AI and the Economy: Will Artificial Intelligence Drive the Next Productivity Boom?

Tony Roth, Chief Investment Officer

Martin Beraja, Economist & Professor, Haas School of Business | UC Berkeley 

Tony Roth: Welcome to Capital Considerations. I'm Tony Roth, Chief Investment Officer for Wilmington Trust and M&T Bank. Artificial intelligence has quickly moved from a technology discussion to a capital allocation question. It's shaping corporate investment, market leadership, productivity expectations, and how investors think about future earnings growth for the entire market.

For investors, the key questions are how much value AI can ultimately create, who captures that value, how quickly it shows up, and whether market expectations have moved ahead of the underlying economics.

Today, I'm joined by Martin Beraja, an economist and professor at the Haas School of Business at UC Berkeley. His research focuses on technology innovation, business cycles, and the economic implications of AI digital technologies. He combines theory and data to examine how firms and institutions adapt to new technologies, including recent work on AI-driven innovation, automation, and the pace of technological change.

Martin earned his PhD from the University of Chicago and has held faculty positions at Princeton and MIT. He's also the recipient of the National Science Foundation's Career Award. So we're incredibly fortunate and excited to have you here today, Martin, to talk to us. And I hope you don't mind if I call you Martin, rather than Professor.

Martin Beraja: No, that's okay. Hi, Tony. I'm, uh, glad to be here.

Can AI Deliver Enough Productivity?

Tony Roth: So when I think about AI, Martin, from an investor's perspective, I think about two questions. Let me lay out how I think about this, and I'd love to try to talk about both of these questions to some degree.

We'll probably start with the big one and run out of time before we spend a lot of time on the second question. But the big question that I have is whether or not AI will deliver the right amount of productivity enhancement to the economy, whether we're thinking about the U.S. economy or the global economy.

When we do our Capital Market Forecast and we think about all the various dynamics and, and risks to the economy, there are a lot of respects in which increasing productivity becomes, as it always is in most investment cycles, but particularly right now in our view, a potential escape hatch, if you will, for a lot of the problems that we have in the economy.

So, for example, we have a massive amount of debt to GDP here in the U.S. We have potentially a shrinking labor force with some of the new policies from immigration and the natural demographics. And having a significant increase in productivity could help us grow our way out of the national debt, for example, as well as to increase the output per person, could really help with the decline in the labor force, or at least the slowdown in the growth of the labor force, that we're experiencing. And so can AI deliver enough productivity? But by the same token, there are also these, if you will, dystopian narratives out there where AI delivers so much productivity that we don't need as many workers.

And we all have people that we know that are young people. I have two daughters that are either going into college or, in college, and we worry about these young people finding jobs, as the labor force, particularly for those less skilled people, coming out of school get, gets hollowed out.

The question is, you know, can too much productivity destroy the labor force and ultimately destroy consumption and cause problems in that way? And so the big question is, can AI allow productivity to hit the sweet spot enough to really be strong and productive, but not enough to be ultimately deeply problematic?

The second question, which is I call the small question from an investing standpoint, is there's been a tremendous amount of money spent by a small number of companies on the infrastructure to deliver AI. And the market has raced forward based on early signs of significant profit growth, on the part of these companies, Mag Seven companies primarily.

And the question is, are the investments going to really pay off , and will the return on investment be there to support, the valuations, and the market caps that we're in the process of achieving on these big companies? And If for whatever reason, either because AI is not successful enough or because AI becomes commoditized too quickly, or the Chinese eat our lunch in the AI space, if you will, these companies are way ahead of themselves, then that could be a real problem for the stock market and, and in and of itself be the trigger for a recession.

So a critical question in and of itself, but conceptually a question that is secondary to that first question. Can AI even be successful? So we're not expecting you to answer these questions for us today, the hope is that you could give us your thoughts around how should we even approach these questions.

With that, I'm going to turn it over to you and invite you to maybe tell us how you'd even approach thinking about that first question.

A Technology of Learning

Martin Beraja: There's a lot to unpack there, so let me start maybe with the question of, productivity and where do I see the biggest gains from AI coming from.

I think there's been a lot of focus on AI as being an automation technology, sort of like, a fancy robot that instead of automating the tasks that robots used to automate, which were more manual tasks, people have now figured out that, okay, well, AI is sort of a robot, but that can automate the tasks that workers involved in cognitive tasks perform like coding and design and more creative tasks.

And I think that view to some extent has, people have tried to put some numbers on that and the numbers that come out of that say that if it's just an automation technology, then this is going to take a long time to take off. There's bottlenecks everywhere and ultimately it may be somewhat disappointing in terms of just delivering productivity There's a second view that is much more optimistic that has to do with the ability of AI to accelerate science and innovation.

And I know we've given actually, you know, the last Nobel Prize in physics went to David Thouless because of, how AI, uh, will, improve and accelerate science. And so in that view, you get into a terrain where it's sort of science fiction. You can think that anything goes, that we get to this, you know, utopia of infinite productivity where machines improve themselves and we get more products, more innovation, and then those products themselves, those new machines improve themselves, and so on and so forth, and we get to a place that is just utopic.

So you know, if you take those two views, you're left with sort of like, yeah, but what's the truth? Either this is just an automation technology or this is really a technology that will accelerate innovation and anything goes in between. So the span of things that can happen is just huge. That's, I think, where a lot of the uncertainty that you're describing around this technology is really coming from.

And I could sit here and speculate about those things too, but I think something that may be more valuable to your audience would be to kind of bring a different view, a third view of what I think this technology is about. And the advantage of this view is something that we can at least put some numbers on, that we can quantify based on some evidence that we already have.

And this view is that the technology at its core, AI, it's a technology of learning. And there's plenty of evidence that the way that people are using this technology, both at home but also at work, is to learn stuff. You ask, you know, the chatbot questions and how you are doing duties or tell me about that and people use it that way to learn.

And so when you take that view you start thinking, well, where is it that learning is really important? And one place where learning is really important is within organizations, firms. In two ways. So first, as organizations grow, they kind of learn a couple of things that take many, many decades to figure out what is the right organization for this company, what is the right product that we need to, focus on and launch, and how is it that we develop.

And all of this takes years of trial and error and experimentation and that's why we see these kind of like steep learning curves in a lot of industries where young firms are two to three times smaller than more mature established firms.

That's one way in which learning is important. And the second way in which learning is often important in industry is sort of at the beginning stage of capital allocation. Banks and, you know, venture capitalists, they really need to learn what are the fundamentals of these companies

And, uh, that learning process in understanding is this a product that will have a market, is this a company that will have a market, and are these the right managers for this company? So that type of predictions is that where we're trying to figure out whether this firm is going to be successful or not, or whether it's just going to exit pretty fast is another form of learning that is being done.

It's Not Just Faster Outcomes, But Better Outcomes

Tony Roth: Can I just summarize what I think I'm hearing, Martin. So it sounds like m- maybe I can say it this way. Tell me if this resonates with you. It's not that the real crux of AI is that it helps us do things more efficiently or faster, although it does, but it may help us do things better.

So in other words, it may help workers learn so they can do their jobs better, so they can deliver better outcomes. Not just faster outcomes, but better outcomes across the economy. And one example might be capital allocators can actually use AI to evaluate which of the multitude of new companies that are created within the AI space, which of those will be successful or not successful.

They can actually use AI to help understand where to allocate that capital by using AI itself to understand which companies are going to be successful.

Martin Beraja: Correct. And in fact, what you just said, there's evidence of that, that when you pair venture capital analysts with AI, they perform better than the typical venture capital analysts without.

Tony Roth: Whether it's in the tech space or whether it could be in any other space actually, for that matter.

Martin Beraja: Correct. So then when, once you start taking this view, then I think this does, well, if AI can indeed accelerate this sort of, uh, organizational learning, both at the capital allocation stage and later as these companies accumulate what we call organizational capital economics, all these things that make firms internally more productive.

Then you start thinking, okay, well, if we can flatten these curves, if we get these learning curves somehow with AI, then that can deliver big gains think about it is like imagine if you start firms with 30 years of experience, and that that whole learning, you know, 30 years of learning could be compressed in one or two years because of the feedback loop that you could get with AI being much accelerated.

If that is true then I can see, you know, AI delivering very big gains. So how big?

Tony Roth: Before you answer how big, let me just try to mediate another example here. Just because I think it's helpful to make sure bring everybody along.

We provide some concrete examples. Another example could be Wilmington Trust and M&T are a bank.

Martin Beraja: Mm-hmm.

Tony Roth: And one of our foundational tasks is to extend credit. And so one of the big functions that we provide to our stockholders, if you will, is we evaluate the credit worthiness of various players in the economy, whether individuals or businesses that want to borrow money from us.

Martin Beraja: Mm-hmm.

Tony Roth: And ultimately we have some kind of success rate with that. I think at M&T we tend to be on the more conservative side of the continuum. We have more opportunities to lend that we turn away than many other banks would, for example, because we like to run a very tight ship and not have too many loan losses.

Now what could AI do to help M&T? Well, number one, AI could probably help our credit team work faster and underwrite those loans more quickly, and maybe we could save some credit analysts because the workload of anyone, given one could increase if they're using…

Martin Beraja: AI And that would be the fancy robot view of, uh AI,

Tony Roth: Right.

Martin Beraja: Yeah.

Tony Roth: But more fundamentally, AI could actually help us better predict which of these potential suitors would pay back their loans, which would not. And i.e., better outcomes, right? So the organizational learning that we've always been engaged in to try to be better at underwriting credit could be accelerated to the point where we could do a better job picking who to lend to, and we could be more confident lending to some of these people we don't lend today because we're not so sure whether they're going to pay the loan back, and the AI could come in and say, "Yeah, that is actually going to be a situation where you probably shouldn't worry too much because you're going to get paid back given what the learning shows."

Does that... Is that a good example?

Martin Beraja: Yeah. And even beyond that, so part of what, uh, I imagine you do is curating products, loan products, different products. And that takes time in that, you know, you need to design the product, then you need to launch the product, then you figure out whether is this product working or not, you get some feedback. You collect some data, you go and iterate and improve the product based on the feedback that you got. And that whole process may take a year. Eventually you say, "This was a bad loan product," or, "This was a good loan product," and you keep it. So that whole trial and error experimentation process that all firms engage in to some extent or another, it is something that that AI can accelerate. And there's evidence of that already in different industries. Actually, the finance industry is one.

I think that's my take on this, that like to view AI not just automation, not just as an innovation technology, but really as a technology that can help firms learn what they were already learning, but faster.

How Much Productivity Could AI Deliver?

Tony Roth: So let's talk about the magnitude of productivity increase, right?

Martin Beraja: Yeah.

Tony Roth: That's how I, how I sort of frame things, and you're answering the question by saying we should focus on determining whether we're going to get enough productivity gains, how the companies in the economy are using AI to accelerate or- organizational learning. So if you think about that, how do we answer that question?

How much increase in productivity will this deliver?

Martin Beraja: Right. So the ideal experiment would be one where you go to all the firms, you give this technology, and then you track them 30 years later.

Tony Roth: Right.

Martin Beraja: But then we're 30 years later, and we want to know the answer now.

Tony Roth: Right.

Martin Beraja: So the problem that we're facing with AI is that to get these type of answers in real time, you cannot bank on small experiments in some firms, which is what a lot of people are doing because it's hard to go from those experiments to what would happen in the whole economy.

So the approach that we've taken in, my work was to say, well, we know something already about how important learning is from the ways that we see firms already learning in the economy. And one telltale sign that, you know, where is it that learning is, is more important are two features of, say, particular industries.

One is particular industries where you see the firms that are young take a long time to grow into their mature routine.

Tony Roth: Right.

Martin Beraja: And that's sort of the telltale sign of, of an industry where, you know, this learning, uh, within, uh, the organization takes time.

The other side is when you see that firms enter and they kind of exit right away while young. So imagine a firm that enters and then it stays for one or two years and then they exit, as opposed to an industry where kind of the firms that enter, you know, they don't tend to exit at much different rates than other firms. And the reason for that is a telltale sign that learning is really important is because if you see an industry where there's a lot of exit fast at the beginning of the firm's life cycle, is because you didn't do a good job at predicting whether that firm was going to be successful or not. And that selection stage when you see fi- firms fail so fast, it means that you if you could have initially predicted that this firm was going to do badly early on, then you would have allocated that capital to a different firm.

Tony Roth: Right.

Could AI Double GDP?

Martin Beraja: And these are two options that we can measure in the US and in many other countries, and in particular across industries to see how important learning is and what are the potential gains from AI. And what we come up with is a number that says, look, given how slow firms are learning in, in the sense how long it takes them to, to grow to their mature level of, of productivity.

Tony Roth: Mm-hmm.

Martin Beraja: And given how much exit we're seeing of young firms in the, in the US economy, a really good version of AI that could accelerate learning a lot, say, by thirty years, could double GDP. Of course, if you get a technology, AI is not as good to accelerate things by ten years in terms of accelerated learning, then you get, you know, a fifty percent increase in GDP instead of, you know, a twofold increase.

Tony Roth: So, so over what period of time are we talking about? So let me just-

Martin Beraja: We don't know.

Tony Roth: So, so we don't know.

Martin Beraja: No, that's the thing. These are exercises that think about the, the technology is sort of here, and we all adopt it, and we just, it just drops from the sky. So this may take thirty years. I, I can't really tell, uh, when it will happen.

Uh, but I can tell you that if it happens, this is the type of, magnitudes that we, that are involved. And, and I would say they are big deal.

Tony Roth: Cause this is all incremental to what's the underlying normal productivity growth that we always get in the economy through…

Martin Beraja: Correct.

Tony Roth: …efficiency and such, right? I mean, most productivity growth is through efficiency, not, probably not innovation. But now we're in a world where this accelerated learning is not about efficiency, it's about just delivering better results, period. And what you're saying is that the best way to understand it is to look at how new companies are faring.

Martin Beraja: Right.

Tony Roth: Are you saying that if the failure rate is lower, or the failure rate may not be lower, but it happens faster?

That's a good thing because it means we're not wasting capital in the economy on approaches that are not going to work.

Martin Beraja: Correct. And, uh, and we could reallocate that capital, and that, that turns out could have big effects in terms of, uh, aggregate productivity.

Tony Roth: And are we actually seeing that? Are we seeing in the US that there's a, a faster failure rate or a, a lower failure rate, or we can't tell, we don't know yet?

Martin Beraja: We don't know yet from AI. What I can tell you that if you compare that to some countries in Europe, the failure rate of young firms in Europe compared to old firms is smaller than in, in the US.

What that is telling is that the technology may have a bigger impact in the US than Europe precisely because that's where we're seeing that the capital allocation could be improved more by better predicting which firms are going to fail.

And that may be just because the European, the people that are allocating capital in Europe are already not that very good at predicting who's going to fail, but they have such a high threshold for, for funding a firm that they're really getting the best firms already.

There's not much improvement left there in that sense, unless they go down the list of firms that they could fund.

Tony Roth: So what can you tell us if you put all this together?

Martin Beraja: Yeah.

Tony Roth: And we need to get over the next thirty years, let's just say, just a half a percent increase in productivity from, you know, the Congressional Budget Office, right? Bipartisan economic apparatus in, in Washington predicts about point nine percent productivity growth without taking into account AI over the next several decades. If we grew that by only fifty basis points more, that would be enough to just transform the whole economy.

So from what everything we've talked about, how can we understand whether or not we're likely or not likely to get that extra fifty basis points of productivity growth per year, over the next several decades?

Are We Thinking About AI the Wrong Way?

Martin Beraja: I think for that we need much more evidence on what AI will actually end up doing in terms of accelerated learning, and we need a couple more years to be able to tell. What I can tell you is this is about the potential of the technology and put some numbers on that.

If, if that potential is realized, we're talking about a different beast. I think the robot view of this technology is the too narrow view. That analogy is, is frail. It, this technology is going to do something else to organizations that robots didn't do, and it's not just automating labor.

For investors, I think the way that people are thinking about which are the industries or activities that are going to be impacted the most is looking at this exposure metrics that have been put on the table.

And those metrics that they capture is whether a given activity or given a- activities in a given industry, the tasks performed by the workers in that industry can feasibly be done, performed by AI. And so you have things like coding and so saying software are things that AI can do well, and things that are very manual, uh, where AI will right now cannot do much.

I think this view of AI as accelerated organizational learning, what it tells you is that there are some industries, not all industries where AI is feasible will deliver much value.

Tony Roth: Mmm-hmm.

Martin Beraja: So you shouldn't expect- much adoption and impact there. And there are some industries where AI looks hard to apply to based on those metrics, but actually they have huge value that you could apply that.

What Industries Could Benefit Most from AI?

Tony Roth: I got so excited that I I thought of my light bulb went off in my head, and I, and I thought about the construction industry.

Martin Beraja: Yeah.

Tony Roth: Because I would've thought inherently that that would be a horrible place to try to apply AI because in order to build a building, you need somebody to hold a hammer and, and hit the nails, right?

And so how are you going to have AI come in a computer can't hold a hammer, right? That's just, that's gotta be just n- knocked out. But- There's probably all kinds of ways that AI can tell you how to stage the construction of a building so that it's much more efficient and you can actually build it with less man-hours, I would imagine.

And maybe that is an interesting use case that you were just alluding to that you wouldn't expect it, but it can actually do a lot. I don't know.

Martin Beraja: Absolutely. That's one example of construction. Another one is, that is not obvious, anything that has to do with, like, food manufacturing, beverage and tobacco, these type of industries are where if you look at these exposure metrics, they tell you that, "Yeah, this is not an industry that's really, really going to be disrupted by AI because it's not a place where you can automate the tasks that involve production.

Tony Roth: Okay.

Martin Beraja: But it is an industry where you may be able to improve a lot on the tasks that involve learning, and that's what we're finding, in fact.

And that learning, for example, has to do with product development. These are industries…

Tony Roth: Right

Martin Beraja: ... you know, food manufacturing is an industry that is built on constant introduction of new products, experimentation, and seeing what works and what doesn't. And that's precisely the type of task that AI can really accelerate and help with.

Tony Roth: We need a couple of years to get the evidence.

Martin Beraja: Yeah.

Tony Roth: Do you have any gut instinct at this point? Like, are you optimistic? You're concerned we're not going to see enough success in this learning arena to underwrite the long-term productivity gains that I'm describing that we need?

Do you have any kind of gut instinct at all?

Martin Beraja: I'm fairly optimistic. Time frames are hard to, I can speculate as, as well as the next guy on that, but, uh, I'm fairly optimistic in that I think once companies start a bit reorganizing around the technology and not just using it, it, to do the exact same things that…

Tony Roth: Right.

Martin Beraja: …they were doing before, I think that's when really we're going to see the gains come in. And what is interesting about that is actually that you may see an initial period that looks really disappointing in that the firms are just sort of automating work, maybe displacing workers, and you don't see much productivity increases.

Followed by another period where they reorganize and they start using this technology in a way that it will really deliver the productivity gains that, that we're talking about, where this is not about displacing workers and automating work, it's really about helping them learn, faster about how to, uh, be-

Tony Roth: Better outcomes.

Martin Beraja: Yeah.

Could AI Lead to Widespread Job Displacement?

Tony Roth: Yeah. And, uh, okay, so let's look at the other side of the spectrum for a second. The other side is this kind of dystopian picture where you can't get a job because the computer's doing it for you. Even if it were to eliminate 20% of the roles in the economy, that would be enough to have a cataclysmic impact on consumption.

So how concerned are you about that, either in the short term or the long term?

Martin Beraja: Let me say a couple of things on this. So I have worked thinking about why we would like to slow down this type of automation technologies when they displace workers and, uh, you don't have other good tools like, you know, providing social insurance transfers, to ensure the workers are displaced by automation.

And, and what we've shown there is that indeed you do have, uh, an economic rationale for maybe slowing down these technologies for maybe five, 10, 15 years to allow the workers that cannot really adjust their skills. Think about middle, mid-career workers that may be displaced by this technology and that are incapable to really switch career, retrain or whatever.

But they're not close enough to retirement that then we just say, "Well, you know, this technology came in, it automated my job, I'm just going to, uh, retire a couple of years earlier." Those are the type of workers that are really at risk and, and may suffer the most from this technology. And based on that, if those workers cannot insure themselves by borrowing or their own, their own savings, then that's when you have a problem and there's a, a rationale for maybe slowing down things a bit to allow these workers to e-eventually retire and be replaced by younger workers that have the right skills to work with, with this. That said, I think what you're alluding to with the dystopia view is more of a very, very short-run view that I would associate with that piece by Citrini that became…

Tony Roth: Yes.

Martin Beraja: ... a bit popular in, in, in finance circles.

Tony Roth: Yeah, that's a big market sell-off, right?

Martin Beraja: Yes.

Tony Roth: You know, earlier in the year.

Martin Beraja: I have a very strong view about this piece, and let me just put it out there. I think it's entirely wrong. The economic logic is entirely wrong, and it rests on the following assumption. Okay? So let me just review the argument, and then I'll tell you why I think it, it's wrong.

The argument goes something like AI is going to come and displace a lot of workers at the same time really fast. Those workers will stop spending. The firms that rely on those workers' spending will cut workers themselves, and you will get this sort of like multiplier effect that will lead into a recession.

So it's a very Keynesian from Keynes logic that we apply to typical recessions where you have a big shock and many workers lose their jobs to the case of automation. Okay?

Tony Roth: Yep.

Martin Beraja: And that's the argument. What does that argument rely on when, when it comes to normal recession?

Let's say that we have, you know, COVID or, or we have, you know, the banking sector collapsing and the housing sector collapsing like in, in 2008. What the argument relies on is that the workers that lose their jobs or become scared and cut their spending and then that triggers a, a recession through this, uh, sort of, uh, multiplier logic.

That relies on the fact that the prices in the economy, wages, prices, are somewhat rigid and don't adjust fast enough. Because imagine if you could say, look, the firms could just reduce wages to keep the workers in, then they wouldn't fire any workers and none of this would happen. If the firms would say, "Oh yeah, we're seeing lower demand from all the workers that are now unemployed," then we can just cut prices and that would stabilize the economy, and you don't get this multiplier logic, and you don't fall into a, a huge demand-driven recession.

Tony Roth: In other words, it's not a dynamic analysis in the way you're describing. It's a very static analysis.

Martin Beraja: No, it is dynamic, but it's wrong in that I think it misses the dynamics that AI or automation in general is not something that looks like a shock, like COVID does or the banking sector collapses.

Tony Roth: Right.

Martin Beraja: It's something that will happen much more slowly, so that yes, maybe we'll leave some workers unemployed, but this will happen over, in the span of decades. It's not like in one year we're going to have 10% unemployment. Just the adoption of the technology is not as fast. And what that means is that then the prices and the wages will have time to adjust.

They don't look rigid the way that they do look in a recession. So you don't get this sort of demand-driven multiplier that you get when you have a shock that concentrates a lot of displacement of workers in a short period of time. I think that's what, like, that whole thing misses. That, that's just wrong.

Tony Roth: Okay. Good. So, so let's go to the other question then, and by the way, I think that what's important at this stage is not to answer these questions, but is to frame them, as we talked about in the beginning, so we know what to look for. And I think that we've learned here that understanding the ability of companies to adopt organizational learning and, and deliver better outcomes, not just outcomes that are more efficient, is how we need to understand the success of AI.

To me, that's the takeaway of this first part of the conversation. And one way to do that is to look at the success rate of small companies.

Martin Beraja: And the learning curves of, of firms to see where, where this can deliver the most, the most gains. Yeah.

Will AI Infrastructure Investments Pay Off?

Tony Roth: Yeah. Okay. So the second question is from an investing standpoint, the market's been very strongly supported by the earnings growth of these companies that are deploying massive amounts of capital, in many cases borrowing the capital.

Companies like Meta that don't have the cash flow that a company like, like Alphabet has, have to go out and borrow some of the money to keep up. And they're trying to become the providers of the AI because as software companies, they monetize this trend.

Will these big investments actually pay off for them? Well, that's going to depend on the answer to the first question. Are If we end up on option two where we have lots of productivity, economy does really well, then there's the opportunity. Doesn't mean that they're going to succeed with these investments, but there's the opportunity, um, to, to pay off.

Will AI Become a Commodity?

And one of the concerns that has come up very recently is not that the AI necessarily won't be successful in efficiency, in organizational learning, et cetera, that it won't add to the productivity, but those things will happen, but that the AI itself will become a commodity.

It will become something that in the vast majority of applications the underlying large language model and the application to the data, et cetera, can be provided by a fancy company like Meta, m- you know, in the Silicon Valley, or it could be provided by a Chinese company pretty much in the same way, such that the amount that companies need to pay for the compute and for the access of, of the compute to the data, et cetera, is going to be much lower than people expect.

And you're already seeing some of that potentially. And, and Meta, for example, is having to start to license some of its capacity to third parties because it's more than they can actually sell themselves. Do you have any w- reaction to this whole se- set of questions…

Martin Beraja: Okay.

Tony Roth: …around this, the so-called small question, if you will, which is how will this specific set of, companies do or do you have any idea at this point?

Martin Beraja: Yeah. So I think the first part relates to are we seeing this investment boom? Uh, is it sort of a bubble or does it have a bubble component or, or not? I really can't say whether it's a bubble or not. Again, it depends on what your view is on the productivity gains.

What I can say, though, is that some of the dynamics that we're seeing in terms of, uh, investment are quite similar to what we in economics we call it patent races. And we call it patent races because the o- original models of this phenomenon had to do with patents. So you think about there's a price.

There's many companies trying to develop a, a drug, let's say If you're the first to develop the drug, then you get a patent for that drug. . So early on, it looked like there's a lot of duplication of investments. So there's multiple companies investing on the same thing.

The one that wins gets the patent.

Tony Roth: The most recent example would be the GLP-1s, right? Everyone wants to get the best GLP-1.

Martin Beraja:Yes. So now when you look at a bit of the dynamics of investment in data centers and infrastructure around AI, I think it has a bit of that feature that there's a lot of companies that are investing in this sort of infrastructure because they think that in the event that we become the company that leads this market- then we'll need all this infrastructure.

Tony Roth: Right.

Martin Beraja: And we need to build it now. We can't wait until then to serve the market. Except it looks like there's a lot of duplication and there's going to be a lot of wasted capacity exposed to serve the market. But it doesn't mean that there's anything inefficient about it.

It really has to do with the dynamics of this industry where you do need multiple pilot investments that eventually will lead to maybe one or two winner, three in the industry as it concentrates. And that's when you will see, yes, uh, probably an investment collapse.

Not necessarily because there have been disappointing news about the technology, simply because it becomes clear that it's not going to be 20 firms that are going to be in this market. It's going to be one, two, or three, three firms.

Tony Roth: So the so-called collapse is not the entire market, it's just the cer- firms that aren't, are not the winners.

Martin Beraja: That are shaken out, yes. And you've seen this phenomena of a shakeout of firms over and over again in, in different industries. The auto industry used to have hundreds of small manufacturers early in the century, and then eventually concentrated in four or five manufacturers, and all the other ones were shaken out.

Tony Roth: Right.

Martin Beraja: But if you look at the early dynamics of that industry, you see a lot of early investments, by all these hundreds of manufacturers.

Tony Roth: Right.

Martin Beraja: And I think the AI industry will just look, uh, like that again.

Tony Roth: So if I think about largely speaking, you have four major companies that are in the game.

You've got Alphabet, Amazon, Microsoft, and Meta. And then they actually in turn are either, in the case of Google, they can create their own model, and Gemini, to a large degree, these are just using OpenAI and Anthropic. They're using third-party models to plug in.

Martin Beraja: Mm-hmm.

Tony Roth: So all four of them probably won't be massively successful in this space.

You might have one or two of them maybe, three possibly, but certainly not four. And so the market will do fine but you'll have to take into account the fact that y- you'll have some spectacular success on the part of one t- one or two or maybe three, but you'll also have some pretty big price declines in one or two of them, and they'll also sort of offset each other, and you'll have to see how that plays out.

Martin Beraja: Yeah. Right. Exactly, yeah.

Tony Roth: Is this something that's imminent, or you think it's going to be a another couple years at least before it's obvious which of these companies are going to be the providers of the tools of choice?

Martin Beraja: I don't think it's imminent, no. I think we're still learning exactly what is the right version of these AIs. What makes it say that it's not imminent is that As new models come out, like I was just trying this morning Fable that I haven't tried before, and I've been blown away by it.

Fable is the latest Anthropic model.

Tony Roth: It's even better than the last one, I guess.

Martin Beraja: Yes, by significant margins. I saw a, kind of a discrete jump in the capabilities of that model compared to what I've been using before.

And that tells me that, again, we haven't really converged yet in any way

Tony Roth: There's no plateauing yet. There's still-

Martin Beraja:Not yet. I'm not seeing that at all. The second thing, you asked a question about where is, is this going to be a commodity or where are these firms are going to get their markup from?

How are they going to differentiate? I think one example of that in the past where people were making similar arguments was the Windows package. You could now get that type of software. You know, Apple has Numbers and…

Tony Roth: Yeah.

Martin Beraja:... Pages that looks pretty much like Word and Excel.

So you would have thought, how is it that Microsoft still keeps their moat with Office, the Office package? And yet firms they still use the Enterprise Windows package because they have been using it forever, and they set up around this package.

There's also a lot of like interoperability issues with other parts of the organization. So I think in general, you know, the firms are really good at finding their moat eventually.

Tony Roth: Right.

Martin Beraja: Yeah. And even we can't say now, I think they will find ways in which they will make the technology profitable, in particular because I think the differentiation will come from each firm's use of their own data.

Right now we're seeing these models being trained on more or less public data that was available as online text. And they all have access to the same data. And so then the differentiation must come from then the firms…

Tony Roth: Yep.

Martin Beraja: …using this product themselves and bringing in their own internal data that no one else has access to, and that way they will be able to generate products that look a bit different.

Tony Roth: You know, they're probably going to have different capabilities. After the data is incorporated, then there's an, a question of agency do you a-append an agent to do something? And where does that capability come from and so there's a lot of different ways to differentiate beside just the raw model.

Martin Beraja: Correct. Correct.

Tony Roth: I would love to, if you'd be willing at some point, to come back and not only give us an update, but I'd love to have a conversation specifically around data, because I think the data is so fascinating to understand so many dimensions of how data plays into this process.

I think that it could be its own conversation in and of itself, given that we've set up this framework now. So perhaps another conversation at some point but thank you so much for, for everything today.

Martin Beraja: Thank you, Tony.

What Should Investors Watch Next?

Tony Roth: Just to summarize, I really think that we as investors should understand that we're still in the very early innings for AI, both in terms of understanding the ultimate productivity increase, how companies will be able to pivot from just an efficiency play to an organizational learning play to deliver not just faster and more efficient results, but better outcomes for clients and customers.

And that as it relates to the second question, which is to say the investments that we've all made in these big Mag Seven companies, they probably won't all be the major players, but it's going to take some time to figure out what the value proposition, what the value chain is for each one of them and really understand which ones are going to be the most successful over time.

So with that, um, I want to invite everybody to go to wilmingtontrust.com for a complete roundup of our thought leadership. And for any other questions, please reach out to your Wilmington Trust Investment Advisor. Thank you all so much for listening today.

DISCLAIMER

This podcast is for educational purposes only and is not intended as an offer or solicitation for the sale of any financial product. Or service or recommendation or determination that any investment strategy is suitable for a specific investor. Investors should seek financial advice regarding the suitability of any investment strategy based on the investor's objectives, financial situation, and particular needs.

The information on Wilmington Trust's capital considerations with Tony Roth has been obtained from sources believed to be reliable, but its accuracy and completeness are not guaranteed. The opinions, estimates, and projections constitute the judgment of Wilmington Trust as of the date of this podcast and are subject to change without notice. The opinions of any guests on the Capital Considerations podcast who are not employed by Wilmington Trust or M& T Bank are their own and do not necessarily represent those of M& T Bank Corporate or any of its affiliates. Wilmington Trust is not authorized to and does not provide legal or tax advice.

Our advice and recommendations provided to you is illustrative only and subject to the opinions and advice of your own attorney, tax advisor, or other professional advisor.

Diversification does not ensure a profit or guarantee against a loss. There is no assurance that any investment strategy will be successful.

Past performance cannot guarantee future results. Investing involves a risk, and you may incur a profit or a loss. Any reference to company names mentioned in the podcast should not be constructed as investment advice or investment recommendations of those companies. Third party trademarks and brands are the property of their respective owners.

Third parties referenced herein are independent companies and are not affiliated with M& T Bank or Wilmington Trust. Listing them does not suggest a recommendation or endorsement by Wilmington Trust. Private market investments are only available to investors that meet the U. S. Securities and Exchange Commission's definition of qualified purchaser and accredited investor.

Facts and views presented in this report have not been reviewed by and may not reflect information known to professionals in other business areas of Wilmington Trust or M& T Bank and may provide or seek to provide financial services to entities referred to in this report. M& T Bank and Wilmington Trust have established information barriers between their various business groups.

As a result, M& T Bank and Wilmington Trust do not disclose certain client relationships or compensation received from such entities in their reports. Investment products are not insured by the FDIC or any other governmental agency, are not deposits of or other obligations of or guaranteed by Wilmington Trust, M& T Bank, or any other bank or entity, and are subject to risks including a possible loss of the principal amount invested.

Wilmington Trust is a registered service mark used in connection with various fiduciary and non fiduciary services offered by certain subsidiaries of M& T Bank Corporation, including, but not limited to, Manufacturers and Traders Trust Company, M& T Bank, Wilmington Trust Company, WTC, operating in Delaware only, Wilmington Trust N.A., WTNA, Wilmington Trust Investment Advisors, Inc., WTIA. Wilmington Funds Management Corporation, WFMC, Wilmington Trust Asset Management LLC, WTAM, and Wilmington Trust Investment Management LLC, WTIM. Such services include trustee, custodial, agency, investment management, and other services. International corporate and institutional services are offered through M& T Bank Corporation's international subsidiaries.

Loans, credit cards, retail and business deposits, and other business and personal banking services and products are offered by M& T Bank, member FDIC.

Copyright 2026 M& T Bank and its affiliates and subsidiaries. All rights reserved.