NOAIURA

ConnectingMarketsEast&West

ANCHORREPORT

GlobalMarketsResearch

9January2026

GlobalAlnetworkingsupercycledrivenbytechupgrade/supplyshortage

AccumulatemarketleadersInnoLight,YOFC,TFCandT&S

WebeievetheglobalAlnetworkingsupercyclewillcontinuein2026F,andevenextendinto2027F,drivenbymultipletechnologyupgraderoadmaps(i.e.SiliconPhotonics

[SiPh],Co-packageoptics[CPO],OpticalCircuitSwitch[OCS],ActiveElectronicCable[AEC],HollowCoreFiber[HCF])andsupplyshortage(i.e.opticallaserchip/material).Weestimate800G/1.6Ttransceivershipmentswillgrowfrom20mn/2.5mnunitsin2025Fto43mn/20mnunitsin2026F,withSiPhtransceiversgaining50~70%marketshare,andexpectInnoLight(300308CH,Buy,oursectortoppick)tocontinueto

dominatethehigh-endmarket.OwingtolikelymoreaggressivebundlingsalesstrategyfromglobalAlleaders,CPOmigrationshouldacceleratefrom2026(fromalowbase),andwethinkTFC(300394CH,Buy)andT&S(300570CH,Buy)willlikelybekey

beneficiariesintheglobalCPOvaluechain.WeupgradeYOFC(6869HK)toBuyfromNeutralasthecompany'sAIDCbusinessenjoysbuoyantdemandwhilethetelecommarketstabilizes.

ResearchAnalysts

ChinaTechnology

BingDuan-NIHK

bing.duan1@+85222522141

EthanZhang-NIHK

ethan.zhang@+85222522157

KeythemesandanalysisinthisAnchorReportinclude:

·DetaileddiscussiononglobalhyperscaleAlplayers'networkingroadmapsfor

scale-upandscale-outnetworks,includingoptics,copperandnetworkingswitches

·Demand-supplyanalysisonkeysegments,includingopticaltransceiversandCPO

·UpgradeYOFCtoBuyandupdatefinancialsforInnoLight,TFC,andT&S.ProductionComplete:2026-01-0916:22UTC

SeeAppendixA-1foranalystcertification,importantdisclosuresandthestatusofnon-USanalysts.

SeeAppendixA-1foranalystcertification,importantdisclosuresandthestatusofnon-USanalysts.

9January2026

GlobalAlnetworkingsupercycleGlobalMarketsResearch

EQUITY:TECHNOLOGY

Whattoexpectin2026F(and2027Fonwards)

UpgradeYOFCtoBuy;maintainBuyonInnoLight,TFC,andShenzhenT&S

Alnetworking:soliddemandthrough2026-2027Fandacceleratedtechmigrationto

underpindominantplayers'growth

WebelievetheAlinfrainvestmentupcyclefromglobalhyperscaleAlcloudcompanieswillcontinuein2026F-27F,ledbytheintensivecompetitionforlargelanguagemodel(LLM)trainingandinference.GPUandASICplayersarestrivingtogainmoremarketsharein

thisLLMenvironment,andtheirshortenedtechnologyupgradecyclewillaccelerate

productandtechnologyinnovationintheAlnetworkingsegment,inourview.We

believeAlnetworkingsolutionssuchasswitches,opticalcable/transceivers,andcoppercableswillplayanincreasinglysignificantroleinAldatacenters,and

leadingplayersintheAlnetworkingsegmentwillbenefitfromtheacceleratedinnovationcycleaswellasanenlargedTAM(totaladdressablemarket)inscale-up,scale-outandscale-acrossnetworks.Meanwhile,wethinkcomponentshortageswillpersistin2026F,duetobuoyantdemandandhighertechbarriers,andthiswillbenefitdominantplayerswithpotentialpriceandmarginuptickintheglobalAlnetworkingsegment.

Opticaltransceivers:1.6TtransceiverandSiPhmigrationarethekeydriversin

2026;NPO/CPOstartcommercialization;opticalchipshortageiskeytowatch

DrivenbyglobalAlleaders'buoyantdemandforhigh-speedopticaltransceivers,aswellaspersistentcomponentshortages(i.e.,opticalchips),webelievetheproductupcyclewillcontinuein2026Fandlikelyextendinto2027F.Meanwhile,productinnovationinthe

opticaltransceivermarketisaccelerating,andfuturetechnologytrendsincludingsiliconphotonics(SiPh),LPO/LRO,andNPO/CPOprimarilytargethigherperformance,lower

energyconsumptionandcosts.Webelieve1.6TupgradeandSiPhmigrationarethe

keydriversin2026F,anddominantplayerssuchasZhongjiInnoLight(300308CH,

Buy)willcontinuetodominatethehigh-endmarket,thankstoitsstrongR&Dand

productionpower,aswellaseffectivesupplychainmanagement.Weestimate800G/1.6Ttransceivershipmentswillgrowfrom20mn/2.5mnunitsin2025Fto43mn/20mn

unitsin2026F,whileSiPhtransceiverswillgain50-60%/60-70%shareinthe

800G/1.6Tsegmentsin2026F.WethinkNPO/CPOtechnologieswillcontinueto

improve,butyettobecomemainstreamsolutionsoverthenext2-3yearsinthescale-outnetwork.Meanwhile,wethinkinvestorsneedtobemindfulifleadingCPOsolution

providersusebundledsalesstrategiesfortheirCPOswitchproducts,whichcould

accelerateadoptionratesinscale-outnetworkinthemedium-term,whilewebelievetherearemoreCPOusecasesinthescale-upnetworkinthelong-term.Toaddresstheopticalchip(EML,CWandVcsel)shortage,wenoteChineseplayersi.e.YuanjieTechnology(688498CH,Notrated;aCWlasermaker),SourcePhotonics(unlisted;anEML

chipmaker)areproactivelytappingintothissegmentwhichiscurrentlydominatedby

globalplayers.Whilethesenewentrantscouldhelptoeasethecomponentshortage

issue,webelievecompetitionintheglobalmarket,whichislikelytointensify,isalsoworthmonitoring.

High-speedcopper:replacedbyoptics?Stillhassignificantroomtogrowinscale-

upnetworks

Wenotethereisagrowingnarrativeinthecapitalmarketthatopticalcommunicationwillreplacetraditionalcopper-basedconnections.However,wethinkcopperwillcontinuetoplayanimportantroleinscale-upandpartiallyscale-outnetworks,givenitsspeedandefficiencyadvantages,whicharecrucialforprocessinglarge-scaleAltasks.Despite

transmissionlimits,coppercableshavecostadvantagesvsoptics,andcertaintypesofcopperproducts,i.e.,AEC(ActiveElectronicCable),canextendthereachtoaslongas10meters.AccordingtomarketresearchfirmLightCounting,thesalesofhigh-speed

ResearchAnalysts

ChinaTechnology

BingDuan-NIHK

bing.duan1@+85222522141

EthanZhang-NIHK

ethan.zhang@+85222522157

Nomura|GlobalAlnetworkingsupercycle9January2026

3

cablescoulddoubleoverthenextfiveyears,reachingUSD2.8bnby2028E.Nvidia(NVDAUS,Notrated)isanearlyadopterofDAC(DirectAttachCable)intheBlackwellNVL72system,andthecompanywilllikelycontinuetousecoppercablesinthe

upcomingRubinplatformin2026.OtherlargeAlcustomerssuchasAmazonAWS(AMZNUS,Notrated),Meta(METAUS,Notrated)andMicrosoft(MSFTUS,Notrated)havealsobeenusingcoppercablesintheirAIDCprojects.Inthecoppercablesupplychain,thekeyplayersincludeAmphenol(APHUS,Notrated),Credo(CRDOUS,

Notrated),aswellasBizlink(3665TT,Buy,coveredbyKennyChen),andWoer

(002130CH,Notrated),whicharealreadyintheglobalAlcustomers'coppercablesupplychain.Meanwhile,wenoteanintegrationtrend,asopticalcommunications

playerssuchasInnoLight(300308CH,Buy)andEoptolink(300502CH,Notrated)havebeentryingtoenterthismarket.

Alswitches:CPOswitchpenetrationhasstartedtoincrease;OCShasbecomea

populartrend;"whitebox"productscontinuetogaintraction

Ethernetandilnfiniband(IB)arethetwomajornetworkprotocolsintoday'sAldata

centers,whicharecompetingwitheachothertoprovidehigherspeedandlowerlatencyinlargeAltrainingclusters.Thankstoimprovedperformance,lowercostsandlower

powerconsumption,wethinkBroadcom(AVGOUS,Notrated)willlikelycontinueto

pushforCPOadoptionbasedonitsTomaHawk(TH)6platform.Meanwhile,Nvidia'sQuantum-X(IB)andSpectrum-X(Ethernet)switchesmayalsolaunchCPOversions,andwethinkNvidia'sGTC2026eventinMarch2026couldshedmorelightonthese

products.Moreover,wenotethatGoogle'sOCS(opticalcircuitswitch)isgaining

popularitythankstoitsstrongperformanceinlargeTPUtrainingclusters,whileother

playerssuchasLumentum(LITEUS,Notrated)andCoherent(COHRUS,Notrated)havebeenproactivelyengagedinthemarket.Wealsonotethegrowingtrendofusing

"whitebox"switchproductsbyglobalCSPs,inordertoreducecostandincreaseflexibility.Inthemeantime,thistrendwilllikelyhaveanegativeimpactonthemargintrendsof

brandedswitchplayers,whichhavetheirownIPandsoftwareintegratedintheirhardwareproducts.

Stocksforaction:UpgradeYOFCtoBuyonbetterAIDCdemandandstabilized

telecomoutlook;accumulateInnoLightasthekeybeneficiaryof1.6TandSiPh

migration,andTFCandT&ScommunicationaspotentialCPOplays

WebelieveChina'sleadingopticalcomponent/transceivercompanieswillcontinueto

benefitfromtheglobalAIDCinvestmentupcycle,andthespeedupgrade(800Gto1.6T),aswellastechnologymigration(frompluggabletoSiPhtoCPO).WemaintainourBuyratingonInnoLight(300308CH),oursectortoppick,aswethinktheglobalNo.1DCtransceiversupplier(source:LightCounting)willbenefitfromthe1.6TtransceiverandSiPhproductupcyclein2026-2027F,anditsvaluationlooksattractivetous.WealsolikeTFC(300394CH,Buy)asabeneficiaryofNvidia'sdemandforhigh-endopticaltransceiverproductsandasapotentialCPObeneficiary.Inaddition,welike

ShenzhenT&SCommunications(300570CH,Buy)asacoresuppliertoCorning

(GLWUS,Notrated),akeyplayerinUSAIDCs,whichprovidesscale-upandscale-outopticalconnections.WeupgradeYOFC(6869HK)toBuyfromNeutral,aswethinkthecompany'sAIDCbusiness(includingHollowCoreFiber,ActiveOpticalCable,Active

ElectronicCable,andMPO)expansionwouldaccelerateintheglobalmarket,whileChina'stelecomnetworkoutlookwilllikelystabilize.

Fig.1:Stocksforaction

Company

Tickar

Ratng

MtCap

{USDmn)

AvgTO

(USDmn)P

Target

rice(LC)

Price

Asof8Jan

UpsideCompanyDescipbion

(%)

NV'srevenueexposure%inFY25F

OherglobalAtech

playersrevenueexposure

%inFY25F

ZhonginnoLight

300308CH

Buy

95,604

2.547

799.001

595.45

34.2%GlobalNo.1DCopticaltransceiversupplier

25%~30%

50%~60%

YOFC

6869HK

Buy

12,727

142

64.501

50.40

28.0%OneofChina'sleadingopticalfibermakersHigh-endopticalcomponentmaker(optical

N.A

10~20%

SuzhouTFC

300394CH

Buy

21,742

1,026

243.00↑

195.25

24.5%engine,FAU)

50%~60%

10~20%

T8SCommunication

300570CH

Buy

3.74

260

148.001

115.99

High-endMPOproductmaker,andkeypartner27.6%toCorning

0~10%

60%~70%

Source:BloombergFinanceL.P.,Nomuraestimates

Nomura

GlobalAlnetworkingsupercycle9January2026

4

Alnetworkingtechnologyroadmap-whatto

expectin2026(and2027onwards)

WebelievetheAlinfrainvestmentupcycleforglobalhyperscaleAlcloudcompanieswillcontinuein2026-27F,thankstotheintensecompetitionforlargelanguagemodel(LLM)trainingandinference.GPUandASICplayersarestrivingtogainmoremarketshareintheLLMera,whiletheirshortenedtechnologyupgradecycleswillaccelerateproductandtechnologyinnovationintheAlnetworkingsegment,inourview.

Aldatacentersneedtoexpandtheircomputingandinterconnectcapabilitiesatdifferentlevelstomeettheever-expandingdemandfromLLMtrainingandinference.Therearethreedifferentlevelsforexpandingcomputingclusters:scale-up,scale-out,andscale-acrosS.

Scale-up:Scale-upnetworkisconnectingmultipleGPUsintoasinglecomputenode.

Scale-upfocusesonwithin-rackinterconnection,requiringextremelyhighbandwidthandlowlatencyconnections.Thechallengesofscale-upincludeadramaticincreaseinpowerconsumptionandthetrade-offbetweenbandwidth,speed,andtransmissiondistance.

Currently,intra-racksystemsprimarilyutilizecopperinterconnects(NVLinkasan

example).Thebottleneckwithcopperliesinthelimitedeffectivetransmissiondistanceofcoppercablesathighspeeds(e.g.,448G),whichislessthan10meters.However,theproductroadmapsofglobalAlleadersindicateever-increasingsizeofscale-upnetworksandGPUclusters,whichwilllikelyexceedcopper'slimit,andnewsolutions,i.e.,opticalcommunicationswillneedtobeintroduced,inourview.ByintegratingopticalI/O(e.g.,siliconphotonictransceivers)withinthechippackage,chipbeachfrontdensitycanbe

significantlyimproved,breakingthroughthephysicallimitsofelectricalI/O,inourview.

Scale-out:Scale-outinvolvesbuildinglargedatacenterAlclusters(typicallywithinthe

sameserverroom/campus)byaddingmoreservernodesandinterconnectingthemviahigh-speednetworks.Scale-outfocusesonrack/clusternetworking,requiringhigh

bandwidthdensityswitchingarchitectures,andlow-latencycommunication.Forscale-upnetworks,opticalinterconnectsarethemainstreamsolutions,with400G/800G/1.6T

opticaltransceiversconnectingserversviaToR(Top-of-Rack)switches,orextendingthereachthroughspine/coreswitches.AsAlclustersexpandandoptical/switchingchipscapacitiesgrowrapidly,thepowerconsumptionandcostsalsoincreasedramatically.Todealwiththischallenge,co-packagedoptics(CPOs)hascomeintofocusforthescale-outnetwork.BydirectlyintegratingtheopticalenginewiththeswitchingASIC,thetotal

bandwidthandportdensityoftheswitchcanbesignificantlyincreased,whilereducingpowerconsumptionperbit.

Scale-across(inter-datacenterexpansion):Scale-acrossinterconnectsmultipledatacentersscatteredacrossdifferentregionsintoalarger-scaleAlfactory,enablingthemtocollaborativelyrunasingleAltask.Scale-acrossrequireswide-area/campusoptical

interconnects,requiringultra-highbandwidthandextremelylowtransmissionlatencyjitteroverlongdistances,suchashollow-corefiberandcoherentopticaltransceivers.

GlobalAlnetworkingsupercycle9January2026

Fig.2:AIDCnetworkingarchitecture:scale-upvsscale-outvsscale-across

SCALEUP

SCALEOUT

SCALEACROSS

GPUGPU

GPU

GPUGPU

GPU

GPU

GPU

CLUSTERS

AlFACTORY

Source:CUBEResearch,Nomuraresearch

Scale-upstandards:NVLinkvsUALinkvsSUE

Inthescale-upcamp,themajorAltechplayershaveself-developedorcustomized

interconnectstandards/protocols,suchasNvidia'sNVLink/NVSwitch,Broadcom'sSUE,AWS'sNeuronLink,Huawei'sUBSwitch,andAlibabaCloud'sALink;inaddition,thereisUALink,whichisbeingdevelopedbyagroupofcompaniesandorganizations,targetingatprovidingindustryopenstandard.Inthisenvironmentofincreasedcompetition,wenotethatkeyplayerswerestrivingtotakealeadingrolebyexpandingtheirecosystems.Forinstance,NvidiaisopeningupitsproprietaryNVLinkprotocolas"NVLinkFusion"and

authorizethird-partyvendorstobuildscale-upsolutionsbasedon“NVLinkFusion”.

Fig.3:Comparisonsamongmainstreamscale-uplinks

Standards

RoCEv2

UECscale-up

SUE

NueronLink

NVLink

PCle

UALink

UB

OISA1.1

Organization

Ethernet

UltraEthernet

Consortium

Broadcomand

others

AWS

NVIDIA

CXL

UALinkAlliance

Huawei

Chinamobileandothers

Usescenario

Server/GPU

interconnection;

Symetric

interconnection

Server/GPU

interconnection;Heterogeneous

interconnection

Server/GPU

interconnection;Heterogeneousinterconnection

Server/GPU

interconnection;Heterogeneous

interconnection

GPU

interconnection;

Symetric

interconnection

Server/GPU

interconnection;symetric/Heterogeneousinterconnection

Server/GPU

interconnection;Heterogeneous

interconnection

Server/GPU

interconnectionsymetric/Heterog

eneous

interconnection

GPU

interconnection

Symetric

interconnection

Memorysemantics

Notsupport

N/A

Support

Support

Support

Support

Support

Support

Support

Flowcontrolmanagement

PFC+ECN

LLR+CBFC

BasedonCredit

N/A

BasedonCredit

BasedonCredit

BasedonCredit

N/A

CBFC/PFC

SerDesbandwidth

112GT/s;224GT/s

112GT/s;224GT/s

112GT/s;224GT/s

112GT/s;224GT/s

112GT/s;224GT/s

32GT/s;64GT/s

112GT/s;224GT/s

N/A

112GT/s

Delay

Microseconds

Hundredsofns

Hundredsofns

Microseconds

Microseconds

Hundredsofns

Hundredsofns

Microseconds

Hundredsofns

Energyefficiency

<100pi/bit

<100pi/bit

N/A

N/A

<10pj/bit

<10pj/bit

<10pi/bit

N/A

<10pi/bit

Interconnection

distance

WithinDC

WithinDC

Multi-racks

Multi-racks

Multi-racks

WithinDC

Multi-racks

Multi-racks

Withinracks

Source:Companydata,Nomuraresearch

Nvidia'sNVLink/NVLinkFusion

NVLinkisaninterconnecttechnologypurposelydesignedbyNvidiaforhigh-performancecomputing,aimingtoenablehigh-speeddataexchangebetweenGPUsorbetweena

GPUandaCPU.Sinceitsdebutin2016,NVLinkhasbeenupgradedtoitssixth

generationnow,tomeetgrowingcomputingdemand.NVLinkhasbecomeanadvancedtechnologysolutionandakeypillarofNvidia'sdominanceintheAlchipindustrythankstothelowlatency,highefficiency,andhighbandwidth.AccordingtoNvidia'sroadmap,

NVLink6willbelaunchedin2026,achievingacommunicationrateof3.6TB/sand

interconnecting576GPUs.NVLink7willbelaunchedin2027,andNVLink8in2028,tofurtherimprovebandwidth,reducelatency,andoptimizetheflexibilityofinterconnect

topologies.Additionally,NvidialaunchedNVLinkFusionatComputex2025,aimingtoopentheNVLinkecosystemtothird-partyCPUandacceleratormanufacturers.This

involvesreleasingIPandhardwaretodriveinteroperabilitybetweenthird-partydesigns

Nomura

GlobalAlnetworkingsupercycle9January2026

6

andNvidiachips.Itallowscustomchips(includingCPUsandXPUs)tointegratewithNvidia'sNVLinkScale-upnetworktechnologyandrack-scaleexpansionarchitecture,enablingsemi-customAlinfrastructuredeployments.CurrentNVLinkFusionpartnersincludeAlchip,AsteraLabs,Marvell,MediaTek,Fujitsu,Qualcomm,Cadence,and

Synopsys.

Fig.4:NVLinkhistoricalroadmap

Version

Year

GPUArchitecture

Bandwidth/link

(Bi.,GB/s)

No.oflink

Total

Bandwidth

(GB/s)

Highlight

NVLink1.0

2016

Pascal

40

4

160

SeveraltimesfasterthanPCle3.0×16(~32GB/s)

NVLink2.0

2017

Volta

50

6

300

IntroducedNVSwitchtosupportfullyconnectedtopology

NVLink3.0

2020

Ampere

50

12

600

MetthetrainingrequirementsofLLMs

NVLink4.0

2022

Hopper

100

18

900

PAM4encodingisused

NVLink5.0

2024

Blackwell

200

18

1800

Supportultra-large-scaleAlclusters

NVLinkFusion

2025

Blackwell

200

18

1800

Opentheecosystemtothirdpartiesandsupportheterogeneouschip

NVLink6.0

2026

Rubin

400

36

3600

SerDeschannelspeedupgradedto400G

NVLink7.0

2027

RubinUltra

400

36

3600

?

Source:Companydata,Nomuraresearch

Ethernetforscale-up:UALinkandSUE

EthernetwasinitiallydevelopedbyRobertMetcalfein1973whileheworkedintheXeroxPaloAltoResearchCenter,andiscurrentlythemostwidely-usednetworkinterconnectionprotocolandthedefactoindustrystandard(maintainedbyIEEEorganization).The

Ethernetswitchesandopticaltransceivershaveamatureindustrysupplychain,which

hashelpedtobuildlarge-scaleon-premiseDCsandcloudcomputingnetworksoverthepastfewdecades,andisnowexpandingtoAlclusters.WhileEthernetwasprimarily

developedforscale-outnetwork,UALinkhasbeendevelopedasanopenprotocolfor

scale-upnetwork,whichiscompatiblewithEthernet,tocompetewithNvidia'sNVLink.

UALinkisanopenindustryinterconnectstandardreleasedbytheUALinkConsortium

includingAMD,AWS,Broadcom,Cisco,andGoogle.AsascalableAlFabric,UALinkcanbedeployedforAltrainingandinferencesolutionstosupportawiderangeofAlmodels.ForLLMtraining,UALinkwillenablescalabledomainstoreachhundredsofGPUs.TheUALink1.0specificationlaunchedinApril2025supportsadatatransferrateofupto

200GTb/sperchannel,enabling200Gextendedconnectivityperchannelforupto1,024acceleratorswithinacomputingpod,withlatencybelow1microsecond.Itfeatureslowpowerconsumption,strongEthernetcompatibility,andhighsecurityandmanageability.Currently,thereisnocommercializedsolutionsbasedonUALinkprotocol,butthekeyenablerssuchasAMD,Intel,AsteraLabs,Marvell,andSynospsysareactively

developingproductsandsolutions.Forexample,AMD'sAlall-in-onemachine,"Helios,"usestheUALoE(UALinkoverEthernet)solution,whichutilizestheEthernetphysical

layerandswitchingchipsattheunderlyingleveltocarryUALinkprotocoldata.

ToprovideanoptimizedsolutionforAlscale-upnetworkwhilealsoremainingcompatiblewithEthernetprotocol,BroadcomhaslaunchedtheScale-upEthernet(SUE)architecture,emphasizingopenness,compatibility,andlowlatencyperformance.UnliketheUALink

whichisstillinthecommercializationprocess,Broadcomhasalreadylaunched

commercialproductsbasedonSUE(tocompetewithNVLink),whichincludethe

Tomahawk6serieschipslaunchedinJune2025andtheTomahawkUltrachiplaunchedinJuly2025.Moreover,theTomahawk6offersaunifiedsolutionthatsupportsboth

scale-outandscale-up.

PCleforscale-up

PCleisahigh-speedstandardusedtoconnecthardwarecomponentsinsidecomputers,andcommonlyusedtoconnectGPUswithCPUs,Wifiandnetworkcards.PCle

(PeripheralComponentInterconnectExpress)wasinitiallydevelopedbyIntelandagroupofothercompanies,andnowit'sanindustryprotocolmanagedbythePCI-SIG

organization.PCle-7.0wasreleasedin2025.PCle7.0bandwidthdoubledcomparedtoPCle6.0,increasingthetransmissionrateperlaneto128GB/s,withabidirectional

bandwidthofupto512GB/sforx16lanesandasingle-lanebandwidthofapproximately16GB/s.Notably,PCle7.0isdevelopedbyPCI-SIGorganizationtointroduceoptical

connectivitytoenhancelong-distancetransmissionperformance.

9January2026

GlobalAl

Nomura

networkingsupercycle

7

Fig.5:PClelinkperformance

Version

Year

Linecode

Transferrate

(perlane)

Throughput(GB/s)

x1

x2

x4

x8

x16

PCle1.0

2003

NRZ

2.5GT/s

0.25

0.5

1

2

4

PCle2.0

2007

5.0GT/s

0.5

1

2

4

8

PCle3.0

2010

8.0GT/s

0.985

1.969

3.938

7.877

15.754

PCle4.0

2017

16.0GT/s

1.969

3.938

7.877

15.754

31.508

PCle5.0

2019

32.0GT/s

3.938

7.877

15.125

31.508

63.015

PCle6.0

2022

PAM-4

FEC

64.0GT/s

7.563

15.125

30.25

60.5

121

PCle7.0

2025

128.0GT/s

15.125

30.25

60.5

121

242

PCle8.0

2028

(expected)

256.0GT/s

30.25

60.5

121

242

484

Source:Nomuraresearch

Scale-outstandards:InfinibandvsEthernetvsOCS

Inthescale-outnetwork,RDMAisacorenetworktechnologythatbypassestheCPUandperformsdatatransferdirectlyatthememorylevel,allowingcomputerstodirectlyaccessthememoryofremotecomputers.RDMAtechnologycurrentlyhastwomain

implementationmethods:InfiniBandandRoCE(RDMAoverConvergedEthernet).

InfiniBandisprimarilysupportedbyNvidia,andhasbeenwidelyadoptedinlarge-scaleAlclusters,asitsupportssubmicrosecondlatencyandintrinsiclosslessnetworking(basedonacreditsignalcontrolmechanism),whichcansignificantlyreducethecommunicationlatencyofGPUclustersinAltraining.Meanwhile,Ethernetasanindustrystandard

communicationprotocolhasastrongcompetitiveedge,thankstostrongcompatibility,anextensiveecosystem,andrelativelylowcost,althoughitisnotpurposelybuiltforAI.TocompetewithNvidia'sInfiniBandnetworkintheAlnetworks,theUEC(UltraEthernet

Consortium)wasfoundedinJuly2023.TheUECprotocolsupportsmultipath

transmissionandmicrosecondlatency,whichissuitableforAltrainingscenarios.The

transportlayerdefinedbyUECcompletelyoverhaulscurrentRoCEtechnology,

optimizingcongestioncontrolandmultipathtransmission.Itisakeytechnologyforsolvingcongestionandunevenloadproblemscausedbylarge-scaleclusters,inourview.InJune2025,UECreleasedUECSpecification1.0,providinghigh-performance,scalable,and

interoperablesolutionsforalllayersofthenetworkstack(includingNICs,switches,fiberoptics,andcabling),therebyenablingseamlessmulti-vendorintegration.

Fig.6:InfinibandandEthernetmarketsizeforecasts

InfiniBandandEthernetMarketSizeForecast

inUSDBillions

2021202320242025(E)2029(日)

Revenue

Note圖indicstosEsirmated.Souce650Croup

Source:650Group,Nomuraresearch

Nomura|GlobalAlnetworkingsupercycle9January2026

8

Scale-acrossnetworks:buildinglargeAlfactorythroughDCI

InAugust2025,Nvidiaproposedthescale-acrossnetwork,integratingdistributeddata

centersintogigawatt-levelAlsuperfactoriesthroughcross-regionalinterconnection,

becomingthethirdpathforexpandingAlcomputingpower.Meanwhile,NvidialaunchedtheSpectrum-XGSEthernetsolution,whichcombinesphotonicstechnologyand

advancedroutingalgorithmstoautomaticallyadjustcongestioncontrolbasedonthe

distancebetweendatacenters,achievingnear-localcommunicationperformancein

geographicallydistributedclusters.Atthescale-acrosslevel,opticalcommunication

demonstratesaclearadvantage,forexample,usingcoherentDWDMopticaltransmissiontechnologytoprovide800Goreven1.6Tsingle-wavelengthcapacityoverdistances

rangingfromtenstohundredsofkilometers.Therefore,webelievethatlong-distance

opticalmodules(coherentopticalmodules),hollow-coreopticalfibers,andotheropticalcommunicationequipmentanddeviceswilllikelyplayimportantrolesinthescale-acrossdomain.

NomuraGlobalAlnetworkingsupercycle9January2026

9

GlobalCSPs'networkingroadmapsNvidia'sscale-upnetwork

TheGB200NVL72utilizesNVLink5.0andPC