超越GPU构建AI基础设施堆栈

MarvellTechnology(MRVLUS) BuildingtheAIInfrastructureStackBeyondGPUs:Initiation InvestmentFocus WeinitiatecoverageonMarvellTechnologywithaconstructiveview,aswebelievethecompanyistransitioningfromadiversifiedcommunicationssemiconductorsupplierintoacriticalAIinfrastructuresiliconplatform.CustomAIsiliconshouldremainthelargestgrowthengine,supportedbyhyperscalers’increasing adoption of workload-optimized accelerators.Meanwhile,Marvell’s leadership in optical DSP provides differentiatedexposuretothegrowinginterconnectbottleneckasAIclustersscaletowardhigherspeedsandlargerdeployments.CXLandAInetworkingfurtherextendthecompany’s positioning across memory infrastructure and clusterconnectivity. Together,thesecapabilitiespositionMarvellacrossmultiplelayersoftheAIinfrastructurestack,spanningcompute,opticalinterconnect,memoryandnetworking.Webelievethismulti-layerexposurecouldimprovegrowthvisibilityandbusinessmix,maysupportdeepercustomerengagementandreducerelianceonanysingleproductcycle.Ashigher-valueAIbusinessesbecomealargershareofrevenue,combinedwithoperatingleveragefromalargerrevenuebase,weexpectsustainedearningsgrowth,marginexpansionandstrongerfreecashflowgeneration.Inourview,thissupportsastructuralre-ratingfromatraditionalcommunicationssemiconductorsuppliertowardabroaderAIinfrastructuresiliconplatform. Valuation&recommendation WeinitiatecoverageonMRVLwithOUTPERFORMandaYE2027targetpriceofUS$249,basedonourDCFvaluation.WebelievethemarketcontinuestounderappreciateMarvell’slong-termrevenueandearningspotentialasitsbusinessmixshiftstowardCustomASIC,OpticalDSP,CXLandAINetworking.ItsexpandingroleacrossmultiplelayersofAIinfrastructureshouldsupportahighervaluation framework than that of a traditional communicationssemiconductorcompany. Risks KeydownsiderisksincludeaslowdowninhyperscalerAICapEx;weaker-than-expectedcustomASICdesignwinsorintensifiedcompetition;opticalDSPmarket-sharelossesandpricingpressure;slowerCXLecosystemadoption;andabroadersemiconductordownturnorprolongedweaknessinMarvell’slegacycarrier,enterprise,automotiveandconsumerbusinesses. Contents 1.CustomAISilicon:TheLargestGrowthEngine.....................................................31.1HyperscalersAreMovingTowardCustomAIAccelerators...................................31.2Marvell’sRoleinCustomAISilicon........................................................................41.3WhyCustomASICCanSupportaHigherValuationFramework...........................52.OpticalDSP:TheUnderappreciatedAssetinAIInterconnect................................62.1AIClusterScalingMakesOpticalInterconnectaStructuralBottleneck................62.2InphiAcquisitionIsBecominganAI-EraStrategicAsset........................................72.3OpticalDSPSupportsBothRevenueGrowthandMarginExpansion....................73.CXLandMemoryInfra:TheNextAIBottleneck....................................................83.1MemoryIsBecomingaKeyConstraintinAIDataCenters....................................83.2CXLEnablesMemoryPoolingandResourceOptimization....................................83.3Marvell’sCXLPortfolioPositionsItAcrosstheEmergingMemoryInfrastructureStack.............................................................................................................................94.AINetworking:TheBackboneofFutureAIClusters............................................104.1EthernetandSwitchingBecomeCriticalforAIScale-Out...................................104.2NetworkingHelpsReinforcethePlatformNarrative...........................................105.FinancialOutlook:AIMixShiftDrivesRevenueGrowthandMarginExpansion...125.1RevenueMixIsShiftingTowardAIInfrastructure...............................................125.2MarginExpansionComesfromHigher-ValueAIBusinesses...............................126.Valuation&Recommendation...........................................................................157.Risks.................................................................................................................17 1.CustomAISilicon:TheLargestGrowthEngine 1.1HyperscalersAreMovingTowardCustomAIAccelerators WebelieveCustomASICswillbetheprimarydriverofMarvell’sfuturerevenuegrowthandvaluationexpansion.AIinfrastructurehashistoricallybeendominatedbyGPUs,butastrainingandinferencedemandcontinuestoscale,cloudserviceprovidersareincreasinglyfocusedonperformance,powerefficiencyandtotalcostofownership.Forstable,repetitiveandlarge-scaleAIworkloads,customASICscanofferhigherenergyefficiencyandlowerlong-termcostscomparedwithgeneral-purposeGPUs. Lookingahead,we do not expect ASICs to fully replace GPUs in AI computingarchitecture.Instead,GPUsandASICsarelikelytocoexist.GPUswillremainwellsuitedforflexible,rapidlyevolvingmodeltrainingandcomplexworkloads,whileASICsarebetterpositionedformature,stableandlarge-scaleinferencedeployments,aswellasspecifictrainingtasks.Therefore,therisinginvestmentbyhyperscalersininternallydevelopedAIchipsshouldbeviewedasanaturaloutcomeofthematurationofAIinfrastructure.AccordingtoPrecedenceRes