A skeptical look at AI investment

Show video

Generative AI is being heralded as one  of the most transformative  innovations in human history,  and AI optimism has become  one of the market's biggest  drivers. Companies are estimated  to devote over $1 trillion  to AI-related spending in  the coming years, so will the  benefits and returns of the  technology justify the cost?  Jim Covello: We're a couple  years into this, and there's  not a single thing that this  is being used for that's cost  effective at this point. Allison Nathan: I'm Allison   Nathan, and this is Goldman Sachs Exchanges. Every month,   I speak with investors, policymakers, and academics   about the most pressing market-moving issues   for our Top of Mind report from Goldman Sachs Research. This month, I looked at  generative AI. 2 

I think we're all familiar  with the bull case by now. My  colleague Joseph Briggs from our Global Economics Research team estimates that generative AI could  ultimately automate a quarter  of all work tasks and boost  US productivity by 9% and US GDP growth by 6.1% cumulatively over the next decade.   And investors have certainly bought into   the theme with the big tech firms at the center of it all, accounting   for over 60% of the S&P 500 Index's year-to-date return.  But given the enormous cost to develop and run the technology with no so-called   application for it yet found, questions about whether the technology will ever  sufficiently deliver on this  investment have grown. I spoke  to two people who are  skeptical. Daron Acemoglu, an  Institute professors at MIT,  the author of several books,  including Power and Progress:  Our Thousand-Year Struggle  Over Technology Prosperity. In  a recent paper, he estimated 

that only a quarter of AI-exposed  tasks will be cost effective  to automate over the next ten  years, implying that AI will  impact less than 5% of all work tasks and boost US productivity by only 0.5% and US GDP about 1%  cumulatively over the next  decade. I asked him to explain.  3 Daron Acemoglu: Part of the reason why I wrote the  paper is because I did see  a lot of enthusiasm, some  quantitative, some qualitative,  from commentators, some  experts, others observers of  the tech industry about the  transformative effects that AI  is going to have very quickly  on the economy. And I think  economic theory actually 

puts a lot of discipline on  how some of these effects can  work once we leave out those  things like amazing new  products coming online, something much better than silicon coming, for example,   in five years. Of course, if that happens, all right, that's big.  But once you leave those out,  the way that you're going to  get productivity effects is you  look at what fraction of the  things that we do in the  production process are impacted  and how that impact is going  to change our productivity or  reduce our costs. So those  are the two ingredients. And  my prior, even before looking  at the data, was that the  number of tasks that are going  to be impacted by gen AI is  not going to be so huge in  the short run because a lot of  the things that humans do  are very multifaceted. So  almost all of the things that we do in transport, manufacturing, mining, utilities has a very   central 4  component of interacting  with the real world. And AI 

ultimately can help with that  as well, but I can't imagine  that being a big thing within the next few years. So therefore, my intuition was that,   from the beginning, it's going to be pure mental tasks   that are going to be affected. And those are not trivial,   but they're not huge either. So the way I go about doing that is   I rely on one of the most comprehensive studies by   Eloundou, Mishkin, and Rock that codes what the current AI   technology -- generative AI technology -- combined   with other AI technologies and computer vision could ultimately do. What they did  seemed fairly, fairly solid,  so I decided to start from what  they did. So if you take their numbers,  

it suggests that something like 20% in terms of   value-added share in terms of the economic importance of the tasks that we do in the  production process could be  ultimately transformed or  heavily impacted by AI. But  that's a timeless prediction  ultimately. But when will that  ultimately be realized? And  in that, there is another  paper by Neil Thompson and  Martin Fleming and coauthors  which looks at a subset of  these technologies, the computer  vision technologies, where  5 we understand how they   work and the cost is a little bit better. And for the computer  

vision technologies, they come up with some numbers as well. But more  importantly, they make an  effort in thinking about how  quickly these things are going  to be cost effective. Because  something being ultimately  done by generative AI with  sufficient improvements doesn't  mean it's going to be a big  deal within five, six, seven  years. They come up with a 

number for computer vision  technologies, again, which  looking into the details,  seemed reasonable, which is at  about 20-25% of what is  ultimately doable can be cost  effectively automated by  computer vision technologies  within ten years.  So then I combined these two  estimates, and I forecast that  about 4.5-4.6% of those 23%  is going to be done in the  short run and within the  10-year horizon. And that's the 

basis of my sort of fairly  uncertain -- of course, we cannot  be certain about any of  these things -- but baseline  estimate of what can be achieved  with gen AI within a 10-  year horizon. Allison Nathan: When we think about applying AI  technology to various tasks  to improve productivity and  6 increase cost savings,   we have seen in the past that over time, as technology evolves,   you end up being able to do harder things and you end up   being able to do them in a less costly way. Do you   expect that to be the case for AI? Daron Acemoglu: Absolutely. Absolutely, I expect  that, but I am less convinced  that we're going to get there  very quickly by just throwing  more GPU capacity. So in  particular, I think there is  one view -- and this is, again,  another open area. That's why  any estimate of what can be  achieved within any time  horizon is going to be very  uncertain. But there's a  view among some people in the 

industry that there's, like,  a scaling law. You double the  amount of data, you double the amount of compute  capacity, say, the number of  GPU units or their processing  power, and you're going to  double the capacities of the AI  model. Now, the difficulty there   is at least three-fold. One is what does it mean to double AI   capabilities? Because what we're talking about here -- for example,   open-ended things like customer service or   understanding and summarizing text -- there isn't a very clear metric that   I get twice as good. So that's one complication.  7 The other one   is "doubling data," what does that mean? If we throw in more data from reddit into the   next version of GPT, that will   be useful in improving prediction of the next word when you are engaged in some sort of informal  conversation, but it won't  necessarily make you much  better at helping customers  that are having problems with  their telephone or with their  video service. So you need 

higher and higher quality  data, and it's not clear where  that data is going to come  from. And wherever it comes,  it's going to be easily and  cheaply available to generative AI.  So the doubling data part is also not very clear. And then the final is I think there   is the possibility that there could be very severe   limits to where we can go with the current architecture. Human  

cognition doesn't just rely on a single mode;   it involves many different types of cognitive processes, different types   of sensory inputs, different types of   reasoning. So the current architecture of the large language models has proven to be more  impressive than many people  would have predicted, but I  think it still takes a big leap  of faith to say that, just on this  architecture of predicting the  next word, we're going to get  something that's as smart as Hal in 2001 Odyssey. 8  So those are all the sort of  uncertainties which, to me, says  if you're thinking about  the next few years, we know  already the models. So  anything that's invented or big  breakthroughs, it's not going  to have a huge effect within  the next few years. Allison Nathan: You've said many  

times it's really about the horizon.  Daron Acemoglu: Right. Allison Nathan: But ultimately,   there are people out there arguing that this   technology is paving the way for superintelligence that can really   accelerate innovation broadly in the   economy. Are you questioning that at all? Daron Acemoglu: Well, again, for the current paper  we're talking about, all I  need to say is that it's a time  horizon issue. So I don't  think anybody seriously is  arguing or can make a serious  argument that, within five to  ten years, we're going to have superintelligence. Now, going beyond the 10-year horizon,   I would also 9  question the premise that we  are on a path towards some  sort of superintelligence,  precisely because of the reasons  that I tried to articulate a  second ago, that I think this one  particular way of understanding and summarizing information is only a small part   of what human cognition does. It's going to be very  

difficult to imagine that a large language model is going to have   the kinds of capabilities to pose the questions for itself,   develop the solutions, then test those solutions, find new analogies,   and so on and so forth.  Like, for example, I am  completely open to the idea that,  within a 20-, 30-year horizon,  the process of science could  be revolutionized by AI tools.  But the way that I would see  that is that humans have to  be at the driving seat. They 

decide where there is great  social value for additional  information and how AI can  be used. Then AI provides  some input. Then humans have to come in and start bringing other types of information and   other real-world interactions for   testing those. And then once they are tested, some of those reliable models   have to be taken into other things like drugs   or new products and another round of testing and so on and so forth   have to be developed. 10 

So if you're really talking  about superintelligence then you  must think that all of these  different things can be done by  an AI model within a 20-,  30-year horizon. Again, I find  that not very likely. Allison Nathan: Your colleague David Autor and  coauthors have shown that  technological innovations tend  to drive the creation of new  occupations. Their stat is 60% 

of workers today are employed  in occupations that didn't  exist 80 years ago. How does  that dynamic factor into your  AI productivity and growth calculation? Daron Acemoglu: I think that dynamic is very  important if you are going to  think historically about say,  for example, how digital  technologies and the Internet have  impacted things. But that is  not a law of nature. So it's  about what types of technologies  we have invented and how  we have used them. So again, my hope is that,   in fact, we can use AI for creating these sorts of new   tasks, new occupations, new competencies, but there is no   guarantee. So we talked about the scientific   discovery process, for instance. So my scenario for doing that right would be  

exactly by creating 11  new tasks for scientists  rather than scientists using  intuition for coming up,  say, new materials which then  they test in various different  ways. If AI models can be  trained to do part of that  process then humans can then be  trained to become more  specialized and provide better  interactions, better inputs  into the AI models. And that's  the kind of thing that I  think would ultimately lead to  much better possibilities for human discovery. Allison Nathan: Right. Well,   the investment in AI is absolutely surging.  Daron Acemoglu: Yes. Allison Nathan: We have   equity analysts forecasting a trillion-dollar spend. Is that  

money going to go to waste? Daron Acemoglu: That's a great   question, and I don't know. So my paper and   basic economic analysis suggests that there should be investment   because many of the things that we are using   AI for is some sort of automation. That means that we are substituting   algorithms and capital for human labor,   and that should lead to an investment boom. And that's the reason why my numbers for GDP 

12 increases are twice   as large as the productivity increases because economic theory suggests   there should be an investment boom.  But then reality supervenes  on that and says, when you  have an investment boom, some  of it is going to get wasted  because some of it is going to  be driven by things that you  cannot do yet but people  attempted. Some of it may be  driven by hype. Some of it  may be driven by being too  optimistic about how quickly  you can integrate AI into your  existing organization. On  the other hand, some of it is  going to be super useful because  it's going to lay the seeds  of that next phase where much  better things can be done. 

So I think the devil is in the detail. So I don't have a very strong   prior [?] as to how much of that big investment boom is   going to get wasted and how much of it is going to   lay the seeds for something much better, but I expect both will happen.  Allison Nathan: I then spoke  to Jim Covello, the head of  Global Equity Research at  Goldman Sachs, and a longtime  watcher of technology trends  as a former well-known  semiconductor analyst. He  drew some pretty surprising 

13 insights from past tech   spending cycles when I asked him whether tech companies are   currently overspending on AI. Jim Covello:: The biggest challenge is that, over  the next several years alone,  we're going to spend over a  trillion dollars developing  AI, you know, around the  infrastructure, whether it's  the data center infrastructure,  whether it's utilities  infrastructure, whether it's the  applications. A trillion dollars.  And that is the issue in my  mind. What trillion-dollar  problem is AI going to solve?  This is different from every  other technology transition that  I've been a part of over the  last 30 years that I've closely  followed the technology industry. Historically, we've always had  

a very cheap solution replacing a very expensive   solution. Here, you have a very expensive solution that's meant   to replace low-cost labor. And that doesn't even make   any sense from the jump, right? And that's my   biggest concern on AI at this point. Allison Nathan: But isn't technology   always expensive in its nascent stage,   and then you improve, you evolve, you iterate, and the cost comes down dramatically?  14 Jim Covello:: Yeah, not always. Let's take  ecommerce and the Internet  as the best example of this.  From the get-go, right, you  had a very cheap technology,  ecommerce, replacing a very  expensive brick-and-mortar  retail solution. Amazon was  able to sell books from the 

first day that they started  selling books on the Internet  because it was cheaper to sell  over the Internet than it was  for Barnes & Noble to have  retail stores. That was cheaper  from the beginning. Like, so  there's a real-life example of  arguably the most important  technology development of  our generation, ecommerce.  That was cheaper from day  one. Fast forward 30 years,  

right? And it's still cheaper. We still have a cheaper solution   replacing a more expensive solution. Take, you know, Uber replacing limousine  services, right? So you started  cheaper, and 30 years later  the Internet is still enabling  things to be cheaper than what  the incumbent solution is. There's nothing about AI   that's cheap today, right? And you're starting from a very high   cost base. So that part I think there's a lot of   revisionist history on about how things always start expensive and get   cheaper. Nobody 15 

started with a trillion  dollars. And there's examples of  when there's a monopoly on the bottleneck of the technology, the technology costs don't   always come down. I'll give you an   example. You know, the main bottleneck in making a semiconductor is lithography,   and there's only one company in the world,   ASM Lithography, that can make advanced lithography   tools. Lithography systems, when I covered semis 20 years   ago, were in the tens of millions of dollars. Now,   a single lithography system can cost in the hundreds of millions   of dollars because there's only one person that can do it. 

And right now, Nvidia is the  only person that can provide  GPUs that power AI, and that's  why AI is so expensive. It's  really the GPU costs, the  number that you have to use in  order to run the data centers  and then how much the chips  cost. I think a big determination   in whether AI costs ever become affordable is going to be   whether there are other players that come in that can provide chips   alongside of Nvidia. If one wants to argue that   we're going to see costs come down significantly, it's going to be because   other providers like 16  Intel and AMD come alongside  Nvidia and are able to make  GPUs that can be used in  data centers and/or the data  center, the hyperscale providers  themselves like Google and  Microsoft and Amazon, are  going to make their own chips.  I think that's a big leap from  where we are today. There's  certain pockets of semiconductors  that those companies  can compete with Nvidia in,  but they haven't been able to  take over the dominant GPU  position that would enable a  more competitive cost  environment where Nvidia would  have to be more accommodative on pricing. And so I think that's a big question  

mark. I think there's a lot of complacency on the part   of the tech world that costs are going to come down,   and I don't think that's a foregone conclusion. Even it does come down,   the starting point of how expensive this technology   is means costs have to come down an unbelievable   amount to get to the point where this is actually affordable to automate some of these  technologies. Allison Nathan: Right.   And ultimately, you don't really have a lot of expectation that it   will be able to perform in terms of cognitive ability   close to humans. Do you see real limits to the technology relative to   the promise that some 17  people are purporting? Jim Covello:: Many people want to say this is the  biggest technology invention  of their lifetime. I just think  that's, to me, almost a silly  starting point for all of this,  right? How can someone say  this is bigger than when we  first put a cell phone in  someone's hand or when we first  put the Internet in front of  someone or, frankly, when we  first put a laptop in front  of someone, right? Like, those  were transformative technologies  that were fundamentally  enabling you to do something  different that you had ever  done before. You couldn't make a phone call from 

wherever you were. You  couldn't compute from wherever  you were. And, you know,  relative to the Internet, you  could buy something over the  Internet that you used to  have to go to a brick-and-mortar store for. Allison Nathan: Got it. But   when we first came up with cell phones, I don't think anyone understood how  transformative it could be.  So why are you so confident 

that this won't be just as or more transformative? Jim Covello:: I think that's revisionist history,  too. Like, I covered  semiconductors when the smartphone  18 was invented, and I sat through hundreds of  semiconductor presentations  where they showed the road  map right away, like, from  day one of the smartphone.  "Here's everything that this is  eventually going to be able to  do." And it was out in the  future that we were going to be 

able to do them, but we had  identified the things that we  were going to be able to do.  Like, immediately upon the  advent of the smartphone, people  said, "Well, we're going to  have our GPS in the smartphone,"  right? Because at the  time, you would have your Hertz  rental cars that had those  old clunky GPS systems, and  they would show that and  they would show your iPhone  and they would say, "Here's  the road map of what this is  eventually going to be able to  do." Same with health   applications. Same with Internet. Same with a lot of these things. But AI   is pie-in-the-sky, big picture, "if you build it,   they will come," just you got to trust this because technology   always evolves. And we're a couple years into this,   and there's not a single thing that this is being used for that's cost   effective at this point. I think there's an   unbelievable misunderstanding of what the technology can do today. The problems  

that it can 19  solve aren't big problems.  There is no cognitive reasoning  in this. Like, people act like  if we just tweak it a little bit,  it's somehow going to -- we're  not even in the same zip code  of where this needs to be. Allison Nathan: Even if  

the benefits and maybe even the returns don't justify the costs,   do the big tech companies that are spending this   money today have any choice but to engage in the AI arms race, given the competitive  pressures? Jim Covello:: Yeah, great question. I really think  that's an important one. And  I think the answer right now  is, no, they don't have a  choice, right? Which is why we're  going to see the build-out  continue for now. That's sort of  what the technology industry  does. Like, look at virtual 

reality. This would not be the  first technology that didn't  meet the hype, right? That's  the other part of historical  context that I think is so important. People act like those of us who think   this might not be as big as some other people   think it is or naysayers on technology. No. There's  

just historical context. There was a period where nobody was ever going   to go need to see a 20  house again in person if they  were buying it because they  were going to use virtual  reality glasses to look at that.  Blockchain was supposed to be a big technology. A metaverse was supposed to be -- all  

the money that got spent on metaverse. Those   things are nonexistent today from a technology use   case standpoint. And just because the tech industry hypes something up   doesn't really mean a lot.  But to your exact point, we're  going to keep building this  for the time being. In the  eyes of the tech industry and,  frankly, in the eyes of a lot  of enterprises, if it does work  and they haven't positioned  themselves for it, they're going  to be way behind. So there's  a huge FOMO element to this  which is powering all the hype.  And I don't think the hype 

is really going to end anytime soon. Allison Nathan: Right. And then   companies outside the tech sector have also   started spending a lot of money on AI capabilities. Well, what do the early   results from that investment show?  Jim Covello:: I think that almost universally it's showing that there's not a lot that AI can do   today. And 21 

again, there are people on  different parts of the spectrum of  anywhere from, "We shouldn't  expect it to do anything  today; it's such early stages.  And technology evolves and  finds a way" to the other side  of the spectrum is, "We're  several years into this and by  this time it was supposed to  be doing something." And  everybody's on a different part of  that continuum. Right now, there's very  

limited applications for how this can be used effectively. Very   few companies are actually saving any money at all doing   this. And that's where I think you get into the   how long do we have to go before people start to really question?  Allison Nathan: So what does  all that we've discussed  mean for investors that are  focused on AI over the near,  medium, and long term? Jim Covello:: Yeah. I think it's all on the  infrastructure side. It gets  back to the point that we were  just talking about where we're  still going to keep building  AI. I don't think we're  anywhere near done building it,  right? The world's so convinced  that this is going to be  something significant that  there's nobody that's even  22 remotely close,   in my opinion, to scaling back on the build. And so what I've been saying for  

two years is what I continue to say. Keep   buying the infrastructure providers. Is it really expensive? Absolutely.   But I've never seen a that stock goes down only   because it's expensive. It got expensive for a reason. People  

believe in the fundamental growth outlook. If the stock   collapses, it's going to be because there's a problem   with the fundamental growth, not because of the valuation.  Allison Nathan: And ultimately,  if you are right, we are  building all this infrastructure  and capacity that at some  point won't really be in demand. Jim Covello:: Right. The very nuanced view,  right? Allison Nathan: What's that look like?  Jim Covello:: Yeah, it looks bad. It looks exactly like 2001, '02, and '03 for the Internet   build-out. Like people, again,   it's a relevant discussion, right, when they want to talk about that "if you build it,   they will come," and when we built the Internet and   then 30 years later we 23  developed Uber and all those  things are true, right? It ends  badly when you build things  that the world's not ready for,  right? And I don't know that  it's as problematic simply  because a lot of the companies  spending money today are  better capitalized than some  of the companies that were  spending money then. But  when you wind up with a whole 

bunch of excess capacity  because you built something that  isn't going to get utilized, it  takes a while. The world has to  then grow back into that supply-demand balance.  So it ends badly if we're right  that this isn't going to have  the adoption that everybody  thinks. But I would say one of  the biggest lessons I've  learned over 25 years here is  bubbles take a long time to  burst. So the build of this  could go on a long time before we see any kind of manifestation of the problem. 

Allison Nathan: What should  investors be focused on to  see a changing of the tea  leaves here that you expect to  come eventually? Jim Covello:: I think it'll be fascinating to see  how long people can go with  the "if you build it, they will  come" approach, right? At some  point in the next 12 to 18  24 months, you would think there has to be a bunch of  applications that show up  that people can see and touch  and feel that they feel like,  "Okay, I get it now; here's how  we're going to use AI." Because again, investors   are trying to use this in their everyday life, and there was   a period a year ago where everybody was pretty excited   about how asset managers could utilize AI. And I   think if you interviewed an asset manager for this, most of them are   going to tell you the same thing, which is,   "We're struggling on how to figure out how to use it. We can't really find applications   that make a ton of sense." 

Again, there's isolated examples,  models, and things in that  nature but nothing significant.  And so I think the longer it  goes without any applications  that are obvious to people or  significant applications that  are obvious to people, the  more challenge. The bigger  thing to me is the corporate  profit issue, right? That's  what I would watch if I were  investors -- corporate profits.  And as long as corporate  profits are great, companies have money to try experiments. But negative  

ROI experiments are the first things to go when corporate profits slow down,   so that's 25  what I would really have my eye on. Allison Nathan: So some   really provocative thoughts from Jim. I should mention that I also spoke to my  Goldman Sachs Research  colleagues Kash Rangan and Eric  Sheridan, who see it very  differently. Eric, our US Internet  equity research analyst, says  that current CapEx spend as  a share of revenues doesn't  look markedly different from  prior tech investment cycles.  And Kash adds that the  potential for returns from  this CapEx cycle 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.  But Eric does warn that, if  AI's killer application fails to  emerge in the next 6 to 18  months, he'll become more  concerned about the ultimate  payoff of all the investment  we're currently seeing.  We'll leave it there for now. 

Thank you for listening to  this episode of Goldman Sachs  Exchanges. I'm Allison Nathan.  If you enjoyed this show,  we hope you follow us wherever  you listen to your podcasts  and leave us a rating and  comment. If you'd like to learn  more, visit GS.com where you  can find a copy of this report 

26 and also   sign up for Briefings, a weekly newsletter from Goldman Sachs about trends spanning markets,  industries, and the global economy.

2024-07-20

Show video