Google CEO Sundar Pichai: AI’s Infinite Possibilities, Compute Scarcity Dilemmas, Future AI Products

We are at early stages of the most consequential technology ever. This is not like fiber. It's not dark. We are using it. AI researchers, for them, it's like the bigger the computer, the better the AI. You give the researchers the biggest possible computer, even if it is to the detriment of today's profits, like the cloud business. So the compute we have today is a function of decisions I made in 2024. Today, obviously, when you sit and look at anything, everyone is short, not just us, other companies, and everyone wishes you had 2x of what you had planned for. Where's Google's Muse or its DOT's product? Users, both in consumer and enterprise, will have access to kind of some multi-agent framework for them working all the time doing things. And so, you know, it feels like a moment of infinite possibilities.

Thanks, Alex. Great to be here.

I mean, very wrong. Look, I think we are so committed to being at the frontier. I think it's in our DNA. We wouldn't do it any other way. We've been thinking about the AI frontier for well over a decade now. We definitely, the frontier is very dynamic. It's very competitive. It's very exciting. I couldn't be more pleased with how Argon has come together. We took our time with it. You know, we made sure we tested it safely and deployed it carefully. We took a bit longer than our other models given its capabilities. But I think across the board, be it for agentic work, knowledge work, coding, creative tasks, I think Argon is, you know, one of the really good frontier models. And, you know, in the first 72 hours since we launched, it's, you know, almost up to 25 trillion tokens used, which is about 3x what Gemini 3.0 had in the first 72 hours. So, and we are capacity constrained as we are serving it. So, excited, but there are always areas, you know, both the frontier is just last... it's very dynamic and you have to keep evolving the models, and there are areas which we can see we need to improve as well. So we'll be at work on all of that as well.

Okay. On any given model, you know, Argon will be across the board better at everything if you just look even two months back. But the space is so evolving. But even on coding, for example, there are areas, pockets where we know we need to do better, and, you know, there's clear feedback, and so we'll make it better.

Look, it's definitely not how I have thought about it at all. I think we are deeply committed to driving the state-of-the-art and actually push what this technology can go, where it can go, all the way from transformers. That's the journey we've been on. So you will see us continue to do that. I think when you're at the frontier, you get an earlier look into the kind of emerging capabilities earlier than others, which helps you better prepare for the future. We are fortunate to have a wide set of opportunities across consumer and enterprise. Some of our cloud customers, for sure, want the model at the frontier. I think the frontier models will be better at developing the next set of models in an era of recursive self-improvement. So I would argue it matters there. Internally, we have definitely seen a steep point of inflection by deploying Argon to our developers. It definitely tackles problems which our previous models could not do. And so being at that frontier puts you at the innovation frontier and helps you drive that as well. There is room for models of all sizes and capabilities. We've always, that's why we did Flash, Flash Lite, Flash, etc. And I agree with you that you can develop very good product experiences with not necessarily the most frontier model because they tend to be slower and more expensive too. So I think, but we care about the parade of frontier, and so we are committed to doing all of that.

You know, the exciting, what makes this moment so exciting is you build something which you think is a culmination of like months of work, and you put it out, you know, you already see clear headroom ahead, right? I think, I mean, the frontier is going to, we've all been working on these agentic long-horizon capabilities, but clearly the frontier is going to have multiple of these agents working together to do things, right? And so that's an example of an exciting paradigm underway. And it's also clear to me that users, both in consumer and enterprise, will have access to kind of some multi-agent framework for them working all the time doing things. And so, you know, it feels like a moment of infinite possibilities.

Well, maybe three parts to your question. Let's first talk about RSI itself. And, you know, you're right, Sergey mentioned RSI being a critical contributor as we are developing Argon. But to be very clear today, you know, in a continuum, we're pretty far from the end state of RSI or something like that, right? So today, the right way to describe it is you have RSI loops along with AI researchers in various aspects of model development, right? So, you know, I would say you almost have agents helping AI researchers in some automated way work together to build these models. So I think it's good to at least understand where we are today, and it's super exciting, and I expect that continuum to keep making progress as we go into next year. The second question is around safety. Look, the way we are, I mean, security comes from defense in depth. You're talking about alignment. Obviously, you know, we all are doing a lot of work to make sure the models are as aligned as possible, but it's an active area of research and there are clear gaps, but that's not the only thing we rely on, right? So you have the work you do on alignment, but beyond that, you know, we put a lot of effort into real-time monitoring as these models are working. So making sure you have good telemetry into what's happening, and so that you can intervene in real time if you see unexpected behaviors, and how you do that, the architecture with which you do that, how do you make sure it's all sandboxed and isolated in a way that it can continue doing its job during real-time monitoring. These are all areas of investment we are doing, and obviously when we deploy these models, giving customers, enterprises control both with real-time monitoring but their ability to decide where they deploy, what type of data these models ever have access to, etc. So all that work is very, very important, and we are doing a lot of that with our all. So I think as an industry we need to constantly up our game and take the time to do all this well, including dedicating the compute required to make sure you can safely deploy these models. The third part is, you know, is it all enough? And, you know, so I think this is why I think, you know, there's a lot of conversations in industry, and I think it's reasonable to be worried, but, you know, I won't... right way to approach something you worry about is early on thinking about what more we can do. And I think all of us are, you know, I think as an industry, I think you're definitely seeing increased levels of work on alignment, testing for safety, etc. As we deploy these models, I think the next step would be developing better standards so there are cross-industry standards of how we do these things, including ability to evaluate using independent evaluators. You know, if you look at areas like financial controls, we have like accounting standards, we have independent auditors, etc. Right? So those are all the next natural steps for us as an industry to work towards this, and regulation will end up playing a role as we go through all of that as well.

You know, I mean, as humans, you obviously perceive the world around you, and, you know, you're taking all those inputs. You know, making sure these models are able to understand the laws of physics, simulate the world in high fidelity, all that I think are key enabling blocks for intelligence. And I think as you're working, we're working across all possible applications of this. You know, take an area like robotics, it's not an accident. Argon, if you look at, you know, embodied reasoning benchmarks, Argon is state-of-the-art, right? I think literally SOTA on many of these benchmarks. And that all comes from, you know, our experience building world models as well. So inherently, I would expect, you know, there will be architectures which will converge all of this in the future. But I think for us, we view this is today, you know, in this jagged frontier that is intelligence, we want to be at the frontier of all the possible approaches, and I think that will help us both drive better research, better insights, and develop a much better, you know, AGI model, you know, is how we think about it.

Look, I think we are, look, the ground zero for us is making sure we are investing in the compute that's needed to be at the AI frontier. So that is, you know, in 2015 I set the company to be AI-first. So AI is that key horizontal ingredient across everything we do, all the way from search to cloud to Waymo to everything we do. So I think that is ground zero for us, making sure both for research and developing the models that we have enough compute needed. Beyond that, look, we have these independent businesses which each have their own long-range plans, and so they're using those long-range plans to project forward, and we plan for compute for all of them. And so the cloud team, you know, they have a business and they have a projection and they have demand signals. They've committed contracts. So with that help, we can plan on a long-term compute trajectory for them. I think we're fortunate to have this wide set of opportunities. I think there's no company in the world which isn't making trade-offs at any given moment in time. But the first principles we operate on is to making sure we have enough compute to invest in the frontier. And that's non-negotiable, and everything else works on top of that, right? I think what you're seeing in these conversations is if you sit now, remember we are building our own compute, our own data centers. So there's long lead times. So the compute we have today is a function of decisions I made in 2024. Even if you think back to a year from today, last year the agent recording takeoff hadn't happened. So today, obviously, when you sit and look at anything, everyone is short, not just us, other companies, and everyone wishes you had 2x of what you had planned for. I think that comes with the territory, right? And like, you know, at any given moment in time, we are planning for what we would need two to three years out in the future. So I think you're going to find some decisions you make may look suboptimal two years out, right, depending on where the state of the art is. So that, you know, I don't spend too much time thinking about it. The way I think about compute allocation is making sure Google DeepMind, both for research and frontier model development, is cutting edge in terms of compute access, right? And after that, we plan for all our businesses. I think it's great we can provide our compute to other people too, like we have done. It allows us to scale that up, get real test for it in the widest application set possible, you know, test it in the real world and make sure we are at the cutting edge of technology at every layer of our AI stack as well.

You know, I think it's a, yeah, it's a reasonable, like, you know, if today a customer is coming out and asking us for compute in '27, that is incredibly difficult to do. You know, there's part of the reason even last year we went out and leased some short-term compute was to accommodate some customers and help them out. But yes, yeah, I think we are all dealing with similar timelines.

It depends on the time frames you're looking at. Are we, you know, this level of build, you know, if I take a long enough time frame, it makes complete sense to me. We are at early stages of the most consequential technology ever, and this is not like fiber. It's not dark. We are using it. We are capacity constrained on a daily basis. The applications, the underlying technology is improving rapidly, right? But running a company, the way I work through the moment is you look at your demand signals. You look at your committed revenue opportunities that you're getting in the forms of contracts. You balance that with the underlying technical capability, how we are using it internally, the model trajectories, and based on all that, you take a disciplined approach, a confident, aggressive, but, you know, a disciplined approach to plan out compute. But when you talk about the broader economy as a whole, you're having people operate with various levels of risk, and so, you know, so it's tough for me to, I focus on what I can control as a company and plan through it. If your question is, can there be an overbuild? Yes, because, you know, you can have some constrained components, and remember what matters is a square set of everything you need to build out that compute.

Yeah, like there are like 2,000 complements or more, right? And so you could get constrained at some point. There could be things which cause bottlenecks, and, you know, and then you could have overbuilds. There could be external shocks, right, to the economy, etc. So there are macro risks, all that is possible at any given time. But, you know, am I confident that the compute we are planning for, we will have uses for it given enough time? Absolutely, yes.

We're all doing some form of that, right? I think our cloud business is at any given moment in time foregoing short-term revenues because we are making compute trade-offs to support frontier AI development. So I think what you're saying is true already, right? And, but, you know, you want to be thoughtful, right? Even our AI researchers, you know, some constraints are also good. That's what inspires creativity, you know, that's what helps you come with a more efficient algorithmic breakthrough, maybe to use less memory, etc. So, you know, so what we are committed is to invest what's needed to stay at the frontier, right? That's how I think about it.

The way I think about it is, I do feel, you know, computing historically has been humans adapting to computing. With AI, it'll be the other way about, the technology will adapt seamlessly to how humans live and go about their day-to-day lives. So there's an aspect of it. You know, you want that AI to be able to speak to it when you want it to. It has context. It can be persistently running somewhere, give you continuity, you're not dealing with a lot of complexity. It's natural. All that makes sense to me. That doesn't mean it won't have an interface, however, to me. Look, today, for example, I mean, nowadays I just, even in my phone, like, you know, the Gemini voice models are so good, I can just press the microphone button, I can just ramble. That's how we call it, rambler, right? I can ramble, it kind of intelligently corrects with my personal context, so even my friends' names, like weird names, it kind of gets it right, even if it first guesses it wrong. It's amazing. It's adapt. I'm not going and backspace editing once I finish saying it, right? It's remarkable how good it's gotten. But there are going to be moments you will need an interface. You know, you may want to entertain yourself. You may want to watch a video, like, you know, so you may ask your AI to create something and you'll want to watch it, consume it, right? So I think the way I think about it is you want it to be as seamless and as natural as possible, and that you are not doing the hard work adapting to whatever the interface is. The technology is adapting to how you're working at any given moment in time.

Look, first of all, I mean, it's a very dynamic moment. We are all pushing each other. I think OpenAI has pushed us hard as a company too. So this is not just, you know, we are doing something to them alone, etc. Far from it.