《Coatue:2023年人工智能革命报告(英文版)(115页).pdf》由会员分享,可在线阅读,更多相关《Coatue:2023年人工智能革命报告(英文版)(115页).pdf(115页珍藏版)》请在三个皮匠报告上搜索。
1、The AI Revolution Sri Viswanath Vibhor Khanna Yijia LiangCoatueNovember 20232DisclaimerDisclaimerTHIS PRESENTATION DOES NOT CONSTITUTE AN OFFER TO SELL OR THE SOLICITATION OF ANY OFFER TO BUY AN INTEREST IN A FUND OR VEHICTHIS PRESENTATION DOES NOT CONSTITUTE AN OFFER TO SELL OR THE SOLICITATION OF
2、ANY OFFER TO BUY AN INTEREST IN A FUND OR VEHICLE LE MANAGED BY COATUE MANAGEMENT,MANAGED BY COATUE MANAGEMENT,L.L.C.(“COATUE”).L.L.C.(“COATUE”).ANY SUCH OFFER OR SOLICITATION WILL BE MADE ONLY AT THE TIME A QUALIFIED OFFEREE RECEIVES A CONFIDENTIAL PRIVATE OFFERING MEMANY SUCH OFFER OR SOLICITATION
3、 WILL BE MADE ONLY AT THE TIME A QUALIFIED OFFEREE RECEIVES A CONFIDENTIAL PRIVATE OFFERING MEMORAORANDUM OR OTHER OFFERING DOCUMENT NDUM OR OTHER OFFERING DOCUMENT(“CPOM”)DESCRIBING THE OFFERING AND RELATED SUBSCRIPTION AGREEMENT.(“CPOM”)DESCRIBING THE OFFERING AND RELATED SUBSCRIPTION AGREEMENT.IN
4、 THE CASE OF ANY INCONSISTENCY BETWEEN THE DESCRIPTIONS OR TERMS IN THIS PRESENTATION AND THE CPOM,THE CPOM SHALL CONTROL.IN THE CASE OF ANY INCONSISTENCY BETWEEN THE DESCRIPTIONS OR TERMS IN THIS PRESENTATION AND THE CPOM,THE CPOM SHALL CONTROL.COATUE HAS PROVIDED THIS PRESENTATION SOLELY IN CONNEC
5、TION WITH ONGOING OR INTENDED DISCUSSIONS WITH THE PERSON TO WHOM IT HACOATUE HAS PROVIDED THIS PRESENTATION SOLELY IN CONNECTION WITH ONGOING OR INTENDED DISCUSSIONS WITH THE PERSON TO WHOM IT HAS BS BEEN DELIVERED.IN THIS REGARD,EEN DELIVERED.IN THIS REGARD,THE PERFORMANCE FIGURES,AND CERTAIN OTHE
6、R INFORMATION INCLUDED IN THIS PRESENTATION,REQUIRE FURTHER EXPLANATION AND MUST BETHE PERFORMANCE FIGURES,AND CERTAIN OTHER INFORMATION INCLUDED IN THIS PRESENTATION,REQUIRE FURTHER EXPLANATION AND MUST BE DIDISCUSSED WITH YOUR COATUE SCUSSED WITH YOUR COATUE REPRESENTATIVES.ACCORDINGLY,PROSPECTIVE
7、 INVETORS SHOULD NOT RELY ON THE INFORMATION IN THIS PRESENTATION ABSENT SUCH DISCUREPRESENTATIVES.ACCORDINGLY,PROSPECTIVE INVETORS SHOULD NOT RELY ON THE INFORMATION IN THIS PRESENTATION ABSENT SUCH DISCUSSISSIONS.ONS.NO SECURITIES OR SERVICES SHALL BE OFFERED OR SOLD IN ANY JURISDICTION IN WHICH S
8、UCH OFFER,SOLICITATION OR SALE WOULD BE UNLNO SECURITIES OR SERVICES SHALL BE OFFERED OR SOLD IN ANY JURISDICTION IN WHICH SUCH OFFER,SOLICITATION OR SALE WOULD BE UNLAWFAWFUL UNTIL THE REQUIREMENTS OF THE UL UNTIL THE REQUIREMENTS OF THE LAWS OF SUCH JURISDICTION HAVE BEEN SATISFIED.WHILE ALL THE I
9、NFORMATION PREPARED IN THIS PRESENTATION IS BELIEVED TO BE ACCULAWS OF SUCH JURISDICTION HAVE BEEN SATISFIED.WHILE ALL THE INFORMATION PREPARED IN THIS PRESENTATION IS BELIEVED TO BE ACCURATRATE,COATUE MAKES NO EXPRESS E,COATUE MAKES NO EXPRESS WARRANTY AS TO THE COMPLETENESS OR ACCURACY NOR CAN IT
10、ACCEPT RESPONSIBILITY FOR ERRORS,APPEARING IN THE PRESENTATION.NEITWARRANTY AS TO THE COMPLETENESS OR ACCURACY NOR CAN IT ACCEPT RESPONSIBILITY FOR ERRORS,APPEARING IN THE PRESENTATION.NEITHERHER COATUE NOR ITS AFFILIATES COATUE NOR ITS AFFILIATES ASSUMES ANY DUTY TO UPDATE THE INFORMATION CONTAINED
11、 HEREIN FOR SUBSEQUENT CHANGES OF ANY KIND AND THERE IS NO ASSURANCE THATASSUMES ANY DUTY TO UPDATE THE INFORMATION CONTAINED HEREIN FOR SUBSEQUENT CHANGES OF ANY KIND AND THERE IS NO ASSURANCE THAT THTHE POLICIES,STRATEGIES OR E POLICIES,STRATEGIES OR APPROACHES DISCUSSED HEREIN WILL NOT CHANGE.APP
12、ROACHES DISCUSSED HEREIN WILL NOT CHANGE.ANY PROJECTIONS,MARKET OUTLOOKS OR ESTIMATES IN THIS PRESENTATION ARE FORWARDANY PROJECTIONS,MARKET OUTLOOKS OR ESTIMATES IN THIS PRESENTATION ARE FORWARD-LOOKING STATEMENTS AND ARE BASED UPON CERTAIN ASSLOOKING STATEMENTS AND ARE BASED UPON CERTAIN ASSUMPTIO
13、NS.OTHER EVENTS WHICH UMPTIONS.OTHER EVENTS WHICH WERE NOT TAKEN INTO ACCOUNT MAY OCCUR AND MAY SIGNIFICANTLY AFFECT INVESTMENT RETURNS OR PERFORMANCE.WERE NOT TAKEN INTO ACCOUNT MAY OCCUR AND MAY SIGNIFICANTLY AFFECT INVESTMENT RETURNS OR PERFORMANCE.ANY PROJECTIONS,OUTLOOKS OR ASSUMPTIONS SHOULD N
14、OT BE CONSTRUED TO BE INDICATIVE OF THE ACTUAL EVENTS WHICH WILL OCCUR.HISTANY PROJECTIONS,OUTLOOKS OR ASSUMPTIONS SHOULD NOT BE CONSTRUED TO BE INDICATIVE OF THE ACTUAL EVENTS WHICH WILL OCCUR.HISTORIORICAL RETURNS ARE NOT PREDICTIVE CAL RETURNS ARE NOT PREDICTIVE OF FUTURE RESULTS.OF FUTURE RESULT
15、S.THE INFORMATION PROVIDED HEREIN,INCLUDING,WITHOUT LIMITATION,INVESTMENT STRATEGIES,INVESTMENT RESTRICTIONS AND PARAMETERSTHE INFORMATION PROVIDED HEREIN,INCLUDING,WITHOUT LIMITATION,INVESTMENT STRATEGIES,INVESTMENT RESTRICTIONS AND PARAMETERS,P,PROCEDURES,POLICIES,AND ROCEDURES,POLICIES,AND INVEST
16、MENT AND OTHER PERSONNEL MAY BE CHANGED,MODIFIED,TERMINATED OR SUPPLEMENTED AT ANY TIME,WITHOUT NOTICE.INVESTMENT AND OTHER PERSONNEL MAY BE CHANGED,MODIFIED,TERMINATED OR SUPPLEMENTED AT ANY TIME,WITHOUT NOTICE.THIS PRESENTATION IS CONFIDENTIAL AND NOT INTENDED FOR PUBLIC USE OR DISTRIBUTION.THIS P
17、RESENTATION HAS NOT BEEN APPROVED BYTHIS PRESENTATION IS CONFIDENTIAL AND NOT INTENDED FOR PUBLIC USE OR DISTRIBUTION.THIS PRESENTATION HAS NOT BEEN APPROVED BY THTHE U.S.SECURITIES AND EXCHANGE E U.S.SECURITIES AND EXCHANGE COMMISSION OR ANY OTHER FEDERAL OR STATE REGULATOR.COMMISSION OR ANY OTHER
18、FEDERAL OR STATE REGULATOR.YOU SHOULD CAREFULLY READ THE IMPORTANT DISCLOSURES IN THE APPENDIX.YOU SHOULD CAREFULLY READ THE IMPORTANT DISCLOSURES IN THE APPENDIX.3Where we are in AI todayKey TopicsWere at Day 1 of AIand riding on top of past waves4Source:World Bank as of September 2023,Statista as
19、of September 2023,Coatue analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Dis
20、closures for important disclosures,including regarding projections and forward-looking statements and trends.%US Technology Adoption200002005205%50%75%100%InternetSmartphonesAISource:Bloomberg;Coatue opinion and analysis as of October 2023.AI Basket represents GSs AI basket.For
21、 illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections an
22、d forward-looking statements and trends.In the first half of 2023,the AI ecosystem exploded!(1/2)AI Basket Performance YTDChatGPT weekly unique visitsNVDA 2Q Earnings512/2201/2302/2303/2304/2305/2306/230M10M20M30M40MGPT-4 ReleasedGPT-3.5 ReleasedOnly 5 months!$7B$11BStreet estimatesActual guide+60%Y
23、TD1001/233/236/239/23202,5003,0004,5006,0007,5009,000Source:GitHub as of June 2023,ThereIsAnAIForT as of September 2023,Hugging Face as of September 2023,and Coatue analysis and opinion as of September 2023.Data visualized begins in 2021.For illustrative purposes only.There is no guarante
24、e that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable statements or trends;do not necessarily represent Coatue investments.See Appendix-Di
25、sclosures for important disclosures,including regarding projections and forward-looking statements and trends.In the first half of 2023,the AI ecosystem exploded!(2/2)Number of models on Hugging Face%of trending Github repos in AI/ML New AI applications over time340K+8K+62021202220230K50K100K150K200
26、K250K300K350K2021202220230%10%20%30%40%50%60%50%Adoption has been twice as fast with each platform shift7Source:Morgan Stanley Research as of September 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or th
27、at any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Halving of penetration time with new technology waves20 Years12 Years6 Years3 YearsInternetPCsGen AI
28、Mobile?#of years to reach 50%user penetration in the USAI has potential to drive the economy for years to come8Source:Bloomberg,CapitalIQ,and Coatue analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future p
29、otential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.#of employees per$1M of revenue(inflation adjusted S&P 500 companie
30、s)7.87.86.46.45.15.15.35.35.15.11 hop)flows of talentINFLOW:25 people joined OpenAI from Google DeepMindOUTFLOW:18 people joined Google DeepMind from FAIRMany talented people are now leaving to start new companies53Source:Coatue analysis and opinion as of November 2023.For illustrative purposes only
31、.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable statements or trends;do not necessarily represent Coatue invest
32、ments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Example AI mafiasModelsCloud,DC,SemisDeveloper ToolsApplicationsFounders from:Number of founders from AI research housesExample companies founded:3632271554Scaling AI mo
33、del performance has been a focus of AI researchers Will scaling parameters Will scaling parameters continue improving performance?More parametersMore parametersGPT-4 still“king”;scaling experiments likely continuingTraining&inference is getting optimized;AI at the edge is emergingLarger models need
34、more data;extending data runway&optimizing quality crucialHow much data is optimal How much data is optimal for training models?Larger datasetsLarger datasetsCan we reduce the compute costs reduce the compute costs of training&inference?More computeMore computeCan training for more epochsCan trainin
35、g for more epochs improve performance?Longer trainingLonger trainingStill an open questionCovered in next section on Cloud,DC,SemisNot covered;waiting for more data points to emergeSource:Coatue analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatue
36、s views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisD
37、eveloper ToolsApplicationsStrategies to scaleOpen research questionsCoatue View?Focus of remainder of sectionThrough 2022:We saw LLM parameters balloon!55Source:Arxiv as of September 2023,Hugging Face as of September 2023,Coatue analysis and opinion as of November 2023.For illustrative purposes only
38、.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Model names listed above are examples of the applicable AI models;do not necessarily represent Coatue investment
39、s.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Illustrative view of parameter sizes of large language models through 2022ModelsCloud,DC,SemisDeveloper ToolsApplications01/201807/1801/1907/1901/2007/2001/2107/2101/2207/22
40、0.010.101.0010.00100.001,000.0010,000.00PartiParameter size(B,log)LaMDAAlexaTM 20BMinerva(540B)GPT-NeoX-20BHuBERTCodexFLANGPT-3 175B(davinci)ALBERT-xxlargePaLM(540B)GPT-2XLNetRoBERTa LargeBART-largeT5-11BGPTELMoBLOOMChinchillaBERT-Large=Relative model parameter sizeMoores Law:doubling every 2 years2
41、023:GPT-4 remains“king”can we continue scaling?56Source:LMSYS,Hugging Face,Semianalysis,Coatue analysis and opinion as of November 2023.Selected benchmarks are examples and are not exhaustive nor are they representative of objective“truth”about which models are better.For illustrative purposes only.
42、There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Model names listed above are examples of the applicable AI models;do not necessarily represent Coatue investments
43、.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Comparison of 2023 models across ELO ranking and MMLU benchmarksGPT-4 is still the best ModelsCloud,DC,SemisDeveloper ToolsApplications=Relative model parameter sizeHigher EL
44、O is better9009501,0001,0501,1001,1501,2001,25000708090100ELO ranking(human feedback)Model performance on MMLU benchmarkGPT-4GPT-3.5LLAMA-2 70BVicuna-33BMPT-30BVicuna-13BVicuna-7BMPT-7BMistral-7B InstructCoatue opinion and analysis as of November 2023.For illustrative purposes only.There
45、is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and t
46、rends.What if there was a$1B training-cost model?Model Training Cost Over Time57ModelTrainingCostTimeNextNext-Gen ModelGen ModelToday Today GPT 4GPT 4GPT 3GPT 3GPT 2GPT 2$50K$100M$1B?$10MWhat would such a model be capable of?ModelsCloud,DC,SemisDeveloper ToolsApplicationsHow AI models are made:Data
47、is a crucial component58Source:Coatue analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Log
48、os listed above as examples of the applicable statements or trends;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Data is upstream in process of developing good modelsModelsC
49、loud,DC,SemisDeveloper ToolsApplicationsData prepData prepData evaluation&curationData evaluation&curationModel trainingModel trainingModel deploymentModel deploymentPublic web dataPublic web dataCustom domain dataCustom domain dataProprietary dataProprietary dataTraining&deployment frameworksTraini
50、ng&deployment frameworksCompute providersCompute providersModel evaluationModel evaluationHosted inferenceHosted inferenceTraining infra(preTraining infra(pre-training&finetraining&fine-tuning)tuning)Select data to train model on,across various sourcesClean&curate to improve dataset qualityCreate cu
51、stom model architecture using training frameworksTrain&iterate models to improve performanceDeploy models&continuously evaluate to fine-tune models Scaling models require scaling datasets59Source:Arxiv,Hugging Face,Meta,Semianalysis,PapersWithCode,CNBC,Coatue analysis and opinion as of November 2023
52、.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Model names listed above are examples of the applicable AI models;do not necessar
53、ily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Number of tokens in training datasets increasing overall04/2007/2010/2001/2104/2107/2110/2101/2204/2207/2210/2201/2304/2307/231001,00010,00010
54、0,000Training token dataset size(log,B)GPT-4GPT-3LLAMA-2LLAMA-1PALMPALM-2BLOOMMPT-30BCode LLAMAGPT-4 rumored to have trained on 13T tokens!ModelsCloud,DC,SemisDeveloper ToolsApplications=Relative model parameter sizeData scarcity is a potential wall to scaling models60Source:Arxiv(Villalobos 2022),C
55、oatue analysis and opinion as of November 2023.For illustrative purposes only;there is no guarantee that data will be exhausted or that image and video data will have a longer runway.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that an
56、y particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.High-quality text data could be exhausted soon,images&video have longer runway Illustrative:Amount of dat
57、a collected relative to overall supplyModelsCloud,DC,SemisDeveloper ToolsApplications200242026202820302032203420362038204020422044204620482050Probable that high quality text exhausted by 2026?Image+video data estimated to last until 2040s,yet will be used to train different types of model
58、sHigh quality textImage+videoOverall data supply(relative for each data type)Overall data supply(relative for each data type)Data quality is just as important as data quantity61Source:Cohere as of September 2023,note that tested model sizes range from 124M to 1.5B.Coatue analysis and opinion as of N
59、ovember 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable statements or trends
60、;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Emerging evidence that training on pruned datasets can result in similar performance in language modelsPerformance on variety
61、of language model benchmarks78%64%65%50%79%66%67%52%0%30%60%90%10%20%40%50%70%80%100%Sentiment analysisSemantic similarityText classificationNatural language inferenceFull datasetPruned dataset(30%of size)For up to a 1.5B parameter model,pre-training on just 30%of the dataset size can result in simi
62、lar,if not better performance,across multiple tests!Can this generalize to even larger models?ModelsCloud,DC,SemisDeveloper ToolsApplicationsSynthetic data can augment fine-tuning62Source:Arxiv(Tang 2023),Coatue analysis and opinion,as of November 2023.For illustrative purposes only.There is no guar
63、antee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Synt
64、hetically generated data can help clean or distill datasets for fine-tuningModelsCloud,DC,SemisDeveloper ToolsApplicationsResults:Results:Synthetic data fine-tuning yields significant improvement compared to zero-shot attempts,though is not yet comparable to real dataPercentile performance on test19
65、3374785568778480858187Named entity recognition(Precision)Named entity recognition(Recall)Relation extraction(Precision)Relation extraction(Recall)+196%+196%+103%+103%Zero-shot ChatGPTFine-tuned on synthetic data(BERT)Fine-tuned on real data(BERT)Zero-shot ChatGPT already decent at some bio tasks!63U
66、ser feedback data is another way to improve performance Midjourney collects user feedback to continue refining models Source:AITuts,DIscord,Coatue analysis and opinion,as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future pot
67、ential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Users choose which image to upscale(or generate a variant of),effecti
68、vely choosing their preferred imagePromptPrompt/dungeons and dragons,female knight,of the rolling plains,full body,dark azure,victorian genre paintings,serene face,realistic depiction of light,golden lightV5(Mar 2023)Midjourney V1(Feb 2022)V2(Apr 2022)V3(Jul 2022)V4(Nov 2022)ModelsCloud,DC,SemisDeve
69、loper ToolsApplicationsV5.2(Jun 2023)Note:Midjourney improvement is partially due to collecting user feedback but is only one part of the equation!There are many other opportunities to improve models!64Source:Coatue analysis and opinion,as of November 2023.For illustrative purposes only.There is no
70、guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.
71、Beyond scaling parameters&data,there are many angles to push on TimeNon-exhaustive,illustrative opportunity setLonger contextsMulti-modalityLocal modelsBetter model evalNew architecturesModelsCloud,DC,SemisDeveloper ToolsApplicationsAI modelsAI appsEnd usersAI developer toolsCloud platformsData cent
72、er/hardware/powerCloud platformsData center/hardware/powerAI has potential to re-accelerate underlying infrastructureSource:Coatue analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are
73、 accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-lo
74、oking statements and trends.This AI wave has been extremely compute hungry66Source:Arxiv,Coatue analysis and opinion,as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Co
75、atue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ML models are doubling compute needs for training in monthsMonths to double FLOPS used in model training21.3 Months21.3
76、Months5.7 Months5.7 MonthsPre-Deep Learning EraDeep Learning Era-73%73%+Training compute needs now grows faster than Moores LawModelsCloud,DC,SemisDeveloper ToolsApplicationsTraining compute has exponentially increased with model sizes67Source:Epoch AI as of September 2023,Coatue opinio
77、n and analysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Model names listed above as examples of the app
78、licable AI models;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Massive inflection in compute needs for AI,70X increase from GPT-3 to GPT-401/1807/1801/1907/1901/2007/2001/2
79、107/2101/2207/2201/2307/230.010.101.0010.00100.001,000.0010,000.00100,000.001,000,000.00Computational Petaflops(M,log scalelog scale)Chinchilla-70BGPT-120MGPT-2-1.5BGPT-3-175BGPT-4-1.8TGPT-NeoX-20BGato-1.2BHuBERT-375MLLaMA-65BALBERT-xxlarge-223MLaMDA-137BMinerva-540BAlphaCode-41BNLLB-55BAlexaTM-20BB
80、ERT-Large-345MSparse all-MLP-10BT5-11BT5-3BPaLM-540B=Relative model parameter sizeModelsCloud,DC,SemisDeveloper ToolsApplicationsLatest executive order requires robust reporting for models trained using more than 100B PFLOPs(1026 FLOPs).This is 5X larger than GPT-4s rumored training compute of 21B P
81、FLOPsInference compute likely to dramatically outpace training68Source:Semianalysis,Coatue opinion and analysis as of November 2023.Assumes 560 TFLOPS/token,150 tokens per query.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of
82、AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Ongoing inference likely to require much more compute than one-time training020
83、406080100Computational Petaflops(B)GPT-4 Training1M GPT-4 users10M GPT-4 users100M GPT-4 users21B1B8B80B+281%+281%Ongoing daily compute cost!Ongoing daily compute cost!1Assuming 100M users at scale making 10 queries per day,thats a 4X increase in computational need incurred PER DAY One time compute
84、cost!One time compute cost!ModelsCloud,DC,SemisDeveloper ToolsApplicationsAI compute has been primarily served by GPUs69Source:Lambda Labs,Inference Benchmark for Stable Diffusion as of 2023.Assuming single precision.For illustrative purposes only.There is no guarantee that Coatues views and project
85、ions regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.GPUs are able to serve AI workloads much
86、 faster than CPUsTime to produce an image with Stable Diffusion(lower is better)6 Seconds8 Seconds8 Seconds9 Seconds12 SecondsA100RTX 3090A6000A5500RTX 8000i7-6850KEPYC 7352459 Seconds530 SecondsGPUs 70 x faster!GPUs 70 x faster!GPUsGPUsCPUsCPUsModelsCloud,DC,SemisDeveloper ToolsApplicationsThe dema
87、nd for GPUs has only begun70Source:Trendforce as of May 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosu
88、res for important disclosures,including regarding projections and forward-looking statements and trends.Estimated ramp up for AI servers growing 25%+Y/Y through 2026Estimated global AI server shipments8551,1831,5041,8952,36920222023E2024E2025E2026E+38%+38%+27%+27%+26%+26%+25%+25%What are the implica
89、tions for this global shift in demand for GPUs?What are the implications for this global shift in demand for GPUs?ModelsCloud,DC,SemisDeveloper ToolsApplications71Follow the GPUs:AI poised to have impacts across our economySource:NVIDIA public filings as of September 2023,Coatue analysis and opinion
90、,as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclaimers for important disclosures,including
91、regarding projections and forward-looking statements and companies and trends.ModelsCloud,DC,SemisDeveloper ToolsApplicationsCloudCloudData CentersData CentersPowerPowerSemisSemis$32B+$32B+NVDA Revenuelast 12 months72The AI wave may stress our power gridSource:NVIDIA public filings as of September 2
92、023,PTR Inc,University of Wisconsin,IEA,Schroders,Coatue analysis and opinion,as of November 2023.For illustrative purposes only.Potential power and price increases are Coatue estimates based on external data sources and are subject to change based on several factors,including demand,supply constrai
93、nts,competition,the pace of AI adoption.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclaimers for important disclosures,including regarding pr
94、ojections and forward-looking statements and companies and trends.$32B+$32B+NVDA Revenuelast 12 monthsMore data center power required 23-24E&3-5 GW50%2-3x3-5 YearsElectrical transformer price increasesPotential increase in power demand by 2026Leadtime to connect to the grid=ModelsCloud,DC,SemisDevel
95、oper ToolsApplications73More GPUs likely means more expensive serversSource:NVIDIA public filings as of September 2023,Semianalysis as of September 2023 showing server bill of materials.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future poten
96、tial of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclaimers for important disclosures,including regarding projections and forward-looking statements and companies and trends.$32B+$32B+NVDA Revenuelast 12 monthsModelsCloud,DC,SemisD
97、eveloper ToolsApplicationsHigher Server Bill of Materials Due to GPUs 23-24E Note:all estimated numbersNon-AI ServerAI ServerCPU$2K10 x$20KGPU$200KStorage&RAM$5k5x$20KPower Supply$1K10 x$5KNetworking&Other$5K3x$15kServer Cost$10K26x$260K2Q3Q4Q1Q2Q3Q4Q1Q2QCould AI workloads drive a cloud reaccelerati
98、on?Y/Y revenue growth rates for AWS,GCP,and Azure from Q2 2022 to Q2 202320222023E2024E36%19%Source:NVIDIA public filings as of 2023,AWS&GCP public filings,Azure sell-side consensus from Bloomberg as of September 2023.Coatue analysis and opinion,as of November 2023.For illustrative purposes only.Ill
99、ustrative future revenue growth rates for cloud providers are solely visuals and should not be interpreted as real figures.See Appendix-Disclaimers for important disclosures,including regarding projections and forward-looking statements.$32B+$32B+NVDA Revenuelast 12 monthsCloud growth rates have dec
100、elerated over last four quartersModelsCloud,DC,SemisDeveloper ToolsApplicationsAI could be a catalyst for growth!Pre 1990Pre 19901990 07 14 22 onwards2022 onwards$50B$50B$100B$100B$150B$150B$250B$250B$500B+?$500B+?$50B$50B$200B$200B$350B$350B$600B$600B$1.
101、5tn+?$1.5tn+?75MainframeMainframePC+InternetPC+InternetSmartphoneSmartphonePublic CloudPublic CloudAIAITime PeriodTotal Semi SpendKey HW playersSource:SIA Industry data as of September 2023,Coatue AI semis market assumptions as of September 2023.For illustrative purposes only.There is no guarantee t
102、hat Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosure
103、s for important disclosures,including regarding projections and forward-looking statements and trends.Incremental Semi SpendAI is the latest tech wave to inflect the semiconductor industryModelsCloud,DC,SemisDeveloper ToolsApplications76Source:Coatue opinion and analysis as of November 2023.For illu
104、strative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily repres
105、ent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.The entire semiconductor supply chain has potential to benefitSemis value chain across manufacturing inputs to device assemblyIC design/embedded softwar
106、eFabrication,pkg.&testingAssemblyCapital equip.MaterialsEx:EUV LithographyEx:Silicon waferDevice Device assemblyassemblySemiSemi design and design and manufacturingmanufacturingSemi manufacturing inputsSemi manufacturing inputsEnd product:NVIDIA H100Ex:NVIDIA chip designEx:3nm chipData center server
107、s powering AI!ModelsCloud,DC,SemisDeveloper ToolsApplicationsDemand for cutting edge GPUs has exceeded supply77Source:Coatue analysis and opinion,as of November 2023.For illustrative purposes only;there is no guarantee that the demand for GPUs will continue to increase.There is no guarantee that Coa
108、tues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.H100 GPU prices on
109、 aftermarket are typically well above list pricePricing of individual H100 80GB GPU units on eBay&Amazon,not reflective of cluster availabilityModelsCloud,DC,SemisDeveloper ToolsApplications$25,000$30,000$35,000$40,000$45,000$50,00005/2306/2307/2308/2309/2310/2311/23List priceAftermarket priceDespit
110、e intense demand,AI compute costs have decreased78Could AI compute resemble cost curves in other sectors?Cost of running GPT-3/3.5 down 90%in 5 months!Source:OurWorldInData,Genome.gov,OpenAI pricing,Coatue analysis and opinion,as of November 2023.Price of Lithium batteries is per kWh,and cost of DNA
111、 sequencing is per megabase.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,inc
112、luding regarding projections and forward-looking statements and trends.02005201020152020$0$3,000$6,000$9,000$1,500$4,500$7,500$10,500-90%90%Price of lithium batteriesCost of DNA sequencing$0.02$0.0020 0.0000.0050.0100.0150.0200.025Cost per 1K tokens($)GPT-3(Jan 2023)ChatGPT-3.5(June 2023)
113、-90%90%ModelsCloud,DC,SemisDeveloper ToolsApplicationsModels on edge could alleviate GPU shortage as well79Source:ExLlama,MLC,OpenAI token generation tracking from GitHub,Nomic,Coatue analysis and opinion,as of November 2023.Performance tracking is from third parties and may not be representative of
114、 performance normally.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company
115、type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Local models on Apple silicon becoming as fast as models running on GPUsModelsCloud,DC,SemisDeveloper ToolsApplicationsTok
116、ens per second(higher the better)of LLMsTokens per second(higher the better)of LLMs1535351924GPT-4GPT-3.5 TurboLLAMA-2 70BGPT4AllCode4AllLLAMA-2 70B10Run on GPUs Run on GPUs Run on Apple siliconRun on Apple siliconExLlamaExLlamaWill we reach AIs Holy Grail?80Source:Coatue analysis and opinion,as of
117、November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do
118、not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Largest,most expensive models are still most performant.for nowModelsCloud,DC,SemisDeveloper ToolsApplicationsSOTA QualityLimited“
119、intelligence”ExpensiveH100s/A100s requiredCheapRuns on Apple SiliconLLAMALLAMA-1 1LLAMALLAMA-2 2OLLAMAOLLAMAHoly Grail:SOTA quality while running locally!GPTGPT-3.53.5GGMLGGMLExisting tradeoff between quality and compute costs we need to break this paradigm!81AI has started to create a new ecosystem
120、 of toolsSource:Coatue analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Discl
121、osures for important disclosures,including regarding projections and forward-looking statements and trends.AI modelsAI appsEnd usersAI developer toolsCloud platformsData center/hardware/powerAI developer toolsAI Ops is a new category of tools for AI engineersML Ops:Focused on creating ML modelsAI Op
122、s:Focused on operationalizing Generative AIData labelingData labeling&curationModel training&developmentModel inferenceMonitoringModel hubFine tuningModel inferenceLLM OpsVector DBsMonitoring&evaluationIllustrative;non-exhaustiveModel hubOutput GuardrailsNomos AINomos AISource:Coatue analysis and op
123、inion as of November 2023.For illustrative purposes only.Not exhaustive of all companies in these categories.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logo
124、s listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplicationsBeyond labeling,dat
125、a curation is becoming more important as engineers try to enhance model performance and alignment.New tools enabling fine tuning and inference have become easier for the AI developer to use3Emerging category of tools enabling developers to use LLMs more effectively;Vector DBs have exploded in popula
126、rity due to embeddings.Securing the last mile deployment of LLMs through model evaluation,guardrails,and ongoing RLHF is supporting a new set of companies421CurateCurateAnalyzeAnalyzePreparationPreparationData curation tools will be critical for improving models83Source:Cohere as of November 2023,Co
127、atue opinion and analysis as of September 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of t
128、he applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Data curation an emerging new category between data collection&labeling,and model trainingData co
129、llecting&labelingDataset curation Dataset curation(New category)(New category)Model trainingLabeled data feeds ML modelsAutomatically clean and scrub data for errorsProfile the dataset across various metrics(e.g.token length distribution,entity label distribution)Filter data to include best examples
130、 for training;discard low-quality samples&identify human errors in labelingPre-trainingFine-tuningRLHFCurating datasets can reduce costs&improve quality.Training on 30%of dataset can be sufficient!(from Cohere study)ModelsCloud,DC,SemisDeveloper ToolsApplicationsSource:Hugging Face,Coatue opinion an
131、d analysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,
132、including regarding projections and forward-looking statements and trends.84Fine tuning models has become much more accessibleHugging Face Autotrain lets you fine tune open-source models on your own data in just a few clicks ModelsCloud,DC,SemisDeveloper ToolsApplications5Select type of task&upload
133、training dataset123Hugging Face suggests suitable modelsSelect model(s)and evaluate model training progressLLM Ops&Vector DBs are a new enabling layer for AI apps85Tools are already being put into productionLLM Ops&Vector DBs enable retrieval augmentationSource:PyPI downloads,Coatue opinion and anal
134、ysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI compa
135、ny type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Cumulative installs(incl.machine)LLM Ops platformsLLM Ops platformsLLMsLLMsVector DBsVector DBsAI ApplicationsAI Applic
136、ationsUnstructured dataUnstructured dataEmbeddings model indexes unstructured dataEmbeddings stored in vector DBs;provides similarity search given query from LLM orchestratorLLM Ops chain and manage prompts+language model responses,and serves this to the end applicationOutputs response to user query
137、,with context provided by LLM orchestratorModelsCloud,DC,SemisDeveloper ToolsApplications10M20M5M15M25M30M11/2201/2303/2305/2307/2309/2311/23Graph DBsReal-time streaming DBsVector DBsOLTP$25Bmkt capOLAPCan VectorDBs be standalone winners?86Only three DB types have significant public outcomesStartups
138、 have tried to create new categories of DBsSource:CapitalIQ,Coatue opinion and analysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will
139、benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Observability$50Bmkt capNo large sta
140、ndalone outcomes yet Will vector DBs be different?$25Bmkt capModelsCloud,DC,SemisDeveloper ToolsApplicationsGeneral purpose databases are being used as VectorDBs too87Source:Supabase data as of June 2023,Coatue opinion and analysis as of November 2023.For illustrative purposes only.There is no guara
141、ntee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.%of S
142、upabase projects using pgvectorModelsCloud,DC,SemisDeveloper ToolsApplications10%2%4%6%8%10%12%01/2302/2303/2304/2305/23Problem:Model evaluation is broken today88Humans prefer Claude 1 to Claude 2But Claude 2 performs better on benchmarksSource:LMSYS,Anthropic,Coatue analysis and opinion,as of Novem
143、ber 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not n
144、ecessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplicationsOverall win rates(human preference)against other modelsPerformance on different benchmar
145、ks77%72%0%10%20%30%40%50%60%70%80%Claude 1Claude 273%56%85%77%71%88%Bar examCoding proficiencyGrade school mathHarmlessnessBase2x betterClaude 1Claude 2LMSYSEvaluation is one of many last mile tools needed for AI in production89Last-mile tooling can improve AI accuracy and alignmentThis may help pre
146、vent hallucinations in critical sectors!Source:LinkedIn Post(Sarah Lukens 2023),Coatue opinion and analysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue inve
147、stment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.EvaluateAI modelsAppsIterateIterateFine-tuning and in-context learningGuardrailsFeedback=Example last mile toolingIndustrials cas
148、e study:AI applied to prognostics&health mgmt.(PHM)Industrials case study:AI applied to prognostics&health mgmt.(PHM)5%10%4%29%Maintenance&Reliability Knowledge ExamPHM Knowledge ExamChatGPTBard%of questions with hallucinated answersModelsCloud,DC,SemisDeveloper ToolsApplicationsDeep Learning EraRis
149、e of ML EngineersEra of AI OpsNatural language programming100K MachineHuman+Machine MachineMachine beats allELO RatingCoatue opinion and analysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accur
150、ate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplicationsAI adoption accelerating across creative mod
151、alities94Usage of Runways AI video tools over timeMidjourney Discord members over timeSource:Runway usage as of June 2023 showing cumulative usage including video generation,AI rotoscoping,etc.Midjourney Discord members as of June 2023.Coatue opinion and analysis as of November 2023.For illustrative
152、 purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coat
153、ue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplications11/2212/2201/2302/2303/2304/2305/2306/231M2M2M3M4M8M11M16M+2,219%12/2201/2302/2303/2304/2305/2311/2206/234M5M7M11
154、M13M14M15M17M+342%+342%AI ramping up for design use cases in professional settings95Source:Aurora Solar and Tome as of June 2023,Coatue opinion and analysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of A
155、I are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplicationsCumulative Tome signups1M2M4M6M8M
156、10M01/2302/2303/2304/2305/2306/23+1,191%+1,191%AI solar designs generated1,0008,00075,000215,000Q2 2022Q3 2022Q4 2022Q1 2023+21,400%+21,400%AI is inflecting within software development tools too96AI projects on Replit(annualized)Cumulative VSCode installsSource:Replit as of June 2023,Coatue opinion
157、and analysis as of November 2023(incl.Codeium installs).For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as exam
158、ples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplicationsQ1 22Q2 22Q3 22Q4 22Q1 23Q2 233K5K8K26K12
159、3K328K+12,042%+12,042%12/2201/2302/2303/2304/2305/2306/2307/2308/2309/23321K3K8K19K42K75K115K159K208K262K+12,105%+12,105%Massive apps created on both sides of S-Curves of prev.cycles97Source:Capital IQ,World Bank as of September 2023,Statista as of September 2023,Coatue opinion and analysis as of No
160、vember 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do no
161、t necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Application tech companies founded and US market cap along S-Curves020200%64200620082001
162、82022202420%40%60%80%100%InternetSmartphones1994 1998$3.2 Trillion$3.2 Trillion2002 2006$900 Billion$900 Billion2008 2011$200 Billion$200 Billion2012 2016$250 Billion$250 BillionModelsCloud,DC,SemisDeveloper ToolsApplications98Incumbents gained most in smartphones:what about AI?Incumbents(PreIncumbe
163、nts(Pre-2007)2007)Native Mobile First(PostNative Mobile First(Post-2007)2007)$6.5 Trillion$500Billion$7 Trillion$500Billion$500Billion2008 Mkt Cap2008 Mkt CapTodays Mkt CapTodays Mkt Cap12x!Smartphone Era GainSmartphone Era GainSource:Capital IQ and Coatue analysis and opinion as of November 2023.Fo
164、r illustrative purposes only.This is showing solely gain in market cap after the smartphone era,and is not claiming causation of mobile on market cap.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or
165、 fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDevel
166、oper ToolsApplications99ModelsCloud,DC,SemisDeveloper ToolsApplicationsAI could drive incumbent price increasesMicrosoft Office Copilot could be an 80%+increase in ARPU for Office users$14$36$36$30On-Prem*CloudCloud+AI$66+157%+157%+83%+83%Per seat per month cost of Microsoft 365 E3 license Source:Mi
167、crosoft pricing as of November 2023(*On-Prem per month pricing takes Office 2010 Box Professional Pricing of$499,and assumes a 3 year usage.Note that Microsoft AI Copilot pricing of$30 per month is based on publicly announced pricing).Coatue analysis and opinion as of November 2023.For illustrative
168、purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-look
169、ing statements and trends.Is AI a game of kings or attackers?Incumbent KingsAI Native AttackersFast moving incumbents AI natives Laggard incumbentsSlow and risk averseDataDistribution&audienceFull support infrastructureLarge capital reservesNeed for external fundingFast and nimbleCreative and risk-o
170、nNo tech debtAttract higher density of tech talentSource:Coatue analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will ben
171、efit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplications100101The biggest mobile startups unlocked new behaviorsOn Demand Access On Demand Access to Underutilize
172、d to Underutilized AssetsAssets+$215B Mkt CapRealReal-time Photo time Photo Sharing&Sharing&MessagingMessaging+$250B Mkt CapMobile TechnologyGPSMobile CameraMobile InternetApp storesMobile Mobile FintechFintech+$80B Mkt CapSource:TikTok valuation taken from last Bytedance fundraise from G42 in Mar 2
173、023.CapitalIQ and Coatue analysis and opinion as of November 2023.For illustrative purposes only.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit fro
174、m the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplications
175、AI startups should aim to create new behaviors tooArtificial Intelligence“Most entrepreneurial ideas will sound crazy,stupid and uneconomic,and then theyll turn out to be right”Reed HastingsYour invention hereAI autopilots?Models that can reason?Infinite gaming worlds?Source:Wired as of 2012,Coatue
176、analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important di
177、sclosures,including regarding projections and forward-looking statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplicationsExample:Character AI creating new virtual companions103Source:Coatue opinion and analysis as of November 2023.For illustrative purposes only.Highlighted company is an ex
178、ample of applicable AI company type;do not necessarily represent Coatue investments.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for
179、important disclosures,including regarding projections and forward-looking statements and trends.Users flocking to Character to chat with AI personas brand new behavior!ModelsCloud,DC,SemisDeveloper ToolsApplicationsMonthly active web usersAverage time spent per visit01/22 03/2205/2207/2209/2211/2201
180、/23 03/2305/2307/2309/230M2M4M6M8M10M12M14M16M18M05511/23Example:AI as an autopilot could transform org structures104Previously,scaling businesses meant scaling headcountWith AI,scaling business means scaling compute!Source:Coatue analysis and opinion as of November 2023.For illustrative
181、purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-look
182、ing statements and trends.ModelsCloud,DC,SemisDeveloper ToolsApplicationsProductSales&Support OperationsSenior managementExecutive leadershipEngineeringGrowth&MarketingFinance&LegalProduct&EngineeringAI autopilotLeadershipSales,support,and marketingAI autopilotFinance,legalAI autopilot105There remai
183、ns huge opportunity across modalitiesCategoryExample opportunitiesOpen challengesSource:(1)Software defined by Statista as Productivity,Application Development,System Infrastructure,and Enterprise software.IFPI,Public Filings,Statista,Zion Market Research,Coatue opinion and analysis as of November 2
184、023.For illustrative purposes only;does not represent an exhaustive list of opportunities or challenges.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appen
185、dix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.AudioVideoCodeSearchOther(e.g.robotics)AI musiciansText-to-videoSoftware synthesisLLM-based searchRobotic perception$600B+size of global software1 market today$26B+size of music indust
186、ry today$162B+Google Search revenue today$35B+size of robotics market today$100B+size of global movie&entertainment industry todayHallucinationsCode security&robustness at scaleIntegrations with existing SW deployment processCosts to use agent-automated searchResponse latencyIncumbent advantages in
187、search data&distributionCopyright infringement with artistsIncumbent distribution effects more pronounced(e.g.Apple audiobooks,Spotify)PhotorealismContent moderationCompute costs at scaleHardware costs and upfront capital outlaysUnproven text-to-action workflowModelsCloud,DC,SemisDeveloper ToolsAppl
188、icationsAI could increase the opportunity size of all these markets!kIQaaS+HardwareUsers$5T$5TMarket CapHardware+OSMobile eraAI eraUsersAI has potential to unlock“the super app of the West”Web browserStandalone Apps$2.5T$2.5TMarket CapAI Super AppAI Super Appone-stop-shop+actionsSource:Capital IQ an
189、d Coatue analysis and opinion as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples o
190、f the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.Super Apps$600B$600BMarket Cap106ModelsCloud,DC,SemisDeveloper ToolsApplications107Coatue view
191、:the best of AI is yet to comeKey TopicsUnimodalityMultimodalityCurrent genNext genCoatue View:The future top coding language will be EnglishSource:Coatue opinion and analysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the fu
192、ture potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding pr
193、ojections and forward-looking statements and trends.EnglishData CenterEverywhereResearchEngineeringPublic dataPrivate data108TodayTodayFutureFutureCopilot systems that suggest codeFoundation models targeted at code generationNatural language to programming,debugging,and deploymentGhostwriterGhostwri
194、terEmerging coding agentsGhostwriterGhostwriterFactoryFactoryFrom idea expressed in English To fully fledged application in minutes!GhostwriterGhostwriterWith help of future AI systemsCursorCursorCursorCursorAnd future companies?UnimodalityMultimodalityCurrent genNext genEnglishData CenterEverywhere
195、ResearchEngineeringCoatue View:On-device AI will become more widespreadSource:MLC,Web Search for Apple AI strategy rumors,Coatue opinion and analysis as of November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are a
196、ccurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for important disclosures,including regarding projections and forward-look
197、ing statements and trends.109Public dataPrivate dataFuture:Rumored Apple AI strategyWill everyone have a powerful LLM in their pocket?Will everyone have a powerful LLM in their pocket?Apple ShortcutsAjax GPTSiriSiri+Today:MLC LLM allows you to run models on your phone+=1B+iPhone users become AI user
198、s139RedPajama 3BVicuna 7BTokens per secondUnimodalityMultimodalityCurrent genNext genEnglishData CenterEverywhereResearchEngineeringPublic dataPrivate dataCoatue View:Scaling AI is an engineering challengeSource:Coatue opinion and analysis as of November 2023.For illustrative purposes only.There is
199、no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and tren
200、ds.Engineering headcount at OpenAI has increased Engineering headcount at OpenAI has increased over 3X over 3X since GPTsince GPT-3 came out3 came outEstimated Y/Y growth rate in OpenAI engineering headcount10%20%30%40%50%60%70%80%90%01/202007/2001/2107/2101/2207/2201/2307/2301/24GPTGPT-3 3130 engin
201、eers130 engineersChatChat-GPTGPT280 engineers280 engineersSep 2023Sep 2023430 engineers430 engineersAug 2023 saw 70%Aug 2023 saw 70%Y/Y engineering Y/Y engineering headcount growth!headcount growth!Estimated#Estimated#of engineersof engineers110Public dataPrivate dataEnglishData CenterEverywhereRese
202、archEngineeringCoatue View:Private datasets can unlock new use casesSource:Coatue opinion and analysis as of November 2023.Companies and logos above are non-exhaustive and should only be construed as examples in a particular category.For illustrative purposes only.There is no guarantee that Coatues
203、views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.Logos listed above as examples of the applicable AI company type;do not necessarily represent Coatue investments.See Appendix-Disclosures for import
204、ant disclosures,including regarding projections and forward-looking statements and trends.UnimodalityMultimodalityCurrent genNext gen111Imagine what AI can do with Proprietary Data!BiotechPersonalizedmedicineMediaGenerate shows on-demand 3D ModelsEngineeranythingGamingInfinite worlds&contenteCommerc
205、eAI shopping conciergeEnglishData CenterEverywhereResearchEngineeringPublic dataPrivate dataUnimodalityMultimodalityCoatue View:Innovation in multimodality is a new frontierSource:Weaviate blog as of June 2023,Coatue opinion and analysis as of November 2023.For illustrative purposes only.There is no
206、 guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends
207、.Future:Lots of innovation within multimodal embeddings space!Today:GPT-4V is already creating new experiences for users+ChatGPT=Step by step written Step by step written instructions if you get instructions if you get confused by IKEA diagrams!confused by IKEA diagrams!Open questions:Does training
208、on multimodal datasets enable better reasoning capabilities for models?How can we keep multimodal models safe to use?What interesting emergent properties can we see from using the multimodal embeddings space?How do we scale multimodal datasets to train better models?Multimodal embeddings map relatio
209、nships across text,image/video,and audio Current genNext gen112EnglishData CenterEverywhereResearchEngineeringPublic dataPrivate dataCurrent genNext genUnimodalityMultimodalityCoatue View:New advancements in architecture coming Source:Stanford Hazy Research Lab,Arxiv,Coatue opinion and analysis as o
210、f November 2023.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potential of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regar
211、ding projections and forward-looking statements and trends.Cutting edge AI research emerging to improve TransformersCutting edge AI research emerging to improve TransformersFuture|New architectures:HyenaToday|Architecture improvements:Flash Attention25%56%72%BaselineFlash Attention(2022)Flash Attent
212、ion 2(2023)Model FLOPS utilization(i.e.efficiency)Could Hyena architecture enable million-token length context windows?113Enterprise Enterprise deploymentdeploymentCompliance guardrailsCompliance guardrailsAI ethics&securityAI ethics&securityAI authentication&AI authentication&permissioningpermissio
213、ningModalitiesModalitiesMultiMulti-modal modelsmodal modelsEmbodied roboticsEmbodied roboticsLongLong-form video form video generationgenerationAvailabilityAvailabilityAbundant hardwareAbundant hardwareOnOn-device AI&lowdevice AI&low-cost deploymentscost deploymentsSOTA models available SOTA models
214、available for allfor allReducing Reducing HallucinationsHallucinationsModel outputModel outputDiscovery&novel Discovery&novel insightsinsightsModel Model explainabilityexplainability+more challenges yet to come+more challenges yet to come114Solving AIs challenges has potential to unlock vast opportu
215、nitiesGenerally useful AI for allNear termLonger termSource:Coatue opinion and analysis as of November 2023.List of challenges are not exhaustive and may change as AI landscape changes.For illustrative purposes only.There is no guarantee that Coatues views and projections regarding the future potent
216、ial of AI are accurate or that any particular Coatue investment or fund will benefit from the AI trend.See Appendix-Disclosures for important disclosures,including regarding projections and forward-looking statements and trends.If you are a founder innovating in this space,we would like to connect w
217、ith you.Please reach out at 115Appendix-DisclosuresThisThis whitepaperwhitepaper reflectsreflects CoatuesCoatues opinionsopinions andand analysisanalysis onon AIAI asas ofof thethe datedate ofof thisthis presentationpresentation andand doesdoes notnot constituteconstitute investmentinvestment advice
218、advice oror a a recommendationrecommendation toto buybuy oror sellsell anyany securitiessecurities.Since AI is an emerging technology,we expect our views may evolve or shift over time.As such,information herein is subject to change at anytime without notice.Unless otherwise indicated,any figures and
219、 calculations herein are for informational purposes only,computed by Coatue or its advisors and not audited by any third party.AlthoughAlthough CoatueCoatue believesbelieves thatthat thethe datadata expressedexpressed in in thisthis presentationpresentation is is accurateaccurate andand reasonable,r
220、easonable,actualactual resultsresults couldcould differdiffer materiallymaterially fromfrom thosethose projectedprojected oror assumed,assumed,andand suchsuch projectionsprojections areare subjectsubject toto change,change,andand areare subjectsubject toto inherentinherent risksrisks andand uncertai
221、ntiesuncertainties.NeitherNeither CoatueCoatue nornor itsits affiliatesaffiliates oror advisorsadvisors guaranteesguarantees thethe accuracyaccuracy oror completenesscompleteness ofof thethe informationinformation.Accordingly,Accordingly,neitherneither CoatueCoatue nornor anyany ofof itsits affiliat
222、es,affiliates,advisorsadvisors oror employeesemployees shallshall bebe liableliable toto youyou oror anyoneanyone elseelse forfor anyany lossloss oror damagesdamages fromfrom useuse ofof thethe informationinformation containedcontained in in thisthis presentationpresentation.Coatue AnalysisCoatue An
223、alysisForwardForward-looking looking Statements&Statements&Projections Projections The companies referenced herein are included for informational purposes only.The information herein does not constitute investment advice or a recommendation to buy or sell any securities.The companies do not necessar
224、ily represent stocks or investments that Coatue owned or owns.In addition,the trends identified and discussed in this presentation reflect the opinions of Coatue.The trends discussed do not reflect the entire universe and could be impacted by market factors,changes in laws and other factors.No third
225、-party firm or company names,brands or logos used in this presentation are Coatues trademarks or registered trademarks,and they remain the property of their respective holders and not Coatue.The inclusion of any third-party firm and/or company names,brands and/or logos does not imply any affiliation
226、 with these firms or companies.None of these firms or companies has endorsed the investment opportunity described herein,Coatue,any affiliates of Coatue,or Coatues personnel.Companies and Companies and TrendsTrendsThis presentation contains forecasts,projections and other forward-looking statements,
227、including(but not limited to)the occurrence or outcome of anticipated events,estimates,future performance and adaption of AI.Due to various risks and uncertainties,actual events,results of these events may differ materially from those reflected or contemplated in such forward-looking statements.Ther
228、e is no guarantee that such forecasts,projections or forward-looking statements will occur and therefore should not be relied upon.Artificial Artificial Intelligence(AI)Intelligence(AI)This presentation contains forward-looking predictions regarding AI and its potential impacts and opportunities,all
229、 of which are subject to a number of factors and uncertainties.Any characterization of AI herein is the opinion of Coatue,is subject to change,and should not be relied upon in making an investment decision.Given that AI is an emerging technology,assessing the future trajectory of the AI industry is
230、inherently challenging,and Coatues views on its success or failure can be subjective and based on incomplete information,limited perspectives,or speculative assumptions.See also the disclosures regarding forward-looking statements.Companies herein are not intended to highlight or represent the Coatu
231、e portfolio,but rather the broader AI theme,which by nature may include Coatue investments.To the extent Coatue portfolio companies or investments are included herein,Coatue makes no suggestion or guarantee regarding the future outcomes or performance of such companies.Even if Coatues characterizati
232、ons and opinions regarding the AI trend were to prove accurate,there is no suggestion or guarantee that Coatue will be able to identify and invest in opportunities presented by AI.The AI industry is multifaceted,encompassing various technologies,applications,and market dynamics.Its complexity makes it susceptible to unpredictable developments,including breakthrough innovations,disruptive technologies,or unexpected challenges.