1、BONDMay 2025Trends Artificial IntelligenceTrends Artificial Intelligence(AI)May 30,2025Mary Meeker/Jay Simons/Daegwon Chae/Alexander Krey2ContextWe set out to compile foundational trends related to AI.A starting collection of several disparate datapoints turned into this beast.As soon as we updated

2、one chart,we often had to update another a data game of whack-a-molea pattern that shows no sign of stoppingand will grow more complex as competitionamong tech incumbents,emerging attackers and sovereigns accelerates.Vint Cerf,one of the Founders of the Internet,said in 1999,they say a year in the I

3、nternet business is like a dog year equivalent to seven years in a regular persons life.At the time,the pace of change catalyzed by the internet was unprecedented.Consider now that AI user and usage trending is ramping materially fasterand the machines can outpace us.The pace and scope of change rel

4、ated to the artificial intelligence technology evolution is indeed unprecedented,as supported by the data.This document is filled with user,usage and revenue charts that go up-and-to-the-rightoften supported by spending charts that also go up-and-to-the right.Creators/bettors/consumers are taking ad

5、vantage of global internet rails that are accessible to 5.5B citizens viaconnected devices;ever-growing digital datasets that have been in the making for over three decades;breakthrough large language models(LLMs)that in effect found freedom with the November 2022 launch of OpenAIs ChatGPT with its

6、extremely easy-to-use/speedy user interface.In addition,relatively new AI company founders have been especially aggressive about innovation/product releases/investments/acquisitions/cash burn and capital raises.At the same time,more traditional tech companies(often with founder involvement)have incr

7、easingly directed more of their hefty free cash flows toward AI in efforts to drive growth and fend off attackers.And global competition especially related to China and USA tech developments is acute.The outline for our document is on the next page,followed by eleven charts that help illustrate obse

8、rvations that follow.We hope this compilation adds to the discussion of the breadth of change at play technical/financial/social/physical/geopolitical.No doubt,people(and machines)will improve on the points as we all aim to adapt to this evolving journeyas knowledge and its distribution get leveled

9、up rapidly in new ways.Special thanks to Grant Watson and Keeyan Sanjasaz and BOND colleagues who helped steer ideas and bring this report to life.And,to the many friends and technology builders who helped,directly or via your work,and are driving technology forward.Seem Like Change Happening Faster

10、 Than Ever?Yes,It IsAI User+Usage+CapEx Growth=UnprecedentedAI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage RisingAI Usage+Cost+Loss Growth=UnprecedentedAI Monetization Threats=Rising Competition+Open-Source Momentum+Chinas RiseAI&Physi

11、cal World Ramps=Fast+Data-DrivenGlobal Internet User Ramps Powered by AI from Get-Go=Growth We Have Not Seen Likes of BeforeAI&Work Evolution=Real+Rapid3123456789-5152-128129-152153-247248-298299-307308-322#323-336OutlineWeekly Active Users,MM4Charts Paint Thousands of WordsSeem Like Change Happ

12、ening Faster Than Ever?Yes,It IsAI User+Usage+CapEx Growth=UnprecedentedDevelopers in Leading Chipmakers Ecosystem12.1Source:Leading ChipmakerDetails on Page 38AI User+Usage+CapEx Growth=Unprecedented2.2Internet vs.Leading USA-Based LLM:Total Current Users Outside North AmericaNote:LLM data is for m

13、onthly active mobile app users.App not available in select countries,including China and Russia,as of 5/25.Source:United Nations/International Telecommunications Union(3/25),Sensor Tower(5/25)0Years InShare of Total Current Users,%Details on Page 56AI User+Usage+CapEx Growth=UnprecedentedLeading USA

14、-Based LLM Users2Source:Company disclosures Details on Page 556MM20052025Number of Developers,MM0%50%100%InternetLLM33Years In90%Year 390%Year 2310/224/25800MMBig Six*USA Technology Company CapEx*Apple,NVIDIA,Microsoft,Alphabet,Amazon(AWS only),&Meta PlatformsSource:Capital IQ(3/25),Morgan Stanl

15、ey20142024CapEx,$B+63%$212BDetails on Page 975Charts Paint Thousands of WordsAI Monetization Threats=Rising Competition+Open-Source Momentum+Chinas Rise5Leading USA LLMs vs.China LLMDesktop User ShareNote:Data is non-deduped.Share is relative,measured across six leading global LLMs.Source:YipitData(

16、5/25)Desktop User Share,%2/242/254/2575%60%10%21%15%0%Details on Page 293USA LLM#1ChinaUSA LLM#2AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage Rising 3Cost of Key Technologies Relative to Launch Year%of Original Price By Year(Indexed to Ye

17、ar 0)Note:Per-token inference costs shown.Source:Richard Hirsh;John McCallum;OpenAIDetails on Page 1380 Years72 Years Electric PowerComputer MemoryAI InferenceAI Monetization Threats=Rising Competition+Open-Source Momentum+Chinas Rise5.1China vs.USA vs.Rest of World Industrial Robots InstalledNote:D

18、ata as of 2023.Source:International Federation of RoboticsIndustrial Robots InstalledDetails on Page 289AI Usage+Cost+Loss Growth=Unprecedented4Leading USA-Based AI LLM Revenue vs.Compute ExpenseNote:Figures are estimates.Source:The Information,public estimates20222024Revenue(Blue)&Compute Expen

19、se(Red)+$3.7B-$5BDetails on Page 1732023ChinaRest of World(excl.China&USA)USA201420236Charts Paint Thousands of WordsAI&Physical World Ramps=Fast+Data-Driven6A Ride Share vs.Autonomous Taxi Provider,San Francisco Operating Zone Market ShareSource:YipitData(4/25)Global Internet User Ramps Pow

20、ered by AI from Get-Go=Growth We Have Not Seen Likes of Before7Leading USA-Based LLM App Users by RegionNote:Region definitions per World Bank definitions.China not included in East Asia figures.Data for standalone app only.Source:Sensor Tower(5/25)5/234/25Mobile AppMonthly Active Users,MMDetails on

21、 Page 315AI&Work Evolution=Real+Rapid8USA IT Jobs AI vs.Non-AIDetails on Page 302+448%-9%1/184/25Source:University of Marylands UMD-LinkUp AIMaps(in collaboration with Outrigger Group)(5/25)Change in USA IT Job Postings,Indexed to 1/18(AI=Blue,Non-AI=Green)Details on Page 33227%19%8/234/25%of Sa

22、n FranciscoGross Bookings34%0%East Asia&PacificSub-Saharan AfricaSouth AsiaNorth AmericaMiddle East&North AfricaLatin America&CaribbeanEurope&Central AsiaRide ShareAutonomous TaxiNon-AI IT JobsAI IT Jobs7OverviewTo say the world is changing at unprecedented rates is an understatement

23、.Rapid and transformative technology innovation/adoption represent key underpinnings of these changes.As does leadership evolution for the global powers.Googles founding mission(1998)was to organize the worlds information and make it universally accessible and useful.Alibabas founding mission(1999)w

24、as to make it easy to do business anywhere.Facebooks founding mission(2004)was to give people the power to share and make the world more open and connected.Fast forward to today with the worlds organized,connected and accessible information being supercharged byartificial intelligence,accelerating c

25、omputing power,and semi-borderless capitalall driving massive change.Sport provides a good analogy for AIs constant improvements.As athletes continue to wow us and break records,their talent is increasingly enhanced by better data/inputs/training.The same is true for businesses,where computers are i

26、ngesting massive datasets to get smarter and more competitive.Breakthroughs in large models,cost-per-token declines,open-source proliferation and chip performance improvementsare making new tech advances increasingly more powerful,accessible,and economically viable.OpenAIs ChatGPT based on user/usag

27、e/monetization metrics is historys biggest overnight success(nine years post-founding).AI usage is surging among consumers,developers,enterprises and governments.And unlike the Internet 1.0 revolution where technology started in the USA and steadily diffused globally ChatGPT hit the world stage all

28、at once,growing in most global regions simultaneously.Meanwhile,platform incumbents and emerging challengers are racing to build and deploy the next layers of AI infrastructure:agentic interfaces,enterprise copilots,real-world autonomous systems,and sovereign models.Rapid advances in artificial inte

29、lligence,compute infrastructure,and global connectivity are fundamentally reshaping howwork gets done,how capital is deployed,and how leadership is defined across both companies and countries.At the same time,we have leadership evolution among the global powers,each of whom is challenging the others

30、competitive and comparative advantage.We see the worlds most powerful countriesrevved up by varying degrees of economic/societal/territorial aspiration 8OverviewIncreasingly,two hefty forces technological and geopolitical are intertwining.Andrew Bosworth(Meta Platforms CTO),on a recent Possible podc

31、ast described thecurrent state of AI as our space race and the people were discussing,especially China,are highly capabletheres very few secrets.And theres just progress.And you want to make sure that youre never behind.The reality is AI leadership could beget geopolitical leadership and not vice-ve

32、rsa.This state of affairs brings tremendous uncertaintyyet it leads us back to one of our favorite quotes Statistically speaking,the world doesnt end that often,from former T.Rowe Price Chairman and CEO Brian Rogers.As investors,we always assume everything can go wrong,but the exciting part is the c

33、onsideration of what can go right.Time and time again,the case for optimism is one of the best bets one can make.The magic of watching AI do your work for you feels like the early days of email and web search technologies that fundamentally changed our world.The better/faster/cheaper impacts of AI s

34、eem just as magical,but even quicker.No doubt,these are also dangerous and uncertain times.But a long-term case for optimism for artificial intelligence is based on the idea that intense competition and innovationincreasingly-accessible computerapidly-rising global adoption of AI-infused technologya

35、nd thoughtful andcalculated leadership can foster sufficient trepidation and respect,that in turn,could lead to Mutually Assured Deterrence.For some,the evolution of AI will create a race to the bottom;for others,it will create a race to the top.The speculative and frenetic forces of capitalism and

36、creative destruction are tectonic.Its undeniable that its game on,especially with the USA and China and the tech powerhouses charging ahead.In this document,we share data/research/benchmarks from third parties that use methodologies they deem to be effective we are thankful for the hard work so many

37、 are doing to illustrate trending during this uniquely dynamic time.Our goal is to add to the discussion.Seem Like Change Happening Faster Than Ever?Yes,It IsAI User+Usage+CapEx Growth=UnprecedentedAI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer

38、Usage RisingAI Usage+Cost+Loss Growth=UnprecedentedAI Monetization Threats=Rising Competition+Open-Source Momentum+Chinas RiseAI&Physical World Ramps=Fast+Data-DrivenGlobal Internet User Ramps Powered by AI from Get-Go=Growth We Have Not Seen Likes of BeforeAI&Work Evolution=Real+Rapid912345

39、678Outline10Technology Compounding=Numbers Behind The Momentum11Technology Compounding Over Thousand-Plus Years=Better+Faster+Cheaper MoreNote:Chart expressed in trillions of real GDP as measured by 2011 GK$on a logarithmic scale.GK$(Gross Knowledge Dollars)is an informal term used to estimate the p

40、otential business value of a specific insight,idea,or proprietary knowledge.It reflects how much that knowledge could be worth if applied effectively,even if it hasnt yet generated revenue.Source:Microsoft,Governing AI:A Blueprint for the Future,Microsoft Report(5/23);Data via Maddison Project&O

41、ur World in DataTechnology Compounding=Numbers Behind The MomentumGlobal GDP Last 1,000+Years,per Maddison Project11001000130012001400150016001700180019002000Printing PressSteam EnginesTelegraphElectrificationMass Steel ProductionMass Production&Assembly LinesInternal Combustion EngineFlightSynt

42、hetic FertilizerTransistorsPCsInternetSmartphonesCloud12Technology Compounding Over Fifty-Plus Years=Better+Faster+Cheaper MoreNote:PC units as of 2000.Desktop internet users as of 2005,installed base as of 2010.Mobile internet units are the installed based of smartphones&tablets in 2020.Cloud&a

43、mp;data center capex includes Google,Amazon,Microsoft,Meta,Alibaba,Apple,IBM,Oracle,Tencent,&Baidu for ten years ending 2022.Tens of billions of units refers to the potential device&user base that could end up using AI technology;this includes smartphones,IOT devices,robotics,etc.Source:Weis

44、s et al.AI Index:Mapping the$4 Trillion Enterprise Impact via Morgan Stanley(10/23)Enabling InfrastructureCPUsBig Data/CloudGPUsComputing Cycles Over Time 1960s-2020s,per Morgan StanleyNote:Axis is logarithmic;i.e.,there are expected to be tens of thousands more AI Era devices than Mainframe devices

45、19601970198019902000201020202030110010,0001,000,000MainframeMinicomputerPCDesktop InternetMobile InternetAI Era1MM+Units10MM+Units300MM+Units1B+Units/Users4B+UnitsTens of Billions of UnitsMM Units in Log ScaleTechnology Compounding=Numbers Behind The Momentum13AI Technology Compounding=Numbers Behin

46、d The Momentum14260%Annual Growth Over Fifteen Years ofData to Train AI Models Led ToNote:Only“notable”language models shown(per Epoch AI,includes state of the art improvement on a recognized benchmark,1K citations,historically relevant,with significant use).Source:Epoch AI(5/25)Training Dataset Siz

47、e(Number of Words)for Key AI Models 1950-2025,per Epoch AITraining Dataset Size Number of Words+260%/YearAI Technology Compounding=Numbers Behind The Momentum15360%Annual Growth Over Fifteen Years ofCompute to Train AI Models Led To*A FLOP(floating point operation)is a basic unit of computation used

48、 to measure processing power,representing a single arithmetic calculation involving decimal numbers.In AI,total FLOPs are often used to estimate the computational cost of training or running a model.Note:Only language models shown(per Epoch AI,includes state of the art improvement on a recognized be

49、nchmark,1K citations,historically relevant,with significant use).Source:Epoch AI(5/25)Training Compute FLOP*Grok 3+360%/YearAI Technology Compounding=Numbers Behind The MomentumTraining Compute(FLOP)for Key AI Models 1950-2025,per Epoch AI16200%Annual Growth Over Nine Years ofCompute Gains from Bett

50、er Algorithms Led ToNote:Estimates how much progress comes from bigger models versus smarter algorithms,based on how much computing power youd need to reach top performance without any improvements.Source:Epoch AI(3/24)Impact of Improved Algorithms on AI Model Performance 2014-2023,per Epoch AIEffec

51、tive Compute(Relative to 2014)+200%/YearAI Technology Compounding=Numbers Behind The Momentum17150%Annual Growth Over Six Years ofPerformance Gains from Better AI Supercomputers Led ToSource:Epoch AI(4/25)AI Technology Compounding=Numbers Behind The MomentumPerformance,16-bit FLOP/s+150%/YearEnabled

52、 by 1.6x annual growth in chips per cluster and 1.6x annual growth in performance per chipPerformance of Leading AI Supercomputers(FLOP/s)2019-2025,per Epoch AI18167%Annual Growth Over Four Years inNumber of Powerful AI Models*As of 4/25,Large-Scale AI Models are generally defined as those with a tr

53、aining compute of 1023FLOPs or greater,per Epoch AI.Source:Epoch AI(5/25)Number of New Large-Scale AI Models(Larger than 1023FLOP*)2017-2024,per Epoch AINumber of New Models Released Each YearAI Technology Compounding=Numbers Behind The Momentum05010020172018201920202021202220232024Includes models f

54、romxAIAnthropicMetaNVIDIAMistralArc Institute&Others+167%/YearBoth models from DeepMind(AlphaGo Zero&Master)Publication Date19ChatGPT AI User+Subscriber+Revenue Growth Ramps=Hard to Match,EverNote:4/25 user count estimate from OpenAI CEO Sam Altmans 4/11/25 TED Talk disclosure.Revenue figure

55、s are estimates based off OpenAI disclosures.Source:OpenAI disclosures(as of 4/25),The Information(4/25)(link,link,link&link)ChatGPT User+Subscriber+Revenue Growth 10/22-4/25,per OpenAI&The InformationAI Technology Compounding=Numbers Behind The MomentumChatGPT Weekly Active Users,MM04008001

56、0/228/236/244/25Users(MM)0102010/228/236/244/25Subscribers(MM)Subscribers,MMRevenue($B)Revenue,$B$0$2$4202220232024Time to 365B Annual Searches=ChatGPT 5.5x Faster vs.Google Note:Dashed-line bars are for years where Google did not disclose annual search volumes.Source:Google public disclosures,OpenA

57、I(12/24).ChatGPT figures are estimates per company disclosures of 1B daily queriesAnnual Searches by Year(B)Since Public Launches of Google&ChatGPT 1998-2025,per Google&OpenAI20AI Technology Compounding=Numbers Behind The MomentumAnnual Searches,BChatGPT Hit 365B Annual Searches in 2 Years(2

58、024)vs.Googles 11 Years(2009)02,5005,000123456789101112131415161718192021222324252627Google SearchChatGPTYears Since Public Launch(Google=9/98,ChatGPT=11/22)21In 1998,tapping emerging Internet access,Google set out toorganize the worlds information and make ituniversally accessible and useful.Nearly

59、 three decades later after some of the fastest change humankind has seen a lot of information is indeed digitized/accessible/useful.The AI-driven evolution of how weaccess and move information is happening much fasterAI is a compounder on internet infrastructure,which allows for wicked-fast adoption

60、 of easy-to-use broad-interest services.AI Technology Compounding=Numbers Behind The Momentum22Knowledge Distribution Evolution=Over Six Centuries23Knowledge Distribution 1440-1992=Static+Physical DeliverySource:Wikimedia CommonsKnowledge Distribution Evolution=Over Six CenturiesPrinting Press Inven

61、ted 144024Knowledge Distribution 1993-2021=Active+Digital Delivery*The internet is widely agreed to have been publicly released in 1993 with release of the World Wide Web(WWW)into the public domain,which allowed users to create websites;however,Tim Berners-Lee invented the World Wide Web in 1989,per

62、 CERN.Source:Google,USA Department of Defense,CERNInternet Public Release 1993*Knowledge Distribution Evolution=Over Six Centuries25Knowledge Distribution 2022+=Active+Digital+Generative Delivery*We define the public launch of ChatGPT in November 2022 as the public release of Generative AI which we

63、see as AIs iPhone Moment.ChatGPT saw the fastest user ramp ever for a standalone product(5 days to secure 1MM users).Generative AI=AI that can create content text,images,audio,or code based on learned patterns.Source:OpenAIGenerative AI Public Launch of ChatGPT 2022*Knowledge Distribution Evolution=

64、Over Six Centuries26Knowledge is a process of piling up facts;wisdom lies in their simplification.Martin H.Fischer,German-born American Physician/Teacher/Author(1879-1962)Knowledge Distribution Evolution=Over Six Centuries27AI=Many Years Before Lift-Off28AI Milestone Timeline 1950-2022,per Stanford

65、University1:AI Winter was a term used by Nils J.Nilsson,the Kumagai Professor of Engineering in computer science at Stanford University,to describe the period during which AI continued to make conceptual progress but could boast no significant practical successes.This subsequently led to a drop in A

66、I interest and funding.Includes data from sources beyond Stanford.Source:Stanford University&Stanford Law School sources,iRobot,TechCrunch,BBC,OpenAI.Data aggregated by BOND.10/50:Alan Turing creates his Turing Test to measure computer intelligence,positing that computers could think like humans

67、6/56:Stanford computer scientist John McCarthy convenes the Dartmouth Conference on Artificial Intelligence,a term he coined1/62:Arthur Samuel,an IBM computer scientist,creates a self-learning program that proves capable of defeating a top USA checkers championAI Winter1(1967-1996)1/66:Stanford rese

68、archers deploy Shakey,the first general-purpose mobile robot that can reason about its own actions5/97:Deep Blue,IBMs chess-playing computer,defeats Garry Kasparov,the world chess champion at the time9/02:Roomba,the first mass-produced autonomous robotic vacuum cleaner that can navigate homes,is lau

69、nched10/05:A Stanford team build a driverless car named Stanley;it completes a 132-mile course,winning the DARPA Grand Challenge4/10:Apple acquires Siri voice assistant&integrates it into iPhone 4S model one year later6/14:Eugene Goostman,a chatbot,passes the Turing Test,with 1/3 of judges belie

70、ving that Eugene is human6/18:OpenAI releases GPT-1,the first of their large language models6/20:OpenAI releases GPT-3,an AI tool for automated conversations;Microsoft exclusively licenses the model11/22:OpenAI releases ChatGPT to the publicAI=Many Years Before Lift-Off29AI Milestone Timeline 2023-2

71、025,per Stanford University*Multimodal=AI that can understand and process multiple data types(e.g.,text,images,audio)together.*Open-source=AI models and tools made publicly available for use,modification,and redistribution.1)4/25 estimate from OpenAI CEO Sam Altmans 4/11/25 TED Talk disclosure.Sourc

72、e:Aggregated by BOND from OpenAI,Microsoft,Google,Anthropic,Meta,Apple,Alibaba,Deepseek,UK Government,US Department of Homeland Security.China data may be subject to informational limitations due to government restrictions.3/23:Microsoft Integrates Copilot into its 365 product suite3/23:Anthropic re

73、leases Claude,its AI assistant focused on safety&inter-pretability3/24:USA Department of Homeland Security unveils its AI Roadmap Strategy5/24:OpenAI releases GPT-4o,which has full multimodality across audio,visual,&text inputs7/24:Apple releases Apple Intelligence,an AI system integrated in

74、to its devices,for developers12/24:OpenAI announces o3,its highest-ever performing model1/25:Alibaba unveils Qwen2.5-Max,which surpasses the performance of other leading models(GPT-4o,Claude 3.5)on some reasoning tests3/23:OpenAI releasesGPT-4,a multimodal*model capable of processing both text&i

75、mages3/23:Google releases Bard,its ChatGPT competitor11/23:28 countries,including USA,EU members&China,sign Bletchley Declaration on AI Safety4/24:Meta Platforms releases its open-source*Llama 3 model with 70B parameters5/24:Google introduces AI overviews to augment its search functions9/24:Alib

76、aba releases 100 open-source Qwen 2.5 models,with performance in line with Western competitors1/25:DeepSeek releases its R1&R1-Zero open-source reasoning models2/25:OpenAI releases GPT-4.5,Anthropic releases Claude 3.7 Sonnet,&xAIreleases Grok 34/25:ChatGPT reaches 800MM weekly users1AI=Many

77、 Years Before Lift-Off30AI=Circa Q2:2531Top Ten Things AI Can Do Today,per ChatGPTAI=Circa Q2:25Source:ChatGPT(5/15/25)32AI=Circa 2030?33Top Ten Things AI Will Likely Do in Five Years,per ChatGPTAI=Circa 2030?Source:ChatGPT(5/15/25)34AI=Circa 2035?35Top Ten Things AI Will Likely Do in Ten Years,per

78、ChatGPTSource:ChatGPT(5/15/25)AI=Circa 2035?36AI Development Trending=Unprecedented37Machine-Learning Model*Trending=In 2015.Industry Surpassed Academia as Data+Compute+Financial Needs Rose*Machine Learning=A subset of AI where machines learn from patterns in data without being explicitly programmed

79、.Note:Academia includes models developed by one or more institutions,including government agencies.Industry-academia collaboration excludes government partnerships and only captures partnerships between academic institutions and industry.Industry excludes models developed in partnership with any ent

80、ity other than another company.Epoch AI,an AI Index data provider,uses the term notable machine learning models to designate particularly influential models within the AI/machine learning ecosystem.Epoch maintains a database of 900 AI models released since the 1950s,selecting entries based on criter

81、ia such as state-of-the-art advancements,historical significance,or high citation rates.Since Epoch manually curates the data,some models considered notable by some may not be included.A count of zero academic models does not mean that no notable models were produced by academic institutions in 2023

82、,but rather that Epoch AI has not identified any as notable.Additionally,academic publications often take longer to gain recognition,as highly cited papers introducing significant architectures may take years to achieve prominence.China data may be subject to informational limitations due to governm

83、ent restrictions.Source:Nestor Maslejet al.,The AI Index 2025 Annual Report,AI Index Steering Committee,Stanford HAI(4/25)2003-2014:Academia Era2015-today:Industry EraGlobal Notable Machine Learning Models by Sector 2003-2024,per Stanford HAIAnnual New Notable Machine-Learning ModelsAI Development T

84、rending=Unprecedented38AI Developer Growth(NVIDIA Ecosystem as Proxy)=+6x to 6MM Developers Over Seven YearsNumber of Developers,MM03620052007200920112013201520172019202120232025Note:We assume negligible developers in NVIDIAs ecosystem in 2005 per this text from an 8/20 blog post titled 2 Million Re

85、gistered Developers,Countless Breakthroughs:It took 13 years to reach 1 million registered developers,and less than two more to reach 2 million.Source:NVIDIA blog posts,press releases,&company overviews+6xAI Development Trending=UnprecedentedGlobal Developers in NVIDIA Ecosystem(MM)2005-2025,Per

86、 NVIDIA39AI Developer Growth(Google Ecosystem as Proxy)=+5x to 7MM Developers Y/YDevelopers Building with Gemini,MMAI Development Trending=UnprecedentedNote:Per Google in 5/25,Over 7 million developers are building with Gemini,five times more than this time last year.Source:Google,Google I/O 2025:Fr

87、om research to reality(5/25)1.4MM7.0MM05101/241/255/245/25+5xEstimated Global Developers in Google Ecosystem(MM)5/24-5/25,Per Google40Computing-Related Patent Grants,USA=ExplodedPost-Netscape IPO(1995).Again+Faster Post-ChatGPT Public Launch(2022)USA Computing-Related*Patents Granted Annually 1960-2

88、024,per USPTO*Uses Cooperative Patent Classification(CPC)code G06,which corresponds to computing,calculating or counting patents.Google patents data changes somewhat between each query so numbers are rounded and should be viewed as directionally accurate.Source:USA Patent&Trademark Office(USPTO)

89、via Google Patents(4/25)05,00010,00015,00019601964196819721976198019841988199219962000200420082012201620202024Number of USA Computing-Related Patents Granted Per Year+6,300more patents granted in 2003 vs.1995(8 years)+1,000more patents granted in 2022 vs.2004(18 years)+6,000more patents granted in 2

90、024 vs.2023(1 year)AI Development Trending=Unprecedented41AI Performance=In 2024Surpassed Human Levels of Accuracy&Realism,per Stanford HAIAI System Performance on MMLU Benchmark Test 2019-2024,per Stanford HAINote:The MMLU(Massive Multitask Language Understanding)benchmark evaluates a language

91、models performance across 57 academic and professional subjects,such as math,law,medicine,and history.It measures both factual recall and reasoning ability,making it a standard for assessing general knowledge and problem-solving in large language models.89.8%is the generally-accepted benchmark for h

92、uman performance.Stats above show average accuracy of top-performing AI models in each calendar year.Source:Papers With Code via Nestor Maslej et al.,The AI Index 2025 Annual Report,AI Index Steering Committee,Stanford HAI(4/25)AI Development Trending=Unprecedented42AI Performance=In Q1:25 73%of Res

93、ponses&Rising Mistaken as Human by TestersNote:The Turing test,introduced in 1950,measures a machines ability to mimic human conversation.In this study,500 participants engaged in a three-party test format,interacting with both a human and an AI.Most discussions leaned on emotional resonance and

94、 day-to-day topics over factual knowledge.Eliza was developed in the mid-1960s by MIT professor Joseph Weizenbaum,It is considered the worlds first chatbot.In January 2025,researchers successfully revived Eliza using its original code.Source:Cameron Jones and Benjamin Bergen,Large Language Models Pa

95、ss the Turing Test(3/25)via UC San Diego%of Testers Who Mistake AI Responses as Human-Generated 3/25,per Cameron Jones/Benjamin BergenDate Released5/241/252/25AI system performance consistently improving over timeAI Development Trending=Unprecedented43AI Performance=Increasingly Realistic Conversati

96、ons Simulating Human BehaviorsTuring Test Conversation with GPT-4.5 3/25,per Cameron Jones/Benjamin BergenSource:Cameron Jones and Benjamin Bergen,Large Language Models Pass the Turing Test(3/25)via UC San DiegoWhat Was Tested:The Turing Test is a concept introduced by Alan Turing in 1950 to evaluat

97、e a machines ability to exhibit intelligent behavior indistinguishable from that of a human.In the test,if a human evaluator cannot reliably tell whether responses are coming from a human or a machine during a conversation,the machine is said to have passed.Here,participants had to guess whether Wit

98、ness A or Witness B was an AI system.Results:The conversation on the left is an example Turing Test carried out in 3/25 using GPT-4.5.During the test,participants incorrectly identified the left image(Witness A)as human with 87%certainty,saying A had human vibes.B had human imitation vibes.A was act

99、ually AI-generated;B was human.AI Development Trending=Unprecedented44AI Performance=Increasingly Realistic Image GenerationNotes:Dates shown are the release dates of each Midjourney model.Source:Midjourney(4/25)&Gold Penguin,How Midjourney Evolved Over Time(Comparing V1 to V6.1 Outputs)(9/24)AI

100、-Generated Image:Womens Necklace with a Sunflower Pendant 2/22-4/25,per Midjourney/Gold PenguinModel v1(2/22)Model v7(4/25)AI Development Trending=Unprecedented45AI Performance=Increasingly Realistic Image GenerationAI-Generated Image(2024)Source:Left StyleGAN2 via The New York Times,Test Yourself:W

101、hich Faces Were Made by A.I.?(1/24);Right Creative CommonsReal ImageAI-Generated vs.Real Image 2024AI Development Trending=UnprecedentedAI Performance=Increasingly Realistic Audio Translation/Generation46Note:China data may be subject to informational limitations due to government restrictions.Sourc

102、e:ElevenLabs(1/24&1/25),Similarweb(5/25)ElevenLabs AI Voice Generator 1/23-4/25,per ElevenLabs&SimilarwebWhen you create a new dubbing project,Dubbing Studio automatically transcribes your content,translates it into the new language,and generates a new audio track in that language.Each speak

103、ers original voice is isolated and cloned before generating the translation to make sure they sound the same in every language.-ElevenLabs Press Release,1/24Global Mobile&Desktop Website Visits,MM010201/23 4/23 7/23 10/23 1/24 4/24 7/24 10/24 1/25 4/25In just two years,ElevenLabs millions of use

104、rs have generated 1,000 years of audio content and the companys tools have been adopted by employees at over 60%of Fortune 500 companies.-ElevenLabs Press Release,1/25AI Development Trending=UnprecedentedElevenLabs Monthly Global Site Visits(MM),per Similarweb 1/23-4/2547AI Performance=Evolving to M

105、ainstream Realistic Audio Translation/GenerationNote:Revenue annualized using Q1:25 results.Source:Spotify,The New York Post,Inside Spotify:CEO Daniel Ek on AI,Free Speech&the Future of Music(5/2/25);Spotify earnings releases;eMarketer,Spotify dominates Apple and Amazon in digital audio(4/25)AI-

106、Powered Audio Translation 5/25,per SpotifyImagine if youre a creator and youre the world expert at somethingbut you happen to be Indonesian.Today,theres a language barrier and it will be very hard if you dont know English to be able to get to a world stage.But with AI,it might be possible in the fut

107、ure where you speak in your native language,and the AI will understand it and will actually real-time translateWhat will that do for creativity?For knowledge sharing?For entertainment?I think were in the very early innings of figuring that outWe want Spotify to be the place for all voices.-Spotify C

108、o-Founder&CEO Daniel Ek(5/25)In Q1:25,Spotify had 678MM Monthly Active Users and 268MM Subscribers and supported16.8B in annualized revenue while hosting 100MM+tracks,7MM podcast titles and 1MM creative artists.AI Development Trending=Unprecedented2/25:Spotify begins accepting audiobooks AI-tran

109、slated into 29 languages from ElevenLabs48AI Performance=Emerging Applications AcceleratingEmerging AI Applications 11/24,per Morgan StanleySource:Morgan Stanley,GenAI:Where are We Seeing Adoption and What Matters for 25?(11/24)AI Development Trending=Unprecedented49AI=Benefits&Risks50AI Develop

110、ment=Benefits&Risks The widely-discussed benefits and risks of AI top-of-mind for many generate warranted excitement and trepidation,further fueled by uncertainty over the rapid pace of change and intensifying global competition and saber rattling.The pros Stuart Russell and Peter Norvig went de

111、ep on these topics in theFourth Edition(2020)of their 1,116-page classic Artificial Intelligence:A Modern Approach(link here),and their views still hold true.Highlights followthe benefits:put simply,our entire civilization is the product of our human intelligence.If we have access to substantially g

112、reater machine intelligence,the ceiling of our ambitions is raised substantially.The potential for AI and robotics to free humanity from menial repetitive work and to dramaticallyincrease the production of goods and services could presage an era of peace and plenty.The capacity to accelerate scienti

113、fic research could result in cures for disease andsolutions for climate change and resource shortages.As Demis Hassabis,CEO of Google DeepMind,has suggested:First we solve AI,then use AI to solve everything else.Long before we have an opportunity to solve AI,however,we will incur risks from the misu

114、se of AI,inadvertent or otherwise.Some of these are already apparent,while others seem likely based on current trends includinglethal autonomous weaponssurveillance and persuasionbiased decision makingimpact on employmentsafety-critical applicationscybersecurityAI=Benefits&RisksSource:Stuart Rus

115、sell and Peter Norvig,Artificial Intelligence:A Modern Approach51Success in creating AI could be the biggest event in the history of our civilization.But it could also be the last unless we learn how to avoid the risks.Stephen Hawking,Theoretical Physicist/Cosmologist(1942-2018)AI=Benefits&Risks

116、Source:University of Cambridge,Centre for the Future of IntelligenceSeem Like Change Happening Faster Than Ever?Yes,It IsAI User+Usage+CapEx Growth=UnprecedentedAI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage RisingAI Usage+Cost+Loss Growth

117、=UnprecedentedAI Monetization Threats=Rising Competition+Open-Source Momentum+Chinas RiseAI&Physical World Ramps=Fast+Data-DrivenGlobal Internet User Ramps Powered by AI from Get-Go=Growth We Have Not Seen Likes of BeforeAI&Work Evolution=Real+Rapid5212345678Outline53AI User+Usage+CapEx Grow

118、th=Unprecedented54Consumer/User AI Adoption=Unprecedented55AI User Growth(ChatGPT as Foundational Indicator)=+8x to 800MM in Seventeen MonthsNote:OpenAI reports Weekly Active Users which are represented above.4/25 estimate from OpenAI CEO Sam Altmans 4/11/25 TED Talk disclosure.Source:OpenAI disclos

119、ures0400800ChatGPT Weekly Active Users,MMConsumer/User AI Adoption=Unprecedented+8xChatGPT User Growth(MM)10/22-4/25,per OpenAI56AI Global Adoption(ChatGPT as Foundational Indicator)=Have Not Seen Likes of This Around-the-World Spread Before0%25%50%75%100%12345678910 11 12 13 14 15 16 17 18 19 20 21

120、 22 23 24 25 26 27 28 29 30 31 32 33InternetChatGPT AppShare of Total Current Users,%Note:Year 1 for Internet=1990;year 33=2022.Year 1 for ChatGPT app=5/23;year 3 for ChatGPT app=5/25.ChatGPT app monthly active users(MAUs)shown.Note that ChatGPT is not available in China,Russia and select other coun

121、tries as of 5/25.China data may be subject to informational limitations due to government restrictions.Includes only Android,iPhone&iPad users.Figures may understate true ChatGPT user base(e.g.,desktop or mobile webpage users).Regions per United Nations definitions.Figures show%of total current

122、users in that year note that as year 3 for ChatGPT has not yet finished,percentages could move in coming months.Data for standalone ChatGPT app only.Country-level data may be missing for select years,as per ITU.Source:United Nations/International Telecommunications Union(3/25),Sensor Tower(5/25)Inde

123、xed Years(Internet 1=1990,ChatGPT App 1=2023)90%Year 2390%Year 3Consumer/User AI Adoption=UnprecedentedInternet vs.ChatGPT Users Percent Outside North America(1990-2025),Per ITU&Sensor Tower57AI User Adoption(ChatGPT as Proxy)=Materially Faster vs.Internet ComparablesNote:Netflix represents stre

124、aming business.Source:BOND,AI&Universities(2024)via company filings,pressYears to Reach 100MM Users 2000-2023Consumer/User AI Adoption=Unprecedented10.34.50.2024681012NetflixLinkedInPinterestUberTwitterTelegramSpotifyFacebookYouTubeSnapchatWhatsAppInstagramDisney+FortniteTikTokChatGPTYear Launch

125、ed:00-0505-1010-1515-2020+58AI User Adoption(ChatGPT as Proxy)=Materially Faster+Cheaper vs.Other Foundational Technology Products*Public launch of ChatGPT=first release to the public as a free research preview(11/22).Note:Per Ford Corporate,the Model T could be sold for between$260 and$850.We use$8

126、50 in 1908 dollars for our figures above.For TiVo,we use the launch of consumer sales on 3/31/99,when TiVo charged$499 for its 14-hour box set.We do not count TiVo subscription costs.We also use the iPhone 1s 4GB entry level price of$499 in 2007.Source:Heartcore Capital,CNBC,Museum of American Speed

127、,World Bank,Ford Corporate,Gizmodo,Apple,Encyclopedia Britannica,Federal Reserve Bank of St.Louis,Wikimedia Commons,UBSDays to Reach 1MM Customers/Users 1908-20222,5001,68074501,5003,000Ford Model T(1908)TiVo(1999)iPhone(2007)ChatGPT*(2022)Days to Reach 1MM Customers/UsersPurchase Price(2024$)$29,33

128、0$945$756$0Consumer/User AI Adoption=Unprecedented59AI User Adoption Time to 50%Household Penetration=Each Cycle Ramps in Half-the-TimeAI Following PatternNote:3 years for AI Era implies that the time to 50%USA Household Adoption is similarly cut in half from the previous cycle.Source:Morgan Stanley

129、,Google and Meta:AI vs.Fundamental 2H Debates(7/23),Our World in Data,other web sources per MSYears to 50%Adoption of Household Technologies in USA,per Morgan StanleyConsumer/User AI Adoption=Unprecedented42 Years20 Years12 Years6 Years3 Years?02550Second Industrial RevolutionPC EraDesktop Internet

130、EraMobile Internet EraAI EraYears60Technology Ecosystem AI Adoption=Impressive61NVIDIA AI Ecosystem Tells Over Four Years=100%Growth in Developers/Startups/AppsNote:GPU=Graphics Processing Unit.Source:NVIDIA(2021&2025)NVIDIA Computing Ecosystem 2021-2025,per NVIDIA2.5MM6MM036Number of Developers

131、(MM)7K27K0153020212025Number of AI Startups(K)Number of Applications Using GPUs(K)1.7K4K02.55+2.4x+3.9x+2.4xTechnology Ecosystem AI Adoption=Impressive62Tech Incumbent AI Adoption=Top Priority63Tech Incumbent AI Focus=Talking-the-TalkSource:Uptrends,Top 15 Companies Mentioning AI on Earnings Calls(6

132、/24),company earnings transcriptsMentions of AI in Corporate Earnings Transcripts Q1:20-Q1:24,per UptrendsTech Incumbent AI Adoption=Top Priority64Tech Incumbent AI Focus=Talking-the-TalkSource:Amazon(4/10/25),Google(4/9/25),TechradarGenerative AI is going to reinvent virtually every customer experi

133、ence weknow and enable altogether new ones about which weve only fantasized.The early AI workloads being deployed focus on productivity and cost avoidance Increasingly,youll see AI change the norms in coding,search,shopping,personal assistants,primary care,cancer and drug research,biology,robotics,s

134、pace,financial services,neighborhood networks everything.-Amazon CEO Andy Jassy in 2024 Amazon Shareholder Letter 4/25The chance to improve lives and reimagine things is why Google hasbeen investing in AI for more than a decadeWe see it as the most important way we can advance our mission to organiz

135、e the worlds information,make it universally accessible and useful.The opportunity with AI is as big as it gets.-Google CEO Sundar Pichai Google Cloud Next 2025 4/25Tech Incumbent AI Adoption=Top Priority65Tech Incumbent AI Focus=Talking-the-TalkNote:On 3/28/25,Elon Musk announced that xAI had acqui

136、red X in an all-stock deal.The deal valued xAI at$80B and X at$33B($45B less$12B debt).Source:Duolingo(5/1/25),DeepMind,Elon Musk(5/2/25),Fox NewsTheres three places where GenAI ishelping us:data creationcreating new features that were just not possibleefficiencies everywhere in the companyI should

137、mention something amazing about the new Duolingo curriculum in chess is that it really started with a team of two people,neither of whom knew how to programand they basically made prototypes and did the whole curriculum of chess by just using AI.Also,neither of them knew how to play chess.-Duolingo

138、Co-Founder&CEO Luis von Ahn Q1:25 Earnings Call 5/25AI with Grok is getting very goodits important that AI be programmed withgood values,especially truth-seeking values.This is,I think,essential for AI safetyRemember these words:We must have a maximally truth-seeking AI.-xAI Founder&CEO Elon

139、 Musk 5/25AI Going Full-Circle:DeepMinds AlphaGo(2014)started with humans training machinesDuolingo Chess now has machines training humansTech Incumbent AI Adoption=Top Priority66Tech Incumbent AI Focus=Talking-the-TalkSource:Roblox(5/1/25),NVIDIA(5/18/25)We view AI as a human acceleration tool that

140、 will allow individuals to do more.I believe long term,we will see people coupled withthe AI they use as the overall output of that person.-Roblox Co-Founder,President,CEO&Chair of Board David Baszucki Q1:25 Earnings Call 5/25Tech Incumbent AI Adoption=Top PriorityI promise you,in ten years time

141、,you will look back and you will realize that AI has now integrated into everything.And in fact,we need AI everywhere.And every region,every industry,every country,every company,all needs AI.AI is now part of infrastructure.And this infrastructure,just like the internet,just like electricity,needs f

142、actories.And these AI data centers,if you will,are improperly described.They are,in fact,AI factories.You apply energy to it,and it produces something incredibly valuable.-NVIDIA Co-Founder&CEO Jensen Huang COMPUTEX 2025 5/2567Traditional Enterprise AI Adoption=Rising Priority68Enterprise AI Foc

143、us S&P 500 Companies=50%&Rising Talking-the-Talk.Source:Goldman Sachs Global Investment Research,S&P Beige Book:3 themes from 4Q 2024 conference calls:Tariffs,a stronger US dollar,and AI(2/25)Quarterly Earnings Call Mentions of AI S&P 500 Companies(2015-2025),per Goldman Sachs Resear

144、ch%of S&P 500 Companies Mentioning AITraditional Enterprise AI Adoption=Rising Priority69Enterprise AI Focus Global Enterprises=Growth&RevenueNot Cost ReductionNote:Survey conducted 5/24,N=427.US-based companies=43%,Japan 15%,UK 14%,France 14%,Germany 14%.Industry mix:18%Technology,18%Financ

145、ial Services,17%Healthcare,17%Manufacturing,15%Industrials,15%Consumer,.Revenue mix:13%$500MM-$750MM,25%$751MM-$1B,36%$1B-$5B,10%$5B-$10B,8%$10B-$15B,3%$15B-$20B,5%$20B+.Revenue-Focused and Cost-Focused categorizations per BOND,not Morgan Stanley.Source:AlphaWise,Morgan Stanley,Quantifying the AI Op

146、portunity(12/24)GenAI Improvements Targeted for Global Enterprises over Next 2 Years 2024,per Morgan Stanley%of Survey Responses0%25%50%75%Hiring CostsHeadcountSG&A/MarketingManufacturing CostsAdmin CostsMarginsMarketing Spend EffectivityROICRevenuesSales ProductivityCustomer ServiceProduction/O

147、utputRevenue-FocusedCost-FocusedTraditional Enterprise AI Adoption=Rising Priority70Enterprise AI Focus Global CMOs=75%Using/Testing AI Tools%of Survey Responses0%25%50%75%Plan on Start TestingWithin 1-2 YearsFully ImplementedPlan on Start TestingWithin 12 MonthsRunning Initial Tests/ExperimentsNote

148、:Survey question asked about the extent to which marketing executives worldwide are using generative AI for marketing activities.Survey conducted 7/24,N=300 marketing executives at companies with 500+employees worldwide.Survey geos:Australia,Belgium,Brazil,Canada,China,Denmark,Finland,France,Germany

149、,Ireland,Italy,Japan,Luxembourg,Mexico,Netherlands,Norway,Poland,Saudi Arabia,Spain,Sweden,UAE,UK,&USA.Source:eMarketer,Morgan Stanley,Quantifying the AI Opportunity(12/24)Global Chief Marketing Officer(CMO)GenAI Adoption Survey 2024,per Morgan StanleyTraditional Enterprise AI Adoption=Rising Pr

150、iority71Enterprise AI Adoption=Rising PriorityBank of America Erica Virtual Assistant(6/18)Note:We assume a start at zero users from Ericas launch in 6/18.Pilot users excluded.Source:Bank of America(2/21,4/24,2/25)Bank of America Erica Virtual Assistant 6/18-2/25,per Bank of AmericaErica acts as bot

151、h a personal concierge andmission control for our clients.Our data science team has made more than 50,000 updates to Ericas performance since launch adjusting,expanding and fine-tuning natural language understanding capabilities,ensuring answers and insights remain timely and relevant.2 billion clie

152、nt interactions is a compelling milestonethough this is only the beginning for Erica.-Head of Digital at Bank of America Nikki Katz,4/24Cumulative Interactions,MM05001,0001,5002,0002,5006/1811/184/199/192/207/2012/205/2110/213/228/221/236/2311/234/249/242/25Cumulative Client Interactions with Erica

153、Virtual Assistant(MM)Note:Erica is a conversational AI built into Bank of Americas mobile app that helps customers manage their finances by providing real-time insights,transaction search,bill reminders,and budgeting assistance.It has handled billions of interactions and serves as a 24/7 digital fin

154、ancial concierge for over 40 million clients.Traditional Enterprise AI Adoption=Rising Priority72Enterprise AI Adoption=Rising PriorityJP Morgan End-to-End AI Modernization(2020)Note:Superscript 2,per JP Morgan,indicates Value is described as benefit in revenue,lower expense,or avoidance of cost maj

155、ority is measured as the lift relative to prior analytical techniques with the remainder relative to a random baseline or holdout control.We indicate 2020 as the start year for JP Morgans AI Modernization(2020 Letter to Shareholders:We already extensively use AI,quite successfully,in fraud and risk,

156、marketing,prospecting,idea generation,operations,trading and in other areasto great effect,but we are still at the beginning of this journey).Source:JP Morgan Investor Day(5/25)JP Morgan End-to-End AI Modernization 2023-2025E,per JP MorganWe have high hopes for the efficiency gains we might get from

157、 AICertain key subsets of the users tell us they are gaining several hours a week of productivity,and almost by definition,the time savings is coming from less valuable tasksWe were early movers in AI.But were still in the early stages of the journey.-JP Morgan CFO Jeremy Barnum,5/25JP Morgan Estima

158、ted Value from AI/MLTraditional Enterprise AI Adoption=Rising Priority+35%+65%73Enterprise AI Adoption=Rising PriorityKaiser Permanente Multimodal Ambient AI Scribe(10/23)Source:Tierney,Aaron A.et al.,Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation(3/24)&

159、;Tierney,Aaron A.et al.,Ambient Artificial Intelligence Scribes:Learnings after 1 Year and over 2.5 Million Uses(3/25)via Nestor Maslej et al.,The AI Index 2025 Annual Report,AI Index Steering Committee,Stanford HAI(4/25)Kaiser Permanente Ambient AI Scribe 10/23-12/24,per New England Journal of Medi

160、cineAmbient artificial intelligence(AI)scribes,which use machine learning applied to conversations to facilitate scribe-like capabilities in real time,have great potential to reduce documentation burden,enhance physician-patient encounters,and augment clinicians capabilities.The technology leverages

161、 a smartphone microphone to transcribe encounters as they occur but does not retain audio recordings.To address the urgent and growing burden of data entry,in October 2023,The Permanente Medical Group(TPMG)enabled ambient AI technology for 10,000 physicians and staff to augment their clinical capabi

162、lities acrossdiverse settings and specialties.-New England Journal of MedicineCatalyst Research Report,2/24Unique Kaiser Permanente Physicians Ever Using AI Scribe&Cumulative Number of Scribe VisitsTraditional Enterprise AI Adoption=Rising Priority74Enterprise AI Adoption=Rising PriorityYum!Bran

163、ds Byte by Yum!(2/25)Note:Yum!Brands names include KFC,Taco Bell,Pizza Hut,&The Habit.Byte by Yum!was officially launched in 2/25.While underlying technologies were previously in-use at restaurants in Yum!s portfolio,the Byte by Yum!product suite had not yet officially been launched;hence,we ill

164、ustratively show zero users in 2/24.Source:Yum!,Introducing Byte by Yum!,an AI-driven restaurant technology platform powering customer and team member experiences worldwide(2/25)Yum!Brands Byte by Yum!2/24-2/25,per Yum!BrandsBacked by artificial intelligence,Byte by Yum!offers franchisees leading te

165、chnology capabilities with advantaged economics made possible by the scale of Yum!.The Byte by Yum!platform includes online and mobile app ordering,point of sale,kitchen and delivery optimization,menu management,inventory and labor management,and team member tools.-Yum!Press Release,2/25Number of Re

166、staurantsYum!Restaurants Using at Least One Byte by Yum!ProductByte is Yum!Brands AI-powered restaurant management platform designed to optimize store operations by automating repetitive tasks like inventory tracking,scheduling,and food preparation alerts.It leverages machine learning to improve dec

167、ision-making at the restaurant level,enhancing efficiency,reducing waste,and supporting staff productivity.Traditional Enterprise AI Adoption=Rising Priority025,00005,00010,00015,00020,00025,0002/242/2575Education/Government/Research AI Adoption=Rising Priority76Source:Arizona State University(8/23)

168、,Oxford University(3/25),University of Michigan(3/25),Launch Consulting(1/25)via AI Advantage Daily News,NPR(1/25)Education&Government=Increasingly Announcing AI IntegrationsArizona State UniversitysAI Acceleration 8/23Oxford Partnership 3/25NextGenAI 3/25$50MM consortium with 15 research univer

169、sities(MIT,Harvard,Caltech,etc.)5-Year Partnership on Research&AI LiteracyNew team of technologists creating artificial intelligence(AI)toolsChatGPT Gov 1/25ChatGPT tailored for USA federal agenciesUSA National Laboratories 1/25Partnering on Nuclear,Cybersecurity,&Scientific BreakthroughsEdu

170、cation/Government/Research AI Adoption=Rising Priority77Source:NVIDIA(2/25&5/25)Government=Increasingly Adopting Sovereign AI PoliciesEducation/Government/Research AI Adoption=Rising PriorityNVIDIA Sovereign AI Partners 2/25,Per NVIDIANations are investing in AI infrastructure like they once did

171、 for electricity and Internet.-NVIDIA Co-Founder&CEO Jensen Huang,5/2578Research=Rapid Ramp in FDA-Approved AI Medical Devices,per Stanford HAINote:FY21,FY22&FY23 USA government budget figures are actuals.FY24 data is enacted but not actual,FY25 data is requested.NIH share of total budget is

172、 requested.Source:Nestor Maslej et al.,The AI Index 2025 Annual Report,AI Index Steering Committee,Stanford HAI(4/25);USA Food&Drug Administration,FDA Announces Completion of First AI-Assisted Scientific Review Pilot and Aggressive Agency-Wide AI Rollout Timeline(5/25);NITRD.gov(5/25)New AI-Enab

173、led Medical Devices Approved by USA Food&Drug Administration 1995-2023,per Stanford HAI&USA FDANumber of AI Medical Devices ApprovedEducation/Government/Research AI Adoption=Rising PriorityIn a historic first for the USA FDA,FDA Commissioner Martin A.Makary,M.D.,M.P.H.,today announced an agg

174、ressive timeline to scale use of artificial intelligence(AI)internally across all FDA centers by June 30,2025To reflect the urgency of this effort,Dr.Makary has directed all FDA centers to begin deployment immediately,with the goal of full integration by the end of June.-USA FDA Press Release,5/25AI

175、-Enabled Medical Devices ApprovedNew USA FDA AI Policy(5/25)10110010011005022336618266480114129160223012525019951999200320072011201520192023Government R&D funding has been a key part of AI development budgets,especially in healthcare:-FY21-FY25 Federal USA AI Budget:$14.7B-FY25 Share Requested b

176、y National Institutes of Health:34%79Research=30%-80%Reduction in Medical R&D Timelines,per Insilico Medicine&CradleNote:Pre-Clinical Candidate Status marks the point at which a lead molecule(or biologic)has satisfied all discovery-stage gates and is officially handed off to the development

177、organization for work related to beginning human clinical trials.Figures collected from 2021-2024.Source:Cradle,Insilico Medicine via BioPharmaTrend,Insilico Medicine Reports Benchmarks for its AI-Designed Therapeutics(2/25)AI-Driven Drug Discovery 2021-2024,Per Insilico Medicine,Cradle&BioPharm

178、aTrendMonths to Pre-Clinical Candidate StatusEducation/Government/Research AI Adoption=Rising PriorityPharma companies that use Cradle are seeing a 1.5x to 12x speedup in pre-clinical research and development by using our GenAI platform to engineer biologics.-Stef van Grieken,Co-Founder&CEO of C

179、radle,5/25Months to Reach Pre-Clinical Candidate Status02550Solid Tumors(QPCTL)Inflammatory BowelDisease(PHD1/2)Idiopathic PulmonaryFibrosis(TNIK)TraditionalApproachesTraditional approaches can take 2.5-4 years80AI User+Usage+CapEx Growth=UnprecedentedAI Usage ChatGPT=Rising Rapidly Across Age Group

180、s in USA,per Pew&Elon UniversityNote:7/23 data per Pew Research study on ChatGPT use,n=10,133 USA adults.Those who did not give an answer are not shown.1/25 data per Elon University study on use of any AI models,n=500 USA adults,.Figures estimated based on overall AI tool usage adjusted for an a

181、verage 72%usage rate of ChatGPT amongst respondents who use any AI tools.Actual ChatGPT penetration may vary by cohort.Note that this chart aggregates data across survey providers and as such may not be directly comparable.Source:Pew Research Center(3/26/24),Elon University(released 3/12/25),Sam Alt

182、man(5/12/25)via Fortune%of USA Adults Who Say They Have Ever Used ChatGPT 7/23 per Pew&1/25 per Elon University 18%33%21%13%4%37%55%44%30%20%0%50%100%All USA AdultsAges 18-29Ages 30-49Ages 50-64Ages 65+7/23 Per Pew1/25 Estimates Per Elon University%of USA AdultsAI User+Usage+CapEx Growth=Unprece

183、dented81A gross oversimplification is:Older people use ChatGPT as,like,a Google replacement.People in their 20s and 30s use it like a life advisor.-OpenAI Co-Founder&CEO Sam Altman(5/25)82Minutes per Day that USA Active Users Spend on ChatGPT App 7/23-4/25,per Sensor TowerNote:Data represents US

184、A App Store&Google Play Store monthly active users.Data for ChatGPT standalone app only.ChatGPT app not available in China,Russia and select other countries as of 5/25.Source:Sensor Tower(5/25)AI Engagement(ChatGPT App as Proxy)=+202%Rise in Daily Time Spent Over Twenty-One MonthsDaily Minutes S

185、pent on App,USA010207/238/239/23 10/23 11/23 12/23 1/242/243/244/245/246/247/248/249/24 10/24 11/24 12/24 1/252/253/254/25+202%AI User+Usage+CapEx Growth=Unprecedented83Note:Data represents USA App Store&Google Play Store monthly active users.Data for ChatGPT standalone app only.ChatGPT app not

186、available in China,Russia and select other countries as of 5/25.Source:Sensor Tower(5/25)AI Engagement(ChatGPT App as Proxy)=+106%Growth in Sessions&+47%Growth in Duration Over Twenty-One MonthsAverage Minutes/Session,USA(Blue Bars)036901237/23 8/23 9/23 10/2311/2312/23 1/24 2/24 3/24 4/24 5/24

187、6/24 7/24 8/24 9/24 10/2411/2412/24 1/25 2/25 3/25 4/25Average Daily Sessions/User,USA(Red Line)AI User+Usage+CapEx Growth=UnprecedentedAverage USA Session Duration(Minutes)&Daily Sessions per User for ChatGPT App 7/23-4/25,per Sensor Tower84AI Retention(ChatGPT as Proxy)=80%vs.58%Over Twenty-Se

188、ven Months,per YipitDataNote:Retention Rate=Percentage of users from the immediately preceding week that were users again in the current week.Data measures several million global active desktop users clickstream data.Data consists of users web requests&is collected from web services/applications

189、,such as VPNs and browser extensions.Users must have been part of the panel for 2 consecutive months to be included.Panel is globally-representative,though China data may be subject to informational limitations due to government restrictions.Excludes anomalies in w/c 12/24/23,12/31/23,12/22/24,12/29

190、/24,1/5/25,potentially due to holiday breaks causing less enterprise usage.Source:YipitData(5/25)Weekly Retention,%Consumer ChatGPT&Google Search Global Desktop User Retention Rates(1/23-4/25),per YipitData50%75%100%ChatGPT RetentionGoogle Search Retention+2,259 bpsAI User+Usage+CapEx Growth=Unp

191、recedented85AI Chatbots Work Tells=72%Doing Things Quicker/BetterNote:N=5,273 USA adults who are employed part time or full time and who have only one job or have more than one but consider one of them to be their primary job were surveyed.Source:Pew Research Center(10/24)%of Employed USA Adults Usi

192、ng AI Chatbots Who Say Tools Have Been _ Helpful When It Comes to 10/24,per Pew%of Employed USA AdultsAI User+Usage+CapEx Growth=Unprecedented0%25%50%75%100%Improving the Qualityof Their WorkAllowing Them to DoThings More QuicklyExtremely/VerySomewhatNot Too/Not at All86AI Chatbots School Tells(Chat

193、GPT as Proxy)=Bias to Research/Problem Solving/Learning/AdviceOpenAI ChatGPT Usage Survey,USA Students Ages 18-24 12/24-1/25,per OpenAINote:Data per OpenAI survey(12/24),n=1,299 USA college and graduate students across a mix of STEM and non-STEM disciplines;only answers from 18-24 year olds used.Sam

194、ple includes both AI users and non-users but excludes“AI rejectors”defined as non-users with little to no interest in adopting AI within the next 12 months.Source:OpenAI,Building an AI-Ready Workforce:A Look at College Student ChatGPT Adoption in the US(2/25)AI User+Usage+CapEx Growth=Unprecedented8

195、7AI Usage Expansion Deep Research=Automating Specialized Knowledge WorkSelect AI Company Deep Research Capabilities 12/24-2/25,per Google,OpenAI&xAISource:Google(5/25),OpenAI(2/25),xAI(2/25),Digital Trends(1/25)Get up to speed on just about anything with Deep Research,an agentic feature in Gemin

196、i that can automatically browse up to hundreds of websites on your behalf,think through its findings,and create insightful multi-page,reports that you can turn into engaging podcast-style conversationsIts a step towards more agentic AI that can move beyond simple question-answering to become a true

197、collaborative partner.-Google Deep Research Overview,launched 12/24Google Gemini Deep ResearchToday were launching deep research in ChatGPT,a new agentic capability that conducts multi-step research on the internet for complex tasks.It accomplishes in tens of minutes what would take a human many hou

198、rsDeep research marks a significant step toward our broader goal of developing AGI,which we have longenvisioned as capable of producingnovel scientific research.-OpenAI Deep ResearchPress Release,2/25OpenAI ChatGPTDeep ResearchxAI GrokDeepSearchTo understand the universe,we must interface Grok with

199、the worldAs a first step towards this vision,we are rolling out DeepSearch our first agent.Its a lightning-fast AI agent built to relentlessly seek the truth across the entire corpus of human knowledge.DeepSearch is designed to synthesizekey information,reason aboutconflicting facts and opinions,and

200、distill clarity from complexity.-xAI Grok 3 Beta Press Release,2/25AI User+Usage+CapEx Growth=Unprecedented88AI Agent Evolution=Chat Responses Doing Work89AI Agent Evolution=Chat Responses Doing WorkA new class of AI is now emerging less assistant,more service provider.What began as basic conversati

201、onal interfaces may now be evolving into something far more capable.Traditional chatbots were designed to respond to user prompts,often within rigid scripts or narrow flows.They could fetch answers,summarize text,or mimic conversation but always in a reactive,limited frame.AI agents represent a step

202、-change forward.These are intelligent long-running processes that can reason,act,and complete multi-step tasks on a users behalf.They dont just answer questions they execute:booking meetings,submitting reports,logging into tools,or orchestrating workflows across platforms,often using natural languag

203、e as their command layer.This shift mirrors a broader historical pattern in technology.Just as the early 2000s saw static websites give way to dynamic web applications where tools like Gmail and Google Maps transformed the internet from a collection of pages into a set of utilities AI agents are tur

204、ning conversational interfaces into functional infrastructure.Whereas early assistants needed clear inputs and produced narrow outputs,agents promise to operate with goals,autonomy and certain guardrails.They promise to interpret intent,manage memory,and coordinate acrossapps to get real work done.I

205、ts less about responding and more about accomplishing.While we are early in the development of these agents,the implications are just starting to emerge.AI agents could reshape how users interact with digital systems from customer support and onboarding to research,scheduling,and internal operations

206、.Enterprises are leading the charge;theyre not just experimenting with agents,but deploying them,investing in frameworks and building ecosystems around autonomous execution.What was once a messaging interface is becoming an action layer.90Source:Google Trends via Glimpse(5/15/24),OpenAI(3/25)AI Agen

207、t Interest(Google Searches)=+1,088%Over Sixteen Months02505001/242/243/244/245/246/247/247/248/249/2410/24 11/24 12/24 12/241/252/253/254/255/25Weekly Google Keyword Searches,K3/11/25:OpenAI Introduces Developer Tools for AI Agents+1,088%AI Agent Evolution=Chat Responses Doing WorkGlobal Google Sear

208、ches for AI Agent(K)1/24-5/25,per Google Trends91Source:Salesforce(10/24),Salesforce Ben,Anthropic(10/24),OpenAI(1/25),Amazon(3/25)AI Agent Deployments=AI Incumbent Product Launches AcceleratingOpenAI Operator(1/25=Research Preview Release)Salesforce Agentforce(10/24=General Release)Anthropic Claude

209、 3.5 Computer Use(10/24=Research Preview Release)Amazon Nova Act(3/25=Research Preview Release)Agent ReleasedSelect CapabilitiesAutomated customer supportCase resolutionLead qualificationOrder trackingControl computer screen directly to perform tasks like pulling data from websites,making online pur

210、chases,etc.Control computer screen directly to perform tasks like pulling data from websites,making online purchases,etc.Home automationInformation collectionPurchasingSchedulingAI Incumbent Agent LaunchesAI Agent Evolution=Chat Responses Doing Work92Next Frontier For AI=Artificial General Intellige

211、nce93Artificial General Intelligence,or AGI,refers to systems capable of performing the full range of human intellectual tasks reasoning,planning,learning from small data samples,and generalizing knowledge across domains.Unlike current AI models,which excel within specific(albeit broad)boundaries,AG

212、I would be able to operatefully flexibly across disciplines and solve unfamiliar problems without retraining.It represents a major milestone in AI development one that builds on recentexponential gains in model scale,training data,and computational efficiency.Timelines for AGI remain uncertain,but e

213、xpert expectations have shifted forward meaningfully in recent years.Sam Altman,CEO of OpenAI,remarked in January 2025,We are now confident we know how to build AGI as we have traditionally understood it.This is a forecast,not a dictum,but it reflects how advances in model architecture,inference*eff

214、iciency,and training scale are shortening the distance between research and frontier capability.The broader thread is clear:AI development is trending at unprecedented speed,andAGI is increasingly being viewed not as a hypothetical endpoint,but as a reachable threshold.If/when achieved,AGI would red

215、efine what software(and related hardware)can do.Rather than executingpre-programmed tasks,AGI systems would understand goals,generate plans,and self-correct in real time.They could drive research,engineering,education,and logistics workflows with little to no human oversight handling ambiguity and n

216、ovelty with general-purpose reasoning.These systems wouldnt require extensiveretraining to handle new problem domains they would transfer learning and operate with context,much like human experts.Additionally,humanoid robots powered by AGI would have thepower to reshape our physical environment and

217、how we operate in it.Still,the implications warrant a measured view.AGI is not a finish line,but a phase shift in capability and how itreshapes institutions,labor,and decision-making will depend on the safeguards and deploymentframeworks that accompany it.The productivity upside may be significant,b

218、ut unevenly distributed.The geopolitical,ethical,and economic implications may evolve gradually,not abruptly.As with earlier transitions from industrial to digital to algorithmic the full consequences will beshaped not just by what the technology can do,but by how society chooses to adopt and govern

219、 it.*Inference=Fully-trained model generates predictions,answers,or content in response to user inputs.This phase is much faster and more efficient than training.Next Frontier For AI=Artificial General Intelligence94AI User+Usage+CapEx Growth=Unprecedented95To understand where technology CapEx is he

220、ading,it helps to look at where its been.Over the past two decades,tech CapEx has flexed upward at points through datas long arc first toward storage/access,then toward distribution/scale,and now toward computation/intelligence.The earliest wave saw CapEx pouring into building internet infrastructur

221、e massive server farms,undersea cables,and early data centers that enabled Amazon,Microsoft,Google andothers to lay the foundation for cloud computing.That was the first phase:store it,organize it,serve it.The second wave still unfolding has been about supercharging compute for data-heavy AI workloa

222、ds,a natural evolution of cloud computing.Hyperscaler*CapEx budgets now tilt increasingly towardspecialized chips(GPUs,TPUs,AI accelerators),liquid cooling,and frontier data center design.In 2019,AI was a research feature;by 2023,it was a capital expenditure line item.Microsoft Vice Chair and Presid

223、ent Brad Smith put it well in a 4/25 blog post:Like electricity and other general-purpose technologies in the past,AI andcloud datacenters represent the next stage of industrialization.The worlds biggest tech companies are spending tens of billions annually not just to gather data,but to learn from

224、it,reason with it and monetize it in real time.Its still about data but now,the advantage goes to those who can train on it fastest,personalize it deepest,and deploy it widest.*Hyperscalers(large data center operators)are Amazon Web Services(AWS),Microsoft Azure,Google Cloud Platform(GCP),Alibaba Cl

225、oud,Oracle Cloud Infrastructure(OCI),IBM Cloud&Tencent Cloud.AI User+Usage+CapEx Growth=Unprecedented96CapEx Spend Big Technology Companies=On Rise for Years asData Use+Storage Exploded97CapEx Spend Big Six*Tech Companies(USA)=+21%Annual Growth Over Ten Years*Note:Big Six USA technology companie

226、s include Apple,Nvidia,Microsoft,Alphabet/Google,Amazon,&Meta Platforms/Facebook.Only AWS CapEx&revenue shown for Amazon(i.e.excludes Amazon retail CapEx).AWS CapEx estimated per Morgan Stanley equals AWS net additions to property&equipment less finance leases and obligations.Global data

227、 generation figures for 2024 are estimates.Source:Capital IQ(3/25),Hinrich Foundation(3/25)Global Data Generation,Zettabytes(Red Line)0306090120150$0$50$100$150$200$25020142015201620172018201920202021202220232024CapEx Spend,$B(Blue Bars)As data volumes rise,CapEx required to build more hyperscale da

228、ta centers,faster network infrastructure,&more compute capacityCapEx:+21%/YearData:+28%/YearCapEx Spend Big Technology Companies=On Rise for Years as Data Use+Storage ExplodedBig Six*USA Public Technology Company CapEx Spend($B)vs.Global Data Generation(Zettabytes)2014-2024,per Capital IQ&Hi

229、nrich FoundationCapEx Spend for Tech Hyperscalers=Mirrored by+37%Annual Cloud Revenue Growth Over Ten Years98Note:Companies do not report“hyperscaler cloud revenue”on like-for-like basis so data represents best estimates and may not align between companies.Oracle Cloud revenue includes Cloud Service

230、s&License Support,as well as Cloud License&On-Premise License.IBM Cloud includes all Infrastructure line items due to reporting standards.Alibaba&Tencent Cloud revenues estimated per Morgan Stanley.Source:Company disclosures,Morgan Stanley(as of 4/25)$0$100$200$300$4002014201520162017201

231、8201920202021202220232024Amazon AWSMicrosoft Intelligent CloudGoogle CloudOracle CloudIBM CloudAlibaba CloudTencent CloudRevenue,$B+37%/YearCapEx Spend Big Technology Companies=On Rise for Years as Data Use+Storage ExplodedGlobal Hyperscaler Cloud Revenue($B)2014-2024,per Company Disclosures&Mor

232、gan Stanley Estimates99CapEx Spend Big Technology Companies=Inflected With AIs Rise100AI Model Training Dataset Size=250%Annual Growth Over Fifteen Years,per Epoch AINote:In AI language models,tokens represent basic units of text(e.g.,words or sub-words)used during training.Training dataset sizes ar

233、e often measured in total tokens processed.A larger token count typically reflects more diverse and extensive training data,which can lead to improved model performance up to a point before reaching diminishing returns.Source:Epoch AI(5/25)AI Model Training Dataset Size(Tokens)by Model Release Year

234、6/10-5/25,per Epoch AITraining Dataset Size,TokensCapEx Spend Big Technology Companies=Inflected With AIs Rise+250%/Year101CapEx Spend Big Six*Tech Companies=+63%Y/Y&Accelerated1ChatGPT WAU data as of 11/23&12/24 due to data availability.*Note:Big Six USA technology companies include Apple,N

235、vidia,Microsoft,Alphabet/Google,Amazon,&Meta Platforms/Facebook.Only AWS CapEx&revenue shown for Amazon(i.e.excludes Amazon retail CapEx).AWS CapEx estimated per Morgan Stanley equals AWS net additions to property&equipment less finance leases and obligations.Source:Capital IQ(3/25),Open

236、AI disclosures(3/25)Global ChatGPT Weekly Active Users,MM(Red Line)070140210280350$0$50$100$150$200$25020142015201620172018201920202021202220232024CapEx Spend,$B(Blue Bars)2023-2024 Change:Big Six CapEx=+63%ChatGPT WAUs=+200%1CapEx Spend Big Technology Companies=Inflected With AIs RiseBig Six*USA Pu

237、blic Technology Company CapEx Spend($B)vs.Global ChatGPT Weekly Active Users(MM)2014-2024,per Capital IQ&OpenAI102CapEx Spend Big Six*Tech Companies=15%of Revenue&Accelerated vs.8%Ten Years Ago*Note:Big Six USA technology companies include Apple,Nvidia,Microsoft,Alphabet/Google,Amazon,&M

238、eta Platforms/Facebook.Only AWS CapEx&revenue shown for Amazon(i.e.excludes Amazon retail CapEx).AWS CapEx estimated per Morgan Stanley equals AWS net additions to property&equipment less finance leases and obligations.Source:Capital IQ(3/25),Morgan Stanley(5/25)CapEx,$B(Blue Bars)CapEx as%o

239、f Revenue(Red Line)0%3%6%9%12%15%$0$50$100$150$200$25020142015201620172018201920202021202220232024+21%/YearCapEx Spend Big Technology Companies=Inflected With AIs RiseBig Six*USA Public Technology Company CapEx Spend($B)vs.%of Revenue 2014-2024,per Capital IQ&Morgan Stanley103CapEx Spend Amazon

240、AWS=Cloud vs.AI Patterns104CapEx as%of Revenue(AWS as Proxy)AI vs.Cloud Buildouts=49%(2024)vs.4%(2018)vs.27%(2013),per Morgan StanleyNote:Figures shown represent AWS only.AWS CapEx estimated per Morgan Stanley equals AWS net additions to property&equipment less finance leases and obligations.Sou

241、rce:Amazon,Morgan Stanley(5/25)Amazon AWS CapEx as%of Revenue 2013-2024,Estimated per Morgan StanleyCapEx/Revenue,%AI/ML Infrastructure Build-OutInitial Cloud Infrastructure Build-Out27%4%49%0%20%40%60%201320142015201620172018201920202021202220232024AWS CapEx as%of revenue decreased as upfront infra

242、structure investments slowed&revenue grew will AI follow?From 2020,AWS began rapidly scaling CapEx(+30%Y/Y)to build AI/ML infrastructure,potentially restarting cycleCapEx Spend Amazon AWS=Cloud vs.AI Patterns105Tech CapEx Spend Partial Instigator=Material Improvements in GPU PerformanceNVIDIA GP

243、U Performance=+225x Over Eight Years1061GPT-MoE Inference Workload=A type of workload where a GPT-style model with a Mixture-of-Experts(MoE)architecture is used for inference(i.e.,making predictions).Note:Annual token revenue assumes a flat per-token cost.Source:NVIDIA(5/25)Performance of NVIDIA GPU

244、 Series Over Time 2016-2024,per NVIDIATech CapEx Spend Partial Instigator=Material Improvements in GPU PerformancePascalVoltaAmpereHopperBlackwell20162018202020222024Number of GPUs46K43K28K16K11K+225xFactory AI FLOPS1EF5EF17EF63EF220EFAnnual Inference Tokens50B1T5T58T1,375T+30,000 xAnnual Token Reve

245、nue$240K$3M$24M$300M$7BDC Power37MW34MW25MW19MW21MW+50,000 xToken Per MW-Year1.3B2.9B200B3T65TPerformance+225x over eight years while requiring 4x fewer GPUs$1B Data Center ComparisonGPT-MoE Inference Workload1Inference token capacity+27,500 x over eight years,implying+30,000 x higher theoretical to

246、ken revenueData center power use down 43%over eight years,leading to+50,000 x greater per-unit energy efficiencyFor a Theoretical$1B-Scale Data CenterNVIDIA Installed GPU Computing Power=100 x+Growth Over Six Years107Note:Analysis does not include TPUs or other specialized AI accelerators,for which

247、less data is available.TPUs may provide comparable total computing power to NVIDIA chips.Source:Epoch AI(2/25)Simultaneous expansion of GPU/computing-related CapEx alongsiderising performance-per-GPU=Exponentially-greater computing capacityTotal Installed Computing Power,FLOP/s+130%/YearTech CapEx S

248、pend Partial Instigator=Material Improvements in GPU PerformanceGlobal Stock of NVIDIA GPU Computing Power(FLOP/s)Q1:19-Q4:24,per Epoch AI108Tech CapEx Spend Beneficiary=NVIDIAKey Tech CapEx Spend Beneficiary=NVIDIA25%&Rising of Global Data Center CapEx,per NVIDIA109Note:NVIDIA data represents J

249、anuary FYE(e.g.,2024=FY25 ending 1/25)vs calendar year for data center CapEx.Data presented by Jensen Huang at NVIDIA GTC 2025(link).Source:DellOro Research for CapEx(3/25);NVIDIA for data center revenue(3/25)0%10%20%30%$0$200$400$600202220232024Global Data Center CapEx,$B(Blue Bar)NVIDIA Data Cente

250、r Revenue as%of Global Data Center CapEx(Red Line)Tech CapEx Spend Beneficiary=NVIDIAGlobal Data Center CapEx($B)vs.NVIDIAs Data Center Revenue as Percent of Data Center CapEx(Global)2022-2024,per NVIDIA GTC110Technology Company Spend=R&D Rising Along with CapEx111R&D Spend Big Six*USA Publi

251、c Tech Companies=13%of Revenuevs.9%Ten Years Ago*Note:Big Six USA technology companies include Apple,Nvidia,Microsoft,Alphabet/Google,Amazon,&Meta Platforms/Facebook.R&D expense shown for Amazon,not AWS,as figures are not broken out in company financials;revenue therefore shown on like-for-l

252、ike basis.Source:Capital IQ(3/25)R&D Expense,$B(Blue Bars)R&D Expense as%of Revenue(Red Line)0%5%10%15%$0$100$200$30020142015201620172018201920202021202220232024+20%/YearTechnology Company Spend=R&D Rising Along with CapExBig Six*USA Public Technology Company R&D Spend($B)vs.%of Reve

253、nue 2014-2024,per Capital IQ112Tech Big Six(USA)=Loaded With Cash to Spend on AI&CapExBig Six*Generating Loads of Cash=+263%Growth in Free Cash Flow Over Ten Years to$389B113*Note:Big Six USA technology companies include Apple,Nvidia,Microsoft,Alphabet/Google,Amazon,&Meta Platforms/Facebook.

254、FCF calculated as cash flow from operations less capex to standardize definitions,as only some companies subtract finance leases and Amazon adjusts FCF for gains on sale of equipment.FCF shown for Amazon,not AWS,as figures are not broken out in company financials.Source:Capital IQ(3/25)$0$50$100Appl

255、eNVIDIAMicrosoftGoogleAmazonMeta201420192024Free Cash Flow,$BTech Big Six(USA)=Loaded With Cash to Spend on AI&CapExBig Six*Public Technology Companies Free Cash Flow($B)2014-2024,per Capital IQBig Six*Generating Loads of Cash=+103%Growth in Cash Over Ten Years to$443B114*Note:Big Six USA techno

256、logy companies include Apple,Nvidia,Microsoft,Alphabet/Google,Amazon,&Meta Platforms/Facebook.Figure measures cash and other equivalents(e.g.,short-term investments and marketable securities)on companies balance sheets.Source:Capital IQ(3/25)$0$50$100$150AppleNVIDIAMicrosoftGoogleAmazonMeta20142

257、0192024Cash on Balance Sheet,$BTech Big Six(USA)=Loaded With Cash to Spend on AI&CapExBig Six*USA Public Technology Company Cash on Balance Sheet($B)2014-2024,per Capital IQ115Tech CapEx Spend Driver=Compute Spend to Train&Run AI Models116Tech CapEx Spend Driver=Compute Spend to Train&Ru

258、n ModelsTo understand the evolution of AI computing economics,its constructive to look at where costs are concentrated And where theyre headed.The bulk of spending in AI large language model(LLM)development is still dominated by compute specifically,the compute needed to train and run models.Trainin

259、g costs remain extraordinarily high and are rising fast,often exceeding$100 million per model today.As Dario Amodei,CEO of Anthropic,noted in mid-2024,Right now,AI model training costs$100 million.There are models in training today that are more like a billionI think that the training of$10 billion

260、models,yeah,could start sometime in 2025.Around these core compute costs sit additional high-cost layers:research,data acquisition and hosting,and a mix of salaries,general overhead,and go-to-market operations.Even as the cost to train models climbs,a growing share of total AI spend is shifting towa

261、rd inference the cost of running models at scale in real-time.Inference happens constantly,across billions of prompts,queries,and decisions,whereas model training is episodic.As Amazon CEO Andy Jassy noted in his April 2025 letter to shareholders,While model training still accounts for a large amoun

262、t of the total AI spend,inferencewill represent the overwhelming majority of future AI cost because customerstrain their models periodically but produce inferences constantly.NVIDIA Co-Founder&CEO Jensen Huang noted the same in NVIDIAs FQ1:26 earnings call,sayingInference is exploding.Reasoning

263、AI agents require orders of magnitude more compute.At scale,inference becomes a persistent cost center one that grows in parallel with usage,despite declines in unit inference costs.The broader dynamic is clear:lower per-unit costs are fueling higher overall spend.As inference becomes cheaper,AI get

264、s used more.And as AI gets used more,total infrastructure and compute demand rises dragging costs up again.The result is a flywheel of growth that puts pressure on cloud providers,chipmakers,and enterprise IT budgets alike.The economics of AI are evolving quickly but for now,they remain driven by he

265、avy capital intensity,large-scale infrastructure,and a race to serve exponentially expanding usage.117Data Centers=Key Beneficiary of AI CapEx Spend118Data Centers=Key Beneficiary of AI CapEx SpendFor one lens into the economics of AI infrastructure,its useful to look at the pace and scale of data c

266、enter construction.The current wave of AI-driven demand has pushed data center spending to historic highs.According to DellOro Research,global IT company data center CapExreached$455 billion in 2024 and is accelerating.Hyperscalers and AI-first companies alike are pouring billions into building out

267、compute-ready capacity not just for storage,but for real-time inference and model training workloads that require dense,high-power hardware.As AI moves from experimental to essential,so too do data centers.Per NVIDIA Co-Founder and CEO Jensen Huang,These AI data centersare,in fact,AI factories.That

268、race is moving faster than many expected.The most striking example may be xAIs Colossus facility in Memphis,Tennessee which went from a gutted factory to a fully operational AI data center in just 122 days.As noted on page 122,at 750,000 square feet roughly the size of 418 average USA homes it was b

269、uilt in half the time it typically takes to construct a single American house.Per NVIDIA Co-Founder&CEO Jensen Huang,What they achieved is singular,never been done beforeThat is,like,superhuman119Data Centers=Key Beneficiary of AI CapEx SpendThese kinds of timelines are no longer the exception.W

270、ith prefabricated modules,streamlined permitting,and vertical integration across electrical,mechanical,and software systems,new data centers aregoing up at speeds that resemble consumer tech cycles more than real estate development.But beneath that velocity lies a capital model thats anything but si

271、mple.CapEx is driven by land,power provisioning,chips,and cooling infrastructure especially as AI workloads push thermal and power limits far beyond traditional enterprise compute.OpEx,by contrast,is dominated by energy costs and systems maintenance,particularly for high-density training clusters th

272、at operate near constant load.Revenue is driven by compute sales whether in the form of AI APIs,enterprise platform fees,orinternal productivity gains.But payback periods are often long,especially for vertically-integrated playersbuilding ahead of demand.For newer entrants,monetization may lag build

273、-out by quarters or even years.And then theres the supply chain.Power availability is becoming more of a gating factor.Transformers,substations,turbines,GPUs,cables these arent commodities that can be spun up overnight.In this context,data centers arent just physical assets they are strategic infras

274、tructure nodes.They sit at the intersection of real estate,power,logistics,compute,and software monetization.The companies that get this right may do more than run servers they will shape the geography of AI economics for the next decade.120Data Center Buildout Construction Value,USA=+49%&Accele

275、rated Annual Growth Over Two YearsNote:All data are seasonally adjusted.Data obtained via USA Census Bureaus Value of Construction Put in Place(VIP)Survey,which provides monthly estimates of the total dollar value of construction work done in USA.Data is annualized to avoid seasonal fluctuations.Sou

276、rce:USA Census Bureau$0$10$20$30$401/141/151/161/171/181/191/201/211/221/231/24Annualized Value of Private Construction,$B12/24+28%/Year+49%/YearData Centers=Key Beneficiary of AI CapEx SpendUSA Data Center Annualized Private Construction Value($B)1/14-12/24,per USA Census Bureau121Data Center New C

277、onstruction vs.Existing Capacity,USA=+16x in New vs.+5x in Existing Over Four YearsNote:Primary USA markets only(Northern Virginia,Atlanta,Chicago,Phoenix,Dallas-Ft.Worth,Hillsboro,Silicon Valley,New York Tri-State.)Source:CBRE,North America Data Center Trends H2 2024(2/25)Data center Capacity,Megaw

278、atts04,0008,00012,00020202021202220232024Net AbsorptionPre-Leased or Under ConstructionExisting capacity but newly-filledNew capacity+16x+5xData Centers=Key Beneficiary of AI CapEx SpendData Center Capacity(Megawatts)by Real Estate Profile,USA Primary Markets 2020-2024,per CBRE122Data Center Build T

279、ime(xAI Colossus as Proxy)=122 Days vs.234 for a HomeNote:Median USA home size shown as of January 2025,per FRED.Colossus was built in a former Electrolux factory in Memphis,TN,USA.Average build time shown for single-unit buildings.Measures time between start of onsite work&completion.Data repor

280、ted in 2024 but measures build times for homes started in 2023.Source:xAI,USA Census Bureau,Federal Reserve Bank of St.Louis,Wikimedia Commons122 Days=A Fully-Operational Data Center 2024750,000 Sq.Ft=Size of 418 USA Homes122 Days=One Half-Built House 2024(Average Build Time=234 Days)750,000 Square

281、Feet1,792 Square FeetWe were told it would take 24 months to build.So we took the project into our own hands,questioned everything,removed whatever was unnecessary,and accomplished our goal in four months.-xAI WebsiteData Centers=Key Beneficiary of AI CapEx Spend123Data Center Compute(xAI Colossus a

282、s Proxy)=0 to 200,000 GPUs in Seven MonthsData Centers=Key Beneficiary of AI CapEx SpendxAI Colossus GPUs 4/24-11/24,per xAINote:We assume 200,000 GPUs as of 11/30/24 per xAIs disclosure that we doubled GPU count in 92 days to 200K GPUs.xAI Colossus ran its first job across 4 data halls on 8/30/24.W

283、e assume zero GPUs as of construction start date(122 days prior to assumed opening date of 8/30/24).Source:xAI(5/25),Memphis Chamber of Commerce(12/24)Were running the worlds biggest supercomputer,Colossus.Built in 122 days outpacing every estimate it was the most powerful AI training system yet.The

284、n we doubled it in 92 days to 200k GPUs.This is just the beginningWe doubled our compute at an unprecedented rate,with a roadmap to 1M GPUs.Progress in AI is driven by compute and no one has come close to building at this magnitude and speed.-xAI Website,5/25GPUs,KxAI Colossus GPUs(K)0K100K200K01002

285、004/248/2411/24xAI ultimately plans on 1MM GPUs,per Memphis Chamber of Commerce 124Data Centers=Electricity Guzzlers125Data Centers=Electricity GuzzlersAI and energy observations/quotes(in italics)here and the two pages that follow are fromWorld Energy Outlook Special Report Energy and AI(link)from

286、IEA(International Energy Agency)*4/10/25To understand where energy infrastructure is heading,it helps to examine the rising tension between AI capability and electrical supply.The growing scale and sophistication of artificial intelligence is demanding an extraordinary amount of computational horsep

287、ower,primarily from AI-focused data centers.These facilities purpose-built to train and serve models are starting to rival traditional heavy industry in their electricity consumption.There is no AI without energy specifically electricity(p.3).Data centers accounted for around 1.5%of the worldselectr

288、icity consumption in 2024(p.14).Energy demand growth has been rapid:Globally,data centre electricity consumption has grown by around 12%per year since 2017,more than four times faster than the rate of total electricity consumption(p.14).As power demand rises,so too does its concentration:The United

289、States accounted for45%of global data centre electricity consumption,followed by China(25%)and Europe(15%)nearly half of data centre capacity in the United States is in five regional clusters(p.14).The flipside is true as well:Emerging and developing economies other than China account for 50%of thew

290、orlds internet users but less than 10%of global data centre capacity(p.18)126Data Centers=Electricity GuzzlersAIs power demands are increasing and its progress is increasingly bottlenecked not bydata or algorithms,but by the grid and strains related to demand.While AI presently places considerable d

291、emands on the energy sector,it is also already unlocking major energy efficiency and operational gainsAI is already being deployed by energy companies to transform and optimize energy and mineral supply,electricity generation and transmission,and energy consumption(p.16).Current AI-driven demand is

292、extremely high.This is forecast to continue,especially as capital gushes into model providers that,in turn,spend on more compute.At some point,these model builders will need to turn a profit to be able to spend more.While demand for both compute and energy will inevitably continue to rise as consume

293、r andbusiness usage does the same,data centers will ultimately only serve those who pay their bills.*IEA member countries include Australia,Austria,Belgium,Canada,Czech Republic,Denmark,Estonia,Finland,France,Germany,Greece,Hungary,Ireland,Italy,Japan,S.Korea,Latvia,Lithuania,Luxembourg,Mexico,Nethe

294、rlands,New Zealand,Norway,Poland,Portugal,Slovak Republic,Spain,Sweden,Switzerland,Republic of Trkiye,United Kingdom,and United States.IEA Association countries include Argentina,Brazil,China,Egypt,India,Indonesia,Kenya,Morocco,Senegal,Singapore,S.Africa,Thailand,and Ukraine.All data shown,unless ot

295、herwise specified,is global.Italicized text is directly quoted from the report.127Data Center Electricity Consumption,Global=+3x Over Nineteen Years,per IEAData Center Energy Consumption by Data Center Type&Equipment,Global 2005-2024,per IEASource:International Energy Agency(IEA),Energy and AI(4

296、/25)Data Centers=Electricity Guzzlers128Data Center Electricity Consumption by Region=USA Leads,per IEAData Center Electricity Consumption by Region 2005-2024,per IEASource:International Energy Agency(IEA),Energy and AI(4/25)Data Centers=Electricity GuzzlersSeem Like Change Happening Faster Than Eve

297、r?Yes,It IsAI User+Usage+CapEx Growth=UnprecedentedAI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage RisingAI Usage+Cost+Loss Growth=UnprecedentedAI Monetization Threats=Rising Competition+Open-Source Momentum+Chinas RiseAI&Physical World

298、 Ramps=Fast+Data-DrivenGlobal Internet User Ramps Powered by AI from Get-Go=Growth We Have Not Seen Likes of BeforeAI&Work Evolution=Real+Rapid12912345678Outline130AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage RisingTo understand wher

299、e AI model economics may be heading,one can look at the mounting tension between capabilities and costs.Training the most powerful large language models(LLMs)has become one of the most expensive/capital-intensiveefforts in human history.As the frontier of performance pushes toward ever-larger parame

300、ter counts andmore complex architectures,model training costs are rising into the billions of dollars.Ironically,this race to build the most capable general-purpose models may be accelerating commoditization and drivingdiminishing returns,as output quality converges across players and differentiatio

301、n becomes harder to sustain.At the same time,the cost of applying/using these models known as inference is falling quickly.Hardware is improving for example,NVIDIAs 2024 Blackwell GPU consumes 105,000 x less energy per tokenthan its 2014 Kepler GPU predecessor.Couple that with breakthroughs in model

302、s algorithmic efficiency,and the cost of inference is plummeting.Inference represents a new cost curve,and unlike training costs its arcing down,not up.As inference becomes cheaper and more efficient,the competitive pressure amongst LLM providers increases not on accuracy alone,but also on latency,u

303、ptime,and cost-per-token*.What used to cost dollars can now cost pennies.And what cost pennies may soon cost fractions of a cent.The implications are still unfolding.For users(and developers),this shift is a gift:dramatically lower unit costs to access powerful AI.And as end-user costs decline,creat

304、ion of new products and services is flourishing,and user and usage adoption is rising.For model providers,however,this raises real questions about monetization and profits.Training is expensive,serving is getting cheap,and pricing power is slipping.The business model is in flux.And there are newques

305、tions about the one-size-fits-all LLM approach,with smaller,cheaper models trained for custom use cases*now emerging.Will providers try to build horizontal platforms?Will they dive into specialized applications?Only time will tell.In the short term,its hard to ignore that the economics of general-pu

306、rpose LLMslook like commodity businesses with venture-scale burn.*Cost-per-token=The expense incurred for processing or generating a single token(a word,sub-word,or character)during the operation of a language model.It is a key metric used to evaluate the computational efficiency and cost-effectiven

307、ess of deploying AI models,particularly in applications like natural language processing.*E.g.,OpenEvidence131AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage Rising132AI Model Training Compute Costs=2,400 x Growth Over Eight Years,per Epoch

308、 AI&StanfordNote:Costs are estimates.Excludes most Chinese models due to lack of reliable cost data.Source:Epoch AI via Nestor Maslej et al.,The AI Index 2025 Annual Report,AI Index Steering Committee,Stanford HAI(4/25);In Good Company podcast(6/24)Estimated Training Cost of Frontier AI Models 2

309、016-2024,per Epoch AI&StanfordTraining Cost,USD(Log Scale)Approx.+2,400 xRight now,AI model training costs$100 million.There are models in training today that are more like a billion.Right.I think if we go to$10 or$100 billion,and I think that will happen in 2025,2026,maybe 2027I think that the

310、training of$10 billion models,yeah,could start sometime in 2025.-Anthropic Co-Founder&CEO Dario Amodei(6/24)AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage Rising133AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Pe

311、rformance Converging+Developer Usage Rising134To understand the trajectory of AI compute,it helps to revisit an idea from the early days of PC software.Software is a gasit expands to fill its container,said Nathan Myhrvold,then CTO of Microsoft in 1997.AI is proving no different.As models get better

312、,usage increases and as usage increases,so does demand for compute.Were seeing it across every layer:more queries,more models,more tokens per task.The appetite for AI isnt slowing down.Its growing into every available resource just like software did in the age of desktop and cloud.But infrastructure

313、 is not just standing still.In fact,its advancing faster than almost any other layer in the stack,and at unprecedented rates.As noted on page 136,NVIDIAs 2024 Blackwell GPU uses 105,000 times less energy to generate tokens than its 2014 Kepler predecessor.Its a staggering leap,and it tells a deeper

314、story not just of cost reduction,but of architectural and materials innovation that is reshaping whats possible at the hardware level.These improvements in hardware efficiency are critical to offset the strain of increasing AI and internet usage on our grid.So far,though,they have not been enough.Th

315、is trend aligns with Jevons Paradox,first proposed back in 1865*that technological advancements that improve resource efficiency actually lead to increased overall usage of those resources.This is driving new focus on expanding energy production capacity and new questions about the grids ability to

316、manage.Yet again,we see this as one of the perpetual a-has of technology:costs fall,performance rises,and usage grows,all in tandem.This trend is repeating itself with AI.*British economist William Stanley Jevons first observed this phenomenon in 19th-century Britain,where he noticed that improvemen

317、ts in the efficiency of coal-powered steam engines were not reducing coal consumption but rather increasing it.In his book The Coal Question,he noted It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to diminished consumption.The very contrary is the truth.AI

318、 Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage Rising135AI Inference Currency=Tokens*Assumes that the average ChatGPT interaction consumes 200 total tokens(input+output),or 150 words.Thus,1MM tokens equates to roughly 5,000 ChatGPT responses

319、.Source:OpenAI(1/25)Additional context:1MM tokens=750,000 wordsroughly 3,500 pages of a standard book(12-point font,double-spaced)5,000 ChatGPT responses*AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage Rising136AI Inference Costs NVIDIA GPU

320、s=-105,000 x Decline in Energy Required to Generate Token Over Ten YearsEnergy Consumption per Token(Joules)AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage Rising025,00050,000Kepler(2014)Pascal(2016)Volta(2018)Ampere(2020)Hopper(2022)Blackw

321、ell(2024)-105,000 xNote:Kepler released in 2012.NVIDIA materials mark performance threshold shown above for Kepler as of 2014.Source:NVIDIA Company Overview(2/25)Energy Required per LLM Token(Joules),NVIDIA GPUs 2014-2024,per NVIDIA137AI Inference Costs Serving Models=99.7%Lower Over Two Years,per S

322、tanford HAISource:Nestor Maslej et al.,The AI Index 2025 Annual Report,AI Index Steering Committee,Stanford HAI(4/25)AI Inference Price for Customers(per 1 Million Tokens)11/22-12/24,per Stanford HAINote:Axis is logarithmic;every axis tick represents a 10 x price changeAI Model Compute Costs High/Ri

323、sing+Inference Costs Per Token Falling=Performance Converging+Developer Usage Rising138AI Cost Efficiency Gains=Happening Faster vs.Prior TechnologiesNote:Price change in consumer goods and services in the United States is measured as the percentage change since 1997.Data is based on the consumer pr

324、ice index(CPI)for national average urban consumer prices.Per OpenAI,100 AI tokens generates approximately 75 words in a large language model response;data shown indexes to this number of tokens.Year 0 is not necessarily the year that the technology was introduced,but rather the first year of availab

325、le data.Source:Electricity Costs Technology and Transformation in the American Electric Utility Industry,Richard Hirsh(1989);Computer Memory Storage Costs John C.McCallum,with data aggregated from 72 primary sources and historical company sales documents;OpenAI,Wikimedia CommonsRelative Cost of Key

326、Technologies by Year Since Launch,per OpenAI,John McCallum,&Richard Hirsh0%25%50%75%100%020406080Electric PowerComputer MemoryChatGPT:75-Word Response%of Original Price By Year(Indexed to Year 0)AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer

327、 Usage Rising139Techs Perpetual A-Ha=Declining Costs+Improving Performance Rising AdoptionNote:FRED data shows Consumer Price Index for All Urban Consumers:Information Technology,Hardware and Services in U.S.City Average.Source:USA Federal Reserve Bank of St.Louis(FRED),International Telecommunicati

328、ons Union(via World Bank)(4/25)0255075100010020030040019891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023USA Internet Users,MM(Blue Bars)USA Consumer Price Index:Information Technology,1/1/89=100(Red Line)AI Mod

329、el Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage RisingUSA Internet Users(MM)vs.Relative IT Cost 1989-2023,per FRED&ITU1989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019

330、202020212022202320240102030405060708090100140Techs Perpetual A-Ha=Prices Fall+Performance RisesNote:A petaFLOP/s-day represents the total computational work performed by a system operating at 1 petaFLOP/s(10 floating-point operations per second)for 24 hours,equivalent to approximately 8.64 10 operat

331、ions.This metric is commonly used to quantify the compute required for large-scale tasks like training machine learning models.FRED data shows Consumer Price Index for All Urban Consumers:Information Technology,Hardware and Services in U.S.City Average.Note that,while training compute is not a direc

332、t measurement of model performance,it is typically closely correlated with performance.Source:USA Federal Reserve Bank of St.Louis(FRED);Epoch AI(5/25)Notable AI Model Training Compute,FLOP(Scatter Plot)USA Consumer Price Index:Information Technology,1/1/89=100(Red Line)+360%/Year AI Model Compute C

333、osts High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage RisingAI Model Training Compute(FLOP)vs.Relative IT Cost 1989-2024,per Epoch AI&FRED141AI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Developer Usage Rising142AI

334、 Model Performance=Converging Rapidly,per Stanford HAIPerformance of Top AI Models on LMSYS Chatbot Arena 1/24-2/25,per Stanford HAINote:The LMSYS Chatbot Arena is a public website where people compare two AI chatbots by asking them the same question and voting on which answer is better.The results

335、help rank how well different language models perform based on human judgment.Only the highest-scoring model in any given month is shown in this comparison.Source:Nestor Maslej et al.,The AI Index 2025 Annual Report,AI Index Steering Committee,Stanford HAI(4/25)LMSYS Arena ScoreAI Model Compute Costs High/Rising+Inference Costs Per Token Falling=Performance Converging+Develo