《高盛：新一代人工智能：花费太多，收益太少？（2024）（英文版）（31页）.pdf》由会员分享，可在线阅读，更多相关《高盛：新一代人工智能：花费太多，收益太少？（2024）（英文版）（31页）.pdf（31页珍藏版）》请在三个皮匠报告上搜索。

1、GEN AI:TOO MUCH SPEND,TOO LITTLE BENEFIT?ISSUE 129|June 25,2024|5:10 PM EDT”$P$Global Macro ResearchInvestors should consider this report as only a single factor in making their investment decision.For Reg AC certification and other important disclosures,see the Disclosure Appendix,or go to Goldman

2、Sachs Group,Inc.Tech gi ant s and beyond ar e set t o spend over$1t n on AI capex i n comi ng year s,wi t h so far l i t t l e t o show for i t.So,wi l l t hi s l ar ge spend ever pay off?MI T s Dar on Acemogl u and GS Ji m Covel l o ar e skept i cal,wi t h Acemogl u seei ng onl y l i mi t ed US eco

3、nomi c upsi de fr om AI over t he next decade and Covel l o ar gui ng t hat t he t echnol ogy i sn t desi gned t o sol ve t he compl ex pr obl ems t hat woul d just i fy t he cost s,whi ch may not decl i ne as many expect.But GS Joseph Br i ggs,Kash Rangan,and Er i c Sher i dan r emai n mor e opt i

4、mi st i c about AI s economi c pot ent i al and i t s abi l i t y t o ul t i mat el y gener at e r et ur ns beyond t he cur r ent “pi cks and shovel s”phase,even i f AI s“ki l l er appl i cat i on”has yet t o emer ge.And even i f i t does,we expl or e whet her t he cur r ent chi ps shor t age(wi t h

5、 GS Toshi ya Har i)and l oomi ng power shor t age(wi t h Cl over l eaf I nfr ast r uct ur e s Br i an Janous)wi l l const r ai n AI gr owt h.But despi t e t hese concer ns and const r ai nt s,we st i l l see r oom for t he AI t heme t o r un,ei t her because AI st ar t s t o del i ver on i t s pr om

6、i se,or because bubbl es t ake a l ong t i me t o bur st.“I NTERVI EWS WI TH:Daron Acemoglu,I nst i t ut e Pr ofessor,MI T Brian Janous,Co-founder,Cl over l eaf I nfr ast r uct ur e,for mer Vi ce Pr esi dent of Ener gy,Mi cr osoft Jim Covello,Head of Gl obal Equi t y Resear ch,Gol dman Sachs Kash Ra

7、ngan,US Soft war e Equi t y Resear ch Anal yst,Gol dman Sachs;Eric Sheridan,US I nt er net Equi t y Resear ch Anal yst,Gol dman Sachs ADDRESSI NG THE AI GROWTH DEBATE Joseph Br i ggs,GS Gl obal Economi cs Resear ch ONCE I N A GENERATI ON,GENERATI ON Car l y Davenpor t,GS US Ut i l i t i es Equi t y

8、Resear ch AI:POWERI NG UP EUROPE Al ber t o Gandol fi,GS Eur opean Ut i l i t i es Equi t y Resear ch AI S CHI P CONSTRAI NTS Toshi ya Har i,Anmol Makkar,Davi d Bal aban,GS US Semi conduct or Equi t y Resear ch FULL STEAM AHEAD FOR AI BENEFI CI ARI ES Ryan Hammond,GS US Por t fol i o St r at egy Res

9、ear ch AI OPTI MI SM AND LONG-TERM EQUI TY RETURNS Chr i st i an Muel l er-Gl i ssmann,GS Mul t i-Asset St r at egy Resear ch WHATS INSIDEofAllison Nathan|al l i son.nat .AND MORETOP MINDJenny Grimberg|jenny.gr i mber Gi ven t he focus and ar chi t ect ur e of gener at i ve AI t echnol ogy t oday.t

10、r ul y t r ansfor mat i ve changes won t happen qui ckl y and fewi f anywi l l l i kel y occur wi t hi n t he next 10 year s.-Dar on Acemogl uSpendi ng i s cer t ai nl y hi gh t oday i n absol ut e dol l ar t er ms.But t hi s capex cycl e seems mor e pr omi si ng t han even pr evi ous capex cycl es.

11、-Kash RanganAI t echnol ogy i s except i onal l y expensi ve,and t o just i fy t hose cost s,t he t echnol ogy must be abl e t o sol ve compl ex pr obl ems,whi ch i t i sn t desi gned t o do.-Ji m Covel l oAI dol l ar s spent pany r evenues ar e not mat er i al l y di ffer ent t han t hose of pr i o

12、r i nvest ment cycl es.-Er i c Sher i danAshley Rhodes|ashl ey.r NoteNote:The following is a redacted version of the original report published June 25,2024 32 pgs.hElGoldman Sachs Global Investment Research 2 Top of Mind Issue 129 Macro news and views US Japan Latest GS proprietary datapoints/major

13、changes in views No major changes in views.Datapoints/trends were focused onFed policy;we expect quarterly Fed rate cuts beginning inSeptember,for a total of two cuts this year.Inflation;we expect core PCE inflation to stand at 2.7%yoyby Dec 2024 before converging toward 2%next year.Growth;we think

14、most of the slowdown from the 4.1%paceof real GDP growth in 2H23 is here to stay given softer realincome growth,lower consumer sentiment,and election-related uncertainty that could weigh on business investment.Labor market,which is now fully rebalanced,likely meaningthat a material softening in labo

15、r demand would hit actual jobs.Latest GS proprietary datapoints/major changes in views We now expect the next BoJ rate hike in July(vs.Octbefore)as the hurdle for the next hike is low given that itwill likely be only 15bp and the BoJ sees the policy rate assignificantly lower than the current nomina

16、l neutral rate.Datapoints/trends were focused on Japanese inflation;sequential core inflation has recentlyshown signs of weakness,but we expect core inflation toremain above the BoJs target this year at 2.6%yoy.Japans rising interest burden,which will likely be manageablefor households and corporate

17、s given that it is occurringagainst a backdrop of solid activity and steady wage growth.Japan financial conditions,which continue to ease.Election uncertainty:a potential growth drag NFIB Small Business Uncertainty Index Japanese households:net interest receivers Interest payments&receipts(lhs,tn),n

18、et interest receipt(rhs,%of 2023 disposable income)Source:Haver Analytics,Goldman Sachs GIR.Source:Goldman Sachs GIR.Europe Emerging Markets(EM)Latest GS proprietary datapoints/major changes in views We raised our 2024 UK GDP growth forecast to 0.9%(from0.8%)following slightly above consensus April

19、GDP data.Datapoints/trends were focused on ECB policy;we expect the next rate cut in Sept,though wethink a pause in the easing cycle is possible if inflation and wage data surprise to the upside over the summer.BoE policy;we expect the BoE to embark on rate cuts inAugust on the back of renewed UK di

20、sinflation progress.French snap elections(Jun 30),which could result in a fiscalexpansion that would lead the debt-to-GDP ratio to rise.UK general election(Jul 4),which will likely deliver relativelysimilar fiscal outcomes irrespective of which party wins.Latest GS proprietary datapoints/major chang

21、es in views We recently pushed back our PBOC easing forecasts byone quarter given ample near-term liquidity,and now expecta 25bp RRR cut in Q3 and a 10bp policy rate cut in Q4.Datapoints/trends were focused on Chinas economy,which remains bifurcated betweenstrength in exports and manufacturing activ

22、ity and weaknessin housing and credit,coupled with very low inflation.EM easing cycle;we think the fundamental case for furtherEM rate cuts remains strong,though the recent unwind inEM FX carry trades following electoral surprises in Mexico,India,and South Africa could impede policy normalization.Fr

23、ench election:upside risk to debt trajectory French government debt,%of GDP China:a bifurcated economy China activity indicator,%change,yoy Source:Goldman Sachs GIR.Source:Haver Analytics,Goldman Sachs GIR.40506070809010011020102012201420162018202020222024Presidential election years 0.00.51.01.52.02

24、.53.03.5024681012202420262028203020322034Additional interest receipt(lhs)Additional interest payment(lhs)Net(rhs)859095100105110115120201220152018202120242027ActualStatus quoDeadlockExpansion-25-20-15-10-5051015Property-New startsProperty-Sales volumeProperty-CompletionsProperty FAICement production

25、Auto sales volumeElectricity productionSteel productionFixed asset investment(FAI)Infrastructure FAIRetail sales(RS)Services Production IndexCatering salesIndustrial production(IP)IP-ManufacturingExportsManufacturing FAIOnline goods salesWe provide a brief snapshot on the most important economies fo

26、r the global markets hEl Goldman Sachs Global Investment Research 3 Top of Mind Issue 129 The promise of generative AI technology to transform companies,industries,and societies continues to be touted,leading tech giants,other companies,and utilities to spend an estimated$1tn on capex in coming year

27、s,including significant investments in data centers,chips,other AI infrastructure,and the power grid.But this spending has little to show for it so far beyond reports of efficiency gains among developers.And even the stock of the company reaping the most benefits to dateNvidiahas sharply corrected.W

28、e ask industry and economy specialists whether this large spend will ever pay off in terms of AI benefits and returns,and explore the implications for economies,companies,and markets if it does,or if it doesnt.We first speak with Daron Acemoglu,Institute Professor at MIT,whos skeptical.He estimates

29、that only a quarter of AI-exposed tasks will be cost-effective to automate within the next 10 years,implying that AI will impact less than 5%of all tasks.And he doesnt take much comfort from history that shows technologies improving and becoming less costly over time,arguing that AI model advances l

30、ikely wont occur nearly as quicklyor be nearly as impressiveas many believe.He also questions whether AI adoption will create new tasks and products,saying these impacts are“not a law of nature.”So,he forecasts AI will increase US productivity by only 0.5%and GDP growth by only 0.9%cumulatively over

31、 the next decade.GS Head of Global Equity Research Jim Covello goes a step further,arguing that to earn an adequate return on the$1tn estimated cost of developing and running AI technology,it must be able to solve complex problems,which,he says,it isnt built to do.He points out that truly life-chang

32、ing inventions like the internet enabled low-cost solutions to disrupt high-cost solutions even in its infancy,unlike costly AI tech today.And hes skeptical that AIs costs will ever decline enough to make automating a large share of tasks affordable given the high starting point as well as the compl

33、exity of building critical inputslike GPU chipswhich may prevent competition.Hes also doubtful that AI will boost the valuation of companies that use the tech,as any efficiency gains would likely be competed away,and the path to actually boosting revenues is unclear,in his view.And he questions whet

34、her models trained on historical data will ever be able to replicate humans most valuable capabilities.But GS senior global economist Joseph Briggs is more optimistic.He estimates that gen AI will ultimately automate 25%of all work tasks and raise US productivity by 9%and GDP growth by 6.1%cumulativ

35、ely over the next decade.While Briggs acknowledges that automating many AI-exposed tasks isnt cost-effective today,he argues that the large potential for cost savings and likelihood that costs will decline over the long runas is often,if not always,the case with new technologiesshould eventually lea

36、d to more AI automation.And,unlike Acemoglu,Briggs incorporates both the potential for labor reallocation and new task creation into his productivity estimates,consistent with the strong and long historical record of technological innovation driving new opportunities.GS US software analyst Kash Rang

37、an and internet analyst Eric Sheridan also remain enthusiastic about generative AIs long-term transformative and returns potential even as AIs“killer application”has yet to emerge.Despite big techs large spending on AI infrastructure,they dont see signs of irrational exuberance.Indeed,Sheridan notes

38、 that current capex spend as a share of revenues doesnt look markedly different from prior tech investment cycles,and that investors are rewarding only those companies that can tie a dollar of AI spending back to revenues.Rangan,for his part,argues that the potential for returns from this capex cycl

39、e seems more promising than even previous cycles given that incumbents with low costs of capital and massive distribution networks and customer bases are leading it.So,both Sheridan and Rangan are optimistic that the huge AI spend will eventually pay off.But even if AI could potentially generate sig

40、nificant benefits for economies and returns for companies,could shortages of key inputsnamely,chips and powerkeep the technology from delivering on this promise?GS US semiconductor analysts Toshiya Hari,Anmol Makkar,and David Balaban argue that chips will indeed constrain AI growth over the next few

41、 years,with demand for chips outstripping supply owing to shortages in High-Bandwidth Memory technology and Chip-on-Wafer-on-Substrate packagingtwo critical chip components.But the bigger question seems to be whether power supply can keep up.GS US and European utilities analysts Carly Davenport and

42、Alberto Gandolfi,respectively,expect the proliferation of AI technology,and the data centers necessary to feed it,to drive an increase in power demand the likes of which hasnt been seen in a generation(which GS commodities strategist Hongcen Wei finds early evidence of in Virginia,a hotbed for US da

43、ta center growth).Brian Janous,Co-founder of Cloverleaf Infrastructure and former VP of Energy at Microsoft,believes that US utilitieswhich havent experienced electricity consumption growth in nearly two decades and are contending with an already aged US power gridarent prepared for this coming dema

44、nd surge.He and Davenport agree that the required substantial investments in power infrastructure wont happen quickly or easily given the highly regulated nature of the utilities industry and supply chain constraints,with Janous warning that a painful power crunch that could constrain AIs growth lik

45、ely lies ahead.So,what does this all mean for markets?Although Covello believes AIs fundamental story is unlikely to hold up,he cautions that the AI bubble could take a long time to burst,with the“picks and shovels”AI infrastructure providers continuing to benefit in the meantime.GS senior US equity

46、 strategist Ryan Hammond also sees more room for the AI theme to run and expects AI beneficiaries to broaden out beyond just Nvidia,and particularly to what looks set to be the next big winner:Utilities.That said,looking at the bigger picture,GS senior multi-asset strategist Christian Mueller-Glissm

47、ann finds that only the most favorable AI scenario,in which AI significantly boosts trend growth and corporate profitability without raising inflation,would result in above-average long-term S&P 500 returns,making AIs ability to deliver on its oft-touted potential even more crucial.Allison Nathan,Ed

48、itor Email: Tel:212-357-7504 Goldman Sachs&Co.LLC Gen AI:too much spend,too little benefit?hEl Goldman Sachs Global Investment Research 4 Top of Mind Issue 129 Daron Acemoglu is Institute Professor at MIT and has written several books,including Why Nations Fail:The Origins of Power,Prosperity,and Po

49、verty and his latest,Power and Progress:Our Thousand-Year Struggle Over Technology and Prosperity.Below,he argues that the upside to US productivity and growth from generative AI technology over the next decadeand perhaps beyondwill likely be more limited than many expect.The views stated herein are

50、 those of the interviewee and do not necessarily reflect those of Goldman Sachs.Allison Nathan:In a recent paper,you argued that the upside to US productivity and,consequently,GDP growth from generative AI will likely prove much more limited than many forecastersincluding Goldman Sachsexpect.Specifi

51、cally,you forecast a 0.5%increase in productivity and 1%increase in GDP in the next 10 years vs.GS economists estimates of a 9%increase in productivity and 6.1%increase in GDP.Why are you less optimistic on AIs potential economic impacts?Daron Acemoglu:The forecast differences seem to revolve more a

52、round the timing of AIs economic impacts than the ultimate promise of the technology.Generative AI has the potential to fundamentally change the process of scientific discovery,research and development,innovation,new product and material testing,etc.as well as create new products and platforms.But g

53、iven the focus and architecture of generative AI technology today,these truly transformative changes wont happen quickly and fewif anywill likely occur within the next 10 years.Over this horizon,AI technology will instead primarily increase the efficiency of existing production processes by automati

54、ng certain tasks or by making workers who perform these tasks more productive.So,estimating the gains in productivity and growth from AI technology on a shorter horizon depends wholly on the number of production processes that the technology will impact and the degree to which this technology increa

55、ses productivity or reduces costs over this timeframe.My prior guess,even before looking at the data,was that the number of tasks that AI will impact in the short run would not be massive.Many tasks that humans currently perform,for example in the areas of transportation,manufacturing,mining,etc.,ar

56、e multifaceted and require real-world interaction,which AI wont be able to materially improve anytime soon.So,the largest impacts of the technology in the coming years will most likely revolve around pure mental tasks,which are non-trivial in number and size but not huge,either.To quantify this,I be

57、gan with Eloundou et al.s comprehensive study that found that the combination of generative AI,other AI technology,and computer vision could transform slightly over 20%of value-added tasks in the production process.But thats a timeless prediction.So,I then looked at another study by Thompson et al.o

58、n a subset of these technologiescomputer visionwhich estimates that around a quarter of tasks that this technology can perform could be cost-effectively automated within 10 years.If only 23%of exposed tasks are cost effective to automate within the next ten years,this suggests that only 4.6%of all t

59、asks will be impacted by AI.Combining this figure with the 27%average labor cost savings estimates from Noy and Zhangs and Brynjolfsson et al.s studies implies that total factor productivity effects within the next decade should be no more than 0.66%and an even lower 0.53%when adjusting for the comp

60、lexity of hard-to-learn tasks.And that figure roughly translates into a 0.9%GDP impact over the decade.Allison Nathan:Recent studies estimate cost savings from the use of AI ranging from 10%to 60%,yet you assume only around 30%cost savings.Why is that?Daron Acemoglu:Of the three detailed studies pub

61、lished on AI-related costs,I chose to exclude the one with the highest cost savingsPeng et al.estimates of 56%because the task in the study that AI technology so markedly improved was notably simple.It seems unlikely that other,more complex,tasks will be affected as much.Specifically,the study focus

62、es on time savings incurred by utilizing AI technologyin this case,GitHub Copilotfor programmers to write simple subroutines in HTML,a task for which GitHub Copilot had been extensively trained.My sense is that such cost savings wont translate to more complex,open-ended tasks like summarizing texts,

63、where more than one right answer exists.So,I excluded this study from my cost-savings estimate and instead averaged the savings from the other two studies.Allison Nathan:While AI technology cannot perform many complex tasks well todaylet alone in a cost-effective mannerthe historical record suggests

64、 that as technologies evolve,they both improve and become less costly.Wont AI technology follow a similar pattern?Daron Acemoglu:Absolutely.But I am less convinced that throwing more data and GPU capacity at AI models will achieve these improvements more quickly.Many people in the industry seem to b

65、elieve in some sort of scaling law,i.e.that doubling the amount of data and compute capacity will double the capability of AI models.But I would challenge this view in several ways.What does it mean to double AIs capabilities?For open-ended tasks like customer service or understanding and summarizin

66、g text,no clear metric exists to demonstrate that the output is twice as good.Similarly,what does a doubling of data really mean,and what can it achieve?Including twice as much data from Reddit into the next version of GPT may improve its ability to predict the next word when engaging in an informal

67、 conversation,but it wont necessarily improve a customer service representatives ability to help a customer troubleshoot problems with their video service.The quality of the data also matters,and its not clear where more high-quality data will come from and whether it will be easily and cheaply avai

68、lable to AI models.Lastly,the current architecture of AI Interview with Daron Acemoglu hEl Goldman Sachs Global Investment Research 5 Top of Mind Issue 129 technology itself may have limitations.Human cognition involves many types of cognitive processes,sensory inputs,and reasoning capabilities.Larg

69、e language models(LLMs)today have proven more impressive than many people would have predicted,but a big leap of faith is still required to believe that the architecture of predicting the next word in a sentence will achieve capabilities as smart as HAL 9000 in 2001:A Space Odyssey.Its all but certa

70、in that current AI models wont achieve anything close to such a feat within the next ten years.Allison Nathan:So,are the risks to even your relatively conservative estimates of AIs economic impacts over the next 5-10 years skewed to the downside?Daron Acemoglu:Both downside and upside risks exist.Te

71、chnological breakthroughs are always possible,although even such breakthroughs take time to have real impact.But even my more conservative estimates of productivity gains may turn out to be too large if AI models prove less successful in improving upon more complex tasks.And while large organization

72、s such as the tech companies leading the development of AI technology may introduce AI-driven tools quickly,smaller organizations may be slower to adopt them.Allison Nathan:Over the longer term,what odds do you place on AI technology achieving superintelligence?Daron Acemoglu:I question whether AI t

73、echnology can achieve superintelligence over even longer horizons because,as I said,it is very difficult to imagine that an LLM will have the same cognitive capabilities as humans to pose questions,develop solutions,then test those solutions and adopt them to new circumstances.I am entirely open to

74、the possibility that AI tools could revolutionize scientific processes on,say,a 20-30-year horizon,but with humans still in the drivers seat.So,for example,humans may be able to identify a problem that AI could help solve,then humans could test the solutions the AI models provide and make iterative

75、changes as circumstances shift.A truly superintelligent AI model would be able to achieve all of that without human involvement,and I dont find that likely on even a thirty-year horizon,and probably beyond.Allison Nathan:Your colleague David Autor and coauthors have shown that technological innovati

76、ons tend to drive the creation of new occupations,with 60%of workers today employed in occupations that didnt exist 80 years ago.So,could the impact of AI technology over the longer term prove more significant than you expect?Daron Acemoglu:Technological innovation has undoubtedly meaningfully impac

77、ted nearly every facet of our lives.But that impact is not a law of nature.It depends on the types of technologies that we invent and how we use them.So,again,my hope is that we use AI technology to create new tasks,products,business occupations,and competencies.In my example about how AI tools may

78、revolutionize scientific discovery,AI models would be trained to help scientists conceive of and test new materials so that humans can then be trained to become more specialized and provide better inputs into the AI models.Such an evolution would ultimately lead to much better possibilities for huma

79、n discovery.But it is by no means guaranteed.Allison Nathan:Will someor maybe even mostof the substantial spending on AI technology today ultimately go to waste?Daron Acemoglu:That is an interesting question.Basic economic analysis suggests that an investment boom should occur because AI technology

80、today is primarily used for automation,which means that algorithms and capital are substituting for human labor,which should lead to investment.This explains why my estimates for GDP increases are nearly twice as large as my estimates for productivity increases.But then reality supervenes and says t

81、hat some of the spending will end up wasted because some projects will fail,and some firms will be too optimistic about the extent of the efficiency gains and cost savings they can achieve or their ability to integrate AI into their organizations.On the other hand,some of the spending will plant the

82、 seeds for the next,and more promising,phase of the technology.The devil is ultimately in the details.So,I dont have a strong prior as to how much of the current investment boom will be wasted vs.productive.But I expect both will happen.Allison Nathan:Are other costs of AI technology not receiving e

83、nough attention?Daron Acemoglu:Yes.GDP is not everything.Technology that has the potential to provide good information can also provide bad information and be misused for nefarious purposes.I am not overly concerned about deepfakes at this point,but they are the tip of the iceberg in terms of how ba

84、d actors could misuse generative AI.And a trillion dollars of investment in deepfakes would add a trillion dollars to GDP,but I dont think most people would be happy about that or benefit from it.Allison Nathan:Given everything weve discussed,is the current enthusiasm around AI technology overdone?D

85、aron Acemoglu:Every human invention should be celebrated,and generative AI is a true human invention.But too much optimism and hype may lead to the premature use of technologies that are not yet ready for prime time.This risk seems particularly high today for using AI to advance automation.Too much

86、automation too soon could create bottlenecks and other problems for firms that no longer have the flexibility and trouble-shooting capabilities that human capital provides.And,as I mentioned,using technology that is so pervasive and powerfulproviding information and visual or written feedback to hum

87、ans in ways that we dont yet fully understand and dont at all regulatecould prove dangerous.Although I dont believe superintelligence and evil AI pose major threats,I often think about how the current risks might be perceived looking back 50 years from now.The risk that our children or grandchildren

88、 in 2074 accuse us of moving too slowly in 2024 at the expense of growth seems far lower than the risk that we end up moving too quickly and destroy institutions,democracy,and beyond in the process.So,the costs of the mistakes that we risk making are much more asymmetric on the downside.Thats why it

89、s important to resist the hype and take a somewhat cautious approach,which may include better regulatory tools,as AI technologies continue to evolve.hEl Goldman Sachs Global Investment Research 6 Top of Mind Issue 129 Joseph Briggs addresses the AI productivity and growth debate,arguing that generat

90、ive AI will likely lead to significant economic upside We have long argued that generative AI could lead to significant economic upside,primarily owing to its ability to automate a large share of work tasks,with our baseline estimate implying as much as 15%cumulative gross upside to US labor product

91、ivity and GDP growth1 following widespread adoption of the technology.A significant boost to US labor productivity from generative AI Effect of AI adoption on annual US labor productivity growth,10y adoption period,pp Source:Goldman Sachs GIR.That said,substantial debate exists around generative AIs

92、 potential macro impacts.Studies that assume generative AI will accelerate the development and adoption of robotics or that view recent generative AI advances as foreshadowing the emergence of a“superintelligence”,for example,estimate even more upside to productivity and GDP than our baseline foreca

93、st.We see such outcomes as possible but premature since they generally assume AI advancements well beyond the frontier of current models.More notably,MIT economist Daron Acemoglu sees much more limited upside to US productivity and GDP than we expect,with his baseline estimates implying that generat

94、ive AI will boost US total factor productivity(TFP)by 0.53%and GDP by 0.9%over the next 10 years(see pgs.4-5).As we take similar approaches to assessing the economic impacts of generative AI,we explore what explains the large differences in our estimates.Breaking down the differences We find two mai

95、n factors that explain the differences in our estimates versus those of Acemoglu.First,Acemoglu assumes that generative AI will automate only 4.6%of total work tasks,1 Our GDP estimate assumes that the capital stock evolves to match increased labor potential,which seems broadly validated by the siza

96、ble investment response aimed at facilitating the AI transition.2 This figured is calculated by multiplying the labor share of output,62%,by our 15%estimate of the AI upside to labor productivity and growth.3 The quantitative contribution of different channels to the discrepancy between Acemoglus an

97、d our estimates depends on the order that they are considered in,with differences in exposure assumptions explaining more of the gap if differences in cost savings assumptions are considered first and vice versa.To reduce this sensitivity,we consider both orderings and present the average contributi

98、ons.as he estimates that only 19.9%of all tasks are exposed to AI and assumes that only 23%of exposed tasks will be cost effective to automate within the next ten years.In contrast,we assume that generative AI will automate 25%of all work tasks following the technologys full adoption.Second,Acemoglu

99、s framework assumes that the primary driver of cost savings will be workers completing existing tasks more efficiently and ignores productivity gains from labor reallocation or the creation of new tasks.In contrast,our productivity estimates incorporate both worker reallocationvia displacement and s

100、ubsequent reemployment in new occupations made possible by AI-related technological advancementand new task creation that expands non-displaced workers production potential.Differences in these assumptions explain over 80%of the discrepancy between our 9.2%2 and Acemoglus 0.53%estimates of increases

101、 in TFP over the next decade3.The remaining 20%of the gap reflects differences in cost savings and marginal productivity assumptions.For instance,Acemoglu assumes 27%cost savings based on two studies that he considers the most representative of AIs real-world impact,but cost savings would rise to 36

102、%if the full set of studies were considered.We are also more optimistic that AI will raise non-displaced workers output,largely because we expect AI automation to create new tasks and products.Differences in macro estimates mostly reflect differences in assumptions around tasks that can be profitabl

103、y automated and the reallocation of labor to new tasks Reconciling estimates of AI impact on GDP:Acemoglu(2024)vs.GS(2023),%Source:Goldman Sachs GIR.More widespread AI automation ahead So,whose estimates regarding the share of automated tasks and new task creationwill more likely prove correct?We ar

104、e very sympathetic to Acemoglus argument that automation of many AI-exposed tasks is not cost effective today,and may not become so even within the next ten years.AI adoption remains very modest outside of the few 0.30.50.70.81.31.52.42.42.9-10123456Much less powerful AISlower adoption(30 years)Slow

105、er adoption(20 years)Slighlty less powerful AINo labor displacementBaselineSlightly more powerful AIMore labor displacementMuch more powerful AIIncreased productivity of non-displaced workersLabor displacementReemployment of displaced workers0.3pp4.3pp0.9pp0.8pp2.4pp6pp15%0.5%9.2%0369121518Baselinet

106、askexposureOnly cost-effectivetasks todayBroadercost-savingestimatesExpandedproductionofemployedworkersAcemoglubaselineExposureassumptionsCost savingassumptionsReallocation/new taskcreationGSbaselineCapitaldeepeningOverallGDPTotal Factor Productivity(TFP)GDPAddressing the AI growth debate hEl Goldma

107、n Sachs Global Investment Research 7 Top of Mind Issue 129 industriesincluding computing and data infrastructure,information services,and motion picture and sound productionthat we estimate will benefit the most,and adoption rates are likely to remain below levels necessary to achieve large aggregat

108、e productivity gains for the next few years.This explains why we only raised our US GDP forecast by 0.4pp by the end of our forecast horizon in 2034(with smaller increases in other countries)when we incorporated an AI boost into our global potential growth forecasts last fall.When stripping out offs

109、etting growth impacts from the partial redirection of capex from other technologies to AI and slower productivity growth in a non-AI counterfactual,this 0.4pp annual figure translates into a 6.1%GDP uplift from AI by 2034 vs.Acemoglus 0.9%estimate.AI adoption remains modest on average across industr

110、ies Share of US firms using AI by sector,%Source:Census Bureau,Goldman Sachs GIR.That said,the full automation of AI exposed tasks that are likely to occur over a longer horizon could generate significant cost savings to the tune of several thousands of dollars per worker per year.The cost of new te

111、chnologies also tends to fall rapidly over time.Given that cost-saving applications of generative AI will likely follow a similar pattern,and that the marginal cost of deployment will likely be very small once applications are developed,we expect AI adoption and automation rates to ultimately far ex

112、ceed Acemoglus 4.6%estimate.Labor reallocation and new task creation on the horizon We also disagree with Acemoglus decision not to incorporate productivity improvements from new tasks and products into his estimates,partly given his questioning of whether AI adoption will lead to labor reallocation

113、 and the creation of new tasks.The historical record provides strong evidence that economic growth stems mainly from technology-driven reallocation of resources and expansion of the production frontier,and we anticipate that AI will raise output both by raising demand in areas where labor has a comp

114、arative advantage and by creating new opportunities that were previously technologically or economically infeasible.This dynamic clearly played out following the emergence of information technologywhich created new occupations like webpage designers,software developers,and digital marketing professi

115、onals and indirectly drove demand for service sector workers in industries like healthcare,education,and food servicesand is visible over a much longer horizon in recent work by MIT economist David Autor and coauthors.Using Census data,they find that 60%of workers today are employed in occupations t

116、hat did not exist in 1940,with their estimates implying that the technology-driven creation of new occupations accounts for more than 85%of employment growth over the last 80 years.Automation of work tasks should generate significant economic value,particularly as costs decline Value of automating w

117、ork task categories per worker,%of time(lhs),$(rhs)Source:Goldman Sachs GIR.Technological creation of new opportunities is a main driver of employment and economic growth Employment by new and pre-existing occupations,millions Source:Autor et al.(2022),Goldman Sachs GIR.Accordingly,while we believe

118、that Acemoglus relatively pessimistic assessment of generative AIs economic potential highlights valid concerns that the macroeconomic impacts could be more backloaded than is commonly appreciated,we maintain that generative AIs large potential to drive automation,cost savings,and efficiency gains s

119、hould eventually lead to significant uplifts of productivity and GDP.Joseph Briggs,Senior Global Economist Email: Goldman Sachs&Co.LLC Tel:212-902-2163 0102030ConstructionAccommodation and Food ServicesTransportation and WarehousingWholesale TradeOther ServicesManufacturingRetail TradeAdmin/Support/

120、Waste ManagementHealth Care and Social AssistanceAll IndustriesArts,Entertainment,and RecreationReal Estate and RentalFinance and InsuranceEducational ServicesProfessional,Scientific,and TechnicalInformationOctober 2023June 2024Next six months0500100015002000250001234Getting informationOrganizing,pl

121、anning,prioritizingworkUpdating and using relevantknowledgeIdentifying objects,actions,eventsProcessing informationMonitoring processes,materials,surroundingsDocumenting/recording informationAnalyzing data or informationEvaluating info to determinecompliance w/standardsScheduling work and activities

122、Performing administrative activitiesInterpreting the meaning ofinformation for othersEst.quantifiable characteristics ofproducts,events,info0255075100125150175200ProfessionalsManagersClerical&AdminProductionConstructionPersonal servicesTransportationTechniciansSalesCleaning servicesHealthFarmingTota

123、lOccupations that did not exist in 1940Occupations that existed in 1940hEl Goldman Sachs Global Investment Research 8 Top of Mind Issue 129 AI investment has surged over the last several years.Actual and forecasted revenues by AI-exposed sector,index,4Q19=100 .and the market has significantly upgrad

124、ed its AI investment expectations across the AI hardware stack.Change in consensus revenue forecasts since March 2023,$bn,annualized Dashed lines in this chart indicate consensus revenue forecasts.Source:FactSet,Goldman Sachs GIR.Source:FactSet,Goldman Sachs GIR.though much less so across the broade

125、r AI space so far Change in consensus revenue forecasts since March 2023,$bn,annualized Source:FactSet,Goldman Sachs GIR.AI-related software investment isnt yet visible in the US or other DMs official national accounts data.AI-related investment in software:national accounts,log index*,3Q22=100 *Sho

126、wn as log index because software investment grows by different exponential rates across countries.Steady growth in investment would appear as a line with a constant slope,while accelerating growth would appear as a line with an increasing slope.Source:Haver Analytics,Goldman Sachs GIR.nor is AI-rela

127、ted hardware investment,suggesting that other factors are currently playing a larger role than AI in shaping the aggregate capex outlook AI-related investment in hardware:national accounts,log index,3Q22=100 Source:Haver Analytics,Goldman Sachs GIR.However,manufacturers shipments for some AI-related

128、 components have surged.US nominal manufacturing sales,AI-related categories,index,Sept.2022=100,3m average Source:Haver Analytics,Goldman Sachs GIR.801001201401601802002202402019202020212022202320242025SemiconductorsHardware Enablers(ex.semis)Software Enablers0501001502002503003504004505001Q242Q243

129、Q244Q241Q252Q253Q254Q25SemiconductorsHardware Enablers(ex.semis)-500501001502002501Q242Q243Q244Q241Q252Q253Q254Q25Cloud providerData centers(real estate)Manufacturing equipmentMemorySecurityServers and networkingUtilities707580859095100105199920022005200820112014201720202023USCanadaJapanUK7075808590

130、95100105199920022005200820112014201720202023USCanadaJapanUK708090100110120130140150Dec-18Sep-19Jun-20Mar-21Dec-21Sep-22Jun-23Mar-24Electronic ComputersComputer Storage DevicesElectronic Components(incl.semiconductors)Communications EquipmentThe state of the AI transition.hEl Goldman Sachs Global Inv

131、estment Research 9 Top of Mind Issue 129.though this increase has not been uniform across the major developed economies,with the US leading the pack DM nominal manufacturing sales,AI-related categories,index,Sept.2022=100,3m average AI adoption remains muted on average across industries,with adoptio

132、n likely to pick up only modestly over the next six months.Share of US firms using AI by sector,%Source:Haver Analytics,Goldman Sachs GIR.Source:Census Bureau,Goldman Sachs GIR.though adoption rates are much higher among technology industries and other digitally-enabled fields.Share of US firms usin

133、g AI,top 15 subsectors,%Source:Census Bureau,Goldman Sachs GIR.and are expected to increase significantly across these sectors over the next six months Expected change in share of firms using AI over the next six months,top 15 subsectors,pp Source:Census Bureau,Goldman Sachs GIR.Despite rising adopt

134、ion rates,little evidence of net labor displacement from AI exists so far.Layoffs attributed to each in corporate announcements,000s Source:Challenger,Gray&Christmas,Goldman Sachs GIR.with unemployment not looking markedly different across jobs Unemployment rate by AI exposure,%,3m average Source:Ce

135、nsus Bureau,IPUMS,Goldman Sachs GIR.Special thanks to GS GIR global economist Devesh Kodnani for these charts,which were originally published in an April 2024 Global Economics Analyst.405060708090100110120Dec-18Sep-19Jun-20Mar-21Dec-21Sep-22Jun-23Mar-24USUKGermanyJapanCanadaFrance0102030Construction

136、Accommodation and Food ServicesTransportation and WarehousingWholesale TradeOther ServicesManufacturingRetail TradeAdmin/Support/Waste ManagementHealth Care and Social AssistanceAll IndustriesArts,Entertainment,and RecreationReal Estate and RentalFinance and InsuranceEducational ServicesProfessional

137、,Scientific,and TechnicalInformationOctober 2023June 2024Next six months05101520Ambulatory Health ServicesInsurance Carriers and RelatedBeverage and Tobacco ManufacturingPerforming Arts/Spectator SportsReal EstateComputer and Electronic ManufacturingRetail TradeTelecommunicationsEducational Services

138、Securities/Commodities/FinancialProfessional,Scientific,and TechnicalMotion Picture and Sound RecordingInformationWeb Search/Libraries/ArchivesComputing/Data/Web HostingOctober 2023June 202402468101214161820Performing Arts/Spectator SportsBeverage and Tobacco ManufacturingRetail TradeAmbulatory Heal

139、th ServicesReal EstateAdministrative and SupportInsurance Carriers and RelatedCredit Intermediation and RelatedProfessional,Scientific,and TechnicalInformationEducational ServicesMotion Picture and Sound RecordingSecurities/Commodities/FinancialWeb Search/Libraries/ArchivesComputing/Data/Web Hosting

140、0246810121416May-23Jul-23Sep-23Nov-23Jan-24Mar-24May-24Technological ChangeArtificial Intelligence3.03.23.43.63.84.04.24.4Jan-22May-22Sep-22Jan-23May-23Sep-23Jan-24May-24Top 20%most AI-exposed jobsAll other jobs.in pics hEl Goldman Sachs Global Investment Research 10 Top of Mind Issue 129 Jim Covell

141、o is Head of Global Equity Research at Goldman Sachs.Below,he argues that to earn an adequate return on costly AI technology,AI must solve very complex problems,which it currently isnt capable of doing,and may never be.Allison Nathan:You havent bought into the current generative AI enthusiasm nearly

142、 as much as many others.Why is that?Jim Covello:My main concern is that the substantial cost to develop and run AI technology means that AI applications must solve extremely complex and important problems for enterprises to earn an appropriate return on investment(ROI).We estimate that the AI infras

143、tructure buildout will cost over$1tn in the next several years alone,which includes spending on data centers,utilities,and applications.So,the crucial question is:What$1tn problem will AI solve?Replacing low-wage jobs with tremendously costly technology is basically the polar opposite of the prior t

144、echnology transitions Ive witnessed in my thirty years of closely following the tech industry.Many people attempt to compare AI today to the early days of the internet.But even in its infancy,the internet was a low-cost technology solution that enabled e-commerce to replace costly incumbent solution

145、s.Amazon could sell books at a lower cost than Barnes&Noble because it didnt have to maintain costly brick-and-mortar locations.Fast forward three decades,and Web 2.0 is still providing cheaper solutions that are disrupting more expensive solutions,such as Uber displacing limousine services.While th

146、e question of whether AI technology will ever deliver on the promise many people are excited about today is certainly debatable,the less debatable point is that AI technology is exceptionally expensive,and to justify those costs,the technology must be able to solve complex problems,which it isnt des

147、igned to do.Allison Nathan:Even if AI technology is expensive today,isnt it often the case that technology costs decline dramatically as the technology evolves?Jim Covello:The idea that technology typically starts out expensive before becoming cheaper is revisionist history.E-commerce,as we just dis

148、cussed,was cheaper from day one,not ten years down the road.But even beyond that misconception,the tech world is too complacent in its assumption that AI costs will decline substantially over time.Moores law in chips that enabled the smaller,faster,cheaper paradigm driving the history of technologic

149、al innovation only proved true because competitors to Intel,like Advanced Micro Devices,forced Intel and others to reduce costs and innovate over time to remain competitive.Today,Nvidia is the only company currently capable of producing the GPUs that power AI.Some people believe that competitors to

150、Nvidia from within the semiconductor industry or from the hyperscalersGoogle,Amazon,and Microsoftthemselves will emerge,which is possible.But thats a big leap from where we are today given that chip companies have tried and failed to dethrone Nvidia from its dominant GPU position for the last 10 yea

151、rs.Technology can be so difficult to replicate that no competitors are able to do so,allowing companies to maintain their monopoly and pricing power.For example,Advanced Semiconductor Materials Lithography(ASML)remains the only company in the world able to produce leading-edge lithography tools and,

152、as a result,the cost of their machines has increased from tens of millions of dollars twenty years ago to,in some cases,hundreds of millions of dollars today.Nvidia may not follow that pattern,and the scale in dollars is different,but the market is too complacent about the certainty of cost declines

153、.The starting point for costs is also so high that even if costs decline,they would have to do so dramatically to make automating tasks with AI affordable.People point to the enormous cost decline in servers within a few years of their inception in the late 1990s,but the number of$64,000 Sun Microsy

154、stems servers required to power the internet technology transition in the late 1990s pales in comparison to the number of expensive chips required to power the AI transition today,even without including the replacement of the power grid and other costs necessary to support this transition that on th

155、eir own are enormously expensive.Allison Nathan:Are you just concerned about the cost of AI technology,or are you also skeptical about its ultimate transformative potential?Jim Covello:Im skeptical about both.Many people seem to believe that AI will be the most important technological invention of t

156、heir lifetime,but I dont agree given the extent to which the internet,cell phones,and laptops have fundamentally transformed our daily lives,enabling us to do things never before possible,like make calls,compute and shop from anywhere.Currently,AI has shown the most promise in making existing proces

157、seslike codingmore efficient,although estimates of even these efficiency improvements have declined,and the cost of utilizing the technology to solve tasks is much higher than existing methods.For example,weve found that AI can update historical data in our company models more quickly than doing so

158、manually,but at six times the cost.More broadly,people generally substantially overestimate what the technology is capable of today.In our experience,even basic summarization tasks often yield illegible and nonsensical results.This is not a matter of just some tweaks being required here and there;de

159、spite its expensive price tag,the technology is nowhere near where it needs to be in order to be useful for even such basic tasks.And I struggle to believe that the technology will ever achieve the cognitive reasoning required to substantially augment or replace human interactions.Humans add the mos

160、t value to complex tasks by identifying and understanding outliers and nuance in a way that it is difficult to imagine a model trained on historical data would ever be able to do.Interview with Jim Covello hEl Goldman Sachs Global Investment Research 11 Top of Mind Issue 129 Allison Nathan:But wasnt

161、 the transformative potential of the technologies you mentioned difficult to predict early on?So,why are you confident that AI wont eventually prove to be just asor even moretransformative?Jim Covello:The idea that the transformative potential of the internet and smartphones wasnt understood early o

162、n is false.I was a semiconductor analyst when smartphones were first introduced and sat through literally hundreds of presentations in the early 2000s about the future of the smartphone and its functionality,with much of it playing out just as the industry had expected.One example was the integratio

163、n of GPS into smartphones,which wasnt yet ready for prime time but was predicted to replace the clunky GPS systems commonly found in rental cars at the time.The roadmap on what other technologies would eventually be able to do also existed at their inception.No comparable roadmap exists today.AI bul

164、ls seem to just trust that use cases will proliferate as the technology evolves.But eighteen months after the introduction of generative AI to the world,not one truly transformativelet alone cost-effectiveapplication has been found.Allison Nathan:Even if the benefits and the returns never justify th

165、e costs,do companies have any other choice but to pursue AI strategies given the competitive pressures?Jim Covello:The big tech companies have no choice but to engage in the AI arms race right now given the hype around the space and FOMO,so the massive spend on the AI buildout will continue.This is

166、not the first time a tech hype cycle has resulted in spending on technologies that dont pan out in the end;virtual reality,the metaverse,and blockchain are prime examples of technologies that saw substantial spend but have fewif anyreal world applications today.And companies outside of the tech sect

167、or also face intense investor pressure to pursue AI strategies even though these strategies have yet to yield results.Some investors have accepted that it may take time for these strategies to pay off,but others arent buying that argument.Case in point:Salesforce,where AI spend is substantial,recent

168、ly suffered the biggest daily decline in its stock price since the mid-2000s after its Q2 results showed little revenue boost despite this spend.Allison Nathan:What odds do you place on AI technology ultimately enhancing the revenues of non-tech companies?And even without revenue expansion,could cos

169、t savings still pave a path toward multiple expansion?Jim Covello:I place low odds on AI-related revenue expansion because I dont think the technology is,or will likely be,smart enough to make employees smarter.Even one of the most plausible use cases of AI,improving search functionality,is much mor

170、e likely to enable employees to find information faster than enable them to find better information.And if AIs benefits remain largely limited to efficiency improvements,that probably wont lead to multiple expansion because cost savings just get arbitraged away.If a company can use a robot to improv

171、e efficiency,so can the companys competitors.So,a company wont be able to charge more or increase margins.Allison Nathan:What does all of this mean for AI investors over the near term,especially since the“picks and shovels”companies most exposed to the AI infrastructure buildout have already run up

172、so far?Jim Covello:Since the substantial spend on AI infrastructure will continue despite my skepticism,investors should remain invested in the beneficiaries of this spend,in rank order:chip manufacturers,utilities and other companies exposed to the coming buildout of the power grid to support AI te

173、chnology,and the hyperscalers,which are spending substantial money themselves but will also garner incremental revenue from the AI buildout.These companies have indeed already run up substantially,but history suggests that an expensive valuation alone wont stop a companys stock price from rising fur

174、ther if the fundamentals that made the company expensive in the first place remain intact.Ive never seen a stock decline only because its expensivea deterioration in fundamentals is almost always the culprit,and only then does valuation come into play.Allison Nathan:If your skepticism ultimately pro

175、ves correct,AIs fundamental story would fall apart.What would that look like?Jim Covello:Over-building things the world doesnt have use for,or is not ready for,typically ends badly.The NASDAQ declined around 70%between the highs of the dot-com boom and the founding of Uber.The bursting of todays AI

176、bubble may not prove as problematic as the bursting of the dot-com bubble simply because many companies spending money today are better capitalized than the companies spending money back then.But if AI technology ends up having fewer use cases and lower adoption than consensus currently expects,its

177、hard to imagine that wont be problematic for many companies spending on the technology today.That said,one of the most important lessons Ive learned over the past three decades is that bubbles can take a long time to burst.Thats why I recommend remaining invested in AI infrastructure providers.If my

178、 skeptical view proves incorrect,these companies will continue to benefit.But even if Im right,at least they will have generated substantial revenue from the theme that may better position them to adapt and evolve.Allison Nathan:So,what should investors watch for signs that a burst may be approachin

179、g?Jim Covello:How long investors will remain satisfied with the mantra that“if you build it,they will come”remains an open question.The more time that passes without significant AI applications,the more challenging the AI story will become.And my guess is that if important use cases dont start to be

180、come more apparent in the next 12-18 months,investor enthusiasm may begin to fade.But the more important area to watch is corporate profitability.Sustained corporate profitability will allow sustained experimentation with negative ROI projects.As long as corporate profits remain robust,these experim

181、ents will keep running.So,I dont expect companies to scale back spending on AI infrastructure and strategies until we enter a tougher part of the economic cycle,which we dont expect anytime soon.That said,spending on these experiments will likely be the one of the first things to go if and when corp

182、orate profitability starts to decline.hEl Goldman Sachs Global Investment Research 12 Top of Mind Issue 129 Kash Rangan and Eric Sheridan are Senior Equity Research Analysts at Goldman Sachs covering US software and internet,respectively.Below,they argue that while AI remains a work in progress,the

183、large sums of money being put toward it should pay off,eventually.Allison Nathan:When we last spoke in July 2023,you were both very enthused about the potential of generative AI.Are you just as optimistic today?Kash Rangan:I am just as enthusiastic about generative AIs long-term potential as I was a

184、 year ago,and perhaps even more.The pace of technological change over the past 12 months has been mind-blowing,with hardly a week going by without reports of a newer,and better,AI model.The infrastructure buildout has also greatly exceeded expectations.Hyperscalerslarge cloud computing companies tha

185、t provide computing and storage services at scalehave spent$60-80bn in incremental capital above regular cloud capex on critical tools for building and training AI models.And rays of hope have emerged across several domains that demonstrate AIs productivity benefits.In the creative domain,generative

186、 AI has produced new design ideas in minutes that previously wouldve taken many hours,shortening the time it takes to bring an idea to market.In the code development domain,AI has automated low-level code writing,freeing up developers to work on more complex and productive tasks.And in the customer

187、support domain,ServiceNowa digital workflow software companyhas reported an 80%reduction in the average time it takes to resolve a customer service problem thanks to AI technology.That said,applications ultimately drive the success of tech cycles,and we have yet to identify AIs“killer application”,a

188、kin to the Enterprise Resource Planning(ERP)software that was the killer application of the late 1990s compute cycle,the search and e-commerce applications of the 2000-10 tech cycle that achieved massive scale owing to the rise of x86 Linux open-source databases,or cloud applications,which enabled t

189、he building of low-cost compute infrastructure at massive scale during the most recent 2010-20 tech cycle.But this shouldnt come as a surprise given that every computing cycle follows a progression known as IPAinfrastructure first,platforms next,and applications last.The AI cycle is still very much

190、in the infrastructure buildout phase,so finding the killer application will take more time,but I believe well get there.Eric Sheridan:I agree that the visibility into what this infrastructure buildout will translate into in terms of AI applications and adoption rates remains relatively low.And sever

191、al notable issues at the application layersuch as AI chatbots“hallucinating”or giving false answers to user promptshave called into question the scalability of generative AI.So,the technology is still very much a work in progress.But its impossible to sit through demonstrations of generative AIs cap

192、abilities at company events or developer conferences and not come away excited about its long-term potential.Allison Nathan:Its well known that Nvidia has benefitted massively in the current“picks and shovels”phase of the cycle.Are firms beyond Nvidia currently monetizing the gains from generative A

193、I technology?Eric Sheridan:Nvidia has certainly garnered significant revenue as its graphics processing unit(GPU)chip has become the nerve center of AI systems.But the semiconductor industry more broadly has benefitted from the voracious need for chips.Cloud computing companies have also performed w

194、ell owing to the enormous computing capacity required to train and run AI models,with the three large hyperscalers of Microsoft,Alphabet,and Amazon seeing an acceleration in revenue growth in the last quarter.So,capital is shifting into the AI theme,the theme and the capital are aligning against the

195、 building,and many companies exposed to semiconductors and computing workloads are monetizing these gains.So,this is not just a Nvidia story.Allison Nathan:Are you concerned that the hundreds of billions of dollars in AI capex big tech firms are estimated to spend in coming years is a sign of irrati

196、onal exuberance,and that the payoff may be low or never come?Eric Sheridan:Those who argue that this is a phase of irrational exuberance focus on the large amounts of dollars being spent today relative to two previous large capex cyclesthe late 1990s/early 2000s long-haul capacity infrastructure bui

197、ldout that enabled the development of Web 1.0,or desktop computing,as well as the 2006-2012 Web 2.0 cycle involving elements of spectrum,5G networking equipment,and smartphone adoption.But such an apples-to-apples comparison is misleading;the more relevant metric is dollars spent pany revenues.Cloud

198、 computing companies are currently spending over 30%of their cloud revenues on capex,with the vast majority of incremental dollar growth aimed at AI initiatives.For the overall technology industry,these levels are not materially different than those of prior investment cycles that spurred shifts in

199、enterprise and consumer computing habits.And,unlike during the Web 1.0 cycle,investors now have their antenna up for return on capital.Theyre demanding visibility on how a dollar of capex spending ties back to increased revenues,and punishing companies who cant draw a dotted line between the two.We

200、saw this with Meta a few months ago when the companys stock fell sharply after it announced plans to spend several billion dollars on AI,potentially disrupting its core business in the process,while offering little visibility into the eventual payoff.So,while I would never say Im not concerned about

201、 the possibility of no payback,Im not particularly worried about it today,though I could become more concerned if scaled consumer applications dont emerge over the next 6-18m.Kash Rangan:Spending is certainly high today in absolute dollar terms.But this capex cycle seems more promising than A discus

202、sion on generative AI hEl Goldman Sachs Global Investment Research 13 Top of Mind Issue 129 even previous capex cycles because incumbentsrather than upstartsare leading it,which lowers the risk that technology doesnt become mainstream.Incumbents have access to deep pools of capital,an extremely low

203、cost of capital,and massive distribution networks and customer bases,which allows them to experiment with how the capital dollars could eventually earn a return.Leading the late-1990s investment cycle,by contrast,were companies that didnt have the financing,reputation,or knowledge to succeed,resulti

204、ng in a tremendous amount of underutilized capacity.The companies spearheading the current investment cycle are also run by very capable managements,with CFOs watching expenses like a hawk,holding companies accountable for the return on investment,and standing ready to tap the brakes on spending if

205、the returns disappoint.Of course,this could all still fail,resulting in the loss of tens of billions of dollars in capex and interest income.But the opportunity costs of pursuing these strategies despite unknown outcomes still seems small compared to the potential opportunity of building the foundat

206、ion for the next big computing architecture.Allison Nathan:Even if corporate investments in AI eventually pay off,could this take longer than expected?Eric Sheridan:Many consumer internet companies have yet to see significant returns on their AI investments,and the timing of these returns remains un

207、certain because the three main channels of payoutadvertising,e-commerce,and subscription feeswill depend on shifting consumer habits.In the Web 1.0 tech cycle,the Netscape web browser debuted in the mid-1990s and the market peaked in March 2000,but the return on capital only turned positive in the l

208、ate 2000s/early 2010s as consumers were slow to embrace the technology.The payback period was much shorter in the Web 2.0 tech cycle that began in 2006,with most companies calling themselves mobile-first companies by 2012/13.But its not just the timing of returns that mattersif firms continue runnin

209、g at current levels of annualized spend over the next several years,the magnitude of returns will need to be outsized to justify the costs.That said,a longer-than-expected payoff process wont kill this tech cycle.Im loathe to use the word“bubble”because I dont believe that AI is a bubble,but most bu

210、bbles in history ended either because the cost of capital changed dramatically or end-demand deteriorated and affected companies ability to deploy capital,not because companies retreated from investing in a technology where the payoff was taking longer than expected.Kash Rangan:Monetization of AI te

211、chnology spend for enterprise software companies,including Salesforce,SAP,and Oracle,will come from customers willing to pay a premium for AI-infused products.The consumer market is a good leading indicator for the enterprise market,so once consumers begin embracing the technology,enterprise softwar

212、e companies should also benefit from a revenue tailwind.However,several such companies recently issued disappointing revenue guidance,leading some to go so far as to question whether the next decade will see“hardware eat software”,in stark contrast to the last decade of“software eating the world”.Bu

213、t the recent disappointments likely owed at least in part to the high-rate environmentsoftware is a$600-700bn industry,which makes it susceptible to high cost of capital.And,again,every computing cycle follows the IPA progression.So,while spending is currently aimed at the infrastructure,it will eve

214、ntually shift to platforms and applications,which is where the software companies will come in.Allison Nathan:Given the competitive pressures,do firms have any option other than competing in the AI arms race?Eric Sheridan:Even if ChatGPT didnt exist,Alphabetthe poster child for this debate in my cov

215、erage universewould probably still invest in AI.I remember sitting in the audience at Google I/O,the companys annual developer conference,in 2017 when Alphabet announced that it was now an AI-first company.But Alphabet has gone through more iteration and innovation since ChatGPT launched in 2022 tha

216、n in 2017-2022,which leads me to believe that the driving force behind AI spend is as much offensive as defensive.Allison Nathan:Some people argue that AI technology is too expensive,isnt actually fixing any real problem,and will likely never approach the cognitive abilities of humans because traini

217、ng on existing/historical data can only go so far.What are they missing?Kash Rangan:AI technology is undoubtedly expensive today.And the human brain is 10,000 x more effective per unit of power in performing cognitive tasks vs.generative AI.But the technologys cost equation will change,just as it al

218、ways has in the past.In 1997,a Sun Microsystems server cost$64,000.Within three years,the servers capabilities could be replicated with a combination of x86 chips,Linux technology,and MySQL scalable databases for 1/50th of the cost.And the scaling of x86 chips coupled with open-source Linux,database

219、s,and development tools led to the mainstreaming of AWS infrastructure.This,in turn,made it possible and affordable to write thousands of software applications,such as Salesforce,ServiceNow,Intuit,Adobe,Workday,etc.These applications,initially somewhat limited in scale,ultimately evolved to support

220、a few hundred million end-users,not to mention the impressive scaling of Microsoft Azure that supported ubiquitous applications such as Office 365.Over the last decade,these applications have evolved and helped create hundreds of billions of dollars in shareholder value,providing even more evidence

221、for the clich that people tend to overestimate a technologys short-term effects and underestimate its long-term effects.Nobody today can say what killer applications will emerge from AI technology.But we should be open to the very real possibility that AIs cost equation will change,leading to the de

222、velopment of applications that we cant yet imagine.Eric Sheridan:Again,I readily acknowledge that the return on invested capital(ROIC)visibility is currently low,and the transformative potential of AI will remain hotly debated until that becomes clearer.But AI skeptics miss three key things.One,trai

223、ning on existing/historical data to inform and drive analytic outcomes in the future sounds exactly like going to universitypeople go to learn and then improve productivity and efficiency for decades after graduation,and machines can absolutely do the same.Two,machines today can do a whole host of t

224、asks more productively and efficiently than humans,and that will remain true for decades into the future.And three,people didnt think they needed smartphones,Uber,or Airbnb before they existed.But today it seems unthinkable that people ever resisted such technological progress.And that will almost c

225、ertainly prove true for generative AI technology as well.hEl Goldman Sachs Global Investment Research 14 Top of Mind Issue 129 A short history of AI developments Note:This does not constitute an exhaustive list of all AI-related developments.Source:BBC,cancers,OpenAI,tech.co,Google,various news sour

226、ces,compiled by Goldman Sachs GIR.hEl Goldman Sachs Global Investment Research 15 Top of Mind Issue 129 Brian Janous is Co-founder of Cloverleaf Infrastructure,which develops strategies to help utilities unlock new grid capacity.Previously,he was Vice President of Energy at Microsoft.Below,he argues

227、 that US power infrastructure is not prepared for the coming surge in power demand from AI and other sources,setting up for a painful power crunch in the coming years.The views stated herein are those of the interviewee and do not necessarily reflect those of Goldman Sachs.Jenny Grimberg:Power deman

228、d is surging across parts of the US,and utilities and grid operators have significantly raised their estimates of US electricity demand growth over the next five years.What role are advances in AI technology playing in the US growing hunger for electricity,and how do data centers fit into that?Brian

229、 Janous:Cloud data centers have grown rapidly since the advent of cloud computing around 2010.However,global data center electricity consumption barely budged over the subsequent decade as these data centers cannibalized on-prem workloads,which used multiples more electricity per unit of compute tha

230、n cloud data centers.So,the migration of data to the cloud resulted in a significant increase in computation with almost no rise in electricity usage.But as the cloud data center capacity of the three large hyperscalers of Microsoft,Amazon,and Google grew from a few hundred megawatts in the early 20

231、10s to a few gigawatts by the end of the decade,power consumption began to rise.And the release of ChatGPT 3.5 in November 2022 ushered in a new layer of AI-related demand,which will likely require adding hundreds of megawattsif not gigawattsof data center capacity annually.So,power demand is set to

232、 continue surging over the coming years.Jenny Grimberg:How much does electric grid capacity have to expand to meet this surge?Brian Janous:Thats the million-dollar question.Utilities are fielding hundreds of requests for huge amounts of power as everyone chases the AI wave,but only a fraction of tha

233、t demand will ultimately be realized.AEP,one of the largest US electric utility companies,has reportedly received 80-90 gigawatts(GW)of load requests.Only 15 GW of that is likely real because many of the AI projects that companies are currently envisioning will never actually see the light of day.Bu

234、t 15 GW is still massive given that AEP currently owns/operates around 23 GW of generating capacity in the US.And even if overall grid capacity grows by only 2%annuallywhich seems like a reasonable forecastutilities would still need to add well in excess of 100 GW of peak capacity to a system that c

235、urrently handles around 800 GW at peak.The increase in power demand will also likely be hyperlocalized,with Northern Virginia,for example,potentially requiring a doubling of grid capacity over the next decade given the concentration of data centers in the area.So,grid capacity will need to expand su

236、bstantially across the US,and likely even more in certain regions.Jenny Grimberg:Are utility companies and the underlying power infrastructure equipped to meet the rapid surge in power demand?Brian Janous:No.Utilities have not experienced a period of load growth in almost two decades and are not pre

237、pared foror even capable of matchingthe speed at which AI technology is developing.Only six months elapsed between the release of ChatGPT 3.5 and ChatGPT 4.0,which featured a massive improvement in capabilities.But the amount of time required to build the power infrastructure to support such improve

238、ments is measured in years.And AI technology isnt developing in a vacuumelectrification of transportation and buildings,onshoring of manufacturing driven partly by the Inflation Reduction Act and CHIPS Act,and potential development of a hydrogen economy are also increasing the demands on an already

239、aged power grid.Regulatory lags and interconnection and supply chain constraints are also impediments to meeting the rising power demand.The total capacity of power projects waiting to connect to the grid grew nearly 30%last year,with wait times currently ranging from 40-70 months,and lead times for

240、 critical electrical components such as transformers and switchgears have substantially increased.Until those issues can be resolved,and the grid can catch up,a significant power crunch will likely force utilities and states to pick and choose who receives power.My concern is that data centers will

241、become an easy target because theyre not perceived as major engines of job creation relative to building the next Hyundai factory,for example.This dynamic has already occurred in Dublin,where EirGrid,a state-owned power operator,enacted a de facto moratorium on new data centers by delaying their gri

242、d connection until 2028.Amsterdam also recently unveiled new rules that would impose fines on data centers that dont switch off idle servers to conserve energy.Its not out of the realm of possibility that something similar could happen in the US.Jenny Grimberg:Didnt similar concerns arise during the

243、 era of hyperscale computing,which many worried would gobble up all the worlds power,only to be proven wrong as data centers became more efficient?Brian Janous:The experience of 2010-20 has provided a false sense of comfort.Most of the efficiency gains over that period owed to the shutdown of ineffi

244、cient on-prem data centers in favor of cloud data centers.Data centers themselves did not become significantly more efficient.To put some numbers on this,the average power usage effectiveness(PUE)a measure of data center efficiency calculated by dividing the total amount of power a facility consumes

245、 by the amount used to run the serversof on-prem data centers was 2-3 vs.around 1.3 for cloud,but the PUE of cloud data centers themselves only declined by around 0.2 over the course of the decade.And the Interview with Brian Janous hEl Goldman Sachs Global Investment Research 16 Top of Mind Issue 1

246、29 average PUE of data centers today is around 1.1,meaning over 90%of the power they consume goes directly to the servers vs.cooling,lighting,etc.So,only limited room exists to extract more efficiency from a data center.And even if new ways were discovered to increase the efficiency of data centers

247、or AI chips themselves,humans capacity to consume data is nearly insatiable.Every time we develop a more efficient chip or process,we find ways to use more of the underlying resource,not less of it,which is known as Jevons paradox.Big tech firms are currently engaging in an AI arms race to create th

248、e most powerful and capable AI model,and until we reach a level of saturation in terms of human capacity to consume data,any amount of efficiency gains will undoubtedly be gobbled up by even more demand.Jenny Grimberg:Whats required to expand the grid?Brian Janous:Expanding the grid is no easy or qu

249、ick task.The electric utility industry is highly regulated,and utility companies must go through a lengthy permitting and approval process before starting to construct new capacity.They then must contend with a supply chain that isnt prepared for every utility company to suddenly double their equipm

250、ent orders,and building up the supply chain to meet the growing demand in itself will take at least months,and even years in some cases.To help ease some of the power constraints in the meantime,utilities will need to find ways to extract more efficiencies from the current system,for example by reco

251、nductoring,or replacing,existing transmission lines to move more power over them,and investing in grid-enhancing technologies.Expanding long-duration storage to deliver electricity when and where its most needed could also ease periods of peak capacity,as can integrating more flexibility into how,an

252、d when,energy is consumed.While there has been discussion about AI workloads being flexible,the reality is they will still need a high level of power availability.However,data centers generation and storage assets can be leveraged for flexibility.That can help ensure that peak capacity isnt increasi

253、ng at the same rate as the overall consumption of electricity.But again,the power constraint issue ultimately cannot be resolved without a significant buildout of electric grid infrastructure.Jenny Grimberg:Big tech companies seem to have endless amounts of capital to throw at AI technology.Couldnt

254、they just spend some of that money securing the power supply they need?Brian Janous:If this was simply an issue of money,it would have already been solved.The big tech companies face the same constraints as the utility industry.They must go through the same regulatory processes and are subject to th

255、e same supply chain issues.Some people have suggested that these firms should just generate all their own power.But the only way to do so today is by using natural gas,which still sits on a grid that comes with its own set of constraints and requires upgrading and building new infrastructure.Nuclear

256、 power also gets tossed around as a potential solution,but building a new nuclear plant within the next decade isnt feasible.Those who believe nuclear is the answer often point to Amazon Web Services recent purchase of a data center from Talen Energy located next to the Susquehanna nuclear power sta

257、tion in Pennsylvania.But nothing new was built to unlock that powerthat nuclear plant has been around for over four decades.And so,if the problem were trying to solve is how to increase power supply this decade,we have to expand the electric grid.No other solution exists.Jenny Grimberg:So,is the US

258、up to the task?Brian Janous:The US has unfortunately lost the ability to build large infrastructure projectsthis is a task better suited for 1930s America,not 2030s America.So,that leaves me a bit pessimistic.That said,utilities and policymakers are starting to take seriously the need to invest in A

259、mericas transmission infrastructure,which isnt designed for todays energy generation mix.The transmission infrastructure was built from the coasts into the country,but today,massive wind resources are located in the center of the country and solar in the southwest.So,the transmission system ideally

260、needs to run from the inside out.Utilities have also begun to recognize that the significant increase in load growth that lies ahead creates a massive economic development opportunity,probably the biggest theyll ever see.The utilities that can find ways to offer more power to more customers sooner w

261、ill attract that economic development and growth.And while data centers arent perceived as major job creators,they still create significant economic activity in the form of construction jobs and large tax revenues.No utility wants to turn away customers or tell their states governor that a new facto

262、ry went elsewhere because they couldnt provide enough power.Thats a significant motivator for utilities to not only invest in grid infrastructurewhich,if done thoughtfully,shouldnt lead to rate increases for consumersbut also find efficiencies in the current system,which they currently lack the ince

263、ntive to do because theres no money in it for them.So,on balance,Im optimistic that America can rise to meet the challenge,though the next decade will likely prove painful as the demand for power outpaces the available supply.Jenny Grimberg:Are the big tech companies spending hundreds of billions of

264、 dollars on AI infrastructure,including data centers,underappreciating the power constraint?Brian Janous:Yes and no.These companies are very optimistic about what they can achieve with AI,but tech firms are starting to realize that power supply will be a significant constraint on the technology.When

265、 generative AI first exploded onto the scene,people debated what would constrain its potentiala shortage of chips or a shortage of power.That debate has now been settled,with everyone agreeing that over the medium-to-longer term the major constraint will be power.Metas Mark Zuckerberg recently state

266、d in an interview that energy constraints are the biggest bottleneck to building out AI data centers.Microsofts Satya Nadella has also spoken about this.And Nvidias Jensen Huang recently addressed the electric utility industry at the EEI conference,which would have seemed crazy even a year ago but n

267、ow makes total sense.So,companies have woken up to the fact that electricity is an incredibly important commodity,and are now hyper-focused on the power constraint.But recognition of the problem is one thingsolving it is a much more difficult challenge.hEl Goldman Sachs Global Investment Research 17

268、 Top of Mind Issue 129 Carly Davenport answers key questions about the coming surge in US power demand from AI technology and data centers The proliferation of generative AI technologyand the data centers needed to feed itis set to drive an increase in US power demand not seen in a generation.Here,w

269、e address key questions about the coming power demand surge,how much new generation capacity will be required to meet it,and the implications for companies and investors.Q:How significant will the power demand growth from AI/data centers be?A:After stagnating over the last decade,we expect US electr

270、icity demand to rise at a 2.4%compound annual growth rate(CAGR)from 2022-2030,with data centers accounting for roughly 90bp of that growth.Indeed,amid AI growth,a broader rise in data demand,and a material slowdown in power efficiency gains,data centers will likely more than double their electricity

271、 use by 2030.This implies that the share of total US power demand accounted for by data centers will increase from around 3%currently to 8%by 2030,translating into a 15%CAGR in data center power demand from 2023-2030.After stagnating over the last decade,US power demand should grow by 2%per year on

272、average through 2030 US power demand growth,%Source:EIA,Goldman Sachs GIR.Q:How much power generation is required to support this demand growth,and where will it come from?A:We estimate the US will require 47 gigawatts(GW)of new power generation capacity through 2030 to support the growth in data ce

273、nter power demand.We expect this capacity to be split 60%/40%between natural gas and renewables generation,reflecting a balance between the reliability needs of data centers and companies green energy commitments.The data centers that power AI models must essentially run 24/7 given the nature of AI

274、workloads,and so require a constant energy source like natural gas that can be dispatched on demand rather than renewables,which are more intermittent in nature.But we still believe renewables will play an important role given that many of the companies building data centersespecially the hyperscale

275、rs of the worldhave committed to green electricity consumption and are unlikely to abandon those commitments to meet growing data center demand.We estimate around 47 GW of incremental capacity is needed to serve data center-driven load growth in the US through 2030 Data center-driven capacity adds,m

276、egawatts(MW)Source:Goldman Sachs GIR.Q:Could nuclear energy also be part of the solution alongside natural gas and renewables?A:The US historically hasnt demonstrated the best track record of building nuclear plants on time or on budget,so we dont think utilities would take on the risk of attempting

277、 to build new capacity.However,data centers could attempt to strike deals with companies that operate unregulated nuclear plantsthose not regulated by a state utility commission and therefore not precluded from striking direct contracts with customersbecause nuclear power solves exactly what the dat

278、a centers are looking for:reliability and no carbon emissions.And some such deals have already occurred.Data centers should contribute 90bp to our 2.4%US power demand CAGR from 2022-2030 Composition of US power demand CAGR,2022-2030,%Source:EIA,Goldman Sachs GIR.-5%-4%-3%-2%-1%0%1%2%3%4%5%2000 2003

279、2006 2009 2012 2015 2018 2021 2024 2027 2030Power demand growth2000-2007 Average10-year AverageEstimated power demand growthOnce in a generation,generation hEl Goldman Sachs Global Investment Research 18 Top of Mind Issue 129 Q:How much capital investment will utilities need to make to provide the n

280、ecessary capacity?A:Approximately$50bn of investment through 2030,or roughly$7bn annually,is needed to facilitate the new power generation alone.But utilities will also need to build out the supporting infrastructure,such as the transmission wires that transport electricity over long distances and d

281、istribution cables that carry electricity to homes,so the overall investment will likely prove much higher.Between generation,transmission,and distribution needs,we expect the utility companies in our coverage universe to spend nearly 40%more from 2024-2027 relative to the prior four-year period,amo

282、unting to roughly$140bn on average annually.So,a significant increase in utility capex likely lies ahead,with this investment already materializing in regions like Northern Virginia,a hotbed for data center growth(see pg.19).Capex across the utility companies in our coverage should increase by rough

283、ly$140bn on average annually from 2024-2027 Capex by year,$bn Source:SNL,Company data,Goldman Sachs GIR.Q:What constraintsif anycould prevent the industry from delivering the required capacity?A:The most significant constraint is the long timelines for infrastructure permitting and construction.Many

284、 power project developers start the process 5-7 years in advance to adequately plan for land acquisition,resource planning,permitting timelines,and interconnection queueswhich currently range from 40-70 months across the countryas well as any potential supply chain constraints.Affordability is also

285、an important constraint for utilities,which is a highly regulated industry.Regulators are focused on ensuring that electricity bills remain affordable for residential customers and that the capital investment necessary to meet data center growth isnt borne by the residential customer.This ultimately

286、 puts a cap on the rates utility companies can charge and still get their project approved.Q:If utilities dont have much leeway to raise prices,where will the funding for the capacity investment come from?A:Utilities dont generate a significant amount of free cash flow,so they will need to add debt

287、capacity or issue equity to facilitate this massive investment.But part of the investment could also come from the data center customers themselves.If a utility needs to build infrastructure that will only support a data center,several ways exist for the utility to structures its contracts and rates

288、 to ensure that the capital is sourced only from that customer and the cost isnt socialized across the broader customer base.Q:What are the main risks to your demand/investment forecasts?A:AI/data center-led demand could be lower than we expect if advancements in energy-efficient hardware materializ

289、e or data center customers overestimate their near-term power needs.However,the increasing proliferation of AI technology and demand for data,as well as a slowdown in efficiency gains,should ultimately drive stronger power demand from data centers.The ongoing electrification of transportation,buildi

290、ngs,and oil&gas operations,as well as increased manufacturing activity in the US due to reshoring,strengthens our confidence that power demand growth will rise to levels not seen since the turn of the century.This growth,together with the energy transition,need to address aging existing infrastructu

291、re,and increased climate risk,should ultimately support a significant rise in grid investments from utilities as well as infrastructure contractors and industrials making products that support the buildout.Q:What companies will benefit the most from the coming surge in power demand?A:We see two broa

292、d categories of beneficiaries:demand growth beneficiaries and supply chain/infrastructure beneficiaries.Demand growth beneficiaries include companies levered to power needs/prices,including unregulated power producers,gas companies,energy storage players,and those that provide power solutions to dat

293、a centers.This category also includes companies involved in building power generation capacity to help meet the growing load,including regulated utilities,merchant power producers,renewables companies,and generation kit suppliers.Supply chain/infrastructure beneficiaries include companies positioned

294、 to invest in infrastructure or equipment to help facilitate the buildout of power infrastructure and support grid reliability.Although some stocks have moved higher on the potential for increased power demand,many of the downstream exposed names have continued to trade on more cyclical factorslike

295、elevated interest ratesversus the secular tailwinds that we believe will contribute to future earnings growth.Carly Davenport,Senior US Utilities Equity Research Analyst Email: Goldman Sachs&Co.LLC Tel:212-357-1914$92$92$105$126$129$143$143$148$0$20$40$60$80$100$120$140$16020202021202220232024202520

296、262027$414bn$563bnhEl Goldman Sachs Global Investment Research 19 Top of Mind Issue 129 Virginia is for data centers Following two decades of stagnation,investors are increasingly focused on the potential boost to US power demand from AI and data centers.While the expected increase in AI-driven powe

297、r demand is in its early days(see pgs.17-18),evidence from Virginiathe likely epicenter of this demand growthprovides a glimpse of the coming US power demand surge.Evidence from the data center capital Virginia is a useful starting point to assess the boost to overall power demand in the US given it

298、s concentration of data centers.Data centers have grown rapidly in Virginia since late 2016 despite a brief pause during the pandemic,with Northern Virginia home to the most data centers in the US.Alongside this explosion of data centers,commercial power consumption in Virginia rose 37%from 2016 to

299、2023,while remaining flat in most other states.And within Virginia,commercial power consumption growth has also outstripped non-commercial power consumption,with both residential and industrial power consumption decreasing over 2016-2023 by 3%and 4%,respectively.Virginia commercial power consumption

300、 growth has outpaced other states since late 2016 Commercial power consumption,12m moving avg,indexed,2010-2016 avg=100 and has also outpaced other power consumption sectors Power consumption across power sectors,12m moving avg,indexed,2010-2016 avg=100 Source:Haver Analytics,EIA,Goldman Sachs GIR.S

301、ource:Haver Analytics,EIA,Goldman Sachs GIR.points to a boost in power consumption We use a statistical“doppelganger”technique to estimate how much data centers have contributed to the observed rise in Virginia power consumption data.The doppelganger method uses the historical relationship(2010-2016

302、)between commercial power consumption in Virginia and the US more broadly to estimate what Virginia commercial power consumption would have looked like without data centers.Taking the difference between actual Virginia power demand and doppelganger demand,we find that data centers boosted Virginia p

303、ower consumption by 2.2 gigawatts(GW)in 2023,accounting for 15%of the total power consumption in the state that year,compared to virtually 0%in 2016 and roughly 3%in 2019.Doppelganger demand for Virginia commercial power consumption points to lower consumption levels Commercial power consumption,gig

304、awatts implying data center power consumption of 2.2 GW in 2023 Implied data centers power consumption of Virginia,gigawatts Source:Haver Analytics,EIA,Goldman Sachs GIR.Source:Haver Analytics,EIA,Goldman Sachs GIR.although only a modest one so far Hongcen Wei,Commodities Strategist Email: Goldman S

305、achs&Co.LLC Tel:212-934-4691 While the evidence suggests that AI and data centers are boosting US power demand,the overall magnitude of the boost remains modest compared to both the current level of total US power demand as well as the level of data center power demand expected later this decade.We

306、estimate the 2.2 GW of Virginia data center power demand in 2023 makes up only 0.5%of the 470 GW of total US power demand and 7%of the roughly 30 GW increase in overall data center demand our equity analysts expect by 2030.But the magnitude of the recent increase in data center power demand in Virgi

307、nia provides a glimpse of the large boost in US power demand likely ahead.809010011012013014015020102012201420162018202020222024VirginiaUS ex.VirginiaCovid-19 recession809000102030405020102012201420162018202020222024Virginia commercial power demandVirginia residential power demandVirginia industrial

308、 power demandCovid-19 recession7.65.44.04.55.05.56.06.57.07.58.020102012201420162018202020222024VirginiaUS DoppelgangerGrowth has slowed amiddata center expansion innearby states and power transmission bottlenecksData center growthtook off2.32.2-0.50.00.51.01.52.02.520102012201420162018202020222024h

309、El Goldman Sachs Global Investment Research 20 Top of Mind Issue 129 US data centers,mapped out hEl Goldman Sachs Global Investment Research 21 Top of Mind Issue 129 Alberto Gandolfi argues that the expansion of AI data centers will boost European power demand over the next decade,which should benef

310、it“Electrification Compounders”Over the past fifteen years,a series of exogenous shocksthe Global Financial Crisis,Covid pandemic,2022 energy crisis,a slower-than-expected electrification process,and the ongoing de-industrialization of the continents economyhave hit power demand in Europe.As a resul

311、t,electricity consumption has declined by around 10%from its 2008 peak.However,this negative trend may be on the verge of reversing.We estimate that the rapid expansion of data centers amid the increasing proliferation of generative AI technology and gradual acceleration of the electrification proce

312、ss could boost Europes power demand by around 40-50%over the coming decade.AI data centers:a new driver of power demand Traditional data centers have rapidly expanded to meet higher demand from retail customers(owing to the increased popularity of cloud storage,social media,and movie streaming),the

313、service industry(on increased computational and storage needs),and the large tech companies such as Google,Amazon,Meta,and Microsoft.However,data centers currently account for only just over 1%of power demand globally.Our conservative base case scenario assumes that the expansion of traditional data

314、 centers could boost European power demand by around 10-15%over the coming ten years.Studies show that AI data centers can consume up to around 10 x more energy than traditional data centers,particularly during their training phase.We estimate that AI data centers and electrification could boost Eur

315、opean power demand by+40%over the coming decade based on our Tech analysts expectations for global AI shipments,conservative assumptions on energy efficiency,and a declining market share in Europe vis-vis the US.However,a bull case for AI data centerswhich assumes a slightly higher data center marke

316、t share of 25%for Europe and no efficiency gains on future server deliveriescould see cumulative electricity consumption growth of around 50%over the next decade.But even in our base case,the incremental power consumption we expect from AI and traditional data centers in Europe over the next decade

317、would be equivalent to the current consumption of the Netherlands,Portugal,and Greece,combined.A highly regional impact We expect this power demand to be highly concentrated in two areas.First,countries with cheap,abundant baseload power(i.e,those that enjoy a higher proportion of wind,solar,hydro,a

318、nd nuclear in their energy mix):the Nordics,Spain,and France.Second,countries with a strong financial services presence and those acting as big tech hubs as well as those willing to offer incentives to attract data centers and to support a faster adoption of electrification technologies:Germany,the

319、UK,and Ireland.Assuming these two groups of countrieswhich currently account for nearly three-quarters of Europes total power consumptionabsorb 85-95%of the total incremental power demand from data centers,electricity consumption in these regions could rise by 10-15%over the next decade.ChatGPT quer

320、ies are 10 x as power intensive as Google searches Power consumption per query/search,Watt-hour(Wh)Source:Google,SemiAnalysis,compiled by Goldman Sachs GIR.AI data centers and electrification could boost European power by over 40%in the coming decade EU-27 power demand scenario analysis,index,2023=1

321、00 Source:EMBER,Goldman Sachs GIR.Meeting higher demand:grids and renewables to the rescue We estimate that the rapid expansion of data centers in these areas,together with the REPowerEU Planwhich is set to kickstart a major electrification process in Europewill lead European power demand to grow by

322、 around 40-50%over the next ten years.Investments in power grids and renewables will likely prove pivotal in meeting this substantial rise in demand.On the power grid front,we expect a secular capex supercycle ahead with European investments in power grids accelerating by 80-100%,depending on the re

323、gion.And on the renewables front,we expect Europe to add nearly 800 gigawatts(GW)of wind and solar over the coming 10-15 years,nearly tripling the amount currently installed in the region.Investment implications:look to electrification compounders For European utilities,an industry with elevated ope

324、rational and financial gearing,the coming inflection in power demand should have significant positive implications for revenues and,in turn,profits.This revenue boost will also likely trigger secular organic growth in power grids and renewablesthe key infrastructure enabling the proliferation of dat

325、a centers and the electrification process.And we see“Electrification Compounders”utilities that mostly grow profits from power grids and renewablesas the main beneficiaries of the trend toward rising power demand given their highly attractive risk/reward profiles.0.3 Wh2.9 WhGoogleChatGPTc.10 x90100

326、1101201301401501602008 2010 2012 2014 2016 2018 2020 2022 2024 2026 2028 2030 2032HistoricalStatus quoElectrificationAIAI bull caseAI:powering up Europe Alberto Gandolfi,Head of European Utilities Equity Research Email: Goldman Sachs and Co.LLC Tel:39-02-8022-0157 hEl Goldman Sachs Global Investment

327、 Research 22 Top of Mind Issue 129 Our US semiconductor team led by Toshiya Hari expects supply constraints in the semiconductor industry to remain a limiting factor on AI growth over the next few years As the popularity of generative AI technology continues to grow,the demand for AI chipsincluding

328、everything from Nvidias GPUs to custom chips designed by large cloud computing companieshas skyrocketed,leading to questions around whether the semiconductor industry can keep up.We expect industry supply,rather than demand,to dictate AI chip shipments through 2H24 and into early 2025 given constrai

329、nts on two key fronts:High-Bandwidth Memory(HBM)technology and Chip-on-Wafer-on-Substrate(CoWoS)packaging.An undersupplied HBM market AI applications use two types of dynamic random-access memory(DRAM):HBM and DDR SDRAM.HBM is a revolutionary memory technology that stacks multiple DRAM diessmall blo

330、cks of semiconducting material on which integrated circuits are fabricatedon top of a base logic die,thereby enabling higher levels of performance through more bandwidth when interfacing with a GPU or AI chips more broadly.We expect the HBM market to grow at a 100%compound annual growth rate(CAGR)ov

331、er the next few years,from$2.3bn in 2023 to$30.2bn in 2026,as the three incumbent suppliers of DRAM(Samsung,SK Hynix,and Micron)allocate an increasing proportion of their total bit supply to meet the exponential demand growth.Despite this ramp-up,HBM demand will likely outstrip supply over this peri

332、od owing to growing HBM content requirements and major suppliers supply discipline.We therefore forecast HBM undersupply of 3%/2%/1%in 2024/2025/2026.Indeed,as Nvidia and AMD recently indicated,updated data center GPU product roadmaps suggest that the amount of HBM required per chip will grow on a s

333、ustained basis.And lower manufacturing yield rates in HBM than in traditional DRAM given the increased complexity of the stacking process constrains suppliers ability to increase capacity.Packaging,bottlenecked The other key supply bottleneck is a specific form of advanced packaging known as CoWoS,a 2.5-dimensional wafer-level multi-chip packaging technology that incorporates multiple dies side-by