl+,-s.Q1*p=x,8V(MW?=*x,#$%:_h+612A?=no_tplZ1j_:4/I1-pAjij4300(ij4(tij4;500(ST?-?j4ss-?j+4-?jJ4nono-j415%F-j45%15%ll-j:4J-15%5%Z-j415%FqH;d-(F“”/d/=no$NNjJNOax,hiQR*d/=no1=/0ETef/0ETefststhi*jklmno

52、Slst*uv41*x,yz|,(qU*91320000704041011Jhi jplmnoSl*uv41O*QRBC|,(qU*AOK809tthi*ptlmnotunpFINRA(SlJt XY|1,(q|UCRD#:298809/SEC#:8-7023146hi jhi jpplmnolmnos+s99ss585808-12+852-3658-6000/+852-2169-http:/.hkhi*hi*ppttlmnolmnot!i#j1041p10001+212-763-8160/+917-725-9702:Huataihtsc-http:/www.htsc-.l2022hi*jklmno+,789:;&+,789:;&$%$%&%&%$%iU()*s+228hi*#j1/,210019&%i-(./012328.456A718/,10003286 25 83389999/86 25 8338752186 10 63211166/86 10 63211275ht-ht-88998i:;(4;+5999610/,5180179i*(*g+18R#jE=23/,20012086 755 82493932/86 755 8249206286 21 28972098/86 21 28972068ht-ht-