The history and future of AI at Google, with Sundar Pichai

27:54You referenced the supply constraints, and I think that's a really interesting defining aspect of twin 26 basically where you said 150 billion in capex 180 we have said it's it'll be between 175 and 185 okay so 180 ish uh billion of capex and what's interesting to me is that google could not spend 400 billion dollars in capex if it wanted to because the memory isn't there and you know the power isn't there and all these components so can you just tick through you can find a number of electricians we would need exactly so i'd love to hear just your overview of the various bottlenecks? Look, at some level, you have to work back to actual wafer capacity or something like that, right?

28:36So there are deeper ground troops, right? I think so wafer starts. It's kind of a fundamental constraint. I think power and energy are more solvable. Permitting and actually working through a regulatory environment might be a constraint, right? It's the pace at which you can do things. Even though there's lots of land in pro-growth, you know, Texas or Nevada or Montana, just maybe not enough. I think we're making tremendous progress. I think for the U.S., I think it's a particularly important thing. You know, you're in awe of, like, the pace in China, how fast they can build things. So I really think we need to learn to build things much faster.

29:21Like, you almost have to shift your mentality to think about what would it take to do things 10x faster, right, in the physical world. Yes. Construct 10x faster. But I would worry about that as a constraint. I think there could be growing resistance, so it's not as simple as a few people deciding you want to build faster. The data center moratoriums and stuff. So I would say way for starts, the ability to permit and do things. And I do think there's a lot of good work being done from the government on. I think people realize you need to do these things better. Then comes critical competence in the supply chain.

30:02Memory is a good one. We are constraining those things in the short term. Everyone will respond to it. But I think all of us running companies, regardless of how AGI pill you are, then comes this error bans. of like, you know, how bullish can you be? What's the margins you can afford? Because there are extraneous factors which can go wrong in the world, right? Which are outside of control. So everyone is making those adjustments. Those are all constraints, right? And constraints. So I think that's where I see the constraints. Is memory the biggest component that you think about? Memory is definitely one of the most critical components now, yes.

30:48And you said in the short term, Do you think just people ramp up supply and so high prices will take care of us? There is no way that the leading memory companies are going to dramatically improve their capacity. So you have those constraints in the short term, but they get more relaxed as you go out. But I do expect all of this to constrain. By the way, I think it'll push a lot of innovations on we will make these things 30x more efficient. like so all that is happening simultaneously as well works. Does that enforce an oligopoly market? So if you actually look on the model side, because if you look at a lot of the views of models and how they're going to improve, a lot of it is going to be both self-improvement.

31:30So the models will start writing more and more pieces of themselves, do more data labeling for themselves, et cetera. So it's a musical chairs game of who has compute right now, basically you're saying? Who has compute right now and how much can you actually scale relative to overall industry capacity? And if everybody is roughly per rata up to some number, you've effectively put a ceiling on how much far ahead somebody can pull versus everybody else. Do you think that's a correct statement or an incorrect statement? I think it's a reasonable framework to think about it that way. But there are things which are, you know, I'm coming here as we just shipped Gemma 4, right?

32:03And it's a really good open source model. I mean, the Chinese models are very good. But I think outside of China, you know, it's a very good open source model. you know, the frontier to Gemma 4 is both huge and not so huge in terms of time. Like Gemma 4 is based on Gemini 3 architecture, right? You know, it's a very weird thing, right? You're talking about a set of weights which can fit on a USB stick. Yeah, yeah. So it's like a really, you know, crazy. It's not like a SpaceX rocket. I'm always shocked that you run a data center for months and months and months, and then your output is a flat file.

32:47Literally, it's like having a Word doc or something, and that's your model. It's amazing. So there are these unique attributes about this, which makes me challenge those frameworks and say, how should we think about this? But I think it's reasonable. At least on the inference side, what you're saying is a very reasonable way to think about it. Think about it. But I do think everyone is trying to figure out how to blow through the capitalist incentive to break through these constraints. It's immense. But as you say, there's only so much memory in the world. So no capitalist incentive will really solve 26 or 27 memory supply.

33:28That may be the era where you see more divergence. Yeah, so, you know, and remember, that has to balance with wafer capacity increasing, you being able to permit those data centers. So this constraint may be less severe than it appears, right? So you have to kind of envision the total square set of, like, all the things that you need and think it through, right? Incurating capital. Yes, yes. But again, what's interesting to me is that plausibly people would invest beyond the current CapEx, but we're now just running against 26 and 27 real world constraints. It's a little about the Strait of Hormuz.

34:10You can have whatever price of oil you want. Ultimately, if you take 20 million barrels a day out of the system, you need to destroy 20 million barrels a day of demand. And it's kind of similar with memory where ultimately some people have to not get the memory they want. But there are other constraints, right? Like, which, you know, take security as a constraint. And these models are definitely, like, really going to break pretty much all software out there. Maybe already we don't know if we sit here and speak. Do you really think all software there? Because, like, SSH, people have been trying to break for a long time.

34:45Do you think like... I'm talking about just thinking... Just regular software. ...open source software, large platforms, right? How many zero days? Yeah. You know, so there are constraints here in the system, right? You just can't wish away. Somebody was telling me the black market price of zero days is dropping because the supply is growing due to AI, which I thought was a really interesting market metric. Not at all surprised, right? And not at all surprised. So, but how does it practically diffuse through society? What are the implications of it? And so there are parallels, I think. So I think there could be hidden constraints.

35:20Yes. and there could be shocks to the system, if you will. But having said that, I genuinely think there's a lot of upside ahead. Some of the constraints maybe are helpful. Yes. Right, I think constraint inspires creativity. Or is it a compaction cycle where we get more efficient? Forces maybe important conversations to be had which otherwise won't happen. Right, I think, you know, just on my security point alone, Like I thought about we are going to need more coordination, which is not happening today. There will be a moment of, you know, it could be a sharp moment, right? And like, you know, and so all those things, I don't think you can wish them away.

36:04Yes, yes. Right. Yeah. Actually, related to that, Google does have an amazing portfolio of things that's both built and bought into. From an ownership perspective, you know, you own a reasonable amount of SpaceX. I think, I don't know the exact amount, but I think it was 10 % way back when. Anthropic, 10-ish percent, the majority of Waymo, which is like an amazing thing. And then internally, obviously, there's this enormous swath of amazing technology that's been developed. We talked about AI and transformers. There's TPUs. Obviously, Waymo is another one of these things. There's quantum. You know, you just released a very interesting result there.

36:36Are there other hidden gems that people should know about or that are especially interesting or that may have very big impact in the future? People may be underestimated. Look, we're constantly trying to take these long-term projects, which when you first announce them slightly marginally looks ridiculous. You know, like we're in the earliest stages of thinking about data centers in space, right? But to your earlier discussion around constraint inspires creativity. But if you take a 20-year outlook, right, where are you going to put most of these data centers? Really hard problems to solve. but those are examples of projects we think about today which are way more in 2010 like 10.

37:22Quantum itself is one of the one of these projects we are like in a deeply committed way making progress there and I'm excited about it. Where do you think quantum will have the biggest impact because mainly people talk about molecular modeling. They talk about cryptography. There's quantum proof sort of cryptography that people have been developing over time. On the molecular modeling side, it actually looks like the deep learning models tend to be very good at that in certain circumstances. I mean, you all pioneered that with AlphaFold. Do you think quantum will actually matter? And if so, where do you think it'll have the biggest impact?

37:58Look at abstract level, to me, it feels like to simulate nature more and more. Like, you know, like given it's inherently quantum, you would need quantum systems to better simulate it. We may get there with classical computing techniques in a surprising way or get at it with enough compression and, you know, abstraction. It may work. But I fundamentally felt like quantum would have an edge there. And I don't know. we still don't understand the Haber process for fertilizer. There are many components. I mean, you know, it's probably your background, going back to what you did in college, more. So my, you know, my instinct tells me there'll be, you know, simulating weather, simulating, you know, reality, all that, I think, quantum 11 advantage.

38:55I think the way the history of technology is, you get something to a scale where it works, and then you use it and people's creativity on the top finds the applications. So, you know, I mean, I always give this example of mobile phones plus GPS enabled Uber. Yeah. Like, there's nobody who was working on phones who would predict that as an outcome of this platform shift. So, you know, confident quantum will have many, many, many applications if you can actually make it work. Yeah. So that's how I think about it. But sorry, we interrupted you. You were talking about kind of your favorite of the Google further afield.

39:36I think we're making, you know, the GM team is deeply thinking through robotics, right? And, you know, robotics is an area where we were too early as a company before. It turned out AI was the missing ingredient for a lot of ideas maybe 15 years ago or 10 years ago. but you know the Gemini robotics models are sort of on spatial reasoning etc so we definitely have state-of-the-art models there and we are partnering uh back in an ironic way with Boston Dynamics and and Agile and a few other companies and and in a determined way making progress and there are extraordinary startups out there as well but so we are investing in you know I spoke about quantum data centers in space, drone delivery with Ving.

40:30You know, I think we are scaling up Ving where in some reasonable time period, like 40 million Americans will have access to a Ving delivery service, right? And I'm not talking years out or something like that. But again, these are all like methodical compounding when you take these long-term projects. So, you know, we are committed. Isomorphic. Yeah, isomorphic is very exciting. Think about being focused on these models in a targeted way to improving all the possible steps in drug discovery. And even though you have long pulls like phase three trials, et cetera, getting there with a much higher probability of success.

41:14Yeah, I think it's definitely the smartest approach I've seen in terms of the different biomodels and really thinking about the broader swath beyond just the molecular design, which is, I think, where most of them are stuck. Yeah. It seems very smart. Can I ask, I'm curious, how capital allocation actually works at Google? And what I mean by that is, you know, the idea good capital allocation is about internalizing the opportunity cost for capital and putting the cash that a business generates towards its highest invest use. And in the toy example in a business school book, you know, maybe you're Boeing, and we can either, you know, we have this cash that our business generates, and we can either go bid on the next defense contract, and we'll invest this much in R &D dollars, and we model this much revenue from the contract, or we go develop a clean sheet commercial airliner, and we'll put in this money, and we model this kind of thing.

42:05It's like a 16 % IRR versus a 19 % IRR. Okay, I prefer the 19%. In Google's case, the projects are extremely heterogeneous, where it's like, okay, We can give the YouTube team more funding so they can go improve the recommender algorithm and therefore time and size increases, and so does monetization. Or we can give the Waymo team more funding so that they can actually get to market faster or scale up faster. Or we can invest in this new AI approach that might pay off in five years' time. And so I'm curious, if you are trying to put capital towards the highest and best use and you're ultimately comparing, how do you compare initiatives that are so different in nature and so different in payoff curve shape?

42:49This is the most John question ever. I need to know the answer. You need to throw an RIC and then it's like... It's a good question. Look, I feel it today more than ever, ironically, because of TPU allocation. So in some ways, I feel it even way more than TPU. That's interesting. Computers made the question, ironically, much more front of mind. By the way, of all the things I do, I'm really looking forward to how AI, as a companion, at least gives inputs to this task. And I think once we can actually get all the data connected and flowing through, models are already capable. It's more getting all the data unlocked.

43:30I think it will be helpful. So I feel it there. historically I think at Google one of the advantages we have had is sometimes we make these decisions very early in the cycle so it's almost like going back to that truth as a deep technology orientation and you know we actually think about the question you were asking a lot a bit ago about like what are those longer term things and so I think thinking at that stage it's easier because your initial funding amounts can be smaller but then like you know you stay committed for the long term, but you're making sure you're making progress in a deep way.

44:08So as long as you're seeing that underlying technology, like take quantum, for example, how do we judge it? Like we're judging the underlying, like, you know, so you have goals around, you know, what logical qubit error corrected, large stable, logical qubit threshold by when you're going to get to and is the team able to do that, right? So I think you assess it that way. So one of the, I won't say advantage, I think one of the ways we have thought about it and we've been disciplined about, or at least to me, matters a lot is to make those early technology bets in kind of a deep way. And so that's helped.

44:41But on a constant basis, look, I always view it as you have to assist the long-term value of these things. Right? So it's almost like in some intuitive way, you're thinking about the option value and the time of something five to 10 years down the line. and you assume like a crazy growth, right? And think through whether those decisions make sense. So the TPU investments have been great that way, right? And, you know, we've steadily invested in that. Waymo was a great example where I think we increased our investment two to three years ago when the rest of the world got pessimistic on it. When others, some of the people were backing off.

45:23It's very magical. It's such a magical experience. I take Waymo now every day to work when I can I think Waymo is a good example of this question I have, which is Google does cut projects. And there's various things you've tried where you said, we're actually not going to fund this part of X all the way. Or we're going to retire this product that's not working. But Waymo, despite the fact that it was a long road from a compelling demo to commercial service in market, you guys didn't lose the faith. And so what was it that you were seeing? Is that a qualitative decision or a quantitative decision?

45:59How do you decide that we're going to cut Loon, but keep Waymo? I think it's to do with that some kind of quantified, you look at the Waymo driver. That's underlying technology, which, you know, how does the software drive the car? And the progress in terms of safety and reliability. So it's a long running task, how safe and how will you do it? And you follow that curve. and you predict or you set goals where you want to be and how you perform against those curves. I think the team has been phenomenal. There have been maybe phases where it didn't progress, but those are the times you need to kind of like, you know, you have confidence in the quality of the team to break through those phases.

46:42But I think the more you're able to evaluate things at that deeper technology level, I think you tend to make those decisions better, or at least that's how I have tried to do it. One argument I've heard or one discussion I've heard made about Waymo is that a lot of the huge gains that have been seen recently because it used to be this hand mapped heuristics of like how do you deal with edge cases of driving or something happens, how do you respond? And a subset of those were almost like hand drawn out for the cars to follow. And so I had kind of a narrow set of things that it could do. And then really the breakthrough was moving to end-to-end deep learning a couple of years ago as this big transformer wave was happening in general.

47:20Do you think if Waymo had been started five years ago, it'd be at the same place as it is relative to having been started 15 plus years ago? Just given that that's the breakthrough that's kind of propelled it forward. Look, I think, you know, we spoke earlier about robotics. You can think about Waymo as a robot, right? I think people who are starting robotics in the last three years, by definition, would be making faster progress maybe. But I think Waymo is such an integrated system. There are aspects of it, not quite like, you know, like you take something complex like TSMC or SpaceX launching things.

47:56You are talking about system integration and these things in a very complex way. I think Waymo has hidden aspects of that, which the time of how you do it, the craft of it matters. But having said that, I do think the end-to-end approaches are going to be an accident. because just having a team arguably was a huge benefit to Alphabet and Google, right? I mean, just the fact that you kept investing in it and then it hit a moment in time where this technology liftoff was more than worth it and was very smart and forward thinking. I just think it's interesting to ask, how does that apply to other domains?

48:31Because to your point on robotics, it seems like with robotics, we'll potentially have a different history where you can move very quickly now. Do you folks think about re-internalizing hardware again, or is it largely going to be a partner-driven model to bringing this stuff to the world? I think we'd keep a very open mind. My lesson from Waymo and on the AI side with TPUs, etc., I think to really push the curve well, particularly in areas where you have safety, regulatory, everything. You want the first-hand experience of the product feedback cycle. So I think having first-party hardware will end up being very important.

49:14That's how I would say right at this stage. Makes sense. So I have two more capital allocation questions. Can you make the case that Google has historically been under-levered, where Google has historically carried a strong-neck cash position? And given that both Google has more ideas than it knows what to do with, like it's just brimming with good ideas, and just the core business grows very durably, And I think Google clearly has a very good understanding of that core business. And it has grown at a higher rate than Google's cost of capital. As you look back on it, should Google have been more leaned in and said, OK, we will be willing to have a leveraged position that's slightly more aggressive than strongly net cash.

49:55And we will put that towards new initiatives or just buy more of this core Google business for Google shareholders or do more minority investing, which again, Google seems to have been best in class at. It's a great question. For example, if Waymo had reached this point earlier, I think I would have invested the capital earlier. So to some extent, I think you were judging it by, like you want to be good stewards of capital. So to the extent you're bullish on ROIC, you want to invest every last dollar you can there. but to the extent you have access where you don't think. I mean, this is why we've invested in other companies too, even if not then, but we've always thought about it with the lens of being good stewards of it.

50:45We felt our investment in Stripe was being a good steward of our capital. SpaceX, right? You know, SpaceX and Anthropik and so on. So I think now with the AI shift, there are more opportunities on which we can deploy capital in a good way. And so we are doing that. Yes, yes. But I think we always had that mindset. Yes. But I would have been glad to invest more capital in Waymo earlier. But we weren't at the level of maturity needed to do that. There was a point in Waymo from a safety standpoint, you know, we did approach Waymo safety first. Yes. And you just, it wasn't the right thing to do. So you feel like you cannot point a project where they would have gone faster had they gone more capital sooner.

51:33They just needed a, they had a natural ramp. I wouldn't say that, but I think in generally at least we might have gotten the decision wrong, but our approach at least was like to say, if we got excited about something and had the conviction, we were willing to come in the capital to see through. My other capital allocation question was, historically at tech companies, the large majority of the R &D expense was the people walking around the building. And, you know, headcount was managed through a very tightly controlled process. And indeed, as you thought about kind of allocating R &D effort, it was really allocating kind of highly paid people to go work on the challenge.

52:12And the tech costs were, unless you were doing something very computationally expensive, which obviously Google did in place, you know, Google Books or something. But broadly speaking, the tech was an afterthought compared to the cost of the people. We're now going to a world where, as you say, that's not the case with TPUs and how you allocate that. Just at a very concrete budgeting level, how does that work inside of Google? Do you have an overall TPU budget for the company? And then when you are giving a project resourcing, previously you gave us a certain headcount budget, and now you give it a headcount and a TPU budget, are they the same budget?

52:52Just how does that work when you're doing a quarterly review or an annual review? Look, we've always had a compute budget. Ask me for a friend. Ask me for a friend. Now, we've always had a compute budget, right? You know, even classic compute. I would say with ML, and we use both TPUs and GPUs, by the way, extensively. But ML compute planning is, we are super thoughtful about headcount planning too, but we've always had to plan that. And ML compute, we've gone through phases where they've been easy. And then there have been phases where we've been constrained as a company. But now it is really acutely constrained, right?

53:32So you spend a lot more time. I at least spend a dedicated hour a week thinking about that question at a pretty granular level. So I will know by projects and by teams the compute units they are using, right? or at least I have that information and I'm looking at it and assessing it. And in some ways, it's a really important thing to be doing right now, I feel. So the scarce resource is compute in a lot of cases. And so you're ensuring that Google's precious compute resources are being spent on the most worthwhile initiatives. How do you think about that in the context of GCP and Google Cloud?

54:20Because there you're actually allocating the compute to others instead of for your own purposes. And given the constraints in the system, how do you think through that differential allocation? Look, Amy, plan ahead, right? So when we do the forward planning, the cloud team is forward planning and they're putting a plan in place. And so you're funding that and you're doing that for our internal needs. You forward plan. And as part of that, but you're also saying long-term commitments to customers. Anything we commit to a customer is sacrosanct, right? So these are contractual commitments. So you solve a lot of it with planning.

55:04And so there are, when you plan, we're all in a constrained world. So I think the cloud team would say they don't have the compute they want, et cetera, et cetera. But you solve it with planning ahead. Speaking of Google Cloud, I have my product requests that I've been saving up for this section that I know you're looking forward to. You're going to post it on that. Exactly, yeah, yeah. We're taking care of it. But no, I'll say one thing that works really well is the GCP MCP is awesome, where your AI can just interact programmatically with Google Cloud. And I guess you guys have exposed almost everything except like the core, you know, permissioning stuff.

55:39And I feel like, in a way, part of the curse of Google Cloud has been there is so much functionality there that I'm sure you occasionally hear from people that it was like a little hard to navigate, that you log in, you have to create an organization, a project and whatever and find the right services or whatever. And now, all that doesn't matter. And so you just say, hey, go add this Google Cloud functionality. And so that is something that actually, it feels like Google Cloud is really benefiting from. It is so broad and there's so much functionality there. I mean, we have a little bit of this problem with Stripe where as we add more functionality to it, just the right way to navigate this big product surface area is an AI that's read all the API docs for you.

56:15So that's working really well. The promise of AI being this orchestration layer, like for anything you think about. To my earlier question, even internally within the enterprise as a CEO, it's not like you don't have all the data, but how do you get it in one place and you see it? In the past that would have meant one more big ERP-ish project to go connect all the data sources, etc. Again, like, you know, AI being this orchestration layer in a way that makes sense for the end user, I think has been delightful to see. And the bigger the product surface area, the more that benefit, you know, hits you.

56:49And again, we've seen that to some extent with Stripe, but I feel like with GCP, it must be just a massive effect. I think we could do a lot better. So, but you're right. It's an immense opportunity, I think, yeah. I've been really happy with it. Okay, and then that gets to my product suggestions. Did you bring product suggestions for us? No, you go first, yeah. I wanted to, but... What's interesting to me about kind of open claw and the product market fit of things like that is they're allowing stateful AI for consumers. And if you want to say, you know, the classic, you know, round up the daily news that I'm interested in and send it to me each morning, or just something that involves persistence that none of the popular, you know, or like mainstream AI apps allow persistence?

57:28Is that common? I think directionally, look, I think you want to give users capability where you have persistent long-running tasks in a reliable, secure way. You know, you have to think through things like identity access, et cetera. But I think that's the future. That's the agentic future. And bringing that for consumers is like a bit of an exciting frontier we are looking at. Yeah, this is what I'm saying. This is Dreamer, which was the former CTO of Stripe's company that just got bought by Meta. I think they did a very good version of this. It's a very early kind of view of... Yeah, they were making custom software, including persistence, but also, you know, you could just kind of spec out.

58:14Kind of make your own little app. Exactly, yeah, yeah. And they made that very easy to use. But I feel like when people have this experience, there's a surprise and delight moment. And it's just interesting to me that... Look, I think effectively the consumer interfaces are going to have full coding models underneath, right? And the right harnesses and the right skills and the ability to persist and run somewhere securely in the cloud, locally and in the cloud. So all those primitives are coming together. And so what developers are, like today, I feel like there's 1 % of the world, maybe not 1%, 0.1 % of the world, who's kind of living this future.

58:56They are building stuff for themselves, but bringing that to mass adoption is a very exciting frontier, I think. Okay, my other product suggestion is, sorry, you have to enjoy this part of the interview. It's a right of passage. Exactly. My other product idea is for some reason, I don't know if this is your lived experience, but certainly my lived experience, that searching Google Docs is so much harder than, say, searching Gmail. And obviously, they're both equally good search engines. But I think what's going on is keyword search works reasonably well for email, because you can probably remember a unique set of keywords for that email.

59:36Whereas what always happens, at least to me, is like, I want to go back and look at the 2026 budget. It turns out if I search Google Slides for 2026 budget, neither of those words is particularly unique in the context of words that exist in PowerPoints at Stripe. And so I can never find the exact right one. And I'm curious, does Sundar Pichai also have this problem? Somehow I haven't felt it as acutely as you're describing it. But when you describe it, it resonates well with my experience. I'm literally playing through the person to whom I'm going to play this segment of the conversation. I know exactly who I'm going to go talk to.

1:00:13The people are working on it. I think we can make it a lot better. I think the AI integration into these services, including Google Docs, I think you will see sharp improvements in the coming months ahead. I think we all did the first versions of it where you just put it in somewhere. But I think, you know, over time, what all can you keep in context? What can you cache? And what can you really bring to bear? I think we can make a lot of progress on. So I think we can do a lot better. Okay, great. We have a good action. A lot of companies that I'm involved with, even ones that were started reasonably recently, have had to dramatically shift their workflows relative to product development, engineering practices.

1:00:55Who they even think of should be on the design team and the capabilities of that. Are you revisiting all that at Google? Are you rethinking it? Has there been big shifts in workflow or other aspects? The way I would say it is, you can think of it as concentric circles. There are some groups that's in Google who are shifting more profoundly. And so for me, a big task is how do you diffuse that to more and more groups, particularly in 2026? Some of it, we couldn't do it early because it breaks so often that like, you know, it's almost like you see this promising new world, but it's kind of semi-broken.

1:01:29But this year, I feel like the curve is shifting pretty dramatically. So I can see groups, particularly I would say GDM and some of the SWE groups really changed their workflows, right? And, you know, they are using, we call this for some strange reason, we have a different name internally than externally of the same product, but it's JetSki internally, which is anti-gravity. And you're living on it, you're living in an agent manager world, your workflows, and you're kind of working in this new way, right? but just last week we kind of rolled it out to the search team right so we're constantly pushing that you know in a large organization i think change management is is a hard aspect of this technology diffusing which may be easy for a small company right okay you know you can quickly switch over can i lay out a few um problems i see when it comes to actual diffusion of ai in industry and I'm curious how and when you think we'll solve them.

1:02:29Because as I see it, we have a big intelligence overhang. Like the AIs are now amazing in terms of what they can do in the abstract. And if you look at how AI native a company is or just kind of how much it uses that intelligence, there'll probably be a shortfall. And the problems that I see are something like, one, it actually takes a while to get good as an engineer at prompting your AI well. And you can prompt an AI better or worse to write code. Then there's a lot of, say, Stripe-specific prompting in our case to know which tools to use. And so there's kind of the general being good at prompting, and then there's the Stripe being good at prompting.

1:03:06And then, of course, you have the fact that it's hard to share an AI-generated code base because you have a blast radius, and you're just changing so much, and the turnover of the code is high enough, or maybe you're rewriting it several times before you ship, that it's kind of hard for many people to collaborate on the code base versus before when the code velocity was slower. And then as you go outside of engineering, the big one I see is access to data, where you'd like to have your agent go, how many times a day do people at companies around the world say, hey, what's the status of this deal?

1:03:37And that is like information that the company knows and should be agentically answerable. And we actually have some cool stuff at Stripe where I was seeing where you can actually answer that pretty well. But with both habits and access to data, and as you get into a bigger company, the permissions engine of who can actually get access to this data, that all needs to be rewritten. And then you get into role definition where kind of like you were saying, NGPM design kind of stems a little bit from a prior year. And you may want to, at least in some cases, merge those roles a little bit as AI gets better at all those since you've got a product door.

1:04:13Anyway, that's kind of my characterization of in 2026, the models are capable of this, but we're only using them so much. What do you think that adoption of the intelligence looks like? Look, a lot of us are working on, like, literally what the Gemini teams, the Gemini Enterprise teams and the anti-gravity teams, they're precisely working on these problems. This is the roadmap you're talking about, right? Like, you know, and that's literally we are using it internally, running into these barriers, kind of working past it. So that's the products that are shipping. We are still diffusing it because what you do is people, as part of using it, like if you're the SRE team at Google, you suddenly find portions which you can create an automated workflow.

1:05:08And so that's happening in like these parts. Right. But doing it more systematically when you develop skills, how does it get centralized? How is it available to the models and for everyone to use? Identity access controls are like real hard problems. And so we are working through those things. But those are the key things which are limiting diffusion to us too. Right. And we take security a lot more seriously. And so we have to. Right. So that is another layer on top of all these things. The cost of mistakes when you're running these services. And so we have to work through it. But I think because of it, when we solve it, I think we will bring it in a more robust way, which will help.

1:05:52So I feel like we're going through that fixed cost right now. but you will see these jumps of what people are able to do when we bring it outside and other others are doing it too and and in a more robust way the models are improving google um re-forecasts its business a few times a year formally i presume at least we do at stripe where we you know we set a budget for the year and then three times a year we produce a formal re-forecast and when you think about it a re-forecast is a moment in time function where you take the state of the business some of which is in people's heads, but most of which is written down everywhere where it's like, how is this product doing?

1:06:27How is that product doing? Will this deal close? Will that happen? Whatever. So there's like the moment in time stage of the business. We put it into a function and out comes the updated numbers for the year. You can imagine an AI doing a fully no human in the loop forecast. What quarter do you think Google's first fully agentic forecast is? I definitely expect in some of these areas, 27 to be an important inflection point for certain things. Even the people doing it, that is the workflow through which they would produce it. And maybe for a while, you would check it in the conventional way, but you kind of switch over, cross over.

1:07:14But I expect 27 to be a big year in which some of those shifts happen pretty profoundly. I think that was Alad's question was Eng is an early adopter, but kind of outside of Eng. And okay, it sounds like you think 27, a lot of these non-Eng processes really start wrapping up. I do think your question earlier on, like, you know, I think you were asking in the context of way more robotics, like companies. I do think companies which are, that's one advantage startups are going to have. More AI native teams. And, you know, you can probably get at it through your interview processes, et cetera. Whereas for us, we would have like retraining, transformation, et cetera.

1:07:52And I think that that's maybe an advantage like the younger companies are going to have. And we have to, you know, kind of like drive the transformation. Last question. We're talking a lot about initiatives that started small at Google, like the Transformer, which is not Google's main priority, you know, when that initiative started. What's a small thing inside Google that you're excited about these days? It probably would surprise people. When we decided to do data centers in space, we started as a very small team. It's literally a few people with a small budget to go to the first milestone. So I think it's important to start small, even if it's a big idea.

1:08:34So that is an example of a small thing. look I literally spent time yesterday who was explaining some improvement in post training like which is like one person talking through the improvement they are doing listening to it I'm like oh that's going to like really show up like a nice jump right so that's the constant power of this moment and so all of that I don't want to be specific about the second one but we'll publish it one day I'm sure you know But those are some of the small gems I'm excited about. Because it did send us in space and new ML techniques. Yeah. Great answer. Sundar, thank you.

1:09:15All right. Real pleasure. Thanks. Take care.