The Exchanges

Q its infrastructure advantage, something that Google's invested in to its core from the beginning. Can you Tell me a little bit about where you view Google's infrastructure advantage playing out in the AI competitive landscape today. How does it translate into cost, speed, product, quality, and where do you guys think about investing the seventy billion of capex this year in the chip layer, in the networking, the data center?

A We can unpack both, right, like, where our CapEx is going, but on your first part, right, like, one of the ways, you know, we look at the Pareto frontier of performance and cost. Google literally is on the Pareto frontier, so we deliver the best models at the most cost-effective price point, right, like, you know, and our Flash series of models are a real workhorse in the industry, right, and And part of why we are able to do that, uh, is because, you know, we train and serve our models on our infrastructure, including TPUs, right? And we are in our seventh generation of TPUs, and, uh, we built our first version in 2017. Um, I remember talking about it at Google I.O. Probably people didn't pay attention to it because, like, you know, why are you building a specific machine learning X-rated chip? Uh, look, it plays out everywhere. To your earlier question on cost per query in search, The reason we feel comfortable we can serve it at that scale is because we are constantly innovating through each generation, including chips which are really, really good at inference, right? And Ironwood, which is our latest in our TPU series, a single part of Ironwood is over 40 exaflops, right? And, and so the scale of these things are incredible, and, and we have thought about it all the way from subsea cables The scale at which we do infrastructure is unparalleled, and I, I've always viewed th…

AI assessment note: “we deliver the best models at the most cost-effective price point”

Q in other bets. Since that time, do you still think about Alphabet as a holding company? Are there still multiple businesses that you want to kind of stand up and foster and have this kind of holding company model? Is that still hold? Or is Google really the core engine that's going to continue to evolve and continue to have ancillary businesses that are, you know, somewhat adjacent to Google?

A I'll answer two ways, right? So I think The way we are not a holding company in the sense that we are, we are not just like looking to invest capital in other attractive businesses. That's, that's not who we are, right? Um, we are, you know, from a foundational technology basis, if we can take that technology, And that R and D we do, and identify problems in which we can innovate and bring a differentiated value proposition, we'll do that, right? So that's the way we approach. And so the, the structure is an outcome of that, right? So, and which means you will have businesses, uh, on, on, on, on paper, they may look like very disparate, But there is a common strand underneath them, right? So, like, I mean, Waymo is going to keep getting better because of the same work we do in Gemini and AI over time as Google Cloud, to Search, to YouTube, to Isomorphic, to Robotics, etc. So that is the unifying layer. Right? And, and then it's a continuum. Is Google Cloud a Google business or an alphabet business? Right? You know, we segment it out. Right? And so, so, you know, so the, the, the, the branding matters less, I think. Right? You know, we'll have a range of companies. Some of them will leave and IPO out because maybe that's the best way they can make progress. So all of that is a possibility. But what I think I, the founders, think about it as, like, the underlying innovation by wh…

AI assessment note: “we are not a holding company in the sense that we are”

Q So tell me about quantum. Because everyone ignores quantum. You've had this investment for some time. Why is quantum so important? Because again, I want to use the historical data that it does. It seems like a small bet. Good luck. But what does quantum evolve to from a compute perspective for humanity? And when does that happen, do you think?

A Obviously, quantum has gotten a lot more attention in the last 12 months or so. Uh, but we have been Just like Waymo, we work through these things, whether there's attention from the outside or not, because we are working on these things out of conviction on the long-term trends, right? So it's, it comes from those first principles, um, Obviously the universe is fundamentally quantum, uh, you know, to, to do any kind of large scale, uh, simulations in a way that truly represent nature, you know, you would need, uh, some versions of quantum computing. I think to me quantum feels like where AI was around, you know, 2015. So I would say in a five year time frame, You would have that moment where some, a really useful practical computation, you know, is done in a quantum way, far superior to classical computers. And that'll be that aha moment, uh, you know, I think which will really show the promise of the industry. Uh, I'm absolutely confident that we will get there when I see the progress, and I can pattern match to progress in the other fundamental areas we have worked on. So it really doesn't feel like, obviously, look, these are very challenging areas. You may hit a constraint. I do think a lot of people are making announcements in quantum, so in some ways it's tough to distinguish them. We had the same scenario in self-driving maybe three years ago. There were so many people …

AI assessment note: “in a five year time frame, You would have that moment where some, a really useful practical computation”

Q a shift in your ability to recruit top talent? A lot of great talent has started other great companies. Other great companies in Silicon Valley have recruited folks. I know there's always a talent war going on, but has there been a tenor shift for Google in the last period of time because of some of the underlying advantages in AI or some of the cultural changes that are underway?

A The talent market, you know, we go through these fierce moments for talent. AI is one of them. And whenever there are these Google, you know, obviously we, we are fortunate to have some of the most talented employees. So we are a source. I'm equally proud of the fact that I think Google has left to start over 2000 companies, right? And, and so, you know, there is a virtuous cycle. I think people come back, we acquire companies. I think all of that keeps the company, uh, fresh, but in the current AI moment, Look, I think we are both holding on to critical talent. We are recruiting. You know, I always look at the tip of the tree of, are we able to attract the best PhD researchers coming out of the top programs? And the answer is yes. Um, you know, and there are people who have left, who have come back, and so I feel good about the position we are, but you work at it hard every week, every month, and so on.

AI assessment note: “in the current AI moment, Look, I think we are both holding on to critical talent.”

Q ad revenue on search is about a two hundred billion dollar run rate out of three hundred and sixty billion of your total revenue, most of your profits. And it seems like Google's in a really challenging quandary where if you disrupt yourself too quickly, all of that revenue can go away, can be really impactful. So is Google being disrupted by AI at this moment, or is Google leading?

A It's a good framework, good question to talk about. Uh, you know, I've definitely, uh, you know, for almost a decade, uh, you know, one of the first things I did was to think of the company as AI first. It was very clear to us. Uh, we had Google Brain underway in 2012. We acquired DeepMind in 2014. 2015, when I became the CEO, I said, look, the technology is really evolving. The reason we were excited to be, approach our work as AI first, Uh, it's because we really felt that AI is what will drive the biggest progress in search. And so, you know, I, I think even the last couple of years, I viewed this as an extraordinary opportunity for search. I think if you look at how much information means to people, I think they're going to, each person is going to have access to information in a way they've never had before. So it feels very far from a zero sum construct to me. And we are seeing it empirically when people are using search. Obviously, there are a couple of major things, ah, we have done with search. Ah, you know, Transformers drove some of the biggest innovations in search with Bert and Mum. Dramatically improved search quality. Um, we, we launched AI overviews about a year ago. It's now being used by over one and a half billion users, uh, in over 150 countries. It's expanding the types of queries people can type in, and we see empirically the nature of queries is expanded,…

AI assessment note: “feels very far from a zero sum construct to me”

Q from now as AI evolves, as computing evolves? Am I looking at a screen? Am I typing in a chat? Am I using an AirPod and just getting audio? Am I doing audio plus a screen? Is it just a personalized interface and there's no even concept of the web? What does the future look like for accessing information And pursuing my interests in life as a human using compute?

A That's a great question. Uh, I do think, you know, the answer has got to be, you know, we've always, you know, humans have adapted to computing, and it's always been that way. But over time, the answer will be that you need to do less of the hard work, less of the adaptation, and computing kind of works for you, right? And that's the holy grail, I think. And And we are making progress, right? Be it touch, be it voice, everything inches us towards this future. Um, for example, when I wear AR glasses, I already wear glasses, so it's not that, you know, but the AR glasses aren't quite as comfortable as my normal glasses, but they're getting there. It's obvious to me that that'll push it to the next level of seamlessness where it kind of is ambiently there and doing stuff for you. So I think that's the arrow there, you know, the arrow of, uh, how it'll, you know, it has to be more seamless and just be there for you. You know, will it be like Neuralink down the line, right? You know, like, you know, when I, when I want to understand something, you know, is it, is it that seamless, right? You know, I, I think all of that is a possibility. But I think in the immediate world, given you're going to have really natively multimodal models, which can take You know, audio, vision, language, uh, all of that, and be there in your, uh, line of view. So I think when AR really works, I think tha…

AI assessment note: “when AR really works, I think that'll wow people.”

Q CEO, the stock has gone up by four and a half X to a two trillion dollar market cap today. You've grown revenue from twenty billion a quarter to nearly a hundred billion a quarter. It's been a really, like, incredible run to see someone that kind of started as a PM and You know, grew your way into this incredible role, so congrats. How have you liked the job?

A No, look, I mean, ah, I love building products, and in some ways, you know, Google was really set up, I think the founder set up this kind of a deep computer science approach, and like, you take that and apply it to build things which can impact people on a day-to-day basis, and so, you know, it's that kind of a product and, ah, technical culture which, you know, is, is the essence of the company. So I love doing that. And, you know, there's not a single week which goes by, but I, I feel like I don't get to do that. So those are the parts I really enjoy. But obviously, you know, running a company of this scale where you impact, uh, so many people, I think it's a privilege. So enjoyed every part of it.

AI assessment note: “So enjoyed every part of it.”

Q Well, let's talk about that, and let's talk about the unknown competitor DeepSeek popped up. Tell me about your impressions of the model, the performance, the rumors about the next model, and what does that tell you about what's going on in China, and what's going on that we're not seeing?

A Look, I think the main moment from DeepSeek was, ah, Look, always. You know, if you, if you kind of follow the AI research and scan through papers and read them, no, nobody who does that would underestimate China, right? Like, you know, so when you look at the amount of research output from China, right, um, they have extraordinary talent, um, And, and, and so, but I do think all of us had to adjust our priors a little bit after the deep seek moment, which was like, wow, they are even closer to the frontier than most people maybe assume, you know, and so I think, I think that was a moment. I think internally for us, I think externally people are very impressed and rightfully so with how efficient their models were. Interestingly for us internally, we benchmarked it to Flash and, you know, Flash was, ah, as efficient or, you know, you could argue in some ways better. So, you know, I think, I think to our earlier conversations, I, I, I do think this is more, ah, maybe internal baseball for us, you know, we were benchmarking and saying, look, it's good to see, you know, because they had to work in a hardware constrained way, I think, which is what drove a lot of their, ah, innovations and efficiency improvements and, you know, ah, and so I was pleased with that. But, you know, it tells you that the frontier is, is, uh, evolving rapidly. There are more players closer to it than peo…

AI assessment note: “externally people are very impressed and rightfully so with how efficient their models were.”

Q Yeah. Do you want to speculate on a business in quantum?

A Look, I, I, we are committed to, you know, in almost all these cases, our goal would be to, you know, demonstrate more and more useful practical algorithms and show progress on that and, and give access to it through cloud, right? And, and I think, you know, I always say it's tough to project innovation on top of a platform, right? Nobody could say just because you had smartphones and GPS and payments, something like Uber would get invented. So you couldn't linearly sit and project Uber from the underlying innovation. That's how the world works. And so, uh, for me, quantum is that foundational, again, just like AI, there's going to be extraordinary innovations on top of it.

AI assessment note: “I always say it's tough to project innovation on top of a platform”

Q And do you have a point of view on ad revenue per AI query?

A You know, we already with AI overviews, um, you know, we, we are at the baseline of, uh, you know, it's, it's the same as without AI overviews. And so we've, we've reached that stage in a, uh, so, but from there we can improve, right? And I think, uh, you know, I've always felt, uh, You know, the reason ads have worked well in search is because commercial information is also information. People in, when they have that intent are looking for that most relevant information. So I don't see any reason why AI, you know, just from a first principle standpoint, why won't AI do a better job there as well, right? And, and so I think, I think, you know, I think we are comfortable that we can work the transition through. Some of it may take time, but all indicators are that we'll be able to do it well.

AI assessment note: “we are at the baseline of, uh, you know, it's, it's the same”

Q consideration for Google. That's the outside kind of narrative. Can you share a little bit about, and then I want to come back to non LLM models where there's other advantages for Google in a minute, but maybe just on this point, how much more, um, of, uh, an opportunity to continue to evolve LLMs is there? Where's Google's advantage lie in maintaining better performance in the models over time?

A I think maybe it was Andrei Karpathy who used the term AJI, which is like he called it artificial jagged intelligence, right? So I think the progress is not going to be always smooth, right? Like you, you go through these periods, it looks like something slow, and then you see a paradigm breakthrough, et cetera. And it's been going like that for a while. Uh, I think, obviously, over the last couple of years, uh, you know, all of us scaled up on pre-training, and then there was a lot of momentum with post-training, and then with inference compute, uh, and, and, and now, you know, there's progress with how do you take all that and stitch together in agentic workflows and, you know, and, and, and so on. So, I do think there's a lot of progress, and it feels pretty continuous to me, right? I think it's both true progress gets harder, Which I think will distinguish the elite teams, at least on the foundational side. Uh, you know, I think, I think, I think that, that might be a factor. Uh, I felt the heart, the heart of the problem is, I think, you know, we are well set up for that. Uh, I think, I think we are well set up for that. I do think we are, uh, pushing the research frontier in a much broader way than most other people. Beyond just LLMs, transformer-based models, I mean, diffusion, you know, do you do diffusion-based models? All those areas we are exploring in a deep, deep w…

AI assessment note: “we are pushing the research frontier in a much broader way than most other people”

Q conversation is NVIDIA has got the real market monopoly in AI is what everyone says. Um, do TPUs provide a wholesale replacement for your need for NVIDIA? In the supply chain, or is NVIDIA still a core part of the mix in the data center for training versus inference in LLMs versus other models? Maybe just share your, your understanding of where, like, the mix evolves to for you guys.

A Look, first of all, at a high level, NVIDIA is a phenomenal company. Uh, you know, Jensen is awesome. We, we have been working with NVIDIA now for a very, very long time, and we continue to do so, right? We serve a lot of the Gemini traffic on GPUs as well, right? And so we give customers choice, et cetera. Internally, we train our Gemini models on TPUs, right? And, and, and we serve it that way across our products, but, uh, we use both. And I do think, look, I, I do think everyone in the industry is going to try and do something like that, but, uh, but, you know, it's, it's, uh, you know, NVIDIA's R&D, uh, and, and their ability to, Drive that innovation. Uh, their software stack is world class, so, you know, they have a lot of, uh, advantages as a company, and I have, uh, extraordinary respect for them. Uh, you know, but we've always had, you know, we are committed, you know, we are actually deploying GPUs internally as well. I think I like that flexibility, and, and, uh, but I, I, I, we are also long-term committed to the TPU direction as well, so I think it's a good combination to have both. And, and I think we push each other, um, you know, and drive the frontie forward.

AI assessment note: “we are also long-term committed to the TPU direction... good combination to have both.”

Q now and 2040. China's going from three to eight, and there's probably upside given all the new electricity production technologies that they're rolling out now, which will be additive to that. How much is electricity generation going to play a role in who is going to economically benefit from AI over the next 10 to 15 years? And where is the U.S. compared to China? And maybe where is Google?

A Well, look, you are, uh, definitely, uh, hitting on, you know, what is, Yeah, when you, when you look at any system, you want to find where the constraint is, because that's what, like, gets the whole system, and you are rightfully identifying, ah, the most likely constraint for, ah, AI progress, and, and hence, by definition, GDP growth and all that stuff, right? So, I do worry about it a lot, um, but, You know, the answers are, you know, sometimes you run into challenges which are, you know, you have to solve, you know, you're running into physics barriers or something like that. This is not a problem like that, right? Like, we already know the technologies that can work to supply the demand we need. So it's more, to me, an execution challenge, right? I would phrase the energy problem as, Uh, it's obviously multifaceted, uh, but I think, I think be it really embracing, we shouldn't have innovators dilemma in the energy sector, right? So we should lean into all the possible innovations ahead, and there are many of them. Obviously, first of all, people perpetually, I think, will underestimate solar, right? You know, solar plus batteries will end up being huge. Uh, you know, obviously, uh, the amount of innovation that's going into nuclear, Geothermal. All of that are, uh, opportunities to, uh, embrace, and it's more, I'm not mentioning, but I think, you know, upgrading the grid…

AI assessment note: “you are rightfully identifying, ah, the most likely constraint for, ah, AI progress”

Q ended up creating a culture that Kind of moved away from more accountability and performance and was much more about coddling employees. Can you just comment on kind of your observations on the evolution of Google over the 20 years that you've been here and what you've tried to do lately as a leader, how you think about the culture you want to foster and what you're doing about it?

A Look, I think it's important to step back and say, you know, the underpinnings of a culture in which you really invest in employees and, and you empower them. And even some of the perks was to create a culture where it's positive, optimistic. You're in an innovation mindset. People are talking to each other. Maybe by giving lunch here, people are all sitting and talking ideas through lunch. You're cross-pollinating. Imagine. So, you know, that is the thesis of it, not that we are trying to give lunch to people, right? And so, I, till today, feel, you know, we still get a lot of innovation in the company, at all levels of the company, and I think people wake up, um, you know, and say, well, I can go do this Notebook LM, et cetera. Great examples, right? And so people do that all the time. So I think empowering employees Has been and is and will be a source of strength for Google, right? I think we can attract higher caliber people who feel like they have agency to do that, right? And, but that doesn't mean, like, you know, I think, I think people shouldn't confuse that with, like today, for example, you can take something like Google DeepMind, I think there is all the way from Demis and others, you know, extraordinary leadership team, be it Cori, Jeff, Oriol, Noam, et cetera. All these leaders have strong opinions on how to drive that frontier forward, and, and that's happening …

AI assessment note: “empowering employees Has been and is and will be a source of strength for Google”