Ain2030Extrapolatina current trendsThis Epoch Al report was commissioned by Google DeepMind. All points ofviews and conclusions expressed are those of the authors and do nonecessarilv reflect the position or endorsement of Gooale DeepMindAlin2030|EpochAl
Table of ContentsExecutivesummaryIntroductionScalingandcapabilitiesScaleComputeInvestmentDataHardwareEnergy and the environmentInterlude:Fromscale tocapabilitiesCapabilitiesHow capabilities are deplovedSoftwareengineeringMathematicsMolecularbiologyWeatherpredictionsDiscussionandconclusionAppendix:Al'spotential toreduceGHG emissionsAppendix:benchmarkextrapolationdetallAlin2030|EpochAl9
Executive summaryHow wil a dvanced Al be developed, and what willitseffects be in the worldat large? What will happen if current trends in scaling up Al developmentpersistallthe wayto2030?This report examines what this scale-up wouldinvolve in terms of comoute. investment. data, hardware, and energy. Weexplore the role of compute across inference and training. the promise oeconomic value that would be necessary to justify such investment, andpotentialchallengesindata availabilitvandenerav,BasedonthesepredictionsforhowAlwillbedeveloped.we turn to oredictscience isthe explicit gpoal of severalleading Al developers, and is likely tobe among the top prioritiesforAldeployment. Scientific R&Dprovides avaluablelensforunderstandinawhatadvancedAlwillachieveCombutescalina has plaved akevrole in Al development, and willikelycontinuetodoso.Computefortraining and inferencedrivesimorovementsin Al capabilities, and much progress in Al research has come fromdeveloninaceneral-ournosemethodsto enable theuseof more comnuteThetraiectoryofAldevelopment canbeforecasted basedoncontinuedcomputescaling.Scalinghassianificantimolications acrossmanvareas ofAI development: training and inference compute, investment, datal hardware, and energy. When we predict that compute scaling will continue,we can then examine the consequences within each of these-and howthey need to scale accordingly to allow compute scaling trends to continueExponentialgrowth willielycontinue to2030 acrossallkey trends,Across trainina and inference compute. investment. data. hardware. andenerav. we araue that acontinuationofexictina tronde ic faacihlaWeand discussinathe most credible reasonsfor slowdown or accelerationbeforethen. We arquethe mostcredible reasonsforadeviation from trendare changes in societal coordination of Al development (e.g. investorsentiment or tight requlation), supply bottlenecks for Al clusters (e.a.chipsAlin2030|EpochAl
oreneravl).orparadiamaticshiftsin Al production le.a.substantialR&Iautomation).On current trends, the largest Al models of 2030 will require investmentsof hundreds of bilions of dollars, and 1,000x the compute of today'slaraestmodels.Investment of this scale ispotentiallv iustifiedifAl canautomate significanttasksinthe economy.The present trendof3xannualAllab revenue growth would lead to revenues exceeding hundreds ofbillions of dollars before 2030. Findina data...
