Speaker 1And so all of a sudden, you're not just attacking enterprise IT software market, you're attacking the trillions of dollar knowledge workflow.
Speaker 2Welcome to The Angle from T. Rowe Price, a podcast for curious investors. Just a reminder that outside of the U.S. and Australia, this podcast is for investment professionals only. I'm your host, Jennifer Martin, a global equity portfolio specialist at T. Rowe Price Associates here in Baltimore, Maryland. The next phase of the AI revolution is here. Chatbots answered questions. AI agents are going a step further and executing digital tasks for us. Here to help us understand the rise of agentic AI and its implications for investors are Dom Rizzo and Frank Shi. Dom has been with T. Rowe Price for more than a decade and is a sector portfolio manager who also co-leads our technology analysts. I like to think of Dom as half man, half GPU. He has a lot of processing power. Frank Shi is an analyst covering global semiconductors and has stayed up late to join us from Hong Kong. And in many ways, Frank is like AI, he's on 24-7. Dom and Frank, welcome to The Angle.
Speaker 1Thanks for having us, Jennifer. It's great to be here.
Speaker 2Thanks, Jennifer. Dom, the last time you were on The Angle, you asserted that AI could become the greatest productivity enhancer since electricity. As we are transitioning from chatbots to agents, what does that mean for this story and the pace and scale of AI adoption?
Speaker 1Well, let's first take a step back, Jennifer. It's amazing to think about all that's changed over the past few years. But what did I mean when we said AI has the potential to be the biggest productivity enhancer since electricity? I think AI is going to result in an unlock of productivity that's going to result in a strong economic boom. And there's only three ways to grow an economy. There's capital, throw more money at the problem, labor, throw more people at the problem, or productivity. Do more with the money and people that you throw at a problem. And I think we're at the early stages of an economic acceleration driven by a productivity boom. And the big change, since I was on the podcast last, is the change from simple chatbots to agentic workflows. A chatbot gives you a response. You're just talking with ChatGPT. But an agent can complete tasks on your behalf. It can break a goal into steps, use tools, check the results, and keep going for hours. That is the economic leap. We're moving from just. Talking about information to doing work. Now, where have we seen that pop up? The first and best use case for agentic workflows is coding, right? And it's the clearest early proof because the feedback loop is just brutally objective. The code runs or it doesn't. Let's look at some of these numbers. I mean, these numbers have been phenomenal. Last summer, Anthropic was just on a $5 billion run rate in revenue. Now, because of the rise of clawed code, they're close to $50 billion.
Speaker 2Dom, that is huge. Put that into context. Five to 50 billion. How many enterprise software companies is that?
Speaker 1It's so funny to think about it that way. It's almost three service now, Jennifer. Wow. So, I mean, I've never seen growth of an enterprise application like this. This is true virality within an enterprise on the back of. This is true virality within an enterprise on the back of coding agents. And it's tools like OpenAI Codex, tools like Clawed Code, tools like Cursor, that are really starting to take off within the enterprise. And I think it's going to result in this economic boon. And if you think about what is coding, coding is the unlock that allows you to get into different workflows, finance, legal, HR, and so all of a sudden, you're not just attacking enterprise IT software market. You're attacking the trillions of dollar knowledge workflow. And how do we make all of these people, whether in finance, legal, coding, more efficient, make them more productive? And will enterprises pay for that productivity? So far, the evidence is absolutely.
Speaker 2Yes, definitely. And you're seeing the adoption, even at our own organization, which I know we're going to talk about, we've had some really good milestones. We've had some really good milestones internally. And so maybe I would like to ask Frank, it's easy to see why the proliferation of agents, which increases the demand for tokens, would be a broad tailwind for infrastructure. However, nuance always matters in these discussions. Could you explain what parts of the hardware supply chain appear well-positioned as AI goes from answering questions to executing complex tasks? And maybe we start with what is a token?
Speaker 3Token, in the technical terms, I would like to measure the basic measurements in AI. But I would like to think of it as a sliver of intelligence. You can use that to address enterprise problems, to provide consumer services, to find cancer-curing drugs. And all of that together contributes to the massive demand we're seeing for token. On the hardware side, it's very obvious to think of the beneficiary as the GPUs. However, what agents has changed is that it's moving the bottleneck from the accelerator to the entire system. We're seeing that CPUs, memories and storage, and networking are the three areas we're seeing massive changes. Starting with CPU. I think of CPU as the operation team around the all-star GPU. And the funny detail here is that the CPU is significantly cheaper than the GPU. So you are very well incentivized to make sure that you have the best CPU to keep your all-star player always fed and ready to play. And here, the CPU's job in the system can be scheduling the work, launching sub-agents, running Python, querying database, all the nitty-gritty details to actually make the agent useful. For memory and storage, this is where the incremental system performance as well as the model performance improves substantially when you add more memory. Think of it this way. When you add one additional unit of memory, it significantly improves the system's throughput, meaning that the model companies can generate more revenue per dollar of AI spend they have down. To the models, memory also increases the amount of context it can have and allows the system to better recall what you have discussed with it, what are the right documents to retrieve when you're trying to address a certain enterprise workflow. And that dramatically increases the value of the model to the end users. Lastly, on networking. The networking is now the computer. There's one way of trying to squeeze everything, every model weights, every context, one piece of memory. There's another way of spreading them out across the system and making sure that every part of the system is properly configured. And that requires networking to manage the traffic, schedule the data movement, and make the system the most efficient. All of them combined makes the system a lot more efficient and makes the model a lot more useful for the customers.
Speaker 1You know, Jennifer, Frank is really getting to this point that Dave raised a few weeks ago on his podcast. AI is not software. It's token manufacturing. And token manufacturing is inherently hardware intensive. And Frank just laid out all those great points. It's the CPU, it's the GPU, it's the memory, it's the networking. All working together in concert to manufacture more tokens.
Speaker 2Both of you have done such a phenomenal job in navigating how hardware needs shift as AI workflows have evolved. The other thing that we've talked about and talked a lot about on this podcast is really about bottlenecks and shortages. This robust demand means that many parts of the AI supply chain are in shortage. And so how do you assess these relative opportunities?
Speaker 1Look, you know my framework. We always stay close to our framework in these hard moments. Lynchpin technologies innovating in secular growth markets with improving fundamentals at reasonable valuations. And right now, probably the thing I think about most is the lynchpin part of the framework. Not every shortage is a durable economic moat. There's a difference between lynchpins and golden spikes. Lynchpins are mission critical, hard to substitute, slow to add, and still very valuable after supply chains catch up. Golden spikes are scarcity and in a moment in time. Frank has done a lot of great work here. Maybe you want to expand, Frank, a little bit on what's happening in memory and why we've seen maybe some of the lynchpin change that's happened
Speaker 3GPUs? Absolutely. If we think of what creates or makes for a durable economic mode, it's the fact that you have technologies that others cannot own, or you have a significant cost advantage that others cannot surpass. Traditionally, people think of memory as a commodity, and they're having their moment in time because of the historic undersupply and the pricing power that has granted them. However, I think a structural change happening to the memory sector is high-bandwidth memory. And what that has changed is that, one, it makes memory more customized than standardized. And that confers better pricing power and margin to the certain segment of the memory products. And secondly, that also gives memory companies a better visibility to the demand and technology roadmap of their customers, turning them from just a memory company to a memory company. A commodity vendor into a real technology partner to their customers. Compare that to, for instance, a CPU or GPU or even networking gear makers. There you have players, especially leaders, who have very differentiated technologies or very differentiated cost structures, and also significant tying and customization with their customers, allowing for what we consider a more durable economic mode.
Speaker 1And it doesn't mean that these golden spikes can't be changed. They often are. What Frank is laying out is knowing the difference between short-term shortages and long-term structural changes in businesses. And probably the most important question that we're asking ourselves regarding the memory debate today, with the rise of high-bandwidth memory, are you actually moving from just a shortage-based business model to more of a linchpin status? What we've seen with CPUs, as we've gone from training to inference, they've become more important to their customers. Their linchpin status has gone up with their customers. And I'll give you an example. In training, the trade ratio between GPUs and CPUs is roughly eight to one. On their last earnings call, Intel talked about the trade ratio between GPUs and CPUs in inference being one to one. That is the definition of linchpin status going up.
Speaker 2Perfect. I mean, we could spend the entire podcast talking about bottlenecks and infrastructure. But I think we have hit on infrastructure. So maybe we should pivot a little bit to the technology sector that struggled the most this year, which is software. And a lot of it is because of the rise of agentic AI, changing the economics and the competitive landscape. So maybe, Dom, why don't you start with some thoughts on that phenomenon?
Speaker 1Well, what's happening in software? Basically, I think you can think about it as software being repriced. It's being repriced around its new user, the agent. It's no longer the human. The user of the software is the agent. And what happens in that world? Three things change. First, the user interface can basically disappear. The agent just talks directly to the software through APIs or direct data integration. They go headless, as Salesforce likes to say. Second, pricing changes radically. In many ways, these software companies were fat and hapity as they used to price based on seats. And guess what? They probably overcharged based on those seats, and they sold a little too many seats in an organization. Now, we're moving to usage and outcome-based pricing, where only the strong survive. And third, the incumbents have to now spend aggressively on their employees. And that's why I think it's so important that we think about pricing. Just to defend their workflow. The cost of manufacturing software is down because of agents. What does that mean? Your traditional workflow is now subject to commodity pricing and commodities quickly coming in and taking you out from underneath. So three things have changed. User interface has changed. Pricing has changed. And the incumbents have to spend aggressively just to keep up.
Speaker 2So the panic around software does remind us of other sell-offs that we've seen, for example, almost, gosh, 20, 30 years ago in brick-and-mortar retailers, where the worries about disruption by Amazon. And a lot of this fear at the time, very well-founded. But there are a lot of big-box businesses, auto part retailers and so forth, that have proved resilient, surviving this disruption risk, and are ultimately thriving. So what might be some of the qualities that position software companies to succeed in the AI era? Because it seems like a really interesting area to look at.
Speaker 1Well, AI is not going to kill all software. I think it's simply going to expose what the software was actually worth to begin with. If you own proprietary context or data, control a mission-critical workflow, have trusted permissions and audibility, or own distribution, there is a real path to you driving an economic growth. I think that's a really good point. I think it's a really good point. It's a really good point.
Speaker 2And do you want to spend 60 seconds explaining why the frontier models are advantaged with their model and harness?
Speaker 1This is probably the thing that most people underestimate when they're looking at token manufacturing. They say, oh, all models are commodities. And Frank can talk about this a bit more. I don't believe models are commodities. I actually think all of the economic value are maybe close to all the economic value, will be derived at the frontier over time. And one of many reasons I believe this is because the marriage between a model and a harness. A model and a harness, in many ways, is like a great spouse. They make up for your deficiencies. The harness makes the model better. The model makes the harness better. And they work in concert to deliver that magical experience. Just look at what SpaceX had to do recently. spacex had a great model in the form of grok but they didn't have a great harness what did they do they had to purchase a harness with cursor open ai has a very strong model and harness anthropic has a very strong model and harness those two things together can create products that delight the user and then you can use that delight to capture user time and aggregate the rest of the enterprise, making everything else a dumb data pipe. Frank, maybe you want to expand on why we don't think the models are commodities over time, because that is probably the most important thing underlying the economics of the build-out.
Speaker 3Absolutely. A token and the model behind the token comes in different shades. In today's world, frontier models provide the best intelligence at the highest cost. and they come at the highest cost for a good reason, because they solve the most economically sensitive tasks at the best ROI for the customers. Obviously, that is going to change as some of the open-weight and less-than-frontier models catch up on capabilities. But what we are seeing at the Frontier Labs is that through various pre-training reinforcement learning techniques, they continue unlocking new capabilities, pushing the frontier of what a model can do. Imagine having one model that can actually develop a cancer-curing drug. How much would you pay for that type of model and that type of token versus the next model that can do your email summary? The massive performance difference was keeping a lot of the economic value at the frontier lapse and as long as the frontier of intelligence and capability continues getting unlocked we think a lot of the economic value will remain with the frontier lapse
Speaker 1jennifer one one question i ask myself all the time is what percentage of economic value in day-to-day life is driven at the frontier yeah right i think what we've seen and it's probably resulted in more inequality over time, is that returns to extraordinary people, returns to extraordinary ideas, keep getting higher and higher. And that's because so much of our economic value and so much of our GDP growth like Frank said, look, for cures of drugs, you know, you're best performing. What is the difference between your average lawyer and the best performing lawyer in the world? What is the difference between your average portfolio manager and the best performing portfolio manager in the world? I think there's a dramatic difference between those two. And now with AI agents, we can arm those people with frontier intelligence, and then we can supercharge them. I think it begs some real strong society questions that we have to answer. What is this going to mean for growing inequality? But I also think it means that the returns to the frontier can be exponentially higher than the rest of the models in the world. Super, super interesting.
Speaker 2And I think most of our discussion so far has centered around the implications of enterprise AI adoption, which makes sense because businesses, as you just highlighted, are willing to pay for productivity. Let's shift a little bit and talk about your views about how agentic AI might evolve for consumers.
Speaker 1Well, Frank, maybe you want to start with how you think that AI can change the smartphone landscape over time, because I know that's something that you've been spending a lot of time on. And then maybe I could talk about our views on the potential for an AI chief of staff in day-to-day life.
Speaker 3For consumer agents, I think smartphone is a natural interface because a phone already contains our identity, location, and identity. And I think that's a really good way to start. I think it's a really good way to start. Preferences, even banking information, that makes for a natural place for an agent to understand the context and take action on behalf of a consumer. Now, a smartphone as an AI agent interface doesn't just mean the hardware. You have to own both the hardware and the operating system such that the agent can have nuanced and privileged access to your information. Here, we think that the operating system and the chip owners and the hardware ecosystem owners, companies like Apple and Google, obviously have an end advantage. But at the same time, the super apps, the meta, etc., who owns significant critical mass of consumer relationships and data and economic incentive, will be looking at owning that interface as well. And that's why I think the landscape is going to look a lot more interesting and lively in the next few years with both the incumbents trying to develop AI agents, leveraging their existing relationships with the clients, and also the new entrants who want to take their super app and turn that into a consumer-facing agent interface.
Speaker 1And probably the biggest question, Frank, when we debate this concept of the AI chief of staff in your life, who can see your calendar, inbox bills, bank accounts, preferences, take actions on your behalf, is what is ChatGPT's role in that world? The long-rumored hardware project, what is that going to be? Is the user interface for AI different than the traditional smartphone user interface? Or is the smartphone user interface actually the perfect one? And the operating systems have an advantage. So this is all something that we're debating. What's really fascinating here is how the economics are going to play out. What have we learned so far? Enterprises are willing to pay for productivity. Consumers? Some are, but not all. So there has to be a different way to monetize this user time. And I think the obvious answer is ads. Ads feel like such a dirty word to so many of us, but they have resulted in more consumer surplus than anything I can think of. The consumer industry is going to be a little bit more competitive. I think the consumer market is going to be a little bit more competitive. And I think the consumer market is going to be a little bit more competitive. And I think the consumer market is free because of the advertising business models of Google and Meta. And whether or not you think Instagram is a good or a bad for society, I can bet you that you've sat there and scrolled through reels on a Tuesday night after a long day at work. And that user attention, that time is very valuable. Those eyeballs are very valuable. Now let's take that to the chatbots. What does that mean for the advertising? Model of chatbots. Similar to how you scroll through reels, many people spend a lot of time on chatbots. And I think the reality is that that's actually a more intent driven search profile than just traditional reels. The difference between searching for video content on the Amalfi Coast and your upcoming trip to Italy versus planning it within the chatbot. And what does that mean for an advertiser? I think that the whole plan of your trip to Italy could be very, very, very valuable.
Speaker 2You're making some of those consumer internet companies very happy with that insight, Dom.
Speaker 1Well, let's see what happens. Whether or not this ends up being a net benefit to their business or a net negative to their business, we actually don't know. And so much of it depends on how much is the agent going to take over the workflow versus how much is the human actually still in the loop?
Speaker 2There's a lot to unpack with the consumer. And I think it's still early days. But what is very obvious to me when I spend time with both of you is the agentic AI is bringing a lot of change. And both of you are using these innovations at work. And I would put you in the leading frontier, I guess. I think we have a group at the firm called AI's Wizards. So you guys have your wizard hat. And they actually handed them out for our audience. And maybe, Frank, we start with you. You did a lot of work over the holiday, Christmas holiday. So what are some of the things that you can say on your, gosh, tech stack, for lack of a better word? So maybe explain all the work that you've been doing to help us.
Speaker 3Absolutely. Now, if we think about what an AI can do for us as investors, it's truly about letting us do a lot more than what we previously have time for, such that we can focus on the most important task of our day, which is finding return-generating insights and act on those insights. So in my case, I have been able to automate an important part of my day-to-day workflow, such as gathering through information, looking at various filings and expert transcripts, and actually picking up new signals in terms of what is changing day-to-day in the tech landscape. The second thing that we typically do after earnings season is actually going through filings and make sure we understand all the important accounting changes, and some of the incentive changes for the management. Those used to be a very labor-intensive task and takes hours of work after each earnings season. Now I can do that in a few minutes and see a time series of how these important factors have changed. Those are just a small part of the things I have tried to automate with my sub-agents or AI agents. And with that, I'm free to actually think a lot more. On what matters to my stock, what's driving the stock price change day-to-day, and how to do my job better as an investor.
Speaker 2That's super energizing. Dom, you got to tell us about your agents, maybe your harness. Did you name your harness?
Speaker 1I haven't named my harness. We've debated a few different names. If we take a second, though, and we think about this moment in time, how many people are actually using true agentic workflows in the investment process? I actually think it's very small. Eventually, it will be everybody. Why have people not used agentic workflows so far is because it's hard work getting them going. Right now, my agents are the equivalent of a junior analyst.
Speaker 2That's impressive. That's a move because a year ago, we didn't have that ability.
Speaker 1That's an incredible move. Yeah. They can build charts. They can build models, do research for me, each of them at different points in my framework. I can make trade recommendations. All sorts of things. However, to get to that point, it has required two people working full-time with me, one we call the plumber, John, who makes sure that the data is clean, easy to integrate, and Albert is another AI wizard who helps me cast spells by helping me build skills. If you think about what we need today to create that junior analyst, it's two full-time people and this is why I'm not a jobs doomer. I think there's going to be so many jobs over time created by AI. So far, all evidence is that AI creates jobs, not loses jobs. I think we're seeing it in our day-to-day workflow. Going forward, this is just going to get better and better and faster and faster, Jennifer. Frankly, if you don't use it, you're going to be left behind. I'm so excited about what we're doing internally.
Speaker 2What I think what is probably underappreciated by our audience is the importance of both of yours investment framework. There's a lot of rigor in it. There's systematic ways of looking at it, which benefits these tools in a really complimentary way. It makes both of you very super-powered. I'm glad you're our wizards today. I just want to say thank you. This has been a really wonderful conversation. Lots of great insights shared by both of you. On behalf of our audience, thank you.
Speaker 3Thanks for having us.
Speaker 2We covered a lot of ground today. What resonated with me was just how large the market for intelligence could be and how increasing demand could benefit different parts of the supply chain, including CPUs and memory. Again, I'm Jennifer Martin. Thank you for listening to The Angle. We look forward to your company on future episodes. You can find more information about this and other topics on our website. Please rate and subscribe wherever you get your podcasts. The Angle, better questions, better insights, only from T. Rowe Price. This podcast episode
Speaker 4was recorded in July of 2026 and is for general information and educational purposes only. Outside of the United States and Australia, it is for investment professional use only. It is not intended to be used by persons in jurisdictions which prohibit or restrict distribution of the material herein. This podcast does not give advice or recommendation of any nature or constitute an offer or solicitation to buy or sell any security in any jurisdiction. Prospective investors should seek independent, legal, financial, and tax advice before making any investment decision. Past performance is not a guarantee or a reliable indicator of future performance. All investments are subject to risk, including the possible loss of principle. Technology companies may be more vulnerable and can be affected by intense competition, regulation, earnings disappointments, and rapid product obsolescence. Discussions relating to specific securities are for informational purposes only and do not represent recommendations and may or may not have been held in any T. Rowe Price portfolio. There should be no assumptions that the securities were or will be profitable. T. Rowe Price is not affiliated with any companies discussed. The views contained herein, including forecasts and forward-looking statements, are of the speakers as of the date of the recording and are subject to change without notice. These views may differ from those of other T. Rowe Price associates and or affiliates. Information is from sources deemed reliable but not guaranteed. Please visit http://www.troweprice.com slash theanglepodcast for full global issue or disclosures. This podcast is copyrighted by T. Rowe Price. 2026.