Pichai Reviews Gemini 4 Argon: Why 3.5 Pro Was Cut

Written by | Key Points

Within 72 hours of launch, throughput approached nearly 25 trillion tokens, with consumption 3 times that of the previous generation Gemini 3.0 in the same period.

This is the debut scorecard delivered by Google's latest flagship model, Gemini 4 Argon. Even Google, with its deep resources, had to once again implement compute throttling in the face of the sudden surge in invocation demand.

In the latest interview on the tech podcast "Sources," Google CEO Sundar Pichai reviewed the release details of Gemini 4 Argon, publicly explaining for the first time why Gemini 3.5 Pro mysteriously disappeared, and the behind-the-scenes reasons for the changes in DeepMind's core management.

1. Cutting 3.5 Pro to sprint directly to Argon, internal R&D has entered the RSI stage

Over the past few months, the outside world has been wondering where Gemini 3.5 Pro, mentioned at the I/O conference, went. Pichai gave a direct answer in the interview: in order to seize the high ground in Agent coding capabilities, the team proactively abandoned it.

In the middle of this year, while training the 3.5 series, Google found that the model had an obvious gap in Agentic Coding. The team judged that using the smaller, more efficient Flash version for high-frequency iteration was the fastest path to close the gap. So Google intensively advanced continuous iterations from 3.5, 3.6, 3.7 to 3.8 Flash over about 3 months.

When this technical approach was proven and ready to be applied to the Pro version, the pretraining of Gemini 4.0 Argon had already been completed. Considering that releasing 3.5 Pro at this point would give it an extremely short lifecycle, and in a compute-constrained environment, the team decided to inject all the technical accumulation validated by Flash directly into Argon's post-training stage.

This decision directly drove Argon's leap in long-horizon tasks. In external tests, Argon autonomously completed "Minecraft" in just over a day at a compute cost of less than $2,000. And internally at Google, it has already been pointed at the Chromium codebase, which has millions of lines of code.

As for the externally rumored Carbon model, Pichai clarified that it was merely a codename for internal frontier model iterations. The team is currently advancing model distillation based on Argon, developing lightweight versions with near-capable performance but lower cost, in order to open them to more users.

As the new model is integrated into the internal development environment, Pichai believes that 2027 will be the golden age of productivity. He revealed that Google internally has already entered the recursive self-improvement (RSI) stage of deep collaboration between researchers and models: the model is directly participating in proposing training recipe suggestions and automatically evaluating the next checkpoint. Pichai himself also uses an internal Agent every day to summarize 24-hour R&D progress across pretraining, mid-training, post-training, and data groups, saving about 1 hour of information synchronization time per day, and uses an Agent connected to telemetry data to monitor the compute utilization efficiency of each department in real time.

2. First comments on DeepMind's senior management reshuffle, the game behind negative free cash flow

Regarding the recently closely watched Google DeepMind personnel shakeup: legendary engineer Jeff Dean leaving to start a company, Demis Hassabis stepping back to Chief Scientist and Chairman, and Koray Kavukcuoglu taking over to lead DeepMind, Pichai gave an official explanation for the first time.

In addition to Jeff Dean, who served for 27 years, wanting to embark on new explorations and found Discovery Loop, the core reason for the internal adjustment is that frontier model R&D has evolved into an ultra-large-scale engineering effort that consumes all one's energy.

Pichai pointed out that from data preparation, compute scheduling, and algorithm evolution to coordinating collaboration between human teams and model Agents, frontier model iteration is getting faster and faster, requiring a dedicated leader to devote all their energy to coordinating daily operations. Koray has worked at DeepMind for over ten years, and as early as 18 months ago moved to the Bay Area to advance cross-departmental integration, so it was natural for him to take over the model R&D pipeline. Demis, meanwhile, was freed up to focus on implementation directions such as AI drug discovery, life sciences, and robotics represented by Isomorphic Labs.

The acceleration of the model race is also directly reflected in the financial statements. This summer, Google saw its first negative free cash flow since its IPO, while continuing to raise its 2027 capital expenditure guidance.

Explaining why he dares to invest so aggressively, Pichai did the time math: the lead times for self-developed chips and data centers are extremely long, and the compute level the market assesses today with perfect information is actually a decision made two years ago, while at this time last year, the industry had not even seen the explosion of Agent coding. Now, Google is already planning land and power reserves for 2028 to 2029.

He emphasized that today's AI infrastructure is not the idle dark fiber bubble of the past; all compute is being genuinely consumed. In addition to demand signals from consumers and enterprises and the long-term contract backlog in the cloud business, breakthroughs in underlying R&D have also given him sufficient certainty. Pichai directly teased in the interview: based on current internal benchmarks and evolution curves, the next generation Gemini 5.0 will be much stronger than the current 4.0 Argon.

3. Underlying shared micro-VMs, space TPU and quantum computing timelines surface

On product architecture, Google is blurring the boundary between enterprise and consumer Agents.

Pichai revealed that whether for enterprise or consumer Gemini Agent, what runs underneath is the same model, the same operating framework, and the same runtime; the only difference is whether enterprise context or personal context is injected. Google's design on the enterprise side—automatically allocating a separate micro-VM for each task and achieving secure 24/7 operation—is being rolled out to the consumer side in parallel. In the interaction interface, Gemini's existing "Chat" and "Tasks" will be merged into a single input box.

Regarding when Google Search and Gemini will merge, which the outside world cares about, Pichai said there is no need to forcibly merge the two products; the cloud-based 24/7 online Agent layer will share the user's personal context, naturally take over cross-device tasks, and let users switch seamlessly among Search, the Gemini app, and Gmail. At the underlying Web interaction level, he believes that Browser Use, which operates interfaces like a human, and native protocols such as MCP and UCP will coexist for a long time.

The following is the original interview text:

1. Argon launch reception and access strategy under compute constraints

Host: It has been about a week since Gemini 4 Argon was widely released. I'd love to hear your thoughts on the early market reception of the new model.

Pichai: We are very excited about it. We had been using it internally for some time before this, and it is absolutely a world-class frontier model. Today's frontier models are very comprehensive in both breadth and depth of intelligence. If you make a heat map of intelligence levels across dimensions such as agentic tasks, knowledge work, and coding, I think it is very strong in many dimensions and extremely competitive. In the first 72 hours after Argon's release, we processed about 25 trillion tokens, roughly 3 times the 3.0 version. Because it is early in the launch, we are indeed somewhat constrained on compute. Customer feedback has been great, especially on coding and agentic capabilities. Of course, I am also aware there are areas that need improvement. In the frontier model space, you keep shipping, gathering feedback, and then moving to the next step—this is the most dynamic period I have ever seen.

Host: What areas do you think you still want to keep improving right now?

Pichai: We want to continue to refine coding capability. This is a very good model, and our engineers really enjoy using it internally, but we know very well that at any moment there is room to do better. After release, we have been continuously pushing forward internally, optimizing behavior based on user feedback.

Host: You just mentioned compute constraints. It is interesting that a company like Google also encounters compute bottlenecks with a new model like this. Argon is a large model, and operating costs are obviously high. Can you talk about how you decided on access restrictions for it? Right now it is only available to the API and AI Pro and Ultra subscribers, and obviously it has not yet reached the broadest user base—for a model of this scale that is indeed a challenge. How do you think about this kind of restriction and release strategy?

Pichai: We accumulated extensive distillation experience during the 3.0 model period, enabling us to extract models that are almost as capable but far more efficient. We are currently developing these versions, which can help us make it broadly available to as many people as possible. At the same time, for this frontier model, safety is crucial. We invested a great deal of effort in external testing, for example through the U.S. government's AI Safety Institute program and by providing early access to cybersecurity researchers, while also conducting extensive internal testing before releasing it externally. As you said, this is a large model, but given its scale, our inference speed is quite fast. Starting now, I will continue to expand access. In the Gemini app we first launched it to all Ultra and Pro tier users, then released the Pro version in the Anti-Gravity app, and we are developing distilled versions of the 4.0 series that can serve more users.

Host: Is that the rumored Carbon model?

Pichai: At any given moment, we have internal versions of frontier models, some of which are just code names; this is not inconsistent with us releasing a more efficient version of Argon.

2. Safety testing, competitive landscape, and long-term frontier positioning

Host: So a more efficient version is coming soon. Regarding the safety testing process, I am curious how you think about it. It is interesting to compare how private frontier labs like Anthropic and companies like Google, Meta, or large cloud providers approach safety and privacy. It sounds like your safety testing for this model was very rigorous and seems to have taken longer than for previous models—can you say that?

Pichai: Absolutely. Given its powerful agentic capabilities and extremely strong coding level, especially when deploying it at scale to enterprise customers, you need to make sure there are no mistakes. For example, beyond model alignment work, it is even more important to build compute and systems for real-time monitoring of the running model. This real-time monitoring is crucial for deployed models, allowing you to detect any behavior that does not meet expectations as early as possible and intervene in time. To that end, we took time to build the corresponding systems, sandbox environments, and real-time monitoring mechanisms, and to provide customers with the protections they need, ensuring they know where the model runs, what permissions it has, and what telemetry data is required.

Host: How closely do you follow this tug-of-war among new models? I mean, I know you certainly follow the dynamics, seeing who is ahead and who is behind. Everyone exclaimed last year when they saw Gemini 3 that Google was back in the frontier ranks. Now looking at benchmarks, Argon is indeed at the top. But it feels like everyone's attention is always shifting to see who is in the lead. How closely do you follow this? How much do you encourage the team to pay attention to it? I'd love to hear your current thoughts.

Pichai: That's a very good question. This is an extremely fast-changing period, and there are a lot of voices outside. Of course I pay attention to what's happening on X. Getting outside voices is very important, because it lets you understand what you're doing well and what needs improvement. Although I watch all of this external feedback, what's more important is that we've built a very rigorous internal evaluation system to assess models. We can clearly see the performance curve the model is on, so we get high-quality signal internally. What I ask of the team is that we are firmly committed to staying at the very forefront of AI technology over the very long term. We didn't just enter this field recently; we've been focused on it for a long time, and we hope to still be at the forefront of intelligence ten years from now. That means we need to do all the necessary work, including safety and security, which is essential for developing cutting-edge intelligence. One of the most important things for us this year is accelerating our product release cadence. The industry is changing extremely fast, and frontier models are evolving rapidly as well. So one of the big adjustments we've made this year is achieving a deeper breakthrough in agentic coding and using that to optimize our release cycle. From 3.5, 3.6, 3.7 to 3.8 Flash, we completed this series of iterations in about three months, so with version 4.0, we also expect to maintain and increase this release cadence.

3. The Leap in Agentic Capabilities and the 2027 Productivity/RSI Outlook

Host: There are some aspects of Argon that interest me. For example, I saw that it independently completed Minecraft in about a day or so, and the compute cost was reportedly less than $2,000, which is very impressive. I'm curious about what this level of agentic capability means. Games are one thing, but you can extrapolate this to many potential things, such as high-level decision-making and long-horizon tasks. When you saw the Minecraft case, what did you think? How does this affect your long-term planning?

Pichai: Internally, we've already witnessed the step-change in productivity that Argon brings. Since its internal deployment, it has been applied to many large codebases. For example, in Chromium, if you want to create an experience, such as dragging a tab and dropping it into an AI experience while keeping it running in real time without reloading, in the past the team would have needed to work through all of Chromium's underlying architecture. But now, you can just have Argon handle it, and it can complete the entire process patiently and persistently. We see examples like this everywhere. What excites me is that I think 2027 will be our golden age of productivity, thanks to the new generation of models and their deep integration with our internal environment. You can also apply these models deeply to the models' own R&D, which the industry calls recursive self-improvement (RSI). These are real areas where we can see continuous progress, and it's very exciting. The capabilities demonstrated in Minecraft show that, as long as there is enough compute and resources, we will be able to advance in many areas faster than ever before.

Host: Any verifiable domain really works, right?

Pichai: Exactly, and in other areas, you can always keep humans in the feedback loop to continuously drive progress. It feels like there are infinite possibilities. This is a moment when, as long as you can think of it, you can invest in it and make progress, and we've recently seen this in mathematics. Things you thought would take a long time to solve can show dramatic progress as long as you put in effort and resources, just as OpenAI did in mathematics with its latest models. So I think we'll see more results like this in 2027.

Host: Yes, I recently heard Sergey Brin talking about RSI and pushing the team to work on it. Where is Google currently in terms of internal RSI progress? How much of R&D now is done with model assistance?

Pichai: The current state is that AI researchers are working deeply in collaboration with models. The models participate, automatically recommending technical directions, evaluating the next checkpoint, and so on, so there is a healthy collaborative relationship between models and researchers. Although some parts of the pipeline may still be some distance from recursive self-improvement, the whole continuum is shifting. That's why I think it's so important to think about how we develop frontier models, and I expect that over the next 12 months, we will see sustained acceleration due to the application of recursive self-improvement.

4. The Fate of Gemini 3.5 Pro and the Programming Capabilities Still to Be Conquered

Host: Before we leave the topic of models, let me ask: I heard you mention Gemini 3.5 Pro at the last annual developer conference (Google I/O), but there have always been questions outside about what happened to that model. Did it just become Argon, or what happened?

Pichai: When we were developing the 3.5 series, we were developing Flash and Pro at the same time. By the middle of this year, we recognized internally that there was a gap in the model's agentic capabilities (especially programming-related capabilities) and in its foundational role for overall progress. So we wanted to accelerate our climb in that area, and the team believed the fastest way to climb was to use Flash. It's a smaller, more efficient model. That way you can get more, denser iteration loops in it. That thinking is what led us to launch the 3.5 to 3.8 Flash series. By the time we finished those, we had an opportunity to apply them to the Pro version, but by then Argon's pre-training had already been completed. So it was because we realized that 3.5 Pro's lifecycle would be very short, so we decided to apply all of that accumulated work to Argon's post-training. In an environment with constrained resources like compute and time, you can only make trade-offs like that. But it also gives us an opportunity to lay the groundwork in advance for where we'll truly need it in the future. So in version 4.0, you'll see different model versions. Argon is our largest model and the version we've currently released externally, but we'll launch more models going forward. At the same time, during this period we're also making Argon available to Pro-tier users.

Host: What difficulties still exist in programming right now? What are the hurdles you're still working to overcome?

Pichai: I wouldn't call them huge difficulties, exactly—it's just that cutting-edge technology keeps evolving. If you look back two months ago, Argon was arguably at the forefront in every aspect of programming. So it's not that we've hit some insurmountable wall; rather, you keep setting more complex end-to-end agentic tasks for it, and you want the model to perform exceptionally well—like a world-class team of excellent programmers, completing tasks with great taste, excellent code reviewability, and a high degree of safety. So this is an ever-evolving technological frontier. Take hallucination, for example—it will always be a frontier area where you need to keep making progress. It's an inherent part of the current state of the art, but we keep working to reduce it. It's both an opportunity and a challenge.

5. DeepMind Leadership Changes and Its Integration with Google

Host: Give me a specific example. We haven't talked since some leadership changes happened at Google DeepMind, and I'd like to ask you about that. CEO Demis Hassabis has stepped back to become Chief Scientist and is also devoting more energy to Isomorphic Labs—I'm very curious about that; at the same time, Corey was promoted to head of DeepMind; and Google legend Jeff Dean left and founded Discovery Loop. When that news came out at the time, it generated a lot of media coverage and all sorts of speculation about what was going on behind the scenes. Since then, I haven't heard you or anyone involved talk about it. I'd love to hear your perspective—maybe you can use this to put some rumors to rest: what was the thinking at the time? What drove these changes? Why did this happen?

Pichai: When you see people in a role for a period of time, some evolution naturally occurs.

Host: I remember Jeff was at Google for 27 years.

Pichai: A full 27 years. I'm very proud of that—Jeff and I worked together for a long time. His contributions at Google have been pioneering, and we're very excited to see what he can accomplish at Discovery Loop. First, I want to say that he really did want, at a certain stage of his life, to do something that excited him and to do it with like-minded people. As for the internal changes, what I'd say is that we're in a period for this industry where developing cutting-edge models is an all-consuming endeavor. In the past we had broad AI research, AGI research, and so on, but now developing cutting-edge models consumes effort across the board. It's advancing rapidly, and it's a massive systems engineering effort—from data, compute, and algorithm work to managing teams and using model agents to complete tasks, it covers everything. So I think this in itself has become a critically important core effort.

Pichai: Demis had actually been thinking about this moment for a long time. He's particularly focused on two things: on the one hand, how to safely navigate this period of explosive cutting-edge technology; on the other, how to apply it to solving other problems. The work he did on AlphaFold, and what we're now doing at Isomorphic Labs, are examples of how to apply AGI to solving other problems—like medicine and drug discovery—and robotics too, which will ultimately be applied across all domains. So I think he wants to spend more time on these things. He's always been passionate about them, and he got involved in them very early on.

Pichai: On the one hand, we've seen model development become a fairly mature process; on the other hand, Corey has already been at DeepMind for more than a decade. So we saw a good moment for Corey to step up. Demis remains involved as Chairman of DeepMind, but he's also spending more time in conversations about AI safety, while thinking about how to more practically apply AI to fields like science and healthcare, and other areas in the future.

Host: Does this mean the relationship between DeepMind and other parts of Google has shifted in some way? Is it becoming more tightly integrated? What follow-on effects does this change naturally produce?

Pichai: Over the years, including during Demis's tenure at the helm, the two sides have gone through a natural process of integration. For example, the Gemini app itself was built inside DeepMind. Corey had already moved to the Bay Area about 18 months ago. The two sides have always been tightly integrated, which is only natural both for DeepMind's compute support and for the impact on Google's other businesses. So this is actually a deep collaboration process that has been going on for a long time; perhaps the outside world just treated this change as a notable moment; but for us, it's a deep coupling that has long been underway.

6. Compute Allocation, Forward Planning, and the Confidence Behind Aggressive Investment

Host: How do you manage and allocate compute to support long-term research investment? DeepMind has historically been known for long-term, forward-looking research, and we're now in an AI wave dominated by near-term frontier models, the Gemini app, various products, and commercialization. Over the past year or two, has compute allocation changed?

Pichai: The issue is that the stakes right now are very high, and we're fortunate to be able to see the extraordinary opportunities presented by frontier intelligence—we have a very successful consumer business across many products and services, and at the same time we're one of the world's top cloud providers, with a vast opportunity to deliver this technology. So the opportunity space is enormous, and we need to plan ahead. We have a fairly rigorous approach to forward-looking planning. These teams all plan independently, and underpinning all of this is the fact that we have long been an "AI-first" company. Accompanying that strategy is a firm commitment to investing in compute, whether for foundational research or for developing models, and this is the cornerstone of how we plan for the future. On top of that, each business unit (such as Thomas and the cloud team) formulates its demand based on its own long-term plans, revenue forecasts, and market opportunities, and we plan compute for them accordingly. In hindsight, you can always find things that could have been done better. Around the world, no company can feel as deeply as we do how fast the technology frontier is advancing. This time last year, the explosion in agentic coding hadn't even happened yet. And don't forget that we make our own compute chips, which requires very long lead times. So at any given moment, you are living with the consequences of compute decisions made two years earlier. But people evaluate you based on the perfect information they have today.

Host: Yet you had to make the call two years ago.

Pichai: Yes, the decision had to be made two years ago. Make judgments based on the signals at the time, stay confident, maintain long-term optimism, and take an aggressive posture on investment. I don't think anyone at any company in the world today would pound their chest and say their decisions from two years ago were 100 percent perfect. You have to live through and deal with these moments; that's simply part of this business. We plan to keep our compute investment at the most advanced level in order to drive foundational research and frontier model development. Beyond that, we'll also have opportunities to apply the technology in more areas, and we plan to do our utmost to make all of that work.

Host: So you're already planning for 2028 and 2029 now?

Pichai: Yes, it depends on... for example, if you're planning resources like land and power supply, then you really are already looking further into the future.

Host: There was a major milestone this summer: for the first time since its IPO, Google had negative free cash flow. Many people on Wall Street noticed this, and when they saw it they were exclaiming: "Google is continuing to ramp up aggressive investment." I recall you also raised your 2027 capex forecast, and now you're looking at 2028 and 2029. What underpins your determination to do this? Especially at a moment like this summer when Google's free cash flow turned negative—this is by no means a small decision. How do you think about this? What gives you that confidence?

Pichai: The strong and durable demand signals we're seeing. To take a step like this, you need confidence on multiple levels. First, we're fortunate to be getting early signals from an extremely diversified portfolio of opportunities—from consumer to enterprise, we're seeing strong demand signals.

Host: You probably have more market signals than any other company on Earth.

Pichai: I hadn't thought about it from that angle before, but what you say makes sense. So we not only have strong external demand signals, but also internal demand from within Google itself—how we, as a company at the frontier, use and adopt AI ourselves. You can see firsthand the progress of the capability frontier, and you can assess the development of the capability frontier in an objective way independent of demand signals, looking at how the performance curves look—for example, scaling laws, new models versus old models, the complexity of tasks we can solve, and internal benchmarks.

Pichai: So you get a firsthand, intuitive feel, and you can also feel it from research breakthroughs: for example, will Gemini 5.0 be much better than the current Argon? The answer is yes. So you get these confidence signals at every layer of the stack. In addition, we're signing durable contracts with long-term committed revenue. This is by no means blind hope; we have actual business orders backing it up. Put all these factors together, and you commit fully in a responsible and disciplined way—and that's exactly what we're doing.

Host: It's impressive to hear you talk about these signals with such confidence, and I believe these judgments. But at the same time, people are trying to grapple with various concerns—for example, Corey's move, concerns about NeoClouds, concerns about the financing of this infrastructure, and their fragility. Looking across the whole market (not just Google, though Google is involved in many of these things), do you see any worrying signs, such as whether the market is overextended?

Pichai: When I answered the previous question, I was obviously speaking from Google's own perspective. As for the broader market, whenever you go through a moment like this—with enormous opportunity, rich returns, and a large number of players rushing in—vitality and volatility naturally appear in the market, and I think that's an inseparable part of such periods. That's precisely why I try hard to focus my energy on the signals I can control and see clearly—such as demand signals and our clear business outlook in terms of business contracts and order backlog—and to plan around those. You need to design the company in such a way that even if there is some volatility in the market, its long-term arc remains favorable for the company. That's why keeping a long-term view is so important. If you look back over the past five years, the pandemic drove a massive shift to digital, and later some of that demand pulled back.

Pichai: But for most companies that do this, the investments to adapt to and drive this shift have proven wise, and they have achieved better results as a result. So you have to understand the long-term, long-term trends and invest to seize the transformation opportunity. In fact, a large part of our headcount growth over the past few years has been for Google Cloud, and we are now enjoying the dividends that come as the cloud business scales. We entered the cloud market relatively late, but over a period of six to eight years we took a very aggressive investment posture, and that is precisely the key that has helped us get through the current phase. I think the same is true of how I view the current AI wave.

7. When will compute constraints ease, AI-assisted decision-making and the personal Agent strategy

Host: Can you see a point in the future when we are no longer constrained by compute? Is there such a moment that you can look ahead to?

Pichai: I'm a bit hesitant to say that, because—

Host: Yes, you have to be careful, because if you say that everyone will be terrified.

Pichai: No, what I mean is that we live in an era where the global economy is intertwined in extremely complex ways, and any isolated event can produce a "butterfly effect" everywhere. That's why we distinguish between the two things. I'm not saying everything will be smooth sailing from now on—life is never like that. The economy will grow, and there will be ups and downs; we need to go through business cycles. All of these are inevitable parts of how the world works.

Host: And I think no one would think this is like past bubbles—this is absolutely not the fiber-optic bubble of those years. We are genuinely using all of these technologies.

Pichai: We are currently in a state of severe compute constraint, but more importantly, we are already using these technologies today in tangible ways that have real value. We see a clear prospect and know that in the future all of this will run better. But that doesn't mean the entire industry or economy won't experience overload, or encounter uncontrollable systemic shocks. So these are two different things. Precisely because of this, I focus my energy on the things we can control: using our long-term perspective and years of experience to better anticipate, model and invest, and applying that in a fairly disciplined way to the vast opportunities we have. I put my energy and focus on these things, rather than agonizing over whether there is excessive exuberance in the market.

Host: I've heard that you have said you hope AI can help you more with compute allocation inside Google. I'd love to know, over the past few months, has AI been involved? Do you feel it has helped you make better decisions? Or, on such major and critical matters, would you hand certain decisions over to models?

Pichai: I'll answer that from two angles. First, given how consequential this is, all decisions are made strictly with humans in the loop, including deep involvement from myself. But second, for me personally, I have also deployed AI Agents to assist me. For example, there is a not-complicated but very practical scenario: at our scale of business, just keeping up with the latest progress in model R&D—there are various communication groups every day for pretraining, mid-training, post-training data and so on—if you tried to follow what happened in the past 24 hours purely on your own, it would easily cost an hour every day. But now I run Agents that automatically summarize and present this information for me, and if needed I can still dig deeper at any time. Likewise, there are Agents that can monitor compute usage, efficiency and utilization across our product lines in real time and provide me with decision inputs and insights. That is how we integrate AI into our daily work. Of course, there are also extremely outstanding people handling this work. AI lets each of us do our own jobs better, and that is how I currently use it.

Host: I was going to help you pass the buck, like saying it was the model that said it, not me.

Pichai: This is a gradual process. By the time models develop to the point where they can fully take on that decision, you yourself have already begun devoting yourself to the next level of problems. That's how I see it.

Host: The industry is currently in a race over personal Agents—for example, Muse and Instinct are both doing well. I remember at the last I/O conference you showed Gemini Spark. Do you think Google needs to take more action in the consumer-facing Agent space? I know you recently launched Gemini Workspace, giving Agents email and Workspace identities and so on, but on the consumer track, although it is still early, I guess you must want Google to occupy a larger share.

Pichai: This is an extremely important area of focus. In fact, the goal we set for Gemini is to make it a proactive, powerful and personalized assistant. That is the core direction in which we are building Gemini. The Spark you saw at I/O is exactly a step in that direction. Of course, at our recent enterprise event, we also discussed Gemini more in the context of Agents. To be clear, enterprise-grade and consumer-grade use the same underlying model, the same architecture and runtime environment; the only difference is whether it is handling enterprise data or personal context. All the enterprise-grade features we announced at the launch event have world-class security governance capabilities—each task essentially gets a microVM (microVM) configured to run securely 24 hours a day without interruption, with real-time monitoring, thereby providing enterprises with comprehensive governance, control and audit capabilities. We are bringing all of these capabilities to ordinary consumers. You can already see Chat and Task features in Gemini. Our near-term vision is to integrate them into a unified entry point, so that the boundary between chat and tasks connects naturally—which tasks need to run persistently, which need scheduled orchestration, and how to feed the results back to you. From the Gemini user usage data we have observed, this trend is already very clear: an increasingly high proportion of queries show clear Agentic characteristics. In the coming months, we will make these capabilities more broadly available to users. Previously we mainly rolled out these features to Ultra and Pro users, and next we will further expand coverage. The all-new Argon model brings a step-change improvement; as a consumer, in the future you will not only be able to converse with it in the Gemini app, but also invoke it in various contexts such as Gmail. We are investing deeply in this at every level, including Search.

Host: I have always wanted to ask you a question, especially given your past experience running Chrome: in the era we are currently in, all these Agents are using browsers like humans to complete tasks, as Muse does. I understand why this phenomenon happens—to complete tasks, this is indeed the effective path at present. But it feels like forcing a triangle into a square; it does not fit that well. This makes me wonder: does the web need to change fundamentally for these Agents? How should it change? I know you and other teams have been working on related protocols, but I feel Google is in a uniquely advantageous position to make a bigger change, and I would love to know whether you have been thinking about this question?

Pichai: I have thought about this deeply. This is a very complex moment, because when we talk about the large-scale adoption of Agents, it means that many different parts of the economy need to collaborate with one another in some way, and that collaboration must make sense for all participants. At this stage, the development of the technological frontier often runs ahead of the coordinated evolution of the entire ecosystem. This is also why MCP (Model Context Protocol), the Agent protocol, and payment protocols such as UCP, which we launched together with multiple companies, are gaining tremendous momentum. In the future there will be a period of uncoordinated evolution in this area, but hopefully it will gradually move toward more universal interoperability standards. I think this is a very elegant solution. At the same time, directly using browsers and computers is also an elegant way of working, because that is exactly what you would do as a user, and it also allows users to efficiently stay in control (Human-in-the-loop). So these two paths will develop in parallel. But whenever you can solve a problem through better protocols, that will be a far more scalable approach, enabling seamless operation anytime and anywhere. However, the adoption of protocols will obviously take longer.

Host: In this Agent era, how do you view protecting existing interaction interfaces? For example, some people think a completely new form of agentic Google Search could be invented. But today's Agents do not need these—for example, Amazon has banned Muse. Google has a huge advertising business that depends on user engagement. What impact will the emergence of Agents have on such interfaces? How do you protect or respond to it?

Pichai: In the long run, the right approach is always to put users first, think about what users really need, and do your utmost to innovate in that direction. I think this is the foundation that drives the product experience. But ultimately, it must also allow all businesses that provide value to survive and profit, so at any point in time a dynamic balance will be reached, and that balance will constantly evolve. There will always be new players who see these nodes as disruptive opportunities to break the mold. Just as when the internet first emerged, people also thought: will the way real estate transactions work change completely overnight? But the actual development path was not entirely as people imagined. It could even be said that the vast majority of real estate transactions today are still not much different from twenty years ago. There are many objective constraints in the real world. AI Agents and AI technology will indeed bring greater disruption, but micro-level dynamic balances will also constantly emerge in the process. For us, it is about embracing innovation, thinking from the user's perspective, and continuously advancing in that direction. This is how I approach product development, and it is also what users expect.

Host: Over the past year, I have talked with Liz and several of your other colleagues: when (or whether) Google Search's AI Overviews and Gemini will fully merge? Because now you have these two large-scale AI products—Gemini has hundreds of millions of users, and Search has billions of users, and they both differ and overlap in certain AI experiences. Do you think they will eventually become completely integrated?

Pichai: For example, when you turn on Personal Intelligence, whether in Gemini or in Search, it can grasp your context and work in a unified way. I expect our Agent to be able to provide that unified experience across these products. Rather than trying to force different products together, let the Agent layer naturally achieve this. Because your Agent runs in the cloud, works 24/7, and holds your context, it can pick up your unfinished work at any time—start a task in one place, and when you get to another place, the Agent still retains that context. The Agent layer gives us the opportunity to provide a better unified product experience than ever before.

9. Space data centers, Waymo, Wing, regulation, and quantum computing

Host: You have also recently started sending data centers into space, such as the first pilot of your "Project Suncatcher" in partnership with Planet Labs and SpaceX. How is that going? I heard it has already launched—is it still up there? Or did it burn up? What has the early feedback been like?

Pichai: It is going very well. I think it should still be operating normally in space. We will soon have the opportunity to say hello to the TPU chips, and I am very excited about that. I even had the chance to watch the launch in person. In my view, advancing projects like this is much the same as launching Waymo back then. I still remember flying to the desert to watch the Stanley self-driving car compete in the DARPA Challenge. If, at that time, you had tried to imagine what Waymo has achieved today, it would have sounded like science fiction. But 10 to 20 years is a long time, especially when we are at this kind of technology inflection point. So, although we have only taken the early steps, we are firmly committed to this journey. Our team is a world-class team, and over a 10-to-20-year time span, they will create amazing milestones—that is what excites me most.

Host: What about places like deserts, oceans, and all the places on Earth where we can put data centers? I mean, why invest in space now?

Pichai: If we start from first principles and consider energy availability, being in a sun-synchronous orbit can provide 24/7, around-the-clock energy—that is indeed a very strong reason. Physical space on Earth is limited. Although our data centers look large now, they are not blindly large compared with 10 years ago. So if we project computing demand 20 years forward, it is hard to imagine what Earth will look like. And we will also face various other constraints on Earth. By thinking about and exploring these frontier technologies—just like the moon landing program and the push into space in the 1960s—you are actually pushing every aspect of the entire technology curve forward. That is the way to drive R&D and future development. When you focus on a very ambitious goal, you will invent many technologies along the way that ultimately prove greatly beneficial.

Host: Just like what you discovered with Waymo.

Pichai: It was precisely because we started building AI very early, when compute was not yet sufficient, that we were prompted to develop TPUs. So thinking about the technology curve early and pushing it forward absolutely brings infinite opportunities.

Host: You have mentioned "moonshot" several times. This is not only literally a moonshot; I also saw that word in the announcement. In a sense, will this be the next Waymo? Will it be Alphabet's next spun-out new company?

Pichai: For the AI endeavor we are engaged in, compute is an indispensable core. Although we have not fully figured out the final form of the future, I think this is precisely how we will fundamentally scale compute in the future. This is absolutely a moonshot, and we are committed to doing it well.

Host: I would love to hear about Waymo's strategy, because I can see multiple directions it could develop in. Do you want to continue maintaining this kind of partnership with companies like Uber, or will you build a fully vertically integrated Waymo fleet? I also want to ask, is it possible that one day Waymo cars will be sold directly to consumers, or that I could rent one myself?

Pichai: At present, this is far from being a zero-sum game where one side's gain is the other's loss. Everything has only just begun, and the application prospects for autonomous driving technology are too broad. I expect Waymo to operate in multiple ways, including growing together with other partners. I am also very optimistic that those partners can achieve great success as independent companies. The team is now fully focused on scaling the business. We have just crossed the threshold of more than 500,000 paid rides per week, and the team's core focus is scaling.

Host: The Toronto launch event is coming soon.

Pichai: Ten new cities will launch this year. So, at the operational and logistics level, we need to expand the fleet, adapt to more environments, and improve customer service and operations, while safety always comes first. That is the focus for the next 12 months: expanding to international markets, more weather conditions, and all scenarios such as highways, while maintaining our very high safety standards (95% fewer serious accidents compared with regular driving). For me, the underlying technology we have built is extremely general-purpose, and it has all the optionality you mentioned.

Host: Everyone is curious about where you will take that optionality. Some people are panicking because they want to protect their own businesses, while others are marveling at its scalability.

Pichai: If you look back at Google's history as a company, you will see that we built the Android mobile ecosystem, and in the AI ecosystem we also partner with all kinds of companies. So we will always work with the entire ecosystem, and there will be enough room for growth for everyone to do well.

Host: So will I have a chance to buy a Waymo car myself in the future?

Pichai: This is not something we are currently pursuing. But never say never.

Host: What about Wing? That is another one of your projects that is maturing. I hear that in the near future, roughly 40 million U.S. households will be able to use Wing's drone delivery service. What kind of new possibilities do you think this will unlock?

Pichai: I just spoke with the outstanding Wing team on Friday night—it was my last meeting of the day. I am very excited because they have an excellent product. They made decisions early on to design a product that is lightweight, easy to manufacture, and easy to deploy, so that it can scale. They have laid the foundation, and they have ambitious expansion plans. So I expect that in 2027 we will significantly increase delivery scale and expand it into newer categories than today—that is our growth opportunity.

Host: Is this, like Waymo, also a partner-led approach?

Pichai: Yes, we will take that approach in '27 as well.

Host: I noticed you have been spending a lot of time in Washington, D.C. recently. I have not heard you talk recently about your views on the current policy environment. Demis Hassabis obviously raised the idea of a FINRA-style regulator, which attracted widespread attention. I believe you are aligned with that in many respects. Given the time you have recently spent and the fact that you signed a White House initiative alongside other tech CEOs, what do you hope to see happen? We are in a period where all parties are trying to establish a regulatory framework.

Pichai: This is a technology with profound implications, containing countless possibilities, including extraordinary opportunities to make society better and the economy more prosperous. We are at a stage where people increasingly recognize the power of this technology, and companies at the frontier must operate with the right safeguards and approaches to deploy it as safely as possible. That is my understanding of this moment. For our company, this is exactly the core of how we think and operate. We did the same with the Argon launch. I think the next step is to unify our testing methods as much as possible through better evaluation standards, which will help us assess frontier technology more effectively. Just as the financial world has clear accounting standards and independent professional audit institutions, these are the right directions for our future evolution. Although at some point in the future we will need to do more, getting the direction right is crucial. Technology evolves rapidly, and the current challenge is that any measure you take now may seem insufficient or completely wrong 12 months later. Therefore, building capabilities and establishing industry standards organizations are good ways to make progress. This is also where I spend a lot of effort and time: as Google, are we genuinely raising our level of investment and raising the standards for safely developing and deploying models? Companies that do this well will have a competitive advantage in the future. That is how I think about it.

Host: Interestingly, we began by talking about how you proactively slowed the pace of the Argon launch and tested internally—conducting your largest-ever internal testing and government review. Yet the outside world is calling for the government to force a slowdown. From what you just said, it seems market incentives are already working. One thing that worries me about these matters is how to leave room for startups to grow and avoid regulatory capture. How do you balance the two?

Pichai: In any case, what I advocate most is establishing unified national regulatory standards rather than a patchwork of rules across states. That is what would truly set the industry back and make it fundamentally unsafe by making the technology environment too complex. I am confident in our ability to adapt to the continued evolution of this field. Even on a voluntary basis, we now submit frontier models to the U.S. government for testing when we develop them. That is already a huge change compared with a year ago. Whether or not this is ultimately codified into formal regulation, these developments have already happened. We will continue to evolve based on what we observe in practice. To ensure that we ourselves, and our country, can stay at the frontier of technology and that the technology's potential is maximized, we must keep looking ahead.

Host: Honestly, I do not understand quantum computing very well. I know you have deep investments in this field. Quantum computing seems to sit above all the technologies we just discussed and could have profound impacts across many fields. This is a hard question to answer, and I would love to hear your views on the practical opportunities in quantum computing, because I see you have made some progress in this area. When it truly takes off, what do you think will happen?

Pichai: First, welcome to the club! When you admit you do not understand quantum computing, you are actually just like everyone here. The more time you spend on it, the more you somehow find that you understand it even less.

Pichai: Even more so. In quantum computing, we have set clear internal milestones, and for nearly the past decade the team has steadily met or exceeded these milestones in order to robustly scale quantum systems and move toward commercial applications. The future potential of quantum is extremely broad, and although there is still uncertainty in some respects, I genuinely see exciting practical commercial applications within five years. Just as we now provide TPU compute, if I were sitting in a chair five years from now, our Google Cloud team would be discussing and finalizing quantum deployment deals at some scale. So this is incredibly exciting. This is a frontier technology to which we have made a long-term commitment and will continue to invest firmly.

Host: Thank you very much.

Pichai: Very happy to, thank you.