《高盛:生成式AI投资框架(2023)(英文版)(61页).pdf》由会员分享,可在线阅读,更多相关《高盛:生成式AI投资框架(2023)(英文版)(61页).pdf(61页珍藏版)》请在三个皮匠报告上搜索。
1、Cloud computing created new investment opportunities by enabling the delivery of software as a utility.Generative AI further unlocks value as it extends this utility and provides new tools for enhancing end-user productivity.While traditional AI has been helpful in making predictions of outcomes,Gen
2、erative AI is about generating content such as text,video,images,or computer code,which was previously not possible.Large Language Models(LLMs)are a key enabler of GAI,with a profound level of proficiency and intelligence.AI has the potential to establish new companies,while providing incumbents wit
3、h new growth avenues by turbo-charging end-user productivity.We estimate a Generative AI Software TAM of$150bn,vs.the global software industry TAM of$685bn.The GS Macro team estimates AI could drive$7tn in global economic growth over 10 years,underpinned by productivity growing 1.5pp faster annually
4、.We raise our PTs on MSFT(to$325,vs$315 prior),CRM(to$325,vs,$320),and ADBE(to$480,vs.$475)to incorporate higher outer year estimates that reflect our conviction in the success/adoption of newly launched GAI products.Microsoft has clearly taken the tech industry by storm by being first out of the ga
5、te with Copilot variations of Microsoft 365,GitHub and Dynamics.Adobe stands to revitalize its growth prospects with the launch of Firefly,while CRM should benefit from a front-office productivity boost.Additionally,we highlight INTU,GOOGL,AMZN,NVDA,and META as companies best-positioned to succeed i
6、n this new AI-driven paradigm.PM Summary In our view,the enterprise software industry is embarking on the next wave ofinnovation(after the rise of cloud computing).The recent attention to AI was largely driven by the release of ChatGPT by OpenAI and the subsequent product announcements by tech giant
7、s Microsoft and Google.In this report,we explore the Generative AI TAM for enterprise software,delve into the broader productivity implications,the future of the IT stack while highlighting our top Generative AI stock picks and the burst of product innovation that can lay the foundation of adoption.
8、We believe Generative AI could drive the next wave ofnKash Rangan+1(415)249-7318 |Goldman Sachs&Co.LLC Eric Sheridan+1(917)343-8683 |Goldman Sachs&Co.LLC Toshiya Hari+1(646)446-1759 |Goldman Sachs&Co.LLC Gabriela Borges,CFA+1(212)902-7839| Goldman Sachs&Co.LLC Salveen Richter,CFA+1(212)934-4204| Gol
9、dman Sachs&Co.LLC Anisha Narayan,CFA+1(212)934-1992| Goldman Sachs India SPL Gili Naftalovich+1(917)343-4286| Goldman Sachs&Co.LLC Ben Miller+1(917)343-8674| Goldman Sachs&Co.LLC Alexandra Steiger+49(69)7532-3097| Goldman Sachs Bank Europe SE Matthew Martino+1(212)902-0695| Goldman Sachs&Co.LLC Alex
10、 Vegliante,CFA+1(212)934-1878| Goldman Sachs&Co.LLC Max Gamperl+1(415)249-7311| Goldman Sachs&Co.LLC Jacob Staffel+1(713)276-3526 |Goldman Sachs&Co.LLC Dan Duggan,Ph.D.+1(212)902-4726 |Goldman Sachs&Co.LLCAmericas Technology Generative AI-Part I:Laying Out the Investment Framework 26 March 2023|6:06
11、PM PDT Goldman Sachs does and seeks to do business with companies covered in its research reports.As a result,investors should be aware that the firm may have a conflict of interest that could affect the objectivity of this report.Investors should consider this report as only a single factor in maki
12、ng their investment decision.For Reg AC certification and other important disclosures,see the Disclosure Appendix,or go to employed by non-US affiliates are not registered/qualified as research analysts with FINRA in the U.S._ 起点财经GPTChatGPT不会淘汰你!先驾驭ChatGPT的人会淘汰你!123不定时分享AI智能、ChatGPT、AIGC、GPT-4等最新研报
13、和相关资讯4不定期邀请行业 大咖 演讲互动交流学习5ChatGPT 初体验,专属微信群与GPT互动提问!(目前开放的API为3.5版本)各路大神 畅聊AI 使用指南和落地应用,分享商业化案例,碰撞思维火花一次性领取 301份 ChatGPT、AIGC 相关资料,赠送 80页 ChatGPT、AI绘画Midjourney保姆级教程,资料持续更新中识别二维码查看详情 PM Summary 1 GS Top Generative AI Picks 3 What is Artificial Intelligence?8 Traditional AI vs Generative AI 15 Genera
14、tive AI Technology Stack and Its Implications 24 Workforce Productivity Likely to Benefit From Adoption of Generative AI 29 Sizing the Generative AI TAM for Enterprise Software 32 Raise PTs for MSFT,CRM,ADBE Given Revenue Tailwinds 34 Impact of Generative AI Likely to Reach Multitude of Workforce Pe
15、rsonas 38 Risks Associated with Generative AI 44 Burst of Product Innovation Lays the Foundation for the Growth of Generative AI 45 Valuation&Key Risks 53 Disclosure Appendix 57 26 March 2023 2Goldman SachsAmericas TechnologyTable of Contents _ innovation/monetization in enterprise software after cl
16、oud computing.Just as the cloud hyperscalers(AWS,Azure,GCP)commercialized cloud infrastructure and platforms,which in turn accelerated the growth of Software-as-a-Service businesses,we believe OpenAIs ChatGPT will drive adoption of Generative AI across software enterprises.In this report,we further
17、explore the opportunities for AI to be integrated into the tech stacks of the future by drawing an analogy between a cloud computing stack and an IT stack layered with AI.Artificial Intelligence(AI)is expected to drive significant productivity gains for nthe economy,as well as the software sector.Th
18、e GS Macro team estimates that AI adoption could drive almost$7tn in global economic growth over a 10-year period,with productivity growing 1.5pp faster annually over the same period.AI is touted to be the next big shift in technology after the evolution of the internet,mobile and the cloud.We belie
19、ve Generative AI can streamline business workflows,automate routine tasks and give rise to a new generation of business applications.We believe Generative AI can contribute an incremental TAM of+$150bn to the nglobal enterprise software TAM.Software companies are arming their product portfolios with
20、 new Generative AI product SKUs.By adopting Generative AI,SaaS companies are opening opportunities for upselling/cross-selling products,increasing customer retention and expansion rates.This can present multiple levers for growth from:new product/application releases,premiums for AI integrated SKUs,
21、and1)2)3)price increases over time as the value proposition and stickiness of existing products grows with the integration of AI.Generative AI tools have far-reaching implications across industries,from nenterprise software to healthcare,financial services and more.While we believe Generative AI is
22、still in the early innings,it has already been able to create an impressionable impact on users with its human-like responses and ability to generate original content.Generative AI is enabling sales and marketing teams generate new content,DevOps engines to write code faster,knowledge workers to imp
23、rove day to day office efficiency,scientists to develop drugs to rare disease,and much more.Despite a new innovation across the tech stack,we expect wave of nincumbents Microsoft,Alphabet,Nvidia,Amazon,Salesforce,Meta,Intuit,and Adobe to lead the industry in Generative AI adoption.By adopting Genera
24、tive AI in their product roadmap,these companies can shape the way the tech industry evolves.We see the aforementioned companies being well positioned to strengthen their competitive moats as they benefit from having copious amounts of data.Evaluating the near and long-term implications this can hav
25、e from a growth standpoint,we raise our PT for MSFT,CRM and ADBE as we take into account the long-tail benefits that can ensue from their recent product announcements.GS Top Generative AI Picks While Generative AI is posed to create a multi-year tailwind for the broader tech industry,we call out the
26、 companies we view as best positioned to benefit.By 26 March 2023 3Goldman SachsAmericas Technology_ layering their existing product portfolios with Generative AI capabilities,tech companies such as Microsoft,Alphabet,Nvidia,Amazon,Salesforce,Meta,Intuit,and Adobe can strengthen their competitive mo
27、ats with new product SKUs.This also expands the opportunity for more seamlesss upsell/cross-sell motions and improved customer retention.For the cloud providers,we expect the growing use of Generative AI will drive demand for compute.Microsoft(Buy on CL,PT$325):Microsoft has taken the early lead by
28、incorporating Generative AI as a foundational technology into its product innovation framework.While the significant announcement around an enhanced Bing/Edge experience introduced in February(here)shined light to the companys renewed vigor in the search and browser markets,the steady flow of impres
29、sive product releases across its portfolio have confirmed Microsofts concentrated focus on incorporating Gen.AI across its platform,particularly its core products.With the release of Microsoft 365 Copilot,Business Chat,Dynamics 365 Copilot,GitHub Copilot,Teams Premium,Azure OpenAI Services,Viva Sale
30、s,and more,Microsoft is in the unique position to transform the way software augments human productivity as it showcases the value it can drive in synthesizing,creating,and sharing information across a variety of different use cases.Given the importance of the data input into the underlying models o
31、f such technology,Microsofts productivity suite,cloud services,developer tools and other platforms create a well-rounded data hub that can solidify its competitive moat,which can translate to ongoing strength within its Microsoft Cloud segment,that comprises Azure,Commercial Office 365,Dynamics 365,
32、and LinkedIn Commercial.Drawing a parallel to the initial introduction of Microsoft Office 365,we believe Microsoft can utilize a similar playbook that would lead to new product SKUs(similar to that of E3,E5 today)that offer a varying degree of Generative AI capabilities,ultimately driving higher AS
33、Ps and a steady growth cadence.The early success seen within the first month of its Bing release(reaching 100mn DAU,with a third being net-new)and the release of new products across all layers of the tech stack(with Microsoft 365 Copilot/Dynamics 365 Copilot/Viva Sales/Team premium in SaaS,the relea
34、se of GitHub Copilot/Power BI in PaaS and Azure OpenAI Services in IaaS among others),exemplify Microsofts ability to successfully expand the reach of AI services in a more deliberate and transformative way than we have seen in the past.Another potential positive implication can stem from within Azu
35、re,which may show stabilizing growth as these offerings drive more end-user and OpenAI workloads.We point to the value created within the Office suite(now Microsoft 365)over the last two decades(OneDrive,OneNote,Teams,SharePoint)as an example of Microsofts ability to execute on a playbook driven on
36、feature expansion.Microsofts prioritization of R&D expenses over the last five years and continued investment in OpenAI(since its initial investment in 2019)underscore the resources Microsoft has accumulated to best position itself as a leader in this next chapter of innovation.Alphabet(Buy,PT$128):
37、Going back 5+years ago,Google began a series of introductions outlining how AI would be the driving force behind many of the companys products alongside a broader computing shift(including Google Assistant,Duplex,Lens,Translate,LLMs integrated within Search,etc.).We see the recent announcement of 26
38、 March 2023 4Goldman SachsAmericas Technology_ Bard as an extension of these efforts to match broad product iteration(the continuing evolution of search)with recent consumer excitement about the conversational AI nature of ChatGPT.At each subsequent Google I/O event(the companys annual developer con
39、ference),we have seen Google introduce and build upon the Google Assistant to widen the input mechanism for search(e.g.,text-to-audio,predictive analytics in the Discover tab,etc.)and to infuse all of its products with elements of AI(e.g.,Maps turn by turn directions,YouTubes recommendation engine,a
40、uto-complete in Gmail and Google Docs,etc.).In addition,we see Google at the forefront of AI-driven automation within digital advertising with Performance Max,its automated end-to-end campaign management service that optimizes spend across inventory from Googles O&O properties and third-party sites
41、via the Google Display Network,which we expect will see increasing advertiser adoption&spend going forward.In our recently published note(link),we detailed recent AI product announcements and provided a framework of the current landscape of current AI/ML initiatives inside Alphabet(across Google and
42、 DeepMind).Within our Internet coverage universe,we see Alphabet as the leading collection of AI/machine learning-driven businesses that is uniquely positioned to capitalize on the blurring of the lines between advertising,commerce and media consumption in the years ahead and rising utility across a
43、 number of computing platforms including consumer desktop,consumer mobile&enterprise cloud computing.In addition to its core advertising business,Google Cloud should be a tailwind for consolidated revenue growth over our 5-year forecast period and we expect Cloud to begin contributing at a demonstra
44、ble rate of change to consolidated operating income margins in 2024 and beyond.Lastly,Alphabet has an established multi-year track record of balancing strong levels of investments for long-term growth with delivering a rising pace of shareholder returns via share buybacks.Nvidia(Buy,PT$275):While we
45、 envision a whole host of semiconductor companies across the Compute,Networking and Memory landscape benefiting from the proliferation of AI,we continue to highlight NVDA a stock we recently upgraded to Buy as one that is most levered,particularly to Generative AI.Importantly,we believe the company
46、is well-positioned through its hardware and software offerings to support the production and deployment of large models whether it be for the major cloud hyperscalers or for enterprise customers.As we stated in our upgrade note(here),we model an acceleration in the rate of Nvidias wallet share growt
47、h within the context of overall cloud capex as a growing percentage of data center compute is addressed by GPUs and the company expands its SAM beyond the GPU into areas like the CPU(i.e.,note Nvidias Arm-based Grace CPU is scheduled to ramp in 2HCY23).Furthermore,to the extent there was any doubt r
48、elated to the sustainability of Nvidias competitive position,we believe the acceleration in AI development/adoption brought by the emergence of Generative AI will,if anything,serve to extend the companys leadership as customers with any sense of urgency will lean on solutions that are competitive an
49、d scalable today.Amazon(Buy,PT$145):While much of the focus around the theme of Generative AI has been dominated by market share dynamics in Search&gross margin implications from higher compute costs against a potential increase in mix of non-commercial search,we believe as enterprises push deeper i
50、nto integrating AI/ML tools into their 26 March 2023 5Goldman SachsAmericas Technology_ tech stacks to drive core businesses,hyperscalers stand to benefit and are underappreciated beneficiaries of this theme(including Amazons cloud computing business,AWS).AWS is exposed to this theme in a number of
51、ways,both directly through AI/ML service offerings(Amazon Lex,Amazon Polly,Amazon Transcribe,Amazon Comprehend,Amazon Kendra,Amazon Translate,Amazon SageMaker,among others)and partnerships with Generative AI companies(including Hugging Face,Stability AI,AI21 Labs,C3 AI,etc.).Looking over a multi-yea
52、r timeframe,we believe that Amazon will compound a mix of solid revenue trajectory with expanding margins as it delivers yield/returns on multiple year investment cycles.After trading in a range(&underperforming the broader market)over the past few years,we see AMZN as well positioned for future out
53、performance as eCommerce margins normalize(even if just back to 2018/2019 levels),as its advertising business continues to achieve scale and as AWS can still benefit from a long-tailed structural growth opportunity in the shifting needs of enterprise customers(while producing a balance of growth and
54、 margins).While the next few quarters will likely remain volatile as an output of macroeconomic volatility,the long-term narratives from Amazon and a compelling multi-year risk/reward should appeal to investors.Lastly,we continue to see Amazon positioned as a leader in all aspects of secular growth
55、within our Internet coverage(eCommerce,digital advertising,media consumption,aggregated subscription offerings&cloud computing).Salesforce(Buy on CL,PT$325):As part of Salesforces path to unlocking value(Part I,II),we outline our view that the incorporation of Generative AI would have long-tailed be
56、nefits to the companys growth outlook.With the announcement of Einstein GPT for Sales,Service,Marketing,Slack and developers(building on the AI-focused Einstein brand released in 2016),Salesforce can leverage the vast amount of data stored on the platform to bring AI to the forefront of end-user exp
57、eriences,which can drive longer-term revenue growth,improved sales rep activity and increased user adoption.As a system of record,Salesforce stores a tremendous amount of data across countless industries&clients.We believe that the marriage of CRMs data with Generative AI has the potential to help u
58、sers extract more value from the CRM platform via more proven,data-driven insights and actionable tasks,which can lift productivity.Additionally,as the CRM platform has grown increasingly complex(with Service,Sales,Marketing,Commerce,MuleSoft,Tableau,Slack,etc)this may be hindering customer adoption
59、 as users struggle to easily leverage the platforms potential.Given Generative AIs capabilities,we believe Salesforce could simplify the growing complexity of the CRM offering.This can lead to lower adoption barriers(via improved onboarding,utilization,and engagement)and enhanced financial performan
60、ce(namely better net revenue retention rates,net new logo additions,and increased deal sizes).The internal use of such technology can also bolster Salesforces current efficiency initiatives as products can automate repetitive or low-value tasks,auto-generate action items and create personalized mark
61、eting/sales materials.Overall,we see the potential implementation of Generative AI technology within the CRM platform as another boon to the long-term growth narrative.Meta(Buy,PT$215):We see Meta as an emerging AI leader as investments toward AI development&compute capacity continue to scale(accord
62、ing to our estimates,META 26 March 2023 6Goldman SachsAmericas Technology_ will spend a cumulative$110bn in capex from 2019-2023,the majority of which going toward its AI efforts).With regard to Generative AI specifically,Meta has launched a number of products including Make-a-Scene and Make-a-Video
63、,a text-to-image and text-to-video model,and,most recently,LLaMA,a series of LLMs available to researchers via direct API access and optimized for smaller computing power.Looking at artificial intelligence more broadly,Meta has infused AI into its core products since its inception,including user-fac
64、ing(e.g.,recommendation engine&discover/interest graph,content moderation,etc.)and advertiser-facing(automated creatives&campaign management including Advantage+,ad targeting,modeled measurement/attribution esp.in light of data privacy,etc.)capabilities.While debates will likely persist around produ
65、ct transitions and industry platform headwinds in the quarters&years ahead,we remain focused on Metas large scaled audience across its Family of Apps against which the company can continue to align evolving consumption habits within short-form video,messaging,commerce,augmented reality&social connec
66、tions.We expect recent topline headwinds(platform policy changes incl.Apples ATT;current volatile macro environment;engagement shift to Reels with low monetization;competition;etc.)to start to abate and/or turn into tailwinds into 2023(including off easier YoY comps)as META returns to more normalize
67、d industry-level growth trends in 2024&beyond.Taking a step back from the recent stock performance,we see platform/infrastructure investments by Meta as both a)continuing to build independence from a volatile range of outcomes from future mobile OS platform changes;and b)aligned with a strategic shi
68、ft toward short-form video and from the social graph to the interest graph.Adobe(Buy,PT$480):Though Adobe made its first announcement under the Generative AI umbrella at its Summit conference last week,we highlight Adobes early investments in AI/ML tech evidenced by the release of Sensei in 2016.Ado
69、bes aggregation of digital assets within its Creative portfolio(with Stock housing+175mn images)and the data it stores in its Digital Experience platform(powering 600b predictive insights annually)allow Adobe to form the new standards of productivity that are likely to be unlocked across knowledge w
70、orkers,particularly those within the CMO purview.The release of Sensei GenAi(focused on Digital Experience use cases)and Firefly(a suite of new Generative AI solutions across its portfolio),further situate Adobe to help shape the role Generative AI will play in driving productivity boosts expected t
71、o be a byproduct of greater adoption.For example,these products can jump start the ideation process,extract more value from the same assets(by allowing for a broad range of seamless editing to take place via Adobe Express),automatically generate campaign analytics or build customer journeys.Adobes s
72、trong directto-consumer go-to-market motion,along with its product led growth will likely allow Adobe to drive net-new user growth and be one of the first benefactors of these investments,especially in its Creative Cloud and wider Digital Media business.We also see this simplifying the user experien
73、ce,making switching costs higher and churn lower.Adobe may leverage this via price increases longer term as the tool set available to existing users increasingly evolves into smarter and more efficient offerings,with the innovation underpinned by Generative AI,automation,and synchronization.The comp
74、anys years of investments,the first-party data flowing through Adobes ecosystem and a strong partner network(with MSFT and NVDA)should support Adobes success without 26 March 2023 7Goldman SachsAmericas Technology_ requiring an outsized investment cycle.Intuit(Buy,PT$575):Intuits utilization of AI i
75、n a large-scale platform for its Virtual Expert Platform(which powers TurboTax Live by personalizing each tax-payers journey and connecting users with Tax experts)makes it one of the first-movers in successfully deploying AI-driven consumer-facing products.The companys natural language processing,AI
76、 tools(that can extract the needed tools from key tax forms,for example)and chatbot services(which use customers text and interactions to generate the best response)are additional examples of Intuits investment in such technology that can improve its product offering and user retention over time.Int
77、uit has built a strong reputation of investing ahead of the curve in terms of next-gen technology and embodying that into its product framework for future innovation.This leads us to expect Intuit to have the same mentality when building new products and growing existing services,such as QuickBooks
78、Live Mailchimp and CreditKarma.Another area of focus will likely be around a greater understanding of a customers needs and potential upcoming life events in order to best predict the best services they may need on the Intuit platform.Successfully deploying such enhancements,specifically in TurboTax
79、 Live and business tax offerings,can be important drivers for retention on the platform(which we expect to be in the low-80%).Given the breadth of the companys network,with over 100mn customers/SMBs serviced on the platform,Intuit has a plethora of data driving its models in a meaningful way.We expe
80、ct the 5-10 years of investments Intuit has allocated toward a unified,integrated data platform to pay off in this growing tailwind as it can allow for better self-learning and cross-references without a heavy investment cycle.While still in very early stages of embedding such technology across the
81、key aspects of the business,we expect this will increasingly become a key driver in Intuits success in capturing engagement and wallet-share across both consumers and businesses.What is Artificial Intelligence?The concept of Artificial Intelligence(AI)was first introduced in the 1950s when scientist
82、 Alan Turing decided to test the intelligence of computers to explore the possibility of machines being able to make decisions and solve problems like human beings.However,limitations in computing power and the expensive feat of the project stunted progress.Fast forward to the 21st century as Moores
83、 law came to fruition,technology platforms benefited from the rise of silicon chips that made heavy processing power faster and cheaper.According to Moores law,speed and compute capabilities can be expected to double every two years,thus reducing the cost per compute instance.With the explosion of d
84、ata collection in todays day and age,skeptics worry that the momentum in Moores law is decelerating.The fear is that there is not enough computational capacity currently available to train large language data models.We believe with the passage of time,technology will adapt and evolve to embrace thes
85、e challenges.Artificial Intelligence(AI)in simple terms is the simulation of human intelligence by 26 March 2023 8Goldman SachsAmericas Technology_ machines.Computer programs that leverage copious amounts of data and compute are trained to perform tasks such as decision-making and problem solving wi
86、th minimal human intervention.AI algorithms are typically rule-based and built through iterative processing to recognize patterns and make predictions.The evolution in technology such as cloud,compute and Big Data have been instrumental in making AI faster,cheaper,and more accessible.Key factors beh
87、ind the surge in AI:The first key piece of the puzzle was the development of Graphical Processing Units(GPU)for deep learning.Deep learning models are resource intensive and require a lot of computational power to train.GPUs are specialized processors that can optimize training of large language mod
88、els that run multiple computations simultaneously.Aside from the faster processing speed,GPUs also have higher memory bandwidth which is critical in training of large data sets.Advancements in GPUs have made it possible to run complex computations in Generative AI models quickly and efficiently.The
89、next big breakthrough in AI was transformer technology,which was introduced in 2017 by a Google Research paper called Attention is all you need.Transformer architecture changed the way neutral network models sequence data;transformers enable data parallelization,a mechanism used to make parallels be
90、tween words in a sentence.Transformer models learn relationships between variables based on sequential data or context.The simple premise is that transformer technology looks at associated words in a sentence and builds patterns over time,invariably forming the idea behind a sentence.This technology
91、 allows AI models to compute the relationship between input and output data without having to sequence it,considerably reducing the time to train models and reliance on structured data sets,overall enhancing the self-learning capabilities of AI(discussed further below).Lastly,with data as the powerh
92、ouse that fuels large language models,Big Data has been key in accelerating the growth of AI.Its the proliferation in data that has enabled the training of AI to be smarter and more efficient.Proliferation in data is Exhibit 1:Evolution of AI Source:Goldman Sachs Global Investment Research26 March 2
93、023 9Goldman SachsAmericas Technology_ best exemplified by the below exhibit,showing the amount of generated per minute of the day AI is a broad concept that encompasses various subsets which include machine learning,neural networks,deep learning,and natural language processing(NLP).Some prominent f
94、ields within AI include:Exhibit 2:Amount of Data generated per minute of the day Source:Domo,Data compiled by Goldman Sachs Global Investment Research26 March 2023 10Goldman SachsAmericas Technology_ Machine Learning(ML):AI enables machines to self-learn by training from model datasets.Algorithms ar
95、e trained to make predictions and identify patterns by learning from historical data inputs.These algorithms are created using ML programming languages.YouTube suggesting a video based on an individuals previous views is a simple example of an ML algorithm.The process of teaching the algorithm how t
96、o identify patterns starts with feeding the algorithm with a training data.The training data can be in the form of labeled or unlabeled datasets.Machine Learning models are trained using one of two techniques;supervised learning or unsupervised learning.Supervised learning is a similar technique to
97、how human beings learn.A computer or machine is presented with a training set consisting of a large volume of labeled datasets.And with unsupervised machine learning,the algorithm is trained to learn how to identify patterns from unstructured datasets.After the data is fed into the ML algorithm,it i
98、s tested to see if the predicted outcomes and the results match each other.If the prediction is not accurate,the algorithm is re-trained multiple times until the desired outcome is achieved.The process of iterative training ultimately enables ML algorithms to continuously learn and produce more accu
99、rate outputs over time.Supervised Machine Learning:Machine learning based on supervised training of a nmodel.The algorithms are trained using structured,labeled datasets.The model is Exhibit 3:AI is a broad concept encompassing ML,Neural Networks,Deep Learning Artificial IntelligenceMachine Learning
100、Neural NetworksDeep Learning Source:Goldman Sachs Global Investment Research26 March 2023 11Goldman SachsAmericas Technology_ first trained with large volumes of corresponding input and expected output values,and is then leveraged to predict outputs based on the test data.For example,to train an ML
101、model to recognize the image of a cat:The model is first trained with millions of images of cats and other animals so that the model can learn to distinguish between features such as color,size,shape,etc.Once the model is well-trained,when asked to identify the said object,it can identify a cat.Appl
102、ications include fraud detection,image segregation,medical diagnosis,etc.Unsupervised Machine Learning:Unlike supervised machine learning,these ML nmodels do not require supervision and are trained using unlabeled datasets.The model identifies patterns,similarities and differences from the input dat
103、a to spin out an output answer.Common applications of such models are anomaly detection.Reinforcement Learning:Similar to unsupervised machine learning but enhanced nwith an automatic feedback loop that improves the performance of the algorithm.There are no classified datasets in the model.The model
104、 is trained based on patterns and trends with an added feedback layer that reinforces the algorithm every time the right output is generated,thus strengthening the model with each trial.Neural Networks:Works similar to neurons in a human brain.Just as human beings process information via a vast comp
105、lex network of neurons,an artificially created neural network works on a system of multiple neural nodes that process and filter through information in multiple stages.The data flows via an artificial neural network(ANN)through a feed-forward process or back-propagation algorithm.Feed-forward neural
106、 networks processes data uni-directionally,passing data from the input to output node.The initial layer in an ANN is akin to the human optic nerve that receives the raw data input.Each subsequent node is an individual knowledge hub,which filters and classifies the data at each stage.The output from
107、the previous layer serves as an input feed to the next node,as each node predicts an outcome the next node assesses if the previous output was correct.Back-propagation allows data to flow through multiple different paths assigning the highest weight to the pathway that produces the most accurate ans
108、wers and lower weights to neural pathwith weaker outputs.By leveraging ways these continuous feedback loops,artificial neural networks improve their predictive analysis.ANN models can be trained using techniques such as supervised learning,unsupervised learning and reinforcement learning.26 March 20
109、23 12Goldman SachsAmericas Technology_ Deep Learning:Subset of machine learning.Traditional machine learning requires a higher degree of human intervention in determining the right data inputs and design of the features to be analyzed by the ML software,which limits the creativity of the algorithm.D
110、eep Learning algorithms,on the other hand,learn through observation of unstructured data and are layered with artificial neural networks which attempts to better simulate human behavior.Deep learning algorithms learn independently to identify features and how to prioritize data attributes when they
111、are fed with input datasets.The key difference between traditional ML and deep learning is the use of artificial neural networks to train the algorithm as opposed to human intervention.Thus advancing the usage of deep learning techniques over traditional machine learning.For example,to train an algo
112、rithm to identify the image of a cat from a large set of animal images,the ML algo would need to be told by a human being to identify features such as the shape of the animals tail,ear,color of the fur,number of legs,etc.However,for the training of a deep learning algorithm,the ANN would process the
113、 set of underlying images,assess the features to be evaluated,and the order of priority they need to be evaluated in to generate the most accurate output.Natural Language Processing(NLP):In the past one could only communicate with a computer programs through code.With the evolution of NLP,machines c
114、an communicate with humans in their natural language.NLP enables computers to Exhibit 4:Artificial Neural Networks Architecture Source:Goldman Sachs Global Investment Research26 March 2023 13Goldman SachsAmericas Technology_ interpret speech,gauge sentiment and read text.For machines to be able to u
115、nderstand human languages,NLP involves two techniques;syntactic and semantic analysis.Syntactic analysis identifies the structure and relationship between words in a sentence.Semantic analysis focuses on the meaning of the words and understanding the context of the sentence.NLP algorithms index huma
116、n language queries and convert them to machine language via the following processes:Tokenization is the processes of breaking down text into smaller semantic units.The next step in the process involves removing words such as prepositions and articles that have no incremental information.Lemmatizatio
117、n and Stemming help categorize and convert words to their root words(e.g.,the word better would be transformed to good).Lastly,part-of-speech-tagging helps tag words according to their grammatical context as nouns,verbs,punctuation,etc.Through these various steps computers are able to understand,ana
118、lyze and translate human text and speech.NLP techniques are used to train traditional machine learning and deep learning algorithms.Exhibit 5:Steps in Natural Language Processing Source:Goldman Sachs Global Investment Research26 March 2023 14Goldman SachsAmericas Technology_ Traditional AI vs Genera
119、tive AI AI tools have been around for a while now,however,the recent excitement around the topic is credited to a new break through category of artificial intelligence called Generative AI.Generative AI has creative capabilities to generate original ideas in the form of text,image,audio,video,code a
120、nd more all based on simple human text queries.Generative AI leverages a deep learning technique called Generative Adversarial Networks(GANs)to generate content.A GAN neural network contains a generator node that enables new data creation and a discriminator node that evaluates each output.The gener
121、ator is pre-trained on large datasets to create new outputs,while the discriminator is trained to distinguish between the real output and the AI generated one.The model is fine-tuned until the generator improves its output up to the point that the generator convinces the discriminator that the new c
122、ontent created is different from the original data.Equipped with the power to create original content,Generative AI is touted to be future of artificial intelligence with the potential to disrupt several industries.Traditional AI models are built on discriminative statistical models,that are predict
123、ive in nature and primarily focus on recognizing patterns from existing data.Generative models,on the other hand,can produce new instances of data based on an underlying set of data inputs.Unlike traditional AI,which is cognitive in nature and typically leveraged in an analytical context,Generative
124、AI is more perceptive.Generative AI goes beyond the usual pattern detection and wry data analysis,it hones the creative aspect of AI.Rather than limiting model outputs to mere analytical answers,Generative AI can generate/create new content(be it music,poetry,audio,video,code,etc)from an underlying
125、data set.What makes the Generative AI trend different from the previous AI waves is its ability to break down communication barriers between humans and machines.With Generative AI,humans can easily communicate with computers in their natural language rather than in a programming language.The advance
126、ments in transformer technology and increased computational power aided the process.Demystifying Transformer Technology Transformers revolutionized the way human beings and machines interact.The concept was first introduced by a team of Google research scientists in 2017 in the paper titled Attentio
127、n is all you need.Transformation is a type of data processing that is performed on a sequence of data points,such as words in a sentence or letters in a particular word.Transformer models are machine learning models that are specifically designed to process sequences of elements.They consist of enco
128、der-decoder blocks where the encoder takes the input and the output is generated by the decoder block.The premise of the model is an attention mechanism,that assigns importance to a few words/elements that essentially form the gist of a sentence.Self-attention helps establish a relationship between
129、each word in the sentence with every other word,to form multiple output sentences for every combination of words.Thus,by focusing on the important aspects of a sentence,the attention layer improves the quality of output from a neural network.Transformer architecture changed the way neutral network m
130、odels sequence data;by enabling data parallelization,a mechanism used to make 26 March 2023 15Goldman SachsAmericas Technology_ parallels between words in a sentence.Transformer models learn relationships between variables based on sequential data or context.The transformer technology looks at assoc
131、iated words in a sentence and builds patterns over time,invariably forming the idea behind a sentence.This technology allows AI models to compute the relationship between input and output data without having to sequence it,considerably reducing the time to train models and reliance on structured dat
132、a sets,overall enhancing the self-learning capabilities of AI.Understanding Large Language Models(LLM)The concept of large trained lends itself to large language models as a result of beingon millions of parametersNatural Language Processing and the use of to communicate with the model introduces th
133、e language aspect to LLMs.True to its name,Large Language Models are a type of machine learning model that are trained on large parameters of inputs while using natural language to process queries.LLMs are based on transformer architecture and use deep neural networks to generate outputs.Transformer
134、 neural networks use the self-attention mechanism to capture relationships between different elements in a data sequence,irrespective of the order of the elements in the sequence.The computational power of transformer models to process data sequencing in parallel on massive data sets is the biggest
135、driving force behind large language models.Breaking down how LLMs work.The first step in training a large language model is building a large training data set.The data typically is derived from multiple sources across websites,books and other public datasets.The model is then trained using supervise
136、d learning,where it is trained to predict output words in a sequence.A LLM learns which words are most commonly used together,the order they appear in and how they relate to each other.These relationships are taught by training the neural network on large datasets.The more data the model is trained
137、on,the better the outputs.The process of training LLMs involves first converting the natural language text data into a numerical representation that can be input into the model.This process of converting the input sequence to a vector representation is called word embedding.The self-attention mechan
138、ism of the transformer model then captures the relationship between the input sequence.The strength of Large language Models lies in two key aspects:1)pre-training of the model and,2)fine-tuning of the model to adapt for specific tasks.Pre-training is a rule-based training process that helps the mod
139、els to learn the basic rules and dependencies within a sequence of data inputs.This initial training is done on a large dataset and requires massive compute power to complete.The model is constantly adjusted and the process is repeated until the model reaches a desired level of performance.Fine-tuni
140、ng involves the training of general purpose LLM models for a specific domain or task,this approach enhances the performance of LLM models.These models ultimately get smarter and better over time.A reward mechanism in the model reinforces positive outcomes by rewarding the algorithm when it generates
141、 a desired outcome,thus improving the quality and accuracy of outputs.26 March 2023 16Goldman SachsAmericas Technology_ ChatGPT is the most popular example of a Large Language Learning Model.In November 2022 OpenAI democratized AI with ChatGPT,an AI language model built on transformer technology tha
142、t can generate human-like responses based on text inputs.Since the release of ChatGPT,we have seen a significant spike in conversations on the topic.We leveraged GS Data Works for an analysis around the mentions in ChatGPT/OpenAI in company earnings transcripts.In the recent 4Q22 earnings season we
143、saw multiple companies call out investment opportunities in the technology.Exhibit 6:Continuous reinforcement mechanism in Large Language Models Source:Goldman Sachs Global Investment Research Exhibit 7:Companies calling out opprotunities in Generative AI Russell 3000 1Q 2019-4Q 2022 051015202530Tot
144、al Mentions per Company Source:Goldman Sachs Global Investment Research,GS Data Works26 March 2023 17Goldman SachsAmericas Technology_ Generative AI Frameworks Humans have been interacting with AI for a few years now,take for example the auto type feature on your search bar or the suggested watch li
145、st on your favorite streaming platform.But the recent buzz around AI can be attributed to the launch of ChatGPT by OpenAI in November 2022.While OpenAI remains the more popular Generative AI platform there are several other interesting players in the ecosystem,we highlight below a few of them.1)Text
146、 based Generative AI Frameworks:ChatGPT,JasperAI a)OpenAIs ChatGPT nGenerative Pre-trained Transformer or ChatGPT,as its commonly known,rose to popularity owing to its intuitive,easy-to-use AI language model that can generate long form text outputs.ChatGPT is based on GPT-3,a large language model al
147、so developed by OpenAI and with 175bn parameters,much bigger than the 20bn parameters ChatGPT has been trained on.Both(GPT-3&ChatGPT)models have natural language processing capabilities that can perform tasks such as text summarization and language translation.However,the key improvements in ChatGPT
148、 are the reinforcement learning component and the conversational tone of the AI model that make it unique from its predecessor.OpenAI made ChatGPT and AI common dinner table conversation for many.The company elevated AI to the next level with its Generative capabilities which opened a plethora of ne
149、w use cases(more below).True to its name,the company made the tool free,publicly available to all and easy to use with its human touch.Thus,making ChatGPT more popular than traditional AI tools which are predictive models with analytical outputs.As of March 2023,OpenAI released GPT-4.The GPT-4 large
150、 multimodal model has more advanced natural language processing capabilities with the ability to accept inputs in the form of both text and images to generate text outputs.In general large multimodal models have the ability to process and generate inputs/outputs in multiple modes be it text,image,an
151、d even audio and video.GPT-4 is larger and more powerful than GPT-3.GPT-4 is trained on 100+trillion parameters compared to GPT-3 which was trained on 175bn parameters.This is expected to make GPT-4 Exhibit 8:Google Trends showing interest in AI and ChatGPT have spiked since November 22 020406080100
152、1---------------------05201
153、4------------------02Worldwide Google Search Trends ChatGPTStar WarsAI/MLiPadCl
154、oud Source:Goldman Sachs Global Investment Research,Google Trends,GS Data Works26 March 2023 18Goldman SachsAmericas Technology_ more creative,accurate and close to human performance.GPT-4 can be customized to generate outputs that are in a particular tone,type of writing style and is also multiling
155、ual.GPT-4 is far superior compared to its predecessor with the ability to score high on tests designed for humans,such as the SATs,bar exams,etc.The below exhibit doesnt have the proper formatting(i.e.,missing exhibit number,incorrect exhibit placeholder);please fix so it matches other exhibits and
156、also to avoid possible issues at publication Exhibit 9:No.of ChatGPT visits since launch 000s 05001,0001,5002,0002,5003,0003,500 Source:Goldman Sachs Global Investment Research,SimilarWeb26 March 2023 19Goldman SachsAmericas Technology_ How does ChatGPT works?The step-by-step process:Input:human use
157、r types a question or command in ChatGPT 1.Tokenization:the input text is tokenized or broken into separate words to be 2.analyzed Input embedding:the tokenized words are inputted in the neural network 3.transformer Encoder-decoder:the transformer converts the text into code and generates a 4.probab
158、ility distribution of outputs.Output:the model output is then de-coded to text that is readable by humans.5.ChatGPT is built on the tenets of supervised learning and reinforcement learning.The first step in training the GPT model involves collection of sample data and training it based on a labeled
159、dataset or supervised programs to fine tune it.To reinforce learning in these models,a reward model is built with a set of comparable data outputs which Exhibit 10:OpenAIs ChatGPT was the fastest application to surpass 1mn users Source:Goldman Sachs Global Investment Research26 March 2023 20Goldman
160、SachsAmericas Technology_ are sorted by quality to train the model to generate the best possible outcome.The model is optimized with each use,thus ensuring continuous learning.Step 1-Involves collection of data and supervised training.First the model is trained with a sample prompt,for example expla
161、in the moon landing to a 6 year old.A data scientist or a human labeler then demonstrates to the model the desired output.The model is then supervised and fine-tuned by a human until it provides an output that satisfies the desired performance level.Step 2-Collection of comparable datasets and rewar
162、d training of the model.For the same prompt,the labeler now demonstrates several outputs to the model,with the outputs ranked from best to worst.The model that was previously trained via supervised learning(in step 1),is now expected to generate an output as close to the desired outcome as possible.
163、Several sample data outputs are used to train the model to generate the highest ranked output.Step 3-Optimizing the model using reinforcement learning.When the model produces an output that is ranked highest,it is rewarded to reinforce this positive outcome.When a new prompt is sampled in the model,
164、say for example write a story about frogs,the trained model generates an output.The reward model(from step 2)then kicks in and calculates the reward for the newly generated output.If the new output for the frog story ranks high in the stack of desired outputs,the model policy is automatically update
165、d.26 March 2023 21Goldman SachsAmericas Technology_ b)JasperAI nSimilar to ChatGPT,JasperAI is a chat based text to text Generative AI platform,like ChatGPT,JasperAI too is based on GPT-3.However,ChatGPT is more chatbot like and is largely leveraged as a conversational search tool to get quick answe
166、rs,synthesize content,write essays and articles.JasperAI on the other hand is like a writing assistant and is viewed as the more creative tool.It is used to create content for websites,videos and marketing templates.OpenAIs free offering of ChatGPT resulted in a broad base user audience such as stud
167、ents,creators,individuals,and even enterprises.JasperAI is a paid service that is mostly used by creative professionals in the field of marketing,publishing,etc.2)AI Frameworks&Applications from Alphabet(GOOGL):Language Models&Frameworks:LaMDA(Language Model for Dialogue Applications)is Googles natu
168、ral language n Exhibit 11:How ChatGPT works Source:Goldman Sachs Global Investment Research26 March 2023 22Goldman SachsAmericas Technology_ processing framework that utilizes a group of neural language models built on transformer architecture.Given the fact that the group of models are trained on d
169、ialogue-based text,LaMDA is capable of more open-ended/conversation-based language understanding and text generation compared to Googles previous models(BERT,MUM,etc.).Since LaMDA has first unveiled at Google I/O in 2021,Google has integrated it into many of its core products,most prominently within
170、 Google Search to better understand/interpret search queries and improve search results.PaLM(Pathways Language Model)is Googles large language model trained with nits Pathstem,a model architecture that enables a single model to generalize ways syand execute across multiple domains&tasks at once(incl
171、uding language understanding&generation,reasoning,pattern recognition,code generation,etc.).Google has recently announced the launch of direct API access to PaLM for third party businesses and developers.Generative AI Products:Bard is Googles recently announced conversational AI service.As a standal
172、one nuser-facing product(separate from Google Search or other products),Bard features a conversation-based interface and inputs to generate responses utilizing real-time data from the open web and powered by Googles LaMDA LLM.Google recently opened up early beta access to Bard,with plans to formerly
173、 launch to the broader public in the near future.Imagen,Imagen Video,Parti&Phenaki Googles suite of text-to-image and ntext-to-video products that utilizes a variety of model frameworks(autoregressive,diffusion,encoder-decoder and transformer models)to generate high-fidelity content from text-based
174、inputs.AudioLM&MusicLM Googles framework for audio generation(speech,music,netc.)utilizing input prompts of similar speech or audio.3)Hugging Face:an API platform for AI Is a community hub for open source AI models with the aspiration of becoming a Git like repository of the AI world.The community c
175、onsists of ML engineers,data scientists and AI researchers to share models and datasets.The HuggingFace repository is categorized into models,datasets and spaces.Models and datasets are stored in a repository,and are similar to programming code repositories where users can share models/datasets publ
176、icly,create new branches and versions.The plug and play platform allows users to select any model from the interface,pick a cloud provider and integrate the model with an application.Spaces give users the opportunity to demonstrated their AI applications to the community.Currently development of AI
177、applications is largely limited to large enterprises.With Hugging Face users can easily build AI applications without having to build a scalable platform and investing expensive infrastructure.With over 120k models,20k datasets and 50k demos the Hugging Face platform is democratizing AI for a much b
178、roader audience.4)Image based Generative AI Frameworks:OpenAIs DALL-E 2,MidJourney,Stability AIs Stable Diffusion 26 March 2023 23Goldman SachsAmericas Technology_ Stable Diffusion is a deep learning model that can convert text to image.These models are built using latent diffusion models which are
179、machine learning models that map the structure of the training dataset in a low dimension latent space(the image compression in latent space is faster).AI tools such as DALL-E 2,MidJourney and Stability AIs Stable Diffusion can create images from natural language description.Based on text inputs,the
180、se AI tools can generate a new image or create a variation of an existing image while keeping intact components of the original image such as the color palate,textures,and shadow.These models are trained on millions of images to form an association between the words and images.DALL-E 2 by OpenAI off
181、ers limited number of free images per user with an option to purchase more.MidJourney was released in July 2022,a few months after DALL-E 2 and can be accessed via an API or through the Discord chat server.The big break in the text-to-image space came with the release of Stable Diffusion in August 2
182、022 by Stability AI.Just as ChatGPT made text-to-chat AI easily accessible to all,Stability AIs Stable Diffusion opened the doors for mass usability of text-to-image AI.Unlike DALL-E 2 and MidJourney,Stable Diffusion is open source and free,with the biggest differentiator being its ability to run wi
183、th limited compute on consumer hardware.Although DALL-E 2 requires higher computing power,it is the preferred choice to run more complex queries.5)Text to Code Generation Frameworks:Codex by OpenAI and GitHub Co-pilot by Microsoft Codex by OpenAI is a natural language processing tool that can genera
184、te programmable code scripts from text prompts.Developers can generate new code from plain text comments,translate code from one programming language to another and interpret code written in a different language.The tool is trained from publicly available source codes and public GitHub repositories
185、with the ability to write code in any programming language(python,JavaScript,Ruby,etc).OpenAI and Microsoft partnered to release GitHub Copilot,which fine-tunes the Codex model to leverage GitHubs vast code repositories as a base.Developers note that with an AI-based coding copilot,time to release c
186、ode is faster,resolution times improve,thus increasing the productivity of developers.Currently,Codex is a free offering and GitHub Copilot is available at$10/month for individuals and$19/month for businesses.GitHub Copilot is favored by the developer community because the tool gives developers more
187、 autonomy and control than Codex.While Codex generates a code script in response to prompt query,Copilot offers multiple suggestions and allowing the developer to pick the best fit.Generative AI Technology Stack and Its Implications Just as cloud computing was instrumental in the growth of IaaS/PaaS
188、/SaaS,computing power will be the key accelerant for widespread adoption of Generative AI.When Salesforce and NetSuite went public in 2004/05,we had limited visibility to predict that in 15+years these companies that are powered by the cloud would be billion-dollar software businesses.The mass avail
189、ability of cloud hosting 26 March 2023 24Goldman SachsAmericas Technology_ services made accessible by AWS was the primary driver behind the acceleration of SaaS businesses.Prior to that,creating and managing cloud infrastructure was difficult and expensive.Following suit in 2022,OpenAI is breaking
190、the barriers for the mass adoption of AI as a service.We believe the next could be driven by enterprise wave of AI SaaS companies.Enterprise software companies have a long-standing history of leveraging intelligence platforms.Before AI,enterprise software companies relied on BI(Business Intelligence
191、)applications for data analysis.Much like AI,BI too is grounded by machine learning principles.However,AI has more far-reaching applications than business data analytics.Going forward,we can expect enterprises to integrate AI into their broader business strategy.We believe Generative AI is still in
192、the early days with a long runway for growth with multiple applications in infrastructure,software and beyond.We further explore the opportunities for AI to be integrated into the tech stacks of the future by drawing an analogy between a cloud computing stack and an IT stack layered with AI.Infrastr
193、ucture Infrastructure Layer:the foundation of a Generative AI tech stack is powered by a strong transformer model.Building an underlying transformer model requires an Exhibit 12:In the future we anticipate the traditional IT stack to be integrated with AI at each layer Source:Goldman Sachs Global In
194、vestment Research26 March 2023 25Goldman SachsAmericas Technology_ abundance of compute infrastructure,and funds for research and development of complex neural network models.The high barriers to entry and investments cost will limit the number of players in this market.Akin to the large public clou
195、d service providers(namely,AWS,Azure and GCP),we are already seeing a few dominant players in this space with OpenAI,Stable Diffusion and Google AI research labs.We can compare this layer of the AI stack to that of the Infrastructure as a Service(IaaS)layer of the cloud computing stack.Similar to Ia
196、aS players delivering computing infrastructure such as servers,storage and networking on demand,we anticipate companies potentially buying transformer models to customize and build AI related business solutions on top of these models.Case in point is Microsoft leveraging OpenAIs Generative AI for Te
197、ams Premium,Bing Search(more below).The infrastructure layer going forward.Based on our learnings from the cloud computing era,we postulate what the AI competitive landscape could look like.In the infrastructure layer,similar to the cloud computing stack where the hyperscalers garner majority market
198、 share,we anticipate a limited number of players operating in this foundational layer of the AI technology stack as well.The heavy capex investments,high compute costs,large dataset requirements and intellectual property involved in developing large language models and transformer architecture makes
199、 it difficult to build and scale.This is comparable to the IaaS stack where hyperscalers spend generously on capex investments in setting up and maintaining data centers,servers,networks,etc.Additionally,unlike code which can be easily upgraded to a new version,Generative AI models cant be rebased w
200、hen the underlying model needs to be updated.For these reasons we believe many enterprises would prefer to buy LLMs rather than build from scratch,thus limiting new entrants into the space.The high barriers to entry and investment costs are compelling reasons why we believe the infrastructure layer
201、of the AI tech stack too will be dominated by the likes of Google,Microsoft and OpenAI.Platform Platform Layer:will consist of AI frameworks and APIs(Application Programming Interface)that will help developers produce AI enabled applications.Like Platform as a Service(PaaS)enables software developme
202、nt by allowing developers to focus on application creation without the hassle of maintaining the underlying infrastructure,we anticipate the growth of plug and play AI APIs will democratize the delivery of AI solutions by enterprise software companies.Platforms such as Hugging Face and GladIA have g
203、ained recent traction for their plug and play AI APIs.They offer a one-stop-shop solution for open-source AI APIs,helping software developers easily find a model that fits their needs.The platform layer going forward.On the platform side,we envision a few strong platforms emerging.This layer will be
204、 crucial in integrating the foundational models(LLMs)with the end application.AI Frameworks and APIs will be the key enablers,opening up opportunities for new companies to rise.We are starting to see signs of similar trends to that of the PaaS cycle emerge in the platform layer of the AI stack,i.e.1
205、)the strong relationship between the PaaS vendors and the cloud providers,and 2)the number of open source players in the market.Within the platform layer of the AI stack we are seeing players like OpenAI(which has a strong relationship with Microsoft)offer 26 March 2023 26Goldman SachsAmericas Techn
206、ology_ APIs that can be leveraged to build AI-enabled applications.Community-based platforms like Hugging Face are democratizing AI with their library of pre-trained models and datasets.On the text-to-image side,Stable Diffusion has an open source offering.Following suit with their cloud computing c
207、ounterparts,these platforms have the opportunities to monetize their offerings.Starting with a freemium model to acquire users,these companies can eventually charge for usage of their APIs/frameworks,be it via a subscription model,pay-per-use,etc.Similar to the PaaS market,gaining developer(AI/ML en
208、gineers)mindshare will be instrumental in establishing market dominance within the AI platform layer.We do expect history to repeat itself on the platform side,whereby the number of platforms with GAI likely come down to a select few.We also envision a scenario where a few platforms become verticali
209、zed and support the distinct needs of end-markets such as Financial Services,Government,Healthcare,etc.Implications of Generative AI for data and analytics.Generative AI is likely be most impactful for app-centric enterprise software companies(and the underlying compute providers)in the shorter-term
210、 given tangible productivity benefits for end-users of business apps,as seen with the recent announcements of GitHub Copilot,Microsoft 365 Copilot,Adobe Firefly,and Einstein GPT.While it remains early,we believe the data and analytics ecosystem should also be a beneficiary of this transformative tec
211、hnology over time,as well as the broader AI/ML theme,considering the fundamental role data will play in developing and iterating on large-language models and enhancing production applications leveraging GAI.We highlight our views on the following companies within data and analytics:Snowflake:An area
212、 where we see potential for Snowflake to leverage Generative AI is integrating ChatGPT or a comparable service to allow end-users to construct data warehouse queries in human text or voice,which,through natural language processing(NLP),can then be converted into the equivalent SQL command to extract
213、 data insights.In our view,this has the potential to further democratize access and use of Snowflakes Data Cloud to a less technical audience,while also driving productivity gains by cutting down on human interactions with the data warehouse.With the emergence of several large-language models(and ot
214、her pre-trained ML models),Snowflake is positioned to benefit,for instance,from its customers scoring data for their ML models(i.e.fine-tuning pre-trained models on new datasets to improve the algorithm).Snowflakes Data Sharing use case could also become increasingly relevant in an AI-driven world a
215、s customers leverage public datasets on the Data Cloud to better inform their ML models.The company already has 23%of its customers with at least one stable edge in F4Q23(i.e.,continuous data-sharing connection between two or more parties),up from 18%in the prior year.We also we highlight Snowflakes
216、 announced initiatives below around data science/ML that underscore its prioritization of investment in the category.We expect to hear more from Snowflake on the role Generative AI will play in its business at Snowflake Summit on June 26th.Snowpark:GA in F4Q23(Python).Snowpark allows data scientists
217、 to work with ntheir preferred programming languages,including Python,and Scala(vs.SQL),Java to enable end-to-end machine learning workload development,deployment,and orchestration.Through Snowpark,customers can ingest,analyze,and transform 26 March 2023 27Goldman SachsAmericas Technology_ their dat
218、a to train ML models that can run predictive analytics to drive better business insights and outcomes(i.e.,identifying high-risk credit card customers based on demographic and credit history).Further,Snowflake recently introduced Snowpark optimized-warehouses that specifically cater to ML training u
219、se cases,with 16x memory per node vs.a standard Snowflake data warehouse.As data continues its rapid growth,we expect enterprises to increasingly look to ML use cases to achieve cost savings and competitive differentiation.While Snowflake is early in its opportunity in data science/ML(20%of customer
220、s have tried Snowpark,as of 4Q23)relative to peers such as Databricks,we believe Snowpark could be a growth vector for the company in F25 as Snowflake expands usage/adoption within its enterprise installed base.Snowpark optimized-warehouses(for ML use cases)require 50%more credits/hour relative to a
221、 standard DW,which could serve as a consumption tailwind for the company longer-term as ML workload adoption accelerates.Streamlit:With Snowflakes acquisition of Streamlit in March 2022($800mn),ncustomers can now develop data-intensive apps with only a few lines of Python code.Streamlit was built to
222、 simplify the process of contextualizing data analytics tasks and machine learning model outputs through front-end web applications,which can improve decision-making for customers.With the emergence of Generative AI,we expect developers will be able to offload some responsibility of generating Pytho
223、n code to construct web-apps for data visualization by instructing ChatGPT(or a competing service)to provide the Python script based on a set of commands.Over time,Generative AI may also be useful in assisting developers with creating the most effective visualization based on the underlying model ou
224、tputs.MongoDB:Akin to Snowflake,we believe that one potential benefit of Generative AI services such as ChatGPT as it pertains to MongoDB is the ability to query the database using natural language,which can then be converted into MongoDB Query Language(MQL).However,we believe that in many instances
225、 MongoDB will be an indirect beneficiary of the rise in Generative AI.For example,as discussed in this report,we expect that Generative AI will not only benefit incumbent enterprise software providers through add-on SKUs but also spawn a new generation of AI-centric applications.MongoDBs Atlas is fi
226、rmly levered to the development/growth of cloud-native apps,which have the potential to see an accelerated time-to-market with the introduction of AI DevOps tools such as GitHub Copilot that can increase developer productivity.Further,as enterprise adoption of ML models proliferates,we could see a p
227、otential acceleration in the re-platforming of legacy databases to modern alternatives such as MongoDB as customers seek better access to mission-critical internal data that can be used to inform their ML models.For MongoDBs Enterprise Advanced self-managed offering,we see the potential for services
228、 like ChatGPT to augment the role of database administration over time.Databricks:Databricks(private co.)is a company within the AI/ML ecosystem whose roots are in processing and transforming significant amounts of data(structured,unstructured,semi-)to develop ML models for predictive analytics.Whil
229、e Snowflake and 26 March 2023 28Goldman SachsAmericas Technology_ Databricks are often bucketed as head-to-head competitors,we note that Snowflake has historically been more focused on running analytical queries on structured and semi-structured data(albeit with a more concerted push into ML with Sn
230、owpark),while Databricks has been geared more towards data science and machine learning workloads due to its support of unstructured data through its Lakehouse architecture(Data Lake+Warehouse).Databricks offers customers the option to run open-source,pre-trained LLMs such as Hugging Face on its Lak
231、ehouse platform for out-of-the-box use cases to generate valuable outputs against stored data.Further,with Hugging Faces Trainer API,Databricks customers can fine-tune existing models to a tailored use case by training against their own internal datasets.While OpenAI brought Generative AI to the for
232、efront of public conversation,Databricks has historically been geared towards data science use cases,including recommendation engines and predictive analytics,with a dedicated machine learning environment that manages the entire ML lifecycle.Application Application Layer:we expect AI to be packaged
233、as a solution,much like Software as a Service(SaaS).We believe Generative AI will increase the competitive moat for incumbent SaaS solutions.We dont anticipate AI companies competing with enterprise SaaS companies but rather collaborating.Successful SaaS companies are built on the back of strong go-
234、to-market motions,iron clad execution and robust product integrations.Layering AI with B2B SaaS solutions should enhance their technological moats.Typically,SaaS companies house large volumes of customer data,HCM,financial data,Vertical SaaS-insurance,healthcare,etc.We believe leveraging their 1P us
235、er data to reinforce and train AI driven large language models is a natural segue for B2B SaaS companies to extract critical insights,automate tasks,increase employee efficiency,and eventually monetize these solutions.The application layer going forward.The application layer is where we expect to se
236、e the most innovation.Here we see a convergence of Enterprise SaaS and Generative AI.We anticipate the emergence of standalone Generative AI companies that are venture-funded to be in a position to leverage new data and language models.We also envision existing SaaS businesses embracing Generative A
237、I and vastly reducing the complexity of the end-user experience.To emerge as a category leader,SaaS companies will have to keep up by integrating AI-enabled solutions in their product DNA.SaaS companies will have to be quick to develop new premium AI SKUs for their existing products as well as add n
238、ew AI product modules.With all the players in the market purchasing AI infrastructure and platform services from the same subset of AI companies,establishing product defensibility becomes imperative.This where SaaS companies will need to leverage their competitive moats,be it their efficient go-to-m
239、arket motions,strong sales execution or existing product strength,to rise as SaaS AI-enabled category leaders.Workforce Productivity Likely to Benefit From Adoption of Generative AI Generative AI frameworks can streamline business workflows by unifying data 26 March 2023 29Goldman SachsAmericas Tech
240、nology_ and apps,eliminating routine tasks.A lot of the value of todays knowledge worker is based on ability to navigate different computer programs,be it on a desktop,tablet or smartphone.Generative AI can unclutter user interfaces of such platforms to streamline the user experience by navigating t
241、he enormous complexity under the hood via code generation.This is powered by transformer technology(discussed above),that intelligently divides what is needed to accomplish and offers the completed results quickly.Potential for significant productivity boost.A lot of what software does today is base
242、d on automating a certain piece of logic or business process.Such progress,when first seen with the invention of the PC,resulted in a tremendous boom in productivity,albeit at a lag(Exhibit 13)that was largely attributable to the time it took to reach mass scale.For example,the output of an analyst
243、in any given field 20 years ago would pale in comparison to the level of deliverables seen today.The delta is driven by advancements in software and hardware technology that has allowed us to squeeze more out of the same unit of time due to the proliferated access to information tools that underpin
244、productivity.It would not be a stretch to say todays workers have access to more information and tools that augment their knowledge base and enable faster processing/passing/analyzing of that information.GS macro team estimates that Generative AI can be a boon to the US labor productivity growth,rai
245、sing it by just under 1pp over a 10-year period following mass adoption(link to macro report).They note,however,that this outcome may range from 0.3-3.0%depending on the caliber of the technology,the base of adoption,and potential replacement of certain jobs.We also point out,that despite all this a
246、utomation,the workforce has not shrunk and the economy has only expanded due to the aforementioned improvements.As the labor force gets more efficient,the group tends to accomplish more within the same hour constraints.We see Generative AI being the lever that drives another step up in such efficien
247、cies,helping eliminate routine tasks and cumbersome procedures.For example,imagine a research analyst issuing a very simple command on their search bar inside their firewall where they have trained a technology like ChatGPT or Bard to study their models.When the analyst is asked which companies are
248、showing an accelerating growth rate or expanding margins,neither the analyst nor their team Exhibit 13:Previous Milestone Technologies Have Led to Labor Productivity Booms 062060800100 741953Electricity:ManufacturingElectricity:Househ
249、oldPersonal Computing:WorkplacePersonal Computing:Household PercentPercentAdoption of Previous Technologies Over TimeProductivityboom begins50%Adoption ThresholdDevelopment of electric motorPersonal computer invented00.511.522.533.544.50.00.51.01.52.02.53.03.54.04.5 5520052025P
250、ercent change,10-year annual ratePercent change,10-year annual rateUS Labor ProductivityGreen=Resulting productivity boomDevelopment of electric motor:1890Personalcomputer invented:1981 Source:Bureau of Labor Statistics,US Census Bureau,Our World in Data,Woolf(1987),Haver Analytics,Goldman Sachs Glo
251、bal Investment Research26 March 2023 30Goldman SachsAmericas Technology_ members will need to sift through various models.That should happen in seconds.They dont even need to type the e-mail with their results-the AI will do that for them.That does not mean less work is done.It just means that the a
252、nalysts will spend their time doing more productive and rewarding work that can lead to additional capacity.You dont need to be a programmer to write code.The benefit of Generative AI is that the computer can just take your text or voice and translate that into code that can quickly run various task
253、s,such as risk analytics,statistical modeling,routine memo drafting,forecasting,collateral generation,financial reporting,etc.Comparing this to a SQL-based approach,where an application has to talk through an API to an underlying database using SQL code,Generative AI is likely to modernize this proc
254、ess by using your natural language input to compose or extract the right answer from the database.Said differently,in the past you could only communicate with a computer program through code.You needed to structure your query in the syntax the program would understand to get the answer to a question
255、 from a database.Now,natural language processing(NLP)reduces the premium placed on SQL programming whereby as the NLP is able to work with the database and negotiate commands in plain English.While this may seem to imply the relevance of code is diminishing,it actually means that the need for code p
256、roliferates as Generative AI frameworks divide and contextualize the inputs provided and convert them into programs that accomplishes tasks.Going off the capabilities displayed by Siri or Alexa,which have become household technologies,Generative AI goes further as it has a much more tangible UX expe
257、rience and can process much more complex tasks.A new class of business applications can be built on Generative AI.The big difference between Generative AI as a platform and prior platforms is that Generative AI learns at a rapid clip and gets better with usage.Generative AIs continuous feedback loop
258、,which allows it to constantly improve on its accuracy and knowledge base,should allow it to lead to the creation of a new class of applications.Domain expertise,such as medicine,are likely to be digitized in a transformational way.For example,it may also be able to scour through users medical recor
259、ds and tests,working to provide deeper insights and more definitive conclusions.Another use case can be in the film industry,where a producer has a vision for a short documentary or movie and can execute it entirely with Generative AI.The field of creative software is going to be vastly reshaped,in
260、our view.While it might be concerning to think that computers can take over the creative process,removing the creative artists touch,the opposite could likely occur when creative artists feel that Generative AI can get them to a starting point that was previously unimaginable.The hybrid nature of ge
261、nerated assets(that allow for editing and adjustments from the users)will also likely allow for the artist to add their visceral touch to it.Generative AI can disrupt incumbent application software.Generative AI has clear benefits in the field of customer relationship management,in particular market
262、ing and customer support.It also has very significant implications for financial planning,business risk analysis,inventory management,production scheduling,logistics and shipping.Yet another facet of Generative AI is the availability of soon-to-be half a dozen or so platforms upon which new applicat
263、ions can be built.This has been viewed by 26 March 2023 31Goldman SachsAmericas Technology_ industry leaders as being an“iPhone moment”when the new mobile platform was launched in 2007;yet,we did not have the first intuitive and comprehensive application(that allowed users to order and deliver food,
264、for example)for several years later.As Generative AI has been around for at least five years,and given the monumental improvements we have in the last few months,Generative AI may be on the brink of sparking a variety of new business ideas over the next few years.Some of these ideas will disrupt exi
265、sting markets and are likely to benefit companies that have large amounts of data(such as Salesforce).That being said,a technology company should not take its data incumbency for granted since data is growing at a rapid pace.When new platforms and new applications end up engaging end user interests,
266、the older data gets less relevant and competitive advantage dissipates.Therefore,software companies that deal with applications must incorporate Generative AI very quickly in their application or face substantial risk in the future as data creation moves to more attractive platforms.Sizing the Gener
267、ative AI TAM for Enterprise Software Generative AI can contribute an incremental+$150bn in the global enterprise software TAM.We see the integration and implementation of Generative AI tools across the application stack adding$150bn to the global enterprise software TAM.This assumption is based on t
268、he expectation that adoption can be as ubiquitous as productivity tools(such as Microsoft Office with+300mn paid seats)or 30%of the total knowledge worker base.As this innovation has business implications wave of across an array of existing end-markets(as opposed to a standalone AI/analytics submark
269、et),we foresee monetization occurring in the form of:price increases over 1)time as innovation,adoption and stickiness increase the value of existing applications,and new premium,add-on SKUs,which will likely be the main driver of near-term 2)growth.Our analysis comprises the following steps:1)Evalu
270、ating the price points for Generative AI apps based on those currently navailable.To start,we leveraged the pricing structure of existing products that are powered by Generative AI tools,such as Adobes Creative Cloud and Intuits Turbo Tax Live,as well as recently released products(Microsoft Teams Pr
271、emium and GitHub Copilot).These newer solutions,which are$10-20/user/month(excluding discounts)provide a sound starting point for more premium add-ons.Platforms that have evolved over the last five years,such as Creative Cloud,also showed a pattern of$10 price increases,with Adobe increasing the pri
272、ce of its Teams bundle by$10 in 2018(vs the initial price of$70/month set in 2014)and announcing an additional$10 step-up in 2022.We point to Adobes launch of Sensei,its AI platform tool,in 2016 as a catalyst for the product innovation that warranted such pricing boosts.This also led us to gain conf
273、idence in future Generative AI releases being able to garner at least$10/month.2)Calculating the ASP uplift to existing apps across North America and ndeveloped international markets.Keeping in mind the international mix of the global workforce,we analyzed the software TAM(excluding IaaS)against the
274、 labor force in Europe+Japan to derive the average software spend per worker vs.that of 26 March 2023 32Goldman SachsAmericas Technology_ North America.As shown in Exhibit 14,international investments relative to their labor force are 30%that of the US.Tying this back to the avg SKU price attributed
275、 of a Generative AI offering,we applied this 70%discount to the US$10 SKU discussed earlier to arrive at a 3 monthly cost per international employee.Annualizing these prices and accounting for the roughly even split between North America/International TAM,we derived a weighted average annual spend o
276、f$78 per person.3)Determining the average number of applications that a worker will pay an nadded cost to leverage Generative-AI capabilities.As this annualized cost represents a per user,per month,per application figure,we also analyzed the average number of applications utilized by a typical knowl
277、edge worker to determine the level of spend that can be driven per employee.While this can range depending on the role and company,we honed in on the applications that are likely to warrant the additional spend for more innovative features.We concluded that applications such as collaboration,product
278、ivity,HR,ERP(i.e.,T&E),which are broadly adopted by most knowledge workers fit into this category given their value proposition and wide use.We also discerned that there are likely to be 1-2 role-specific solutions(i.e.,CRM platform,developer tools)that may also garner investment to induce productiv
279、ity.This led us to conclude that Generative AI tools can unlock value in 5-6 apps per knowledge worker.4)Conduct sensitivity analysis around assumed penetration rate of the+1bn nglobal knowledge workers.As outlined in Exhibit 15,assuming 100%of the 1.1bn knowledge workers in the labor force today ad
280、opt 5 GAI applications at an average monthly cost of$78,this would create a$429bn TAM.Given we are still in the early stages of the adoption curve,our base case TAM is predicated on 30-40%of knowledge workers meeting the above characteristics and amounting to a$150bn TAM.Exhibit 16 shows the sensiti
281、vity analysis we conducted that factors in the number of apps/SKUs the average user will pay for Generative-AI capabilities as well as the level of penetration among the 1bn base Exhibit 14:Average software spend(excl.IaaS)per employee(derived by TAM/labor force)indicates international spending is 7
282、0%lower than that of North America$292$907$336$1,037$0$200$400$600$800$1,000$1,200 InternationalNorth AmericaInternationalNorth America 20202021 International Includes:Europe+Japan Software market includes:Application software,app development,data management,security,CPaaS,event-stream processing,RP
283、A,and digital experience platforms Source:Gartner,World Bank,Goldman Sachs Global Investment Research26 March 2023 33Goldman SachsAmericas Technology_ of global knowledge workers.As we expect this to be the next innovation wave of within software,extending the utility model and enhancing productivit
284、y at the end user level,our base case assumptions are underpinned by the level of adoption seen by Microsoft for its commercial Office 365 suite,which has 345mn paid commercial seats.We note that we see upside to this figure as this level of penetration does not fully account for the broader commerc
285、ial user base,that may leverage other platforms,such as Google Workspace.Raise PTs for MSFT,CRM,ADBE Given Revenue Tailwinds We raise our outer-year revenue estimates and price targets for MSFT,CRM,ADBE as we believe Generative AI presents a strong tailwind that can define the next decade of growth
286、for software.In light of the TAM framework we outlined in this report(Exhibit 16)and recent product announcements across a number of companies in our coverage,we set out to determine the possible lift Generative AI can have to our current growth expectations.While likely to have positive implication
287、s across our coverage,we honed in on those companies with tangible product offerings and a certain level of visibility into monetization.This resulted in us being able to attribute revenue growth to companies such as MSFT,CRM,and ADBE.As we are still very early in this cycle(with most products still
288、 not generally available),the changes to our top-line Exhibit 15:Assuming 100%adoption,GAI can add$430bn to the global enterprise software TAM(excluding IaaS)#of Knowledge Workers Globally(mn)1,100 Average Price for AI SKU-US$10.0 Average Price for AI SKU-International$3.0 Weighted Average Price$6.5
289、0 Annualized Price$78.0 Avg#Apps with Paid GAI Capabilities per Employee5 Annualized Spend per Knowledge Worker$390.0 Calculated TAM(assumes 100%adoption)$429 bn Base Case TAM(assumes 30%adoption)$150 bn Source:Gartner,World Bank,Goldman Sachs Global Investment Research Exhibit 16:Generative AI can
290、drive+$150bn in software spend(excluding IaaS)assuming the utilization of 5-6 apps across 300-400mn knowledge workers$Annualized spend on Generative AI SKUs per knowledge worker 3 apps4 apps5 apps6 apps7 apps8 apps$234$312$390$468$546$624100$23$31$39$47$55$62200$47$62$78$94$109$125300$70$94$117$140$
291、164$187400$94$125$156$187$218$250500$117$156$195$234$273$312600$140$187$234$281$328$374#of Knowledge Workers Paying for Generative AI Tools Source:Gartner,World Bank,Goldman Sachs Global Investment Research26 March 2023 34Goldman SachsAmericas Technology_ estimates largely began in 2024 and extended
292、 to the outer years as we account for market release and adoption timelines that can take up to 12-24 months for large deployments.For companies such as MSFT who already have in-market products with defined monetization strategies,(i.e.,GitHub Copilot,Teams Premium),we see such these products increa
293、singly becoming a growth driver as users already begin to adopt the offerings.We also assume a growing impact in later years as larger deployments in more core parts of the business(i.e.,Microsoft 365,Azure,Dynamics)take place.In some cases,we expect the impact to surface later as next-gen products
294、have been announced but are still in pilot/beta testing(i.e.,CRM Einstein GPT,ADBEs Firefly).Though Adobes Firefly services are still in beta,we took into consideration the ease of adoption based on end-market user and the average length of a sales cycle.Adobes exposure to consumers/SMBS coupled wit
295、h its strong direct-to-consumer selling channel for its Creative and Document Cloud offerings led us to assume possible revenue uplift from Generative AI services beginning in 2024.Given the growth estimates incorporated into our models prior to these changes assumed a level of growth from new produ
296、ct releases,we remind investors that the incremental growth we are factoring in does not represent the full underlying benefits of this adoption.Microsoft:We raise our PT to$325 from$315 as we see the recent slew of product announcements translating to revenue over time.While the company has already
297、 begun monetizing add-on features to existing applications,such as Teams Premium($10/user/month)and GitHub Copilot($19/user/month on the business plan),we see the enhancements to its core offerings,such as Microsoft 365 Copilot,being the key driver of future revenue growth given its+345mn paid subsc
298、riber base and prior success in raising ARPU.We reference Microsofts success in introducing and migrating users to premium E1/E3/E5 tiers after the addition of key product enhancements offers a playbook that can succeed in this product cycle.To account for the more near-term uplift from the release
299、of Teams Premium and GitHub Copilot,we slightly raise our Productivity and Business Processes and Intelligent Cloud estimates,mainly via its Commercial Office segment and Azure,with a marginal lift to Dynamics 365.Anticipating pricing plans to be released for Microsoft 365 Copilot,Dynamics 365 Copil
300、ot and other Copilot offerings over the next 12 months and factoring in time to adoption,we revise our revenue expectations of these segments in a more pronounced manner in FY25 and outer years.At a high level,we raise our Microsoft Cloud(which encompasses Azure,Commercial Office 365,Dynamics 365,an
301、d LinkedIn Commercial)yoy growth estimates by 1%and 2%in FY25 and FY26,expecting 20%and 18%yoy growth respectively.Exhibit 17 further outlines these estimate revisions.26 March 2023 35Goldman SachsAmericas Technology_ Salesforce:We raise our PT to$325 from$320 given the potential for Einstein GPT to
302、 drive greater adoption across the CRM ecosystem.As we expect the addition of Generative AI to simplify the Salesforce platform,this can drive improved onboarding,engagement and expansion rates.With multi-cloud customers spending anywhere from 3-300 x a single cloud user and this group holding the h
303、ighest retention rates,the successful deployment of Generative AI and analytics within their ecosystem could lead to numerous benefits,from stronger Data Cloud demanddurable growth in other 1)2)segments,such as Sales,Service and Marketing Clouds as data silos are broken down and improved unit econom
304、ics as churn eases and sales efficiency improves.As 3)Einstein revenue is typically recognized in the segment of the product it sits on(Sales,Service,etc),we see the various use cases across sales,service,marketing as providing tailwinds across the portfolio.Still,we call out potential acceleration
305、in Data Cloud given it comprises Tableau and Mulesoft,which are likely to power the data analytics and integration needed for such offerings.We incorporate this into our estimates by marginally raising our FY25(CY24)subscription revenue in the 2H of the year,with more upside in later years,once larg
306、e deployments begin to be implemented.Exhibit 17:Raise MSFT revenue estimates to account for a lift from the adoption of Gen.AI Copilot SKUs FY23(E)FY24(E)FY25(E)All figures in$mnsPrior Est.GS Est.ConsensusPrior Est.GS Est.ConsensusPrior Est.GS Est.Consensus Azure Growth Rate(YoY)29%22%29%0%29%23%1.
307、1%26%24%1.3%26%23%Productivity and Business Process$68,585$76,420$68,585$0$68,389$76,638$218$74,844$86,509$887$81,913$85,622 YoY8%8%0.0%8%11%12%0.3%9%13%1.0%9%12%Intelligent Cloud$87,926$88,020$94$87,594$98,511$749$101,651$97,762$111,676$113,475$1,799$119,472 YoY17%17%0.1%16%12%0.8%16%15%1.6%18%11%1
308、4%More Personal Computing$53,369$57,159$59,539$53,369$0$52,966$57,159$0$54,632$59,539$0$58,186 YoY-11%-11%0.0%-11%7%7%0.0%3%4%0.0%7%4%Total revenue(non-GAAP)$209,880$231,342$256,837$209,974$208,847$94$232,309$232,095$967$259,523$261,609$2,686 YoY6%6%0.0%5%10.2%11.0%10.6%0.4%11.1%11.7%1.0%12.7%Cost o
309、f revenue$65,906$71,074$65,910$4$65,749$71,248$174$73,366$76,849$77,965$1,116$83,322 YoY5%8%8%5%5%8%12%9%14%Gross profit$143,974$160,268$179,987$144,063$89$143,099$161,061$793$158,729$181,558$1,571$178,287 YoY6%6%6%11%12%12%11%13%12%Gross margin 69%69%70%68.6%$069%69.3%$468%70%($13)68%Operating expe
310、nses$57,816$60,678$64,499$57,816$0$57,965$60,678$0$60,956$64,499$0$65,432 YoY11%5%6%11%11%5%5%6%7%Operating income$86,158$86,247$89$85,134$97,773$99,590$100,383$793$115,489$117,059$1,570$112,855 YoY3%16%16%3%2%16%15%17%15%Operating margin41%41.1%+0 bps 41%43%43.2%+14 bps 42%45%45%+12 bps43%EPS(non-G
311、AAP)9.429.430.019.3111.1013.0411.140.0410.7913.140.1012.46YoY2%2%1%18%16%18%15%Microsoft Cloud Revenue$112,135$133,201$112,135$0$133,692$490$159,597$160,831$1,234 YoY23%19%20%23%19%20%Gross margin73%74%75%73%74%74%OCF$92,157$92,210$88,997$53$102,703$114,701$102,975$104,429$272$115,476$120,718$775 Yo
312、Y3%4%0%12%17%12%16%11%12%Cash capex$26,474$27,995$29,955$26,474$0$26,411$27,995$0$29,806$29,955$0$31,968 YoY11%6%7%11%10%6%13%7%7%FCF$65,683$74,708$84,786$65,737$61,432$54$74,980$74,615$272$85,522$90,317$736 YoY1%1%-6%14%13%14%21%14%21%Source:Company data,Goldman Sachs Global Investment Research26 M
313、arch 2023 36Goldman SachsAmericas Technology_ Adobe:We raise our PT to$480 from$475 after attending the companys Summit conference,where management announced various new features and products that can drive the next leg of growth across its creative,document and marketing offerings.Primarily,we see
314、the companys Firefly suite,which will be available across many user touch points,as underpinning future innovation.Drawing a parallel to the release of its AI/ML technology,Sensei,which Adobe was able to build on to drive two price increases,we assume a similar strategy to be used going forward.We a
315、lso consider the potential for some of these releases to be complimentary to existing offerings.In either scenario,we see this supporting durable growth in the future.Creative Cloud will likely drive the initial step-up in adoption with the benefits to Digital Experience being a later cycle benefit
316、due to its longer-sales cycle motion.We expect to see a more pronounced boost to this segment over time as Adobes DX platform is relatively new(with Adobe ramping investments in this are around 5 years ago),and has many components to adoption(AEM,CDP,etc).Exhibit 18:Raise CRM revenue estimates to re
317、flect potential lift from Einstein GPT FY24(E)FY26(E)All figures in$mnsPrior Est.New Est.ConsensusPrior Est.New Est.ConsensusPrior Est.New Est.ConsensusSubscription/Support Revenue$32,113$36,101$41,539$32,113$0$31,852$36,149$48$35,367$41,715$176$39,977 YoY10.7%12.4%15.1%10.7%9.8%12.6%11.0%15.4%13.0%
318、Professional Services Revenue$2,553$2,859$3,145$2,553$0$2,571$2,859$0$2,831$3,145$0$3,234 YoY9.5%9.5%10.3%12.0%10.1%10.0%14.2%12.0%10.0%GAAP Total Revenue$34,666$38,960$44,684$34,666$0$34,545$39,008$48$38,449$44,860$176$43,523 YoY10.6%12.4%14.7%10.6%10.2%12.5%11.3%15.0%13.2%Total Gross Profit$27,230
319、$30,696$35,302$27,230$0$26,762$30,737$41$29,852$35,449$148$33,366%margin78.6%78.8%79.0%78.6%77.5%78.8%77.6%79.0%76.7%Operating Income$9,360$11,628$14,182$9,360$0$9,191$11,669$41$11,327$14,330$148$13,043%margin27.0%29.8%31.7%27.0%26.6%29.9%29.5%31.9%30.0%Expansion450 bps280 bps190 bps450 bps405 bps29
320、0 bps285 bps200 bps50 bpsNon-GAAP EPS$7.15$8.95$10.91$7.15$0.00$7.10$8.98$0.03$8.89$10.96$0.04$10.45Current RPO$26,568$30,288$34,376$26,568$0$30,288$0$34,376$0 YoY8.0%8.0%14.0%13.5%14.0%13.5%CFO$8,249$8,249$0$8,334$10,311$10,330$19$10,955$12,991$13,112$121$12,955 YoY16.0%25.0%26.0%16.0%17.2%25.2%31.
321、5%26.9%18.3%Free Cash Flow$7,412$9,435$12,075$7,412$0$7,422$9,454$19$9,678$12,195$121$10,393 YoY17.4%27.3%28.0%17.4%17.6%27.6%30.4%29.0%7.4%FY25(E)Source:Company data,Goldman Sachs Global Investment Research26 March 2023 37Goldman SachsAmericas Technology_ Impact of Generative AI Likely to Reach Mul
322、titude of Workforce Personas Since the introduction of ChatGPT and the subsequent popularization of Generative AI,ample use cases spanning a wide range of applications have emerged,from business to technology to healthcare.Using recent product announcements as a launch pad,we hypothesize how Generat
323、ive AI could evolve software categories(CRM,HR,Cybersecurity)as well as various verticals(i.e.healthcare).Office Productivity Tools:Some of the most apparent applications of Generative AI nrevolve around office productivity tools like Microsoft 365 and Google Workspace.When assessing the future use
324、cases of Generative AI with respect to this use case,we expect these scenarios to be underpinned by broader real-time connectivity between data streams,data models,applications and the end-user.We anticipate deeper application integrations resulting in seamless cross product functionality.This shoul
325、d encourage increased efficiency as employees can expeditiously reference&leverage previously disparate data while working within one application.For example an employee creating a presentation could leverage Generative AI to quickly pull notes from a Word file to populate a PowerPoint slide while n
326、ever leaving PowerPoint.This type of technology is on the horizon based on demos of Microsoft 365 Copilot&Google Workspace.Future iterations could potentially prompt users to notify the right stakeholders of the documents with a draft of an email that can let them know its ready.The real time data f
327、low could reduce the time spent needed to synthesize and summarize data.Finally,Generative AI can unlock enhanced utilization within applications.Most users of Word,Excel,PowerPoint,etc.are likely only scratching the surface of these tools potential.We Exhibit 19:Expect Creative Cloud revenue to sho
328、w first signs of gen AI adoption with Digital Experience a longer-tailed benefit given deal nature FY23(E)FY24(E)FY25(E)All figures in$mnsPrior Est.New Est.ConsensusPrior Est.New Est.ConsensusPrior Est.New Est.Consensus Net New DM ARR$1,707$2,156$1,707$1,704$2,176$20$2,170$2,092$2,067$25$1,853 Digit
329、al Media$14,017$15,867$14,017$0$13,985$15,886$19$15,647$18,366$18,39831.94$17,038 YoY9.1%9.1%8.9%13.2%15.7%13.3%11.9%15.8%8.9%Creative Cloud Revenue$11,354$12,785$11,354$0$11,292$12,803$19$12,590$14,827$14,85628.65$13,730 YoY8.5%8.6%8.0%12.6%16.0%12.8%11.5%16.0%9.1%Document Cloud revenue$2,662$3,083
330、$3,539$2,662$0$2,700$3,083$0$3,079$3,5423.29$3,308 YoY11.7%15.8%14.8%11.7%13.3%15.8%14.0%14.9%7.4%Digital Experience$5,005$5,688$6,269$5,005$0$4,973$5,688$0$5,642$6,33162.07$6,439 YoY13.2%13.7%10.2%13.2%12.5%13.7%13.5%11.3%14.1%Total revenue$19,358$21,875$19,358$0$19,278$21,894$19$21,545$24,929$25,0
331、2394.01$24,041 YoY9.9%9.9%9.5%13.0%14.0%13.1%11.8%14.3%11.6%Gross Profit$17,314$19,543$17,314$0$17,144$19,561$18$19,176$22,323$22,39874.60$21,133%margin89.4%89.3%89.5%89.4%88.9%89.3%89.0%89.5%87.9%Operating Income$8,557$9,623$8,557$0$8,642$9,641$18$9,686$11,494$11,56874.60$10,899%margin44.2%44.0%46.
332、1%44.2%44.8%+0 bps44.0%45.0%+5 bps46.2%45.3%12.49 Diluted EPS(ex ESO exp)$15.50$18.13$23.09$15.50$0.00$15.46$18.17$0.03$17.56$23.240.15$20.02 YoY12%13%9%10%13%13%14%14%14%$CFO$8,121$9,296$8,121$0-$9,312$16$11,005$11,07166.02 YoY4%4%-14%18%15%19%FCF$7,674$8,845$7,674$0$7,778$8,861$16$8,855$10,527$10,
333、59366.02$9,870 YoY4%15%19%4%5%15%14%20%11%Source:Company data,Goldman Sachs Global Investment Research26 March 2023 38Goldman SachsAmericas Technology_ believe Generative AI is primed to simplify the complex capabilities of these applications,resulting in more in-depth and thoughtful insights,and better-quality end products.CRM:Digging into potential Generative AI use cases within sales and market