2026物理AI与机器人未来研究报告:塑造下一代硬件技术的驱动力_第1页
2026物理AI与机器人未来研究报告:塑造下一代硬件技术的驱动力_第2页
2026物理AI与机器人未来研究报告:塑造下一代硬件技术的驱动力_第3页
2026物理AI与机器人未来研究报告:塑造下一代硬件技术的驱动力_第4页
2026物理AI与机器人未来研究报告:塑造下一代硬件技术的驱动力_第5页
已阅读5页,还剩23页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

PhysicalAIandthe

FutureofRobotics

Alookattheforcesshapingthenextgenerationofhardwaretechnology,fromsilicontosteel.

Featuringawarehouseautomationspotlightpoweredby:

July2026

Contents

3LetterFromtheAuthor

5Macro

8ChipsandCompute

12FundraisingandInvestment

16Spotlight:WarehouseAutomation

svb》SBanck。nvaley

ADivisionofFirstcitizensBank

PHYSICALAIANDTHEFUTUREOFROBOTICS20262

svb》SBanck。nvaley

ADivisionofFirstcitizensBank

PHYSICALAIANDTHEFUTUREOFROBOTICS20263

EverythingIsHardwareEverywhere,AllatOnce

Whatmakesthishardwarecycledifferentfromthe

lastisintelligence.Forthefirsttime,robotshavethecapabilitytogeneralizeinnovelways,operatingin

unfamiliarenvironments.ThatshiftintophysicalAIisexpandingthe

addressablemarketforrobotsinunprecedentedways.”

SiliconValleyisre-embracingsilicon.Beforetheappstore,chipsanddevicesweren’tthesideshow;theywerethemainevent.Cisco,IntelandQualcommwerethemostvaluablecompaniesintheworldat

theturnofthemillennium.Fromtheearly2000sthroughtheearly2020’shardwarewassidelined.Itrepresentedlessthanone-fifthofallUSventurecapital(VC)investment.

VCsthisyearareonpacetodeploy$120B,roughlyone-thirdofallinvestment,intoUScompaniesdevelopingphysicaltechnologies.

OneinfiveglobalVCfirmsinvestingtodayhaveatleast10%ofits2026investmentsinUShardwarecompanies.

Whatmakesthishardwarecycledifferentfromthelastis

intelligence.Forthefirsttime,robotshavethecapabilityto

generalizeinnovelways,operatinginunfamiliarenvironments.ThatshiftintophysicalAIisexpandingtheaddressablemarketforrobotsinunprecedentedways.Virtuallyeverywarehouse,factoryand

logisticshubmaysoonbewithinreachofahumanoidhand—atleastifcostsaren’tanissue.

OneofthebiggestopportunitiesforphysicalAI’sdisruptivepotentialisakeyopportunity:supplychains.Moderntechsupplylinesarea

bewilderingmeshofcrisscrossingsuppliersfromaroundtheglobe.Noproductshowsthisbetterthanamodernacceleratorchip.

Just10countriesprovidenearlyallofthecomponentpartsneeded

formodernAIchips.AndwhileTaiwanrightfullyreceivesthefocusasalynchpinproducerofchips,otherAsiancountriesareequally

critical.Outofmorethan100primarysuppliersthatwecounted,57%comefromJapan.

Industrialsettingssuchasfactoriesandwarehousesareanideal

placeforrobotstoprovetheirmettle.Tounderstandhowlogisticsfacilitiesareadoptingandexperimentingwithautomation,we

partneredwithPrologisVentures—thecorporateinnovationarmoftheworld’slargestowneroflogisticsandindustrialrealestate.Withmorethan1.3billionsquarefeetofwarehousespaceunder

management,Prologishasaclearviewoftheinnovationsshapingthewarehousefloor.Inasurveyofroughly400Prologiscustomers,83%haveyettoautomateanyoftheiroperations.Industrial

automationisagreenfield—amassivemarketopportunity.

Executivesinthespaceofferedskepticismaboutthepotential

benefitsofbigswingslikehumanoids.Thegapbetweenthis

skepticismandthetechnology’strajectoryreflectstheuphillbattleroboticscompanieshavefaced.Whilethiscycleofinnovationisnodifferentinintermsofchallenges,theacceleratedcomputebehindthemachinesis.Companiesthatcanimprovetheeconomicsmayhavetheupperhandinawide-opennewmarket.

MarkHarris

HeadofFrontierandClimateTechnologySiliconValleyBank

mark.harris@

ClickThroughtoKeyTakeaways

Tradedisruptionshavestrainedsupplychainsandpushedinflationtoafour-yearhigh.

JumptoPage

ImportsareshiftingawayfromChinaandtowardfriendlier,chip-ladenmarkets.

JumptoPage

Over90%ofhigh-bandwidthmemory(HBM)is

consumedbyfourcompaniesasshortagegrows.

JumptoPage

USVCinvestmentinedgecomputinganddatacentertechareatarecordpace.

JumptoPage

svb》SBanck。nvaley

ADivisionofFirstcitizensBank

UShardwareVCinvestmentisonpaceto

doublein2026,reachinganewrecordhigh.

JumptoPage

Automationisstillearlyatwarehouses.Amedianof40%oflargewarehousesreportsnoautomation.

JumptoPage

AIadoptionisgreatestamonglargersupplychain

facilities(46%)comparedtosmalloperators(28%).

JumptoPage

Onceimplemented,warehouseautomations

oftenexceedsavingsandupkeepexpectations.

PHYSICALAIANDTHEFUTUREOFROBOTICS20264

JumptoPage

Macro

⇨RETURNTOTABLEOFCONTENTS

PHYSICALAIANDTHEFUTUREOFROBOTICS20265

StraitTalk:Supply

LinesUnderStrain

Iran’sclosureoftheStraitofHormuzisripplingacross

worldwidetradingchannels.TheGlobalSupplyChain

PressureIndex(GSCPI),aUSFederalReservemetricthattracks27factorsimpedingtrade,isatitsmost

constrictedlevelinfouryears.WestTexasIntermediate(WTI)crudeisup70%sinceJanuary,andthecosttoshipgoods—byair,seaorland—isrisingwithit.

Adisruptionintheflowofrawmaterialslike

aluminumandresinsiscreatingadownstream

shortageinfinishedproductsforeverythingfrom

electricvehiclebatteriestofertilizer.Asaresult,pricesarerisingacrosstheboard.USinflationspikedto4.2%inMay,thehighestlevelsince2022.

Thestraittypicallyfacilitates$3.5Bworthofdailytrade,includingone-fifthoftheworld’spetroleumproducts.

Withoutthatsupply,manycountrieshavetappedinto

strategicenergyreserves,whichareabundantinwealthycountriesbutnotendless.TheclosureposesriskstothesupplyofsemiconductorsthatispoweringtheAIboom.

Japan,aprimarysupplierofchipinputs,sourcesover

90%ofitscrudeoilfromtheStraitandhasbeendrawingfromitsenergystoressinceMarch.1Taiwanismost

exposedthroughitsrelianceonLNGimports—criticalforpoweringfabs.Powerprovidershavebidaggressivelyin

thespotmarkettosecureLNG,whichhaskeptreservesintact,butatacostoftwicethepriorratepercargo.1

svb》SBianck。nvaley

ADivisionofFirstcitizensBank

SupplyChainsFeelMostPressureinFourYearsGlobalSupplyChainPressureIndex

GSCPI

(StandardDeviationsFromAverage)

5

4

3

2

1

0

-1

-2

-3

-4

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

Recalibration:

PeakBuying:

COVID-19Disruption:Globallockdowns

impedeflowofgoods.

Surgingdemandforconsumergoods

outrunssupply.

Supplypressuresstaynear

RelativeBalance:

thehistoricalaverage.

Spendingsubsidesand

supplypressureloosens.

IranWar:

Strikesonoil

refineriesandthe

closureoftheStraitofHormuzhave

slowedtheflowofgoodsandcausedaspikeinoilprices.

HoisttheSales:ShippingCostsRiseonOilSpike

USProducerPriceIndex:ShippingCostsvs.OilPrices

ShortonCircuitBoards

ProducersPriceIndex(Jan2019=100)

AllCommoditiesBarePrintedCircuitBoards

Trucking

Freight

Crude

Oil(WTI)

Air

Freight

DeepSeaFreight

180

170

160

150

140

130

120

110

100

90

PriceChange

SinceDec2025

+17%

CircuitBoards

+12%

AllCommodities

6

ShippingCostIndex

225

200

175

150

125

100

20192020202120222023202420252026

$140

$120

$100

$80

$60

$40

$20

$0

CrudeOilPrice

2019

2020

2021

2022

2023

2024

2025

2026

Notes:1)AccordingtoreportingfromS&PGlobalandothercommoditynewsoutlets.

Source:S&PGlobal,Reuters,USFederalReserve,USBureauofLaborStatisticsandSVBanalysis.

PHYSICALAIANDTHEFUTUREOFROBOTICS2026

ShiftingTradeWinds

BringChipsAshore

Arethereshoredsectorsintheroomwithus?IntheyearssincethetradebottlenecksofCOVID-19,theterms

“nearshoringandreshoring”haveneverstrayedfarfrompublicview.Googlesearchesforthetermshitsevenall-timehighsinthelastthreeyears,thelatestinJanuary.

Judgingbytherhetoric,youmightexpecthundredsofforeignfactoriestohavedecampedtoplaceslike

AlabamaorFlorida.Butthefactoryopeningshavenotfollowedthesearchtraffic.

Instead,whatwe’reseeingisashiftawayfromChinaandtowardother(mostly)distantmarkets.China’sshareofUSimportshasdropped9percentagepointssince

2019,whileplaceslikeTaiwan(achipmanufacturerfuelingthedatacenterboom),Vietnam(arising

computermaker),Ireland(apharmaceuticalshub)

andMexico(anautoassemblycenter)aredominatingtheirownspecializedpocketsofcommerce.

Onenotableexceptionissemiconductormanufacturing.

ConstructionofUSmanufacturingfacilities(primarilyforcomputerchips)hasjumped6xinthelastfouryears.Yet,thisinvestmentisstillyearsfrombearingfruit.For

example,outofsixfabplantsthatTSMChasannouncedforArizonasince2020,threehavebrokengroundandonlyoneisinproduction.Untilmoresuchplantscomeonline,theUSwillremainanetimporterofadvancedtech.

svb》SBianck。nvaley

ADivisionofFirstcitizensBank

Friendshoring:TradeShiftsAwayfromChinaLargestSwingsinImportShareforTopUSTradePartnersandProducts(2019-2025)1

PercentagePointChangeinImportSharebyCountry

AdvancedTechStillComesFromOverseasBalanceofUSTradeforAdvancedTechnologyProducts2

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

$60B

$40B

$20B

Exports

$0B

Imports

-$20B

-$40B

-$60B

-$80B

-$100B

-9.0

-1.6

-1.5

-0.8

-0.6

-0.5

China

Canada

Japan

Russia

UK

Germany

Taiwan

Vietnam

Ireland

Mexico

Thailand

Switzerland

3.7

3.0

1.4

1.4

1.3

1.3

HowTSMCIsInvesting$165BinArizonaTimelineofTSMCInvestmentsinPhoenixSince2020

HalfaTrillionDollarsonFabsandDataCentersUSConstructionSpendingonComputerandChipManufacturingandDataCenters(12-monthtotal)

ChipManufacturingDataCenters

Producing

Announced

Fab1

(4nm3Chips)

---$65B

$100B

$120B

$90B

$60B

$30B

$0B

Since2022:

$390B

ChipPlants

$120B

DataCenters

Fab2

(3nmchips)

Fab3G-●

(2nmchips)

Fabs4-6,R&D,

PackagingFacilities

2020

2021

2022

2023

2024

2025

2026

2027

2028

2029

2015

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

Notes:1)Trademeasuredindollars.2)Censusdefinitionsofadvancedtechnologiesinclude500commoditycodes10core

groups:biotechnology,lifescience,optoelectronics,informationandcommunications,electronics,flexiblemanufacturing,advancedmaterials,aerospace,weaponsandnucleartechnology.3)Nanometer.

7

PHYSICALAIANDTHEFUTUREOFROBOTICS2026

Source:GoogleTrends,USCensusBureauandSVBanalysis.

ChipsandCompute

⇨RETURNTOTABLEOFCONTENTS

PHYSICALAIANDTHEFUTUREOFROBOTICS20268

AIChipsWeaveaGlobalTapestryofSupplyLines

FlowofMajorAIChipSuppliersbyManufacturingCountryandInputComponent1

ComponentGroup(Suppliers)

Components(Suppliers)

PrimaryManufacturingLocation(Suppliers)

RDLMaterial:8

UpstreamInputs:40

Photoresist:4

Underfill:5

Japansupplies57%ofinputsforsomeAIacceleratorchips.

HBMMoldCompound:3

SubstrateCoreMaterials:4

SolderMask:3

Japan:68

HeatSpreader:3

Lid/ModuleCover:3

C4Bump:3

Non-ConductiveFilm:4

Microbump:4

GlassFiber:4

LogicDie:26

CopperMaterials:2

WaferHandling:3

CleaningChemicals:3

Build-UpFilm:2

SiliconWafer:5

CMPSlurries:3

PackageSubstrate:5

TestSockets:2

TSVMaterials:7 ThermalInterface:4

US:19

HighBandwidthMemory:17

Taiwan:10

HBMTesting:2

CopperFoil:2

CMPPads:2

SouthKorea:8

Germany:7

Belgium:1

France:1

UK:1

HBMMemoryStack:5

SolderBall:5

BarrierMetalTargets:3

PackageTesting:2

SpecialtyGases:5

CoWoS2

Packaging:36

HBMandCoWoS2

packagingarethe

primarychokepointsforcurrentchip

production.

QuartzComponents:2

GraphiteComponents:2

SiliconInterposer:4 LogicFoundry:3

Austria:1

Singapore:1

Notes:1)AnalysisbasedonSVBreviewofprimarymanufacturinglocationsformorethan100suppliersacrosstheAIacceleratorsupplychain,including

upstreammaterials,logicfabrication,HBM,andCoWoSadvancedpackaging.Suppliercountsreflectprimarymanufacturinglocation.RDLisredistributionlayer.C4iscontrolledcollapsechipconnection.CMPischemicalmechanicalplanarization.TSVisthroughsiliconvia.2)Chip-on-wafer-on-substrate.

PHYSICALAIANDTHEFUTUREOFROBOTICS20269

Source:SVBanalysis.

2025

HungryHungry

Chippos

ThesupplychainforAIchipsisperhapsthemost

complicatedandcriticalofanyproducttoday.InsatiabledemandforcomputefromtherapidgrowthofAI

inferenceispushingcurrentdatainfrastructuretoits

limits.Advancedchipmakershaveadded13gigawatts(GW)ofannualsemiconductorpowertothemarket

since2022,withNVIDIAaccountingfortwo-thirdsofthatsupply.ThecumulativedrawfromthoseAIchipsisequivalenttoacountrythesizeoftheNetherlands.Morethan830datacentersareunderwayintheUS.1Major

marketslikeDallasandNorthernVirginiahavelessthan1%vacancyatexistingfacilities.Powerprojectsarelininguptomeettheexpecteddemand.Themedianwaitforaproposedprojecttobeapprovedforconnectiontothe

gridisnow45months,up36%fromtwoyearsago.

Beyondthepowerneeds,thechipsrequireacomplicatedwebofcomponentsthatcomefromaroundtheworld,yettherearesomevulnerablebottlenecks.WhileTaiwanis

theepicenteroffabrication,ouranalysismorethan100

suppliersfoundthat57%ofinputsforAIaccelerators

comefromJapan.ManufacturingseveralkeycomponentssuchasphotoresistsandHBMmolds,Japanhasanear

monopolyoncertainparts.ThatchokepointcouldslowthegrowthofAI-enableddevicesastheycompetewithdatacentersforscarcegoods.Thetopfourchipmakerscurrentlyconsume90%ofHBM.

svb》SBianck。nvaley

ADivisionofFirstcitizensBank

ChipsAddPowerDemandtotheGrid

CumulativeChipSalessince2022byTotalChipPower(GW)2

NVIDIAGoogleAMDHuaweiAmazon

TopFourChipmakersControl90%ofKeyInputsGlobalShareofSemiconductorComponentsConsumedin20253

NVIDIAGoogleAMDAmazonOther

14

12

10

8

6

4

2

0

8.67.0

5.74.7

3.7

1.81.01.21.5

0.6

2.2

2.8

0.8

2.2

1.6

0.4

Q1

Q2Q3

2024

Q4

Q1

Q2Q3

2025

Q4

CoWoS

Packaging

HighBandwidthMemory(HBM)

AdvancedLogicDies

9%

60%

69%

88%

10%

8%

0%20%40%60%80%100%

ApprovalTimesStretchforNewPowerProjectsMedianMonthsFromInterconnectionRequesttoAgreement4

TopDataCenterMarketsHave<1%VacancyColliers:DataCenterVacancyRatesbyUSMarket

2010

2011

2012

2013

2014

2015

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2020

2025

60

45

30

15

0

9622061

MedianMiddle50%

PowerCapacity

inQueue(GW)

SaltLakeCityReno,NVNorVA

Columbus

Austin

Dallas

Phoenix

Atlanta

NewYork

ChicagoNorCal

Houston

LosAngeles

0.2%

0.4%

0.4%

0.4%

0.6%

0.8%

1.1%

1.1%

2.0%

2.1%

3.5%

4.0%

5.3%

Notes:1)AccordingtoBloombergNEF.2)CumulativeannualpowerdrawequivalentofchipssoldaccordingtocalculationsbyEpochAI.AMDisAdvancedMicroDevices.3)EstimatesanddefinitionsfromEpochAI.4)Interconnectionrequestsarerequiredforpowerplantstosellpowertothegridbutarenotrequiredforplantsusingpowerforprivateprojects.

10

Source:Colliers,BloombergNEF,EpochAI,LawrenceBerkeleyNationalLabandSVBanalysis.PHYSICALAIANDTHEFUTUREOFROBOTICS2026

HavingYourAIandUsingIt,Too

IfAIisbecomingthenucleusofcomputing,tokensaretheatomicunitofAIcost.Yettheeconomicsofinference

differdramaticallydependingonhowyoudeployit.Cloudinferenceforknowledgeworkisgettingbothsmarterandcheaper,whichisdrivingadoptioninwhite-collarjobs.

HavingAIwriteanemailcostsabout1,000tokens,or

aboutfivecents.Summarizinga60-minuteteam

meeting?That’s100ktokens,or$5.Dollarcostsfor

individualtaskslikethesearealreadydropping

quickly.Thechallengeishowtostringthesetasksintofullyautomatedroles.

AgenticAIcandothat,butatasteeptokenmultiplierthatishardertoquantify;agentsuseupto1,000xmoretokensthansimplequeries.1GoldmanSachsestimatesagenticworkflowswilldrivea24xincreaseintokenconsumptionby2030.DatacentersrunningAIreasoningmodels

dependonHBM,whichisinshortsupply.Onegroup

competingforHBMisAI-embodiedmachines,which

mustcontendwithanotherfactor:latency.Real-worldAIneedshighmemoryitcan’toutsourcetothecloud.Lag

thatistolerableinachatbotmightbecatastrophicforahumanoid,makingHBMshortagesasupplychain

constraintforrobots.FoundersinwarehouseautomationtodaymustthereforenavigatebothtokenpricingandthebottleneckofsupplyforHBM.

svb》SBianck。nvaley

ADivisionofFirstcitizensBank

InferenceWillBecometheWayWeComputeWhatOfficeWorkCanYouAffordToInfer?

EstimatedGlobalDataCenterWorkloadsbyUseType3EstimatedTokenCostsforVariousInferenceTasks2

wTraditionalWorkloadsAITrainingwAIInference

Task

Tokens

2026Cost

Draftanintroductoryemail

~1K

$0.05

Review100linesofcode

~10K

$0.50

Summarizea5K-wordarticle

~25K

$1.25

Comparetwo50contracts

~100K

$5

Summarizea90-minutevideolecture

~180K

$9

Analyzea10-KFiling(Agentic)

~300K

$15

Route100CustomerTickets(Agentic)

~350K

$17

Screen200Candidates(Agentic)

~800K

$40

37%

ofdatacenterworkloadswillbeforAIinference

by2030.

9%

14%

13%

77%

50%

37%

20252030

EventheWorstAIModelsArePrettySmartNowPerformanceofAIModelsontheEpochCapabilitiesIndex

AIModelsAreGettingCheaper,butUnevenly

AIApplicationsbyApproximateMemoryandLatencyRequirements

160

140

120

100

80

60

40

USDper1milliontokens

40

35

30

25

20

15

10

5

0

2022202320242025

2023202420252026

Notes:1)StanfordDigitalEconomyLab“HowareAIagentsspendingyourtokens?”2)Estimatesbasedonblendedinferencerateoffivecentsper1,000tokensandpublishedbenchmarksoftokencostandperformance.3)Colliersestimate.

PHYSICALAIANDTHEFUTUREOFROBOTICS202611

Source:Colliers,EpochAIandSVBanalysis.

Fundraisingand

Investment

⇨RETURNTOTABLEOFCONTENTS

PHYSICALAIANDTHEFUTUREOFROBOTICS202612

WhatHaveYouDone

forMe,Latency?

Asedgecomputeandcloudinferencebothscale,theyarecompetingforthesameconstrainedsupplychains.

Yet,theeconomicsdifferinkeyways.

MachinesrunningphysicalAIhavestrictlatency

requirementscomparedtoinference.Adriverlesscarora

pick-and-packrobotmakingsplit-seconddecisionscan’t

toleratethesamelagthatisitacceptableforAIchatbots.Asaresult,thesedevicesmustcarrytheircomputehardwareonboard.Ratherthanbenefitingfromtheeconomicsof

tokenpricingthroughdatacenters,next-generationAIdevicesrequirethesameexpensivecomponentsthatarebeingchanneledtodatacenters.Thisstrainsthe

economicsandusecasesforadvancedrobots.

Asrobotsmovetowardgeneralizedtasks,theirlatency

andmemoryrequirementsgrow.Machinescapableofthekindofdexterityneededforassembly,forexample,

requireashiftfromtraditionalDRAMmemorytoHBM.

VC-backedcompaniesaresteppinginwithspecializedproductsforedgecompute.CompanieslikeHailoand

Rebellionsaremanufacturinginferencechipspurpose-builtforon-devicephysicalAI.Thespaceisearlyinits

evolution.Asthesechipsmatureandproliferate,the

economicsofon-deviceinferencecouldfollowthesamedeflationarycurvethatreshapedcloud-computing.

svb》SBianck。nvaley

ADivisionofFirstcitizensBank

Edgevs.Cloud:TwoAIPathsDivergingintheWorkingWorldAIApplicationsbyApproximateMemoryandLatencyRequirements1

CloudInferenceEdgeCompute

HighLatencyToleranceLow

PhysicalAIcompaniesfaceaconstraintthatcloudAI

companiesdonot:Theycannotaffordtolag.Toavoidcriticalglitches,deviceslikehumanoidrobotsmustcarrytheir

computewiththem.Companiesarerespondingwithhardwaretoimprovetheperformanceandeconomicsofedgecompute.USVCinvestmentinedgecomputeisontrackfor$3.4B

across100dealsin2026,whichwouldbothbeall-timehighs.

Traditional

HighBandwidth

Memory

Memory

Pick/Pack

Humanoids

Assembly

Robots

AVs

Robots

On-PremCompute

Video

Monitoring

CloudCompute

LLM

Inference

LowMemoryBandwidthRequiredHigh

Cloudinferenceapplicationslikedocumentprocessingcantoleratelatency,butthechallengeisaboutvolume.

Stringingtasksintoautonomousworkflowsconsumesmuchmorecomputethansimplequeries.Investorsaredeployingrecordamountsofcapitalintocompaniesmakingcomputemoreefficientfordatacenters.USVCinvestmentintodatacentercompaniesisonpacetoeclipse$12Bin2026.

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

USVCInvestmentinEdgeComputeCompanies

DealsCapital

$4B

$3B

10196100

89

8078768171

65

$2B

$1B

$0B

USVCInvestmentinDataCenterCompanies DealsCapital

283

$15B

232

$12B

181

158

149

138

$9B

114

128127113

$6B

$3B

$0B

Notes:1)VCinvestmentinedgecomputingincludescompanieswithkeywordsintheirdescriptionthatareassociatedwithedgecomputingsuchasedgecomputing,edgeai,aiofthings,embeddedai,neuromorphiccomputing,neuromorphicchips,andedgedevices.VCindatacentercompanieskeywordssuchas:datacenterinfrastructure,aiinfrastructure,cloud

computinginfrastructure,gpucloud,liquidcooling,anddatacenter.2)DRAMisdynamicrandom-accessmemory

PHYSICALAIANDTHEFUTUREOFROBOTICS202613

Source:PitchBookData,Inc.andSVBanalysis.

EarningsCall,

AutomationAnswers

VCsrecognizethecomingwaveofreal-worldautomation,andtheyareshiftingtheirinvestmentstomatch.In2022,just9%ofVCinvestorshadatleast10%oftheirdealsinUShardwarecompanies;in2026,thathasjumpedto22%.ManyVC-backedsoftwarefoundersworryabout

obsolescenceinthefaceofAIdisruption.Becauseofits

physicalIP,hardwareismoreresilienttodisruption.Infact,itismadepowerfulbynewAImodelswithpotentialto

improvehardwaregeneralizability.

Amidlaborshortagesandoperationalpressure,executivesincreasinglyseeautomationasacoregrowthstrategy.

Mentionsofautomationoncorporateearningscallshave

jumpedthreefoldsince2020.AIimplementationanda

broaderfocusoncorporateefficiencyhaveonlysharpenedthefocusonautomation.Formostcompanies,automationmeansnotjustimplementingAIsoftware,butalso

implementingphysicaltechnologiesinmanufacturing,

warehouses,andtransportation.VCinvestmentin

hardware-focusedcompaniesthisyearhasalready

eclipsedthefull-yeartotalin2025andisonpaceto

reach$120Bbyyear-end.Excludinginvestmentsin

AnthropicandOpenAI,hardwarecompaniesaccountfor

nearlyathirdofVCinvestmentthisyear.Oneareaof

particularinterestisdefensetech.Inasignofthe

enthusiasmforgovernmentapplications,defense-orientedcompaniesarecommandinga31%premiumtoall

hardwarecompanies.

svb》SBianck。nvaley

ADivisionofFirstcitizensBank

2010

2011

2012

2013

2014

2015

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

2015

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

100in2010

Corporatefocusonautomation

outstripsinvestorfocusafter

COVID-19laborandsupplychain

disruptionsbegin.

Indexedto

CorporateAmericaObsessedWithAutomationAutomationMentionsinEarningsCallsandbyFormerly

VC-BackedCompanies

EarningsCallMentionsofAutomation

VCDealsintoCompaniesthatMentionAutomation

1600

1400

1200

1000

800

600

400

200

0

HardwareInvestmentReachesNewHeights

USVCInvestmentinHardwareCompaniesandShareofTotalVC

Capital

Year-End

%ofVCTotal

Extrapolation

(excludingAnthropic

andOpenAI)

30%

$160B

$120B

$80B

$40B

$0B

17%16%20%17%19%18%15%16%18%20%21%

2015

2016

2017

2018

2019

2020

2021

2022

2023

2024

2025

2026

InvestorsTurntoHardware

PercentageofGlobalVCInvestorsWithatLeast10%ofDealsinUSHardwareCompanies1

22%

15%

13%11%

9%9%9%9%

20192020202120222023202420252026

Dual-Use?VCsPayaPremiumforDefenseTech.Pre-MoneyValuationPremium/DiscountforUSVC-BackedDefenseTechCompaniesComparedtoUSVC-BackedHardwareCompanies

72%

31%

16%

11%

22%17%24%18%

6%

-3%

-17%

-23%

Notes:1)Numeratorisinvestorswithatleast10%ofVCdealsinUShardwarecompanies;denominatorisinvestorswithatleast10%ofVCdealsinUScompanies.

14

Source:PitchBookData,Inc.,CBInsightsandSVBanalysis.PHYSICALAIANDTHEFUTUREOFROBOTICS2026

Let’sGetPhysical

AI

Oneofthecorechallengeswithautomatingwarehousesisthatrobotshavehistoricallybeendevelopedfora

specifictask.Often,thesetasksarehighlystandardizedwithfewobstacles(smoothfloors,consistentparts,etc.).Inmanywarehouses,thissimplyisn’tthereality.As

physicalAIcommercializes,thepromiseismodels

thatallowrobotstogeneralizeandperformawide

rangeoftasksinavarietyofenvironments.Thisnot

onlydrasticallyincreasesthepotentialtoautomatemoreofthesupplychain,butitalsoincreasesthetotal

addressablemarketthatroboticscompaniescanserve,makingthemmoreattractivetoventureinvestors.

ThescalinglawsthatunlockedLLMslikeClaudeandChatGPTcatapultedusintotheAIage.Butthose

scalinglawsdon'tapplyinthesamewaytophysicalAI.PhysicalAIfaceshardconstraintsli

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论