Meta拟布局云基础设施业务 或为前沿AI研发保留核心算力

BuyNorth America Company Update United States Internet A Silver-Lining Benjamin Black, CFA Research Analyst+1-212-250-9218Kunal Madhukar, CFA infrastructure business that will sell access toAl computingpowerand models.ambitions to build frontier models.2) If Meta has potentially overbuilt capacity,what does this imply for Capex going forward? 3) How much revenue could Metaexpect to generate from their excess capacity? All in all, our read is not that Metais backing awayfrom frontier Al, rather,we think Meta maybereserving its moststrategic capacity for internal model development, while selling non-core, older-generation,ortemporarilyunderutilized capacity into a still-constrained computemarket.Bigger picture, we contend this could pivotthe Al capex debate from"Capexwithlimitedrevenueupside”into“high-flow-throughrevenueoptionality." Research Analyst+1-212-250-9357RaymondWong Research Associate+1-212-250-2060 Benjamin Hui, Esq.Research Associate+1-212-250-2064 The report suggested that Meta is evaluating a cloud infrastructure business thatcould sell bothraw Al compute and hosted model access,includinga Bedrock-likeAPllayerwhich inourviewcouldhighlightthatthecompanyappearstobelookingforamoredirectrevenuebridgebetweenthecurrentsizableAlcapexinvestmentsand the still-unclear Al product monetization. All in all, we contend that Metapotentially moving into cloud infrastructure sales improves the investor narrativepreserve frontier Al optionality while monetizing less strategic or temporarilyunderutilized compute. That said, we still expect near-term capex to remainelevated, though we do contend that the risk profile of these investments wouldimprove if the company establishes a direct3P cloud salesrevenue stream.OurscenarioanalysisforFY27suggeststhepotential for$9-s30bnin incrementalrevenue(3-10%upsidevsStreet estimates)and~s1.40-$7.50 in incrementalGAAPearningspershare (4-21%upsidevsStreetestimates).Inadditiontothegreaterearningspower,wewouldexpectinvestorstocapitalizeMeta'searningsatamodestlyhigherrate ontheback of the establishment of a cloud business,as it de-risks thebroaderinfrastructurebuildoutthe companyis currentlypursuing.Moreover,wedocontendthata Bedrock-likeAPlbusinesswouldalsocreateamoredurable, higher-quality revenue stream than raw compute resale alone. ImplicationforAlambitions:wedonotviewthisasaretreatfromfrontiermodels Meta has consistently framed superintelligence as a strategic priority and hasdiscussedmultiplegigawatt-scaleclusters,includingPrometheus comingonlinein InternetMeta potentialthird-partycomputesalesasMetaopportunisticallymonetizingaportionof its existing server assets,and de-risking the large infrastructure buildout vs.suggesting the company has reduced its frontier-model ambition. In ourview thekeyquestion is which capacity Meta sells.We think Meta couldbe preparing tomonetizeolderNvidia capacity,GB200/GB300-class systems,inference-orientedclusters, or non-core capacity whilereserving the newest Rubin-era systems forstate-of-the-art training, implying this emerging revenue opportunity could be aresultof chip-mix.Totheextentthat Meta isplanningto sell capacitythatnolongersits atthefrontierofits training roadmap, we view the move as disciplined capacityutilization and growing financialoptionalityratherthana strategic pivot. At the same time,we do not view this report as a signal for lower gross capex (though,giventhepotentialforincremental revenue,itcould have implicationsforcapital intensity). Just last quarter the company raised capex guidance to s125-145bn.ThatAl infrastructure spendhas longlead times across chips, land,power,cooling, data centers, networking, and leases. Our interpretation is that thispotential compute resalecould servetolowertheprobabilitythat Meta is saddledwithstrandedcapacity.Ifthecompany,however,sellsexcesscapacityatattractiverates,the market would not needto underwrite all Al capex solelyagainst indirectbenefits to ads, recommendations, engagement, assistants, and future modelproducts, given the potential addition of a direct revenue stream againstinfrastructure costs that arealreadyembedded in Street/Buyside estimates. As such, we do not think that this implies Meta will cut capex. In fact, it likelysupports a continued buildout if the company secures scaled, longer-termcommitments. The company's recent actions highlight a willingness to minimize balance sheetintensity.Withinthe Hyperion/Blue Owl JV,Blue Owl funds owns 80% andMetaretained 20% while leasing the facilities; we believe that third-party computerevenue would serve to make that financing strategy more defensible. 2027revenueopportunity:Low-SD-to-Low-DDupsidepotential,with45-360bpsbetterOlmarginspossiblieWebuild thepotential revenueopportunityaroundFY27below,based onseveral simplifyingassumptions. Meta'sAl-related capacitywas~2GWatFY25-end.ForFY26andFY27,weestimate the company may be adding 2-3.5GW and4-6GW of additionalcapacity,basedonthecapexguideforFY26andourexpectations