Talking Markets Podcast: Still Early Innings for Artificial Intelligence

Global Legal Information

Podcast transcript

Host/John Przygocki: Welcome to Talking Markets with Franklin Templeton. I'm your host John Przygocki from the Global Marketing Organization at Franklin Templeton. Today I'm joined in the studio by Chris Galipeau, Head Market Strategist with the Franklin Templeton Institute, and Andy O'Brien and Bobby Gray, portfolio managers at Putnam Investments.

The Institute is a research-centric organization here at Franklin Templeton focused on delivering unique macroeconomic and capital market insight to our clients.

Putnam is a diversified equity asset manager of Franklin Templeton, serving investors worldwide, committed to active strategies driven by proprietary research, rigorous risk management, and client-centered innovation. Gentlemen, welcome to the show.

Chris Galipeau: Thanks, John.

John Przygocki: Thank you for joining me today in the studio. Today our conversation topic is AI, artificial intelligence, from the perspective of two professional investors. Before we jump into our topic of discussion, let's start with some brief introductions. I mentioned your titles. Can you share with our audience your core responsibility at the firm? Chris, I'll go to you to kick us off.

Chris Galipeau: Okay, John. I'm the Head Market Strategist for the Institute's capital markets team. And so, our function as a group is to provide market commentary, insight on four general areas: macro being the first, equities, fixed income and then private markets. My job really is to communicate our opinions and thoughts on all of that to clients. And, in addition, we do a lot of primary research that we think will be helpful to investors generally.

John Przygocki: Andy?

Andy O’Brien: Thanks for having me here today, John. So, I am the portfolio manager at Putnam. I'm also the research analyst responsible for coverage of semiconductors and software. I've been in the industry and at Putnam for 15 years.

I started my career as an associate on the consumer team, and one of the first groups that I was able to take over responsibility for covering was the video game sector. And that was right as that sector was going through this really powerful technological transformation, where they went from selling packaged good discs to stores like Best Buy and GameStop to selling digital products to the end gamer. And it was this really powerful transformation that drove a structural change in their gross margins and the economics of those businesses. And so that's what first attracted me to tech investing.

And over the last ten years, I've picked up coverage of semiconductors and software. And when I picked up coverage of semiconductors, Nvidia was a gaming company, and the semiconductor sector represented about 2 to 3% of the S&P 500. Today, Nvidia is an AI company, and semis represent almost 20% of the S&P 500. So it's been a really incredible journey over the last decade plus.

John Przygocki: Thanks, Andy. And, Bobby, a little bit about yourself?

Bobby Gray: Thanks, John. I'm also portfolio manager and an analyst on the research team here focused on covering internet, media and telecom. Been in the industry for about 20 years. Started out on the private side of the business in private equity, and then migrated over to the public side and have been really focused on these subsectors almost my entire career.

John Przygocki: Fantastic, guys. And, again, welcome and thank you for joining me in the studio. Andy, let's start with a couple questions for you. First question is what's driving the urgency behind the current AI investment cycle? And then maybe you could follow that up with why are so many different types of companies racing to participate in it?

Andy O’Brien: So, I guess I would start with what is probably, you know, obvious to most folks at this point, which is that Gen AI's a really true technological breakthrough. You had ChatGPT that launched in the fourth quarter of 2022, and it really brought front and center for consumers and enterprises, the incredible ability for this technology to really take vast amounts of structured and unstructured data and turn it into this usable output that had contextual understanding of what the consumer or the enterprise is asking for or looking to get from the product. And for the first time, really, we started to see this relationship between compute and intelligence.

But, really beyond that, what I would argue has been even more unprecedented is the improvement that's been going on in these models since we saw the launch of ChatGPT. And certainly, for listeners who I'm sure were experimenting with ChatGPT when it originally came out as a chat bot and are probably still using the technology today, I'm sure that they can notice the incredible improvement that's gone on in the experience and the output of these tools.

But even more impressive to me is the fact that, you know, there's been no clear ceiling here. As scaling laws have held, the original benchmarks that were used to determine the quality of the LLMs that were coming out—there were these benchmarks like HumanEval and GSM8K, which would basically ask the LLMs grade school-level math questions—those benchmarks, they're fully saturated. They're not even usable anymore to determine the output of these LLMs, because they've become so good. So, we've actually moved the bar in terms of the benchmarking that we're doing. We need totally new tests to determine the quality of these LLMs. And, so on one hand, you have this incredible new technological breakthrough, and on the other hand, it's improving faster than any technology we've seen before.

And then you asked about the urgency behind the spend. Well, I would say it fits into two buckets, both offensive and defensive. Offensively, all of these companies, whether you're talking about the AI labs, or the hyperscalers or the AI native startups, they all see this incredible revenue opportunity in front of them as AI increases its penetration. On the other hand, many of these players who are spending aggressively are also acting defensively. Nobody here wants to be caught without enough compute or without enough AI IP as the market makes its transition into AI becoming the next platform here.

And I think Mark Zuckerberg had a really telling quote when he was on a podcast several months ago where he stated, “AI capex for Meta is an asymmetric bet.” If they overbuild, they'll waste some capital. But the risk of underbuilding is that they get left completely behind and it becomes an existential risk. And so you have these two offensive and defensive drivers behind the urgency of all of the spending that we're seeing.

And then there's the ability of all of these companies to fund it. And so, you have the incredible free cash flow of the hyperscalers, which now they've largely exhausted. But you're seeing them start to hit the debt and equity markets. And we think there's a lot of capacity there for them to continue to drive the spending. So, you have the combination of ample amounts of capital capacity, combined with a demand and supply that have triggered this real frenzy of spending.

John Przygocki: Sounds like some really powerful forces there. Chris, I can see it in your face. You've got a follow-up.

Chris Galipeau: Well, I was thinking, as Andy was saying, number one, that's a great synopsis. And it is really unbelievable to see the scope and the size of what's happening. But it made me think about Amazon's bond offering this morning. So, Bobby, you cover the hyperscalers. What are some of the bigger unanswered questions around AI business models, the monetization of the effort and, critically, the return on investment?

Bobby Gray: I'll start with the return on investment. Andy used a Zuckerberg quote about wasted capital. I think the single most debated question in kind of our corner of the world throughout 2025 and in ’26 has been, will the big spenders, the Microsofts, the Googles of the world ever earn a decent return on all of this spend? And frankly, the skepticism I think was warranted.

I'll pick on Amazon, but the numbers are similar for all these players. If we were to look at 2027 consensus estimates for revenue today versus two years ago, they're basically unchanged. However, if we were to look at the change in capex expectations over that same time period, they’re up about 120%. So, really, you saw this enormous spending ramp without much to show for it in the way of increased expectations for revenue growth or profitability.

Andy and I did a ton of work on this over the last year, and I think if there was one thing that we came into 2026 with conviction on is that the revenue ramp was going to come in 2026 and that we'd see a significant acceleration. I think we got a glimpse of that with Q4 results earlier this year—and then I think a much clearer signal in Q1 where we just saw a significant acceleration in revenue and revenue backlogs. Again, maybe using Amazon as an example here, their backlog increased by more in Q1 than it had in the 11 prior quarters combined. This was a real step change in revenue growth and contracted revenue.

Do I think the ROI debate is settled for good? No. I think that was a critical first step, but I think there still remain a number of key unanswered questions that we're going to have to address here over the next couple of years if investors are going to have real conviction in the long-term, durable ROI on this capex spend.

Andy O’Brien: Yeah, and I would just jump in, Chris, and add one comment to Bobby's point, which is that, you know, one of the metrics that we track closely is this monetization level per gigawatt, which has become a common nomenclature in the industry. And what we've noticed is that even as the component costs to build out a gigawatt of data center capacity have increased quite a bit, so has the ability to monetize it.

And, for example, we recently saw these big blockbuster deals between xAI and Google and xAI and Anthropic. And even last week we saw AWS raise GPU prices by 20%. And so, one of the things that we feel like the market may be missing a little bit, because a lot of the focus has gone to the infrastructure layer and the beneficiaries of all of that spend is what this means for the hyperscalers that are sitting on all of these large compute fleets that are now increasing significantly in value. The cost to build Microsoft's or Amazon's existing compute fleet has gone up dramatically, but so has their ability to monetize their existing fleet.

And so, we think, actually, this is a very positive backdrop from a monetization standpoint for the hyperscalers. It's also a positive backdrop from an investment standpoint for the hyperscalers, as it will give them increased confidence to continue to spend.

Chris Galipeau: Andy, that's a good point. I was thinking about Q1 earnings, and I know, as Bobby just referenced Amazon, I think we heard something similar with Google. But what you guys both just hit on really is kind of the center point of a lot of the questions that we get. Bobby, back to you for just a second. We just hit the monetization and the ROI part of it. What are some of the unanswered questions that you might have or you and Andy might have still?

Bobby Gray: I'll start with the hyperscalers and then kind of broaden it out to the AI labs and AI more broadly. You know, on the hyperscalers, one thing that we keep circling back to is just this question of industry structure and how it's changing. Historically, the kind of hyperscale or cloud service provider market was a pretty cozy oligopoly, three big players: Microsoft, Amazon and Google.

We've seen a number of new entrants into this space. Obviously, you have a number of neoclouds. Oracle reaching real critical mass. And the news just last week that Meta was looking to rent out some of its potentially excess capacity, xAI and SpaceX doing a similar thing not long ago. So it seems like the space is getting more crowded.

How viable some of those kind of more marginal competitors are over the long term, I think remains an open question. There's also, I think, a question of customer concentration. Historically, the Amazons of the world sold into a big, fragmented market, you know, a long tail of enterprises, which is a pretty favorable dynamic. And to the extent these AI labs, either directly or indirectly become increasingly large and important customers for the cloud providers, it does potentially change the dynamics around customer leverage, margins, and just the risk profiles of the business.

There are certainly some open questions. They're seeing revenue growth at levels we really haven't seen since the peak COVID days. Margins are at the high end of their historical ranges. And this shift has opened up some new opportunities, right? You know, Google and Amazon are likely to build multibillion-dollar, if not trillion-dollar businesses selling custom silicon semiconductors to their customers. So, areas of opportunity but also some risks that we're certainly monitoring very closely.

And maybe just then to widen it out to kind of AI more broadly in the labs, you know, the OpenAIs, the Anthropics of the world, because, you know, what happens with them will certainly ripple through this entire ecosystem. This is where the list of questions is pretty long. I'll start with the most basic one. How many frontier models end up mattering? Do we have two or three? Do we have five or six? Does one player pull ahead? Does someone reach what's called recursive self-improvement, where the model continues to improve itself and kind of pull away from the pack? What will be the impact of open-weight or open-source models? Right now, a significant percentage, potentially a majority, of tokens consumed globally are coming from Chinese open-weight models. We haven't really seen a US open-source champion emerge. I think people thought that might be Meta a couple of years ago, and then they kind of had some stumbles on their AI model development trajectory, but I think are making good progress there.

I think we've also heard a lot about token maxing. We're seeing, I think, enterprises focus much more heavily on optimizing how they use models, potentially routing different queries or tasks to different models to optimize for cost, for latency, for performance. And ultimately, these are all questions that get at who captures the value and what's referred to as the token path. Is it going to be the models? Is it going to be applications that are built on top of those models? Might they be the incumbent applications who fold in AI features and kind of commoditize the models underneath them? Will it be new native applications that are being built right now? It's really hard to say. And the answers to those questions I think will have significant ramifications for the entire ecosystem.

To date, the cleanest way to express our excitement about AI has really been through the infrastructure side of the business, the picks and shovels, where we've had the most conviction.

Chris Galipeau: Could you define what a token is?

Bobby Gray: The simplest way of thinking about it is almost, it's an output. It could be just as simple as a word. So, if you type in an AI query, think of a token as kind of a unit of intelligence or unit of output that you get back that helps you accomplish a task or gives you the answer that you're looking for.

Chris Galipeau: Got it. Okay. Andy, over to you. How should investors think about the scale of this AI infrastructure spend, and is it sustainable over the next three to five years?

Andy O’Brien: We think it is. And we think directionally it moves higher. If you look at the top five what we'll define as hyperscalers, they spent just over 100 billion in capex in 2023. That number this year will be well above 700 billion and we think on a trajectory to be above a trillion next year.

Jensen [Huang], the CEO of Nvidia, a couple of years ago put out a target of 3 to 4 trillion in AI spend by the end of the decade. At the time, it seemed almost implausible, and now it seems a lot less crazy. And certainly, as I mentioned earlier, the capacity to spend between the free cash flow and the access to the capital markets is certainly there.

A couple points that we track that give us conviction in the forward trajectory of the spend, the first is just, AI penetration remains really shallow today. So, a lot of companies, including our own, are using AI somewhere within their business in their workflows. But very few have rebuilt truly deep agentic workflows around AI that are at scale today. There's a number of different estimates out there, but one of the most widely quoted adoption figures is that agentic AI is at less than 1% of potential monetization across enterprise and consumer users.

And so, if you just look at the cumulative AI investment that's gone on to date, kind of post- ChatGPT that I just laid out, we're looking at something like a one trillion range next year. If you assume the penetration of AI doubles from 1% to 2% by the end of the decade, which feels pretty reasonable and arguably really conservative, compute demand is not going to just double. These workloads are getting more and more compute-intensive. So, a doubling of penetration could be equivalent to a three-, four-, even five-times increase in the need for compute. So we just see the compute need as really insatiable right now.

But let's step away from that for a second and talk a little bit about the economic side of it, which is equally, if not more, important. So, one of the things Bobby and I spent a lot of time on is just tracking what the AI revenue base looks like, and it's become really meaningful, particularly over the last three to six months. The leading model companies have moved, in my opinion, well beyond the definition of a lab. If you look at just Anthropic and OpenAI, together, we think they're running somewhere in the range of a $60- to $80-billion annualized revenue run rate. And that's with revenue growing in the triple digits. Microsoft sporadically kind of shares its own AI business. And they said in the last quarter that that had reached a $37 billion ARR, growing in the triple digits. We see similar growth from Google, Meta and others. And so, it's not hard to see AI revenues that are in the several hundred billion-dollar range this year, growing significantly into next year, which kind of underlies the conviction behind a lot of this investment.

John Przygocki: So, Andy, let me follow that up with a slightly different angle here. There's so much focus on GPUs. In the infrastructure space, how is the buildout broadening beyond the GPU, and are there trends or themes that you're identifying and following?

Andy O’Brien: Yeah, that's a great question. You know, one of the most exciting things as an active investor has really been the broadening of AI demand that's played out, particularly over the last 12 months.

If we rewind a little bit and we look at the beginning of this AI buildout, it was really dominated by Nvidia. And certainly everybody knows the stock performance that played out in 2023 and 2024. But one of the things we track is the change in the expectations of the fundamentals of these businesses.

In 2023, Nvidia’s revenue expectations for the following year increased by just over 100%. So, estimates were revised higher on the revenue line by 100%. And we looked at a broader basket of AI infrastructure companies over the same time period. And those estimates only saw a 30% revision higher. So they still improved, but not nearly to the extent of Nvidia. And what we've seen more recently is we've seen that get flipped on its head. And so that same group of companies that saw a 30% estimate revision throughout 2023 saw over 200% over the last 12 months. And that includes memory companies and networking and optical companies who have benefited from some of these, you know, bottleneck and pricing benefits. But then Nvidia continued to see positive estimate revisions, but their estimate revisions were only up 50%. And so, correspondingly, we've seen similar kind of performance in those different groups of stocks, where Nvidia has more recently lagged a little bit and some of the memory, networking and optical and CPU-focused companies have performed better.

You asked a little bit about different themes we're tracking. I'd highlight a few for you. The first is, as I mentioned earlier, bottlenecks has been a real focus for investors. And you've just had such an increase in demand so quickly that a number of companies have been caught offsides with their supply, and particularly the ones that sell more commoditized products have been able to really benefit through pricing. Memory’s the most acute pocket of tightness. But optical networking and analog companies have also seen pricing come through and benefited from this environment. A 100% pricing increase has a really disproportionate impact on earnings growth, because you don't have a corresponding increase in the cost structure of your business. And we're talking to companies in the semi supply chain that are seeing 200 and 300%-plus price increases. And so, we think there are pockets here within this bottleneck theme where tightness is going to remain for quite a while.

Another area or theme that we're tracking is this idea of custom silicon, what we call ASICs in the industry. And what we're tracking is we're really seeing compute silicon broaden out beyond Nvidia GPUs. And it's driven by the leading hyperscalers and the leading labs focusing on designing their own chips. And this has been a real battleground for investors over the last few years as this debate about, can the hyperscalers build their own chips, or will Nvidia always have the superior TCO, has kind of raged on.

And what's become a little bit more clear over the last six to 12 months is that Google has emerged with their TPU as a standout and successful, you know, silicon venture. Amazon’s Trainium has started to emerge as a successful accelerator in the market. And so, it's becoming clear that because the hyperscalers are able to design their silicon and optimize it for their specific workloads, they can actually build out their own TCO advantage relative to Nvidia. And so I think we're going to see more and more companies design their chips. And we're also going to see companies design more and more chips. And so we view this as a durable theme in the market.

And then the last one I would highlight for you is the semiconductor equipment companies. These companies create the equipment that is necessary to build out silicon capacity. And the need to invest behind silicon capacity, whether it's compute or memory, is going to be an incredibly durable theme, I expect, for the next decade. And so, expectations for this group have moved quite a bit higher over the last few months, but they still look really muted compared to the success that their customers have seen. And so, we think there are a number of ways that these companies are going to find to enjoy pricing benefits and margin tailwinds that their customers in the memory and in foundry and silicon compute areas have been able to experience. And so, you know, while WFE, that's what we call wafer fabrication equipment, spending expectations have moved higher, we think it has a long way to go.

So those are the three themes I would say we're focused on today.

Chris Galipeau: Andy, that was great. My summation of that is “accretive to margins” in a lot of what you said. Bobby, let's go back to you for a second. What—a  lot of use cases here. In your estimation, what's the most important use case or cases that are driving AI adoption now? And how do you think that evolves?

Bobby Gray: Yeah. Great question. You know, I think it depends a little bit whether you're talking about enterprise versus consumer.

I mean, on the enterprise side, you know, the killer use case has clearly been on coding. Anthropic really saw an inflection in revenue growth earlier this year with the success of Claude Code. I think the run rate revenue went from about nine billion exiting last year to almost 50 billion by the middle of this year, which is just kind of mind blowing to think about the scale and the rate of growth there. You know, what we expect is, as the models keep improving and keep getting better at tasks, you know, reasoning, computer use, the number of tasks that they'll be able to handle will just keep expanding, and the use cases will increase accordingly. Things are moving so quickly we forget we're still so early on in the turning these models into genuinely useful products for kind of the average business user. And I think, you know, again, we're just months into, you know, creating more user-friendly software interfaces to kind of make the models actually useful. And I think as we do that, you'll see continued incursion of AI into the enterprise in terms of where and how it can be useful.

On the consumer side, I'd say after kind of the big breakout product, which was chatbots—right? ChatGPT Gemini, Claude—I'd say progress has maybe been a little bit slower. Again, the usage is there on chatbots. I think there's some open questions around modernization. Our best guess is this ends up actually looking a lot more like the old world than you might expect—with kind of ad-supported business models, sponsored answers, and kind of monetizing conversational search not that much unlike the way the historical search has been monetized. And I think Google will probably lead the way there.

I think, you know, the second kind of important consumer use case is maybe a little less direct, which is just improving products that we already use. Take Meta as an example. They've been able to use accelerated compute (not necessarily LLMs, but AI compute) to drive better engagement, higher monetization, and have seen revenue growth at levels that I think most folks would not have expected they would be able to achieve just a couple of years ago.

Same thing goes for search. We're enjoying a bit of an expansionary moment, and Google is seeing users find new ways to search, new use cases, deeper engagement with searches. That's again driving engagement. And I think monetization will follow.

I think the real breakthrough, though, that we're all waiting for is personalized AI agents. The key point is, it's still early, and penetration is still really low. So, the potential to continue to just drive usage of these products over the next couple of years, I think is massive. And I think we're excited about the consumer side maybe catching up with the enterprise.

Chris Galipeau: Bobby, if you were to put that in baseball terms, what inning are we in?

Bobby Gray: Oh, gosh. I mean, it feels like we're in the first or second inning. I won't get into the top or bottom of the innings, but it feels like we're in the early, early innings. But I think the sophistication of the models has outrun the kind of the usage and sophistication of the actual products that are put into the hands of consumers and their ability to turn them into useful outputs and to complete useful tasks. And so I think that's really exciting as we think about what the next two or three years could look like, both as investors, also as heavy users of the product.

Chris Galipeau: Interesting.

Andy O’Brien: The only thing I would just add to that is that, you know, the pace of the improvement in the models and the continued reliability of scaling laws makes this a very difficult technology to reliably predict what inning we're in.

As the models continue to improve, so does their addressable workloads and use cases. And so, today as we think about what a model can do based on the existing frontier, it could be totally different than what a model can do in six months or a year. And so framing the TAM or the addressable opportunity is uniquely difficult for this technology.

Bobby Gray: Yeah. I referenced the Anthropic inflect earlier this year. That was not only enabled by the kind of harness that they built on top of the model, but it was actually their release of the latest generation of model that really just unlocked a new set of use cases for coding, where we went from just kind of basic autocomplete to actually writing and testing, you know, entire chunks of software. So I totally agree with Andy’s point.

Chris Galipeau: Amazing. It feels like it's moving at light speed, and it might continue to amplify, right? Andy, speaking of use cases, I know you and Bobby are obviously in the weeds here. As analysts and portfolio managers, how is the Putnam organization using AI in your workloads?

Andy O’Brien: Yeah, we're using a tremendous amount of AI, and this has been a key priority of Putnam's. And we've worked really closely within FT [Franklin Templeton] to really get a lot of tools into our team's hands. And if I look across like currently or in the past, because we've trialed some tools and then moved on, we've deployed ChatGPT, Copilot, Claude, Perplexity, AlphaSense and more.

And, you know, what we've really done is make sure that our investors can utilize the best tools at any given time. And then we've really encouraged sharing as loudly as possible. And for me, you know, one of the most rewarding parts of this process and fascinating parts of this process has been the different use cases and the different ways investors are utilizing AI across the team.

John Przygocki: Gentlemen, as I listen to what was just stated there, that quote of “Share loudly” really resonates with me. I want to thank Chris, Andy and Bobby for your time and wonderful insight today. To all of our listeners, thank you for spending your valuable time with us for today's conversation on artificial intelligence. If you're interested in learning more, please visit franklintempleton.com.

If you'd like to hear more Talking Markets with Franklin Templeton, please visit our archive of previous episodes and subscribe on Apple Podcasts, Spotify or just about any other major podcast provider.