全球 AI 推理部署的經濟性、影響和替代路徑.pdf

《全球 AI 推理部署的經濟性、影響和替代路徑.pdf》由會員分享，可在線閱讀，更多相關《全球 AI 推理部署的經濟性、影響和替代路徑.pdf（43頁珍藏版）》請在三個皮匠報告上搜索。

1、AI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe Gigawatt GambleThe Economics,Impact,and an Alternative Path for Global AI Inference DeploymentThomas SohmersCEO,Positron AIthomaspositron.aiAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsIntroductionIntroductionForbes 3

2、0 under 302013 Thiel FellowMIT ResearcherCEO&Co-FounderDirector of Technology StrategyPrincipal Hardware ArchitectCEO&Co-FounderStudent&EntrepreneurThomas SohmersCEO,Positron AIthomaspositron.aiAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsAre we in an AI bubble?A brief economics dis

3、cussionUnderstanding the coming economic upheavalWhat are the paths to the future?1 12 23 3Three things we will cover Three things we will cover AI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsBubble andBubble andEconomiesEconomiesAI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiand

4、systemsAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsAre we in a bubble?AI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsMy thesisMy thesisThe economic value of AI isnt transient or a ramp to to a point of required capacity,like the build out of the internet and wired,an

5、d wireless connectivity,or the adoption social media.These technology trends had predictable end-limits.The value of AI fundamentally changes the fundamentals on which the global economy is structured:land,capital&labor.AI will turn labor into a near limitless resource-injecting an exponential facto

6、r in the basic equations of wealth creation.This in turn will consume capital at unprecedented levels,and will rapidly test the limits of the land(e.g.power&water).This is not a bubble;This is not a bubble;its the leading indicator of the exponential potential of free labor.its the leading indicator

7、 of the exponential potential of free labor.(pssstand the hyperscalers know this.)AI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsAI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiandsystems1776177620262026AI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiandsystems1776177620262026

8、LandLandFarmsData CentersAI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiandsystems1776177620262026LandLandFarmsData CentersLaborLaborHumansMachinesAI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiandsystems1776177620262026LandLandFarmsData CentersLaborLaborHumansMachinesCapitalCapitalSteam E

9、ngineGPUsAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe Factors of The Factors of Production:LandProduction:LandIn 1776,lands value was what could be directly farmed from it,leading to expansion.500MW Google Datacenter(source:Semianalysis)TSMC Arizona Fab 21(source:TSMC)550MW Dese

10、rt Sunlight Solar Farm(source:TIME)Like land expansion,Data Like land expansion,Data Centers are the embodiment Centers are the embodiment of new wealth creation.of new wealth creation.New Data centers are being built now but also have limits to expansion,particularly access to power and water.AI Ha

11、rdware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe Factors of The Factors of Production:CapitalProduction:CapitalAI Infrastructure spend is expected to hit over$200B$200Bper year per year by 2030,with total spend between 2024 and 2030 approaching$1T.$1T.The Manhattan Project,between 1942 a

12、nd 1946 spent$32B$32B in 2024 dollarsThe Interstate Highway System,between 1956 and 1992,spent$600B$600B in 2024 dollars.And for some reason analysts And for some reason analysts expect direct ROIexpect direct ROIAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe Factors of Production

13、:LaborThe Factors of Production:LaborBusiness will very soon Business will very soon have the ability to to spin up have the ability to to spin up 100K expert employees on 100K expert employees on demand.demand.It cant be just be the biggest and It cant be just be the biggest and wealthiest companie

14、s who can do this!wealthiest companies who can do this!“Labour was the first price,the original purchase-money that was paid for all things.It was not by gold or by silver,but by labour,that all wealth of the world was originally purchased.”Adam SmithImage Source:LifeArchitect.ai/gpt-4-5AI Hardware&

15、SystemsAI Hardware&SystemsaiandsystemsaiandsystemsEconomics LessonEconomics Lesson(Simplified)Cobb-Douglas Production Function:Output(or Wealth)is created using two main factors:Capital(machines,investments)and Labor.Q=A*(K*L)OutputTechnology Productivity FactorCapitalLaborAI Hardware&SystemsAI Hard

16、ware&SystemsaiandsystemsaiandsystemsOur own variantOur own variantOutput/WWealth=T Technology*(R Resources*C Capital*L Labor)W=T*(R*C*L)W=T*(R*C*L)Land/Natural ResourcesAssets(Equipment)Humans(.Machines?)AI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe Previous Industrial Revolution

17、sThe Previous Industrial RevolutionsEarly Mechanized Production+Steam Power1780-18801880-19801980-20202020Electrification+Mass Transportation+Long Distance CommunicationW=T2*(3 3R R*2C*1L)W=T1*(1R*1C*2 2L L)W=T3*(3R*3 3C C*2 2L L)W=T4*(xR*xC*xL Ln n)Increase in labor through specialization+mechaniza

18、tionIncrease in resources through mass extractionDigitization+Fiat CurrencyEmergence of Artificial Intelligence1st Industrial 1st Industrial RevolutionRevolution2nd Industrial 2nd Industrial RevolutionRevolution3rd Industrial 3rd Industrial RevolutionRevolution4th Industrial 4th Industrial Revolutio

19、nRevolutionLabour becomes exponential,effectively limitless.Use of capital re-aligns to the value of information.Labour shifts to knowledge.AI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiandsystems1st order effects we can see in front of us today:today:AI based skilled labour;AI based skilled l

20、abour;AI writing codeAI writing codePower consumptionPower consumptionAI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiandsystems4.5%of the worlds total powerAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe The Consolidation Consolidation ProblemProblemData source:Semianalysis80%

21、7%23.5GW28.1GWAI Specific ComputeAll other computeAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe Unpriced ExternalitiesThe Unpriced Externalities20242030H100 Equivalents Deployed2.25M135MGigawatts Deployed8.5144TWatt/hour Consumed731160Million Metric tons of CO251810160 million ho

22、mestotal number of homes in US+CAN91 billion gallons of gas8 months of US gasoline usage400 billion gallons of water for coolingequivalent to 4 million households2030 Projected Power 2030 Projected Power Usage EquivalentUsage EquivalentGlobal Datacenter Global Datacenter DeploymentsDeploymentsAI Har

23、dware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsWhy it is totally justifiedWhy it is totally justifiedLlama 3.1 405B can currently replace junior engineers Llama 3.1 405B can currently replace junior engineers and/orand/oraugment senior engineers at 30 tokens per second on 8xH100saugment sen

24、ior engineers at 30 tokens per second on 8xH100sGeneration:30 tokens/sec=2.5 million tokens/day=75 million tokens/month=900 million tokens a yearAssume 10:1 input to output with$3/Million Tokens and it only costs a company$30,000/yr for a“Llama employee”$30,000/yr for a“Llama employee”AI Hardware&Sy

25、stemsAI Hardware&SystemsaiandsystemsaiandsystemsSo what isAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsPositron FactsPositron FactsFounded April 2023Goal=Change the underlying economics of applied AI,starting immediately with making inference affordable to more people$12M in seed fu

26、nding18 months from idea to shipping production hardware18 months from idea to shipping production hardware21 employeesWere hiringAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsArithmetic IntensityArithmetic IntensityAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsWhat m

27、akes Transformers different?What makes Transformers different?AI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsNVIDIA Source data:DatabricksPositron high level memory storyPositron high level memory storyAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsNVIDIA Source data:Da

28、tabricksPositron high level memory storyPositron high level memory storyAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsPositron high level memory storyPositron high level memory storyAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsMore density per watt=more economic valu

29、eMore density per watt=more economic value10K Watt Data Center Rack LimitPositronNVIDIADGX H100PositronPositronPositronPositron5,900W1,850Wunusable6x more applied AI per 10KW footprint the costAtlas Server(=1 SW engineer of applied AI)AI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsPos

30、itron shipping productPositron shipping productAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsSpecial ThanksSpecial Thanksto our investorsto our investorsAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsCome talk to us at Booth 35/36Come talk to us at Booth 35/36andandPro

31、duct Demo Stage 4:15 todayProduct Demo Stage 4:15 todayAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsBackupAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsEdit Master title style Edit Master text styles Secon

32、d level Third level Fourth level Fifth levelAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe Previous Industrial RevolutionsFirst Industrial RevolutionFirst Industrial Revolution(Early Mechanized Production+Steam Power)1780-18801880-19801980-20202020LandLaborCapital+Second Industria

33、l RevolutionSecond Industrial Revolution(Mass Transportation+Long Distance Communication)+land limits reached;size,environmental etc,no more being created,resources constraints over timeLabor goes expenetialtipping point where labour can grow indefinitelybiggest capital outlay so farAI Hardware&Syst

34、emsAI Hardware&SystemsaiandsystemsaiandsystemsAI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsThe 2028 Problemcagrcagr20282028202820282024202320222021AI Hardware&SystemsAI Hardware&SystemsaiandsystemsaiandsystemsAI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiandsystems28.3GW51.6GW83.4GW202420262028The 2028 ProblemAI Hardware&SystemsAI Hardware&Systemsaiandsystemsaiandsystems