【ABIResearchGTC】2025通證經(jīng)濟重新定義AI算力疆界白皮書

GTC2025:

REDEFININGAI’S

COMPUTATIONALHORIZONWITHTOKENECONOMY

ABiresearch?

THETECHINTELLlGENCEEXPERTSSM

NVIDIA,sGPUTechnoIogyConference(GTC2025)markedawatershedmoment,within-personattendancesurgingto25,000attendeesrepresenting800globalcompanies—upfrom16,000thepreviousyear—whilevirtualparticipationheldsteadyatapproximately

300,000.Theexhibitionfloorexpandedtonearly400companies,withnotablegrowthin

thestartuppavilion.TheevidentexcitementthroughouttheeventevokedmemoriesoftheearlyMobileWorldCongressgatheringsofthe2000sorMicrosoft’sProfessionalDevelopersConferencesintheirmid-2000sheyday.

JensenHuang’skeynotearticulatedaparadigmshifttranscendingMoore’sLaw,presentingacomputationalarchitecturevisionoperatingatunprecedentedscaletosupportnext-gen-erationArtificialIntelligence(AI)frameworks,includingagenticsystems.Despiteitstechnicalcomplexity,themessagewasclear:wearewitnessingafundamentaltransitionfromgeneralpurposetoacceleratedcomputing,drivenbyexponentialAIdemand.

ThistransformationredefinesAIcomputationalefficiencyaroundqualitytokengeneration—thefoundationofGenerativeArtificialIntelligence(GenAI)andAgenticAIsystems.Thenewparadigmintegratesthreecriticalelements:rawGraphicsProcessingUnit(GPU)perfor-

mance,sophisticatedGPU-GPUinterconnectcommunications,andoptimizedstorage-GPUinteractions,allenhancedbyworkload-specificlibraries(CUDA-X)tailoredtotargetedver-

ticalsandspecificusecases.WhilecompetitorsremainfixatedonbenchmarkingindividualchipsusingtraditionalmetricslikeTrillionsofOperationsPerSecond(TOPS)orFloating

PointOperationsPerSecond(FLOPS),NVIDIAhasshiftedtheevaluationparadigmfromdataprocessingpipestotokengenerationefficiency—measuringperformancebythequalityandquantityoftokensproducedpersecond,whilereducingthecostandpowerconsumptionofeachtokengeneratedbythesystem.TrueAIcomputationalperformancenowreflectshowtokengenerationscaleswithadditionalGPUsthroughhardwareinnovationandsoftware

optimization.

Followingthisstrategicvision,HuangunveiledtheBlackwellUltrachip,deliveringa50%per-formanceincreaseoveritspredecessor,beforerevealinganambitiousroadmapfeaturing

theforthcomingVeraRubinandRubinUltrachips,andrelatedsystemsslatedfor2026and2027—thelatterpromisingastaggering14Xperformanceleap.Thistransparentroadmap

demonstratesNVIDIA’sconfidence,whileprovidingecosystempartnerswithunprecedent-edstrategicvisibility.ForOriginalEquipmentManufacturers(OEMs),softwaredevelopers,

andendusers,thisclarityenablesstrategicalignment,resourceallocationoptimization,andacceleratedplatformadoption.Forcompetitors,itsimultaneouslyestablishesanintimidatinginnovationbenchmark,whileexposingNVIDIAtoexecutionrisksthatrivalsmightexploit.

Thisstrategictransparency,whilefosteringdeeperindustrycollaboration,alsoamplifiescompetitivepressuresandexecutionstakes—particularlyamidgeopoliticaluncertaintiessurroundingsovereignAIinitiatives,manufacturingshifts,andevolvingregulatoryland-scapes.

Inanutshell,GTC2025mayberememberedasreshapingthevisionofcomputing—redefin-ingitthroughtokengenerationefficiency,ratherthanfasterprocessingpipelinestypicaloftraditionalsequentialcomputing.Throughthisvision,NVIDIAestablishesitselfasthefounda-tionoftheAIeconomy.Organizationsshouldrecognizethisshifttoacceleratedcomputingasrevolutionary,necessitatingstrategicrealignmentoftechnologyroadmapsandinvest-

mentpriorities.

GTC2025:REDEFININGAI’SCOMPUTATIONALHORIZONWITHTOKENECONOMY2

BusinessesmustevaluateAIinfrastructurebasedontokengenerationcapabilities.AscomputationaldemandsforAgenticAIscaleexponentiallybeyondinitialprojections,organizationswillneedtoadoptmoresophisticatedforecastingmethodologies,acknowledgingthatcurrentestimateslikelyunderstatefuturerequirements.Withthisinmind,ABIResearchiscurrentlydevelopingcomprehensiveforecaststhatincludevaluecreationmetricsforAgenticAIsystemsandassociatedCapitalExpenditure(CAPEX)assessmentstodetermineReturnonInvestment(ROI)tippingpoints.

Tofullyleveragethismomentum,ecosystempartnersmuststrategicallyaligntheirinvestments,inno-vationpipelines,andsupplychainplanningwiththisnewtransformationwherequalitytokengenera-tionbecomescentraltotheAIfabric.NVIDIAcompetitorsshouldrecognizethistransformationasbothachallengeandcatalyst,emphasizingthenecessityofacceleratingtheirowninnovationcyclesand

strategicpartnerships.

However,industryplayersmustalsocarefullyanalyzeinherentrisks,chiefamongthemtheover-de-pendenceonasinglevendorecosystem,potentialexecutionfailuresagainstNVIDIA’sambitiousroad-map,escalatinginfrastructurecosts,andthestrategicvulnerabilityofbettingtooheavilyonspecific

AIarchitecturalapproachessuchasAgenticAIbeforethemarketfullymatures.Organizationsshouldproactivelyadoptriskmitigationstrategies,includingdiversifiedsupplychains,strategicpartnerships,scenarioplanning,andadaptableoperationalframeworks,toremainresilientamidthesedynamic

marketconditions.

MalikSaadi

VicePresident

MalikKamal-Saadiishead oftheStrategicTechnology GroupatABIResearch focusingontransformative technologiesandinnovation acrossTelecommunicationsandConnectivityTechnologies, EnterpriseITandOTTech-nologies,Cloud,Edge,andDis- tributedComputing,ArtificialIntelligence,DataWarehouses,RoboticsandAutomation,and otheradjacenttechnologies. Inhisrole,MalikleadsABI Research’sthoughtleader- ship,consultancyservices, syndicatedservices,strategicpositioning,marketforecasts, competitiveassessments,andmarketanalysis.

NVIDIADYNAMO—REVOLUTIONIZINGTOKENGENERATION,BUTWITHIMPLEMENTATIONCHALLENGES

AtGTC2025,NVIDIAunveiledDynamo,agroundbreakingopen-sourcedistributedinferenceframeworkdesignedtodramaticallyenhancetokengenerationefficiency—thenewbench-markforAIcomputationalperformance.Thisinnovationdirectlyaddressestheescalating

computationaldemandsofAgenticAIsystemsbyoptimizinglarge-scaleGPUdeploymentsformaximumtokenthroughput.Operatingasahigher-levelabstractionaboveCUDA-Xli-

braries,DynamoextendstheircapabilitiesfromindividualGPUoptimizationtosophisticatedclustermanagementacrossdistributedenvironments.

Dynamo’sarchitectureintroducesthreetransformativeoptimizations:disaggregatedinfer-

ence,hierarchicalKVcachemanagement,andintelligentcacherouting.Bysplittinginferenceintoseparateprefill(calculatingKVpairsfrominput)anddecode(generatingoutputtokens)phasesoperatingondedicatedGPUresources,Dynamosignificantlyincreasesthroughput—withpreliminarydatashowingmodelslikeLlama70Bachievingovertwicetheperformanceusingdisaggregationalone.TheframeworkemploysasophisticatedhierarchicalKVcache

systemthatoffloadscachesfromexpensiveGPUHighBandwidthMemory(HBM)tomore

economicalhostRandomAccessMemory(RAM)orNon-VolatileMemoryExpress(NVMe)

storage,whileitsintelligentroutertrackscachelocationsanddirectsqueriestoGPUsalreadycontainingrelevantdata,dramaticallyreducingredundantcomputations.

Theseinnovationsproveparticularlyvaluableforspecificusecasesrequiringlargecontextwindows,suchascomplexchatbotinteractions,deepresearchapplications,andcoding

tasksinvolvingextensiverepositories.Dynamonotonlyoptimizescurrentworkloads,butdy-namicallyadaptsGPUrolesbetweenprefillanddecodephasesbasedonreal-timedemand,ensuringresourcesarecontinuouslyalignedwithactualneeds.

3

GTC2025:REDEFININGAI’SCOMPUTATIONALHORIZONWITHTOKENECONOMY

HuanggivespresentationonCUDA-X

ThebusinessvalueofDynamoemergeswhenexaminingitsimpactoninferenceefficiencyandre-

sourceutilization.DynamotargetsinferenceefficiencythroughoptimalGPUutilization,minimized

redundantcomputation,intelligentcaching,anddynamicresourceallocation.ThismeansthatusingDynamohelpsimprovetheeconomicsoflarge-scaleinferenceworkloads,particularlyforapplicationsrequiringextensivecontexthandling.AscomputationaldemandsforAIcontinuetogrow,Dynamo

providesorganizationswithaframeworkthatmakestokengenerationmoreeconomical,efficient,andenvironmentallysustainableatindustryscale.

Beyondcostreduction,Dynamoenablesorganizationstosignificantlyenhanceservicequalitythroughimprovedthroughputandreducedlatency,supportingmoreconcurrentuserswithexistinginfrastruc-ture.ThisscalabilityadvantagebecomesparticularlycriticalasAIadoptiongrowsexponentially—allow-ingbusinessestoaccommodateincreasingdemandwithoutproportionalinfrastructureexpansion.ForAI-as-a-Serviceproviders,thisdirectlyimpactsrevenuepotentialandcompetitivepositioning.Addition-ally,byextendingthepracticalusabilityoflargecontextwindowswithoutprohibitivecosts,Dynamo

enablesentirelynewclassesofAIapplicationsthatwerepreviouslyeconomicallyunfeasible,creatingopportunitiesforproductdifferentiationandnewmarketentry.AscomputationaldemandsforAgenticAIcontinuetogrowexponentiallybeyondinitialprojections,Dynamo’soptimizationapproachprovidesacrucialpathwayforsustainablescaling—helpingorganizationsmaximizereturnonAIinvestments,

whileminimizingoperationalcostsandenvironmentalimpact.

Despiteitspromisingbenefits,implementingDynamopresentssubstantialchallenges.Theframework’sdistributednatureintroducescomplexityinorchestrationandresourcemanagementacrossmultiplenodes.Ensuringoptimalhierarchicalcachemanagementwithoutlatencypenalties,balancingcompet-ingfactorsoflowlatencyandhighthroughput,andaccommodatingheterogeneousworkloadsrequiresophisticatedreal-timedecision-making.Organizationsmustensuresufficienthigh-performance

networking,GPUresources,andfaststoragesolutions,whileaddressingdataconsistencyconcerns

indynamicmulti-nodeenvironments.Additionally,theinitialinfrastructureinvestmentandscarcityofspecializedexpertiseindistributedsystems,GPUarchitecture,andAIinferenceoptimizationmaypres-entbarriers,particularlyforsmallerorganizations.

4

GTC2025:REDEFININGAI’SCOMPUTATIONALHORIZONWITHTOKENECONOMY

Nevertheless,DynamoexemplifiesNVIDIA’sstrategicshifttowardopeninnovationinsoftwareto

complementitshardwareleadership,withsourcecodeavailableonGitHubforcommunitycontribu-tionandadaptation.ThisframeworkdirectlysupportsNVIDIA’svisionofredefiningcomputingaroundtokengenerationefficiency,dramaticallyreducingthecostandpowerconsumptionrequiredforeachhigh-qualitytoken—essentialforscalable,economical,andenvironmentally-friendlyAIinferenceas

computationaldemandscontinuescalingexponentiallybeyondinitialprojections.

EXPANDINGDGXPORTFOLIO:FROMEDGETOENTERPRISE

AtGTC2025,NVIDIAintroducedanewlineinitsDGXfamily,AIsupercomputingsolutionsdesigned

specificallyforDeepLearning(DL),AI,andadvancedanalyticsworkloads.Atthehighend,NVIDIA

introducedtheDGXSuperPODequippedwithBlackwellUltraGPUs(specificallytheDGXGB300and

DGXB300systems),offeringunprecedentedcomputationalpowerforenterprisesrequiringscalableAIsupercomputingresources.ThisflagshipsolutionaddressesthegrowingdemandforinfrastructurecapableofhandlingincreasinglycomplexAIworkloadsatscale,eliminatingtheneedforremoteaccesstocentralizedresourcesformanyAIdevelopmentandinferencetasks.

Complementingtheselarge-scaledeployments,NVIDIAintroducedDGXSpark,previouslyknownasProjectDIGITS.ThiscompactAIsupercomputer,developedincollaborationwithMediaTekandpow-eredbytheGB10Blackwellchip,istailoredspecificallyfordevelopers,datascientists,andresearch-ers.ThecollaborationwithMediaTekbringsexpertiseinenergy-efficientchipdesign,enhancingDGXSpark’scapabilitiestoeffectivelysupportinferenceworkloadsattheedge.DGXSparkaimstodemoc-ratizeaccesstocutting-edgeAIcomputingpower,allowinguserstoprototype,fine-tune,andperforminferenceonlargeAImodelswithoutrelianceontraditionaldatacenterresources.Thissolutionis

particularlysuitedfordecentralizedlocations,enablingenhancedAIcapabilitiesinenvironmentswithspace,power,andlatencyconstraints.

ThecollaborationwithEquinixtolaunchtheInstantAIFactoryprovidesamanagedserviceoffering

fullyprovisionedAIfactoriespoweredbyBlackwellUltraDGXSuperPODs.Thisinnovativesolution

significantlyreducesimplementationbarriersfororganizationsseekingimmediateaccesstoenter-

prise-gradeAIinfrastructurewithoutthecomplexitiesofbuildingandmaintainingtheseenvironmentsthemselves.

WhiletheseadvancementspresentcompellingopportunitiesforacceleratingAIinnovationandreduc-inginfrastructurecomplexity,theycomewithimplementationchallenges.Organizationsmustnavi-

gatethecomplexityandcostofinitialdeployment,acquirespecializedskillsindistributedAIandedgecomputing,addressdatasecurityconcernsatdecentralizedlocations,andensureconsistentreliabilityacrossdiverseoperationalenvironments.Despitethesechallenges,NVIDIA’scomprehensiveportfoliodemonstratesastrategicvisionofAIinfrastructureasacohesiveecosystem,ratherthanisolatedde-ploymentmodels,positioningthecompanyasthefoundationalplatformfortheAIeconomy.

5

GTC2025:REDEFININGAI’SCOMPUTATIONALHORIZONWITHTOKENECONOMY

QUANTUMSUCCESS:THEHYBRIDCOMPUTINGIMPERATIVE

AtNVIDIA’sGTC2025QuantumDay,industryleadersrepresentingmajorquantumorganizationsconvergedonacommonmessage:quantumcomputing’ssuccessfuladoptionwilloccurthroughhybridintegrationwithclassicalcomputing,ratherthanasastandalonereplacement.Thishybridmodelisviewedbyleadingplayers,including

AmazonWebServices(AWS)andMicrosoft,asthemostfeasiblepathtowardcommercialsuccess.Huangadvisedthequantumcommunitytopositionquantumcomputingasacomplementaryextensionofclassicalcomputingtopreventstrategicpitfallssimilartothosefacedhistoricallybytechnologieslikemassivelyparallelcomputing.Thisstrategicapproachwillenableindustryplayerstotargetimmediateopportunitieswithpracticalapplications,whilesimultaneouslygeneratingsufficientrevenuetosustainlong-termResearchandDevelopment(R&D)investments,creatingasustainableinnovationecosystemthatcanmethodicallyadvancequantumcapabilitieswithoutfallingvictimtohypecyclesorunrealisticexpectations.

Thequantumcomputingindustryistransitioningfromfoundationalresearchintopracticalapplications,withnota-bleadvancementsinperformancebenchmarks.IonQrecentlyachieveda12%performanceenhancementinbloodpumpdynamicssimulation,whileQuantinuumcontinuessettingquantumvolumerecords.Nevertheless,substan-tialchallengespersistinerrorcorrection,scalability,andintegrationwithinexistingcomputationalinfrastructure.

Industryconsensussuggestsspecializedquantumapplicationscouldreachpracticaladoptionwithin5years,

whilecomprehensiveindustrytransformationmayrequireatleastadecade.Organizationsshouldidentifyspecificcomputationaltaskswherequantumoffersclearadvantages—suchasmaterialsscience,chemicalsimulations,op-timizationchallenges,andscientificresearch—whilecontinuingtouseclassicalcomputingforbroaderworkloads.

Forward-thinkingenterprisesshouldinitiatequantumreadinessassessments,identifyquantum-alignedusecasesrelevanttotheirbusinessgoals,anddevelophybridcomputingframeworks.Buildingcross-disciplinaryteamswithexpertiseinquantumphysics,high-performancecomputing,andtargetedapplicationswillbeessential.Strategicpartnershipswithquantumhardwarevendors,cloudplatforms,andsoftwaredeveloperswillenablecompaniestosecureearly-adopteradvantagesasthetechnologyevolves.

Thenextdecadewillbedefinednotbyquantumreplacingclassicalcomputing,butbyhoweffectivelyorganizationsintegratethesecomplementaryparadigms.Thosethatpositionthemselvesatthisintersection—understanding

boththelimitationsandpossibilitiesofquantumtechnologies—willgainsignificantcompetitiveadvantagesincomputation-intensiveindustries.

—MalikSaadi,VPofStrategicTechnologies

GTC’sinaugural

QuantumDay

6

GTC2025:REDEFININGAI’SCOMPUTATIONALHORIZONWITHTOKENECONOMY

NOTABLEHARDWAREANNOUNCEMENTS

NVIDIA’sdominanceintheAIdatacenterhasnotbeenthrownintoquestion,andthe

ambitiousroadmappresentedatGTCfurtherentrenchesthisposition.Theperformance

improvementstocome,andclarityaboutthecadenceofhardwarereleases,isappreciatedbyitsnetworkofcustomersandGTMpartners,whocanexpectannualhardwareupdates

PaulSchell

IndustryAnalyst

PaulSchell,IndustryAnalystatABIResearch,isresponsible forresearchfocusingon ArtificialIntelligence(AI)hardwareandchipsetswith theAI&MachineLearningResearchService,which

sitswithintheStrategic Technologiesteam.The burgeoningactivityaroundAImeanshisresearchcoversbothestablishedplayersandstartupsdevelopingproductsoptimizedforAIworkloads

forcloud,enterprise,androbotics.WhatevercomesnextintheworldofAIandacceleratedcomputing,weknowthatNVIDIA’ssupercomputers—soldasawholeanddisaggregatedformorepickycustomers,namelyhyperscalers—madeupofGPUs,interconnect,andnetwork-ingsolutions,willremainatthecuttingedge.Nonetheless,itwasinterestingtohearHuang’semphasisontheprogrammabilityofGPUs,whichformspartofthepicturewhenconsideringtheefficiencytrade-offwithApplication-SpecificIntegratedCircuits(ASICs)suchashyperscal-ers’in-housedesigns(thinkGoogle’sTensorProcessingUnit(TPU)andAWS’Inferentia).Whileapplication-specificsiliconhasbeenhonedtotackletransformernetwork-basedGenAI

workloads,tomorrow’sneuralnetworkcouldlookdifferent.Hardwaredevelopmentcycleslagbehindsuchdevelopments,whereastheinherentflexibilityofGPUsrequire“only”anupdat-edsoftwarelibrarytooptimizeforanothertypeofparallelprocessing—somethingNVIDIA’ssoftwareengineershavebeendoingforyears.

GPUS,CPUS,ANDSUPERCHIPSWEREONFULLDISPLAY

ConsciousofGPUs’energyappetite,efficiencyandcostpertokenunderscoreNVIDIA’smes-sagingaboutBlackwell’sreportedlypositiveramp-up,theexpectedimprovementsofBlack-wellUltralaterthisyear,andthe2026Rubinarchitectureinconjunctionwiththenextline

ofcustomArmNeoverseCentralProcessingUnits(CPUs)namedVera.Alsonotableisthe

doublingoftheGPUdensityofRubinUltra,plannedfor2H2027,whichwillhavefourreti-cle-sizedGPUdiesinthesamepackage.UnderpinningNVIDIA’scolossalUS$1trillionofdatacenterCAPEXby2028isthescalingofAIreasoningmodelsandtheconsiderablecomputeneededtocatertomoreintelligentAgenticAIsystems.

Whenunpackingwhat’sunderthehoodofthenewreferencedesigns,suchasNVL144,itisimportanttoconsiderthatNVIDIAcountsthenumberofGPUdies,notthenumberof“pack-aged”GPUs.ThefirstgenerationVeraRubinsystems,duein2026,willbecalledNVL144,

althoughthesystemissimilartoGB200NVL72—thesamerackand72GPUpackages.ThesamerackwillalsobeusedforthefortheGB200andGB300NVL72systems.Separately,

theupcomingB300NVL16willtakeoverfromtheB200HGXformfactorandcontinuetobeinterconnectedwiththein-houseNVLinkprotocol.ThenextrackdesignwillcomewiththeRubinUltraandwillfurtherincreasedensitybyintroducing144GPUpackagesintoasinglerackwithatotalof576GPUdies.

—PaulSchell,IndustryAnalyst

7

GTC2025:REDEFININGAI’SCOMPUTATIONALHORIZONWITHTOKENECONOMY

NVIDIAROLLSBACKSOFTWAREMONETIZATIONTOFOCUSONINFRASTRUCTUREDEMANDGENERATION

In2024,NVIDIAannouncedNVIDIAAIEnterprise,asoftwareplatformaimingtosupport

AImodelinferenceandacceleratedeploymentofapplicationsonNVIDIAhardware.This

ReeceHayden

PrincipalAnalyst

AspartofABIResearch’s strategictechnologiesteam,PrincipalAnalystReeceHayden leadstheArtificialIntelligence (AI)andMachineLearning (ML)researchservice.His primaryfocusisuncovering thetechnical,commercial, andeconomicopportunities inAIsoftwareandAImarkets. ReeceexploresAIsoftware acrossthecompletevaluechain,withacross-verticaland globalviewpoint,toprovidestrategicguidancefor,amongothers,enterprises,hardware andsoftwarevendors,hyper scalers,systemintegrators, andcommunicationservice providers.Reecepreviously workedinthedistributed& edgecomputeteam,where hesupportedclientsacrossvariousareas,includingenter- priseconnectivity(includingnetwork-as-a-service),

edgeAIplatforms,andthesemiconductormarket.

softwareplatformwaspositionedasademanddriverforhardwarebyreducingbarrierstoapplicationdeployment,butalsoamonetizationopportunityforNVIDIAasitisdelivered

asaSoftware-as-a-Service(SaaS)-basedsolutionwithacostofUS$4,500perGPUperyear.

OneyearonanditisclearthatmonetizingNVIDIAAIEnterprisehasbeende-prioritizedwithenergyfocusedonusingthisplatformandothersoftwaretoolstoacceleratedemandgen-

erationforNVIDIA’sAIinfrastructure.ThisisareturntoNVIDIA’straditionalbusinessmodelthathasleveragedtoolslikeCUDAXlibrariesandSoftwareDevelopmentKits(SDKs)tomakeiteasiertooptimizeAIprocessingonhardwareandbuildamoataroundtheirhardware

solution.NVIDIAAIEnterprisewillstillbemonetizedwiththeexistingSaaSmodelstillinplace,butitwillmostlybechanneledtocustomersaswhite-labeledsolutionsthroughpartners

likeAccenture,Amdocs,HewlettPackardEnterprise(HPE),andServiceNow.Althoughthesepartnerswill,mostlikely,takealotof“margin”throughthesewhite-labeledsolutions,NVIDIAwillbetheultimatewinner,asitwillfueldemandforitshighmarginAIinfrastructureandac-celeratedcompute.NVIDIA’smajorsoftwareannouncementsaligncloselywithitstraditional“infrastructuredemandgeneration”strategy.AstheAIhardwaremarketbecomeschalleng- ingmovingforwardwithmorecompetitivesolutionstargetinginferenceworkloads,itmakessenseforNVIDIAtoreturntoitstraditionalmodelfocusedonusingsoftwaretogenerate

demandforitsAIinfrastructure.

NVIDIAFENDSOFFCOMPETITORSBYPLACING“PARTNERS”FRONTANDCENTER

NVIDIA’scommercialmodelhasalwaysfocusedonhorizontalinnovationandverticalcom-mercializationthroughamaturechannelpartnershipecosystem.Thishasbeenasuccessfulapproach,todate,withOEMs,OriginalDesignManufacturers(ODMs),SystemIntegrators

(SIs),IndependentSoftwareVendors(ISVs),andverticalexpertsallplayingacriticalrole.

NVIDIA’sroadmapannouncementclearlysignalsthatvisibilityforitspartnersoutweighsthethreatofcompetitionfromtherestofthemarket.ThisroadmapvisibilityisafirstofitskindintheGPUmarketandisindicativeofhowNVIDIAseesitselfanditspartnersinthemarket.AlthoughthismovewillinvitecompetitiveR&Dfromtheusualsuspectsandnewentrants(especiallytargetinginferenceworkloads),itwillalsoactasabarriertomarketfor

competitors,aschannelpartnerswillmaptheirroadmaptoNVIDIAtoensureguaranteesof“best-in-class”solutions.Beyondhardware,NVIDIAshowcaseditsongoingcommitmenttosoftwarepartners,highlightingtheraftofchannelpartnersthathaveco-developedandwhite-labeledNVIDIAAIEnterprisesolutions.ThisapproachwillenableNVIDIA’spartnerstoleadthemarketinAgenticAI,byleveragingNIMsandNeMotosupportthebuildoutof

optimizedagenticsolutions.

8

GTC2025:REDEFININGAI’SCOMPUTATIONALHORIZONWITHTOKENECONOMY

GeForceRTX5090GPU,whichrunsontheBlackwellarchitecture

INFERENCETAKESCENTERSTAGE,ASAGENTICAIANDREASONINGMODELSCOMETOTHEFOREFRONTOFNVIDIA’STECHNOLOGYR&DAND

EVOLVINGCOMMERCIALSTRATEGY

GenAItechnologyandcommercialmaturityarepushingthemarketawayfromafocusontraining

bettermodelstowardusinginferencetocreateROI.Onthesurface,thistrendcreatesachallengeforNVIDIA.Marketsentimentwidelyviewsinferenceas“easy,”meaningthatitcanberunonlesscom-

pute-dense(expensive)AIinfrastructure,whichposesasignificantrisktoNVIDIA’sleadingAIinfrastruc-ture.Tochangethisnarrative,andplaceNVIDIA’sfull-stacksolutionatthecenteroftheAIinference

market,ittargetedtwokeyareas.

ThefirstwasfocusedonchangingthenarrativearoundAgenticAIandreasoningmodels.Market

hypearoundAgenticAIhasquicklygrown,fueledbytheemergenceofnovel(andhighlyefficient)

reasoningmodels(e.g.,DeepSeek-R1).Thesesmall,moreefficientmodelsthatarereportedlytrainedforafractionofthecostof“traditionalLargeLanguageModels(LLMs)”(e.g.,GPT-4.5)createdwide-

spreadskepticismthatlargeAIdatacentersorfactorieswithleading-edgeinfrastructurewillnotbenecessary,astrainingandinferencingdemandswillnotgrowasquicklyasfirstexpected.Tochangethisnarrativeandattempttorepositionitshardwaretoaddressinference(aswellastraining),NVIDIAshowcasedthatreasoningmodels(comparedtotraditionalone-shotLLMs)usefarmoretokensto

processpromptsgiventheirmulti-stagereasoning.ThissuggeststhatscaledAgenticAIframeworks(evenifbasedonsmallerindividualmodels)willrequireasmuch,ifnotmore,computecapacity.Thesecondwasseveralsoftwareannouncementsthatseektosupportinference,ratherthantraining

workloads.Thesemessages(alongsideothers)