Google researchers invented Transformers to solve practical problems like better translation and scaling speech recognition to billions of people. While the technology is often associated with products built outside the company, Google applied it immediately to its core business. Models like BERT and MUM represented some of the most significant improvements in search quality because they transformed how Google understood web pages and user queries.

Internal teams were also building early versions of conversational AI, such as LaMDA. At one point, an engineer famously thought the model was sentient, showing that an early version of ChatGPT existed internally long before the public launch. However, Google had a higher bar for product quality and safety. Early models could be toxic, and the company lacked the end-to-end reinforcement learning from human feedback needed to make them safe for the public.

The version I saw was a lot more toxic at a level we couldn't have possibly put it out at that time. As a company with a search quality bias, we had a higher bar for what we thought was an acceptable product quality to go out.

In the consumer internet space, surprises are inevitable. Small teams will always prototype and release new ideas that gain sudden traction, similar to how YouTube emerged alongside Google Video or Instagram alongside Facebook. Sundar notes that larger companies must accept that they will not invent every successful product variation, even when they created the underlying technology.

Google has long used speed as a core way to differentiate its products. This focus started with search results and continued through Gmail and Chrome. Currently, the focus is on making Gemini fast by using TPUs. Sundar sees latency as a sign that the technical foundation of a product is solid. While the speed of shipping new features is important, the actual performance speed of the product is a key feature of a great user experience.

I have always internalized speed, or latency, as one of the distinguishing features of a great product. It almost always reflects that the technical underpinnings of the product have been done well.

Managing latency is a rigorous process involving strict budgets measured in milliseconds. For example, if a search team saves three milliseconds through optimization, they earn half of that time back to use for new features, while the other half is given back to the user to improve performance. This system helps balance the desire for more features with the need for speed. Even as capabilities grow, Google has improved search latency by 30% over the last five years. In AI, this same balance is achieved by offering Flash models that provide about 90% of the capability of larger models while being significantly faster and more efficient.

It is easy to say you want latency, but you are constantly adding capabilities. The capability frontier is progressing, so there is a sense of how you balance that.

Search adapts to match how people interact with technology. In the mobile era, search shifted to help people find locations while on the move. Today, search is moving toward completing entire tasks through agentic flows. This means search could become an agent manager that runs multiple threads to get things done for the user.

Search would be an agent manager in which you're doing a lot of things. I think in some ways I use anti gravity today and you have a bunch of agents doing stuff and I can see search doing versions of those things and you're getting a bunch of stuff done.

The old model of a single-line prompt returning a list of links is expanding. People already use AI in search for deep research, which requires more than a simple answer. Future tasks might be asynchronous, meaning they run in the background over time. Sundar notes that it is difficult to plan ten years ahead because the progress curve is so steep right now. It is more exciting and practical to focus on the next year because the models change so much in a short time.

I think you can paralyze yourself thinking 10 years ahead. But we are fortunate to be in a moment where you can think a year ahead and the curve is so steep. It's exciting to just do that year ahead.

This shift is not a zero-sum game where one product must die for another to live. Instead, it is an expansionary moment. The total value of what people can do is increasing. Products like search and Gemini will overlap in some areas and diverge in others, but both will continue to evolve at the cutting edge.

A year ago, many people felt that Google's core search business was under threat. The sentiment was negative, but this ignored the fact that the company had been preparing for an AI-first world for years. Sundar points out that Google began announcing TPUs and building AI data centers as far back as 2016. While they might have been behind on frontier large language models for a moment, they already had the research teams and infrastructure in place to catch up quickly.

I felt like the company was built for that moment. It was very intentful. We were in the seventh version of TPUs. We were thinking about the company operating in an AI-first way. So we had deeply internalized this shift.

The real power of this shift is that AI serves as a common technology that accelerates every part of the business, from YouTube to Waymo. It is a highly leveraged way to make progress across different platforms at once. Sundar does not see the current AI landscape as a zero-sum game. Just as Amazon and Facebook grew alongside Google in the past, he expects the total market to scale significantly, leaving room for many players to succeed.

Some critics suggest that Google is not as AGI pilled as newer research labs. They argue that the company might not believe in the rapid approach of artificial general intelligence. Sundar counters this by pointing to Google's massive increase in capital expenditure, which has scaled from 30 billion to roughly 180 billion. A company does not spend that kind of money without a deep belief in the technology's trajectory. Furthermore, the pioneers of the field, like Demis Hassabis and Jeff Dean, have been at Google for years, proving that the foundation for AGI has always been part of their DNA.

I think even within the company, there is a set of us living on the bleeding edge, seeing what these things can do. We see the agents pick up skills and do stuff. We are living that exponential internally.

The path toward artificial general intelligence is marked by specific moments where the technology suddenly feels visceral. Sundar identifies early milestones as far back as 2012, when Google Brain first recognized a cat in a video. Other significant moments included attending the DARPA challenge with Larry to see autonomous vehicles and watching Demis demo early models that exhibited a sense of imagination. These experiences made the rapid progression of the technology obvious long before the current AI boom.

My first AGI moment was 2012 when Jeff demoed the earliest version of Google Brain. This is when the neural networks recognized a cat. Then there was seeing Demis demoing the earliest versions of the models, having what we would call imagination. So it was obvious the technology is progressing.

Today, that visceral feeling of intelligence often appears in the workflow of software development. Using agent managers allows for the completion of complex tasks without ever interacting with a traditional coding environment. The experience is so seamless that the specific programming language used becomes a minor detail to be asked about after the project is functional. While these individual moments are powerful, the most striking aspect of current AI is the sheer steepness of its progress curve across multiple paradigms.

Tech products are often abstract, which makes it easy for leaders to lose touch with the actual user experience. Managing purely through slide decks and reports can lead to a disconnect. John explains that his team has a recurring segment called Walk the Store where they click through their dashboard together. This helps them find confusing parts of the product and see where users might trip over certain features. Other CEOs use similar tactics, such as working as delivery drivers to stay connected to the ground level experience.

I block time to use it intensely, like kind of focus time to do it. Even just two weeks ago I was stretching in the gym and I had the phone with Gemini Live and so I'm like, I'm going to talk to it for like thirty minutes on one topic. You do those things and some of it works, some of it is frustrating, but you learn a lot. I force myself to use it in those power user modes to stay in touch.

Sundar maintains this connection by blocking dedicated time to use Google products as a power user. He recently spent thirty minutes talking to Gemini Live while at the gym to find its points of friction. He also monitors social media platforms like X for raw feedback and uses an internal AI tool called Anti Gravity. This tool allows him to bypass polished reports and ask directly for the five best and five worst things people are saying about a new launch. This direct access to raw sentiment provides a clearer picture of product performance than filtered corporate updates.

AI is changing how we process information and conduct research. AI agents now assist in tasks that previously required significant manual effort to understand a topic. This shifts the balance between spending time to get a firsthand feel for something and leveraging tools to reach the same goal. Even leaders at major tech companies are still learning how to adapt to this new future where productivity gains are constant.

I am trying to adapt to this future. In the past I would have to spend a lot more time trying to get a sense for it. Now an AI agent is helping me in that journey.

The economic impact of AI is distinct from previous cycles like the internet or mobile. While those technologies took years to show up in GDP numbers, AI is already contributing through massive data center expansions. While critics question if the massive capital expenditures will eventually reconcile with actual returns, the current reality is a market that is deeply supply constrained. The demand for these tools across all sectors suggests a massive outcome for the global economy.

At some point it has to reconcile. To be very clear, we are supply constrained. We are seeing the demand across all the surface areas we offer. I actually don't have any doubt that this is a massive market and outcome.

The market for software engineering and coding is likely much larger than current estimates suggest. While observers often focus on the cost of AI tokens compared to developer salaries, the industry has historically been limited by a lack of supply. Increasing the supply of engineering power through AI could expand the market tenfold rather than simply replacing existing roles. This shift reflects a broader trend where technology creates immense value that traditional economic metrics might not fully capture.

The compute build out is a different curve than the rate at which we can improve the models. You are already dealing with a more constrained curve there. Then how do you diffuse it into society? How do you diffuse it through society responsibly? There are constraints in all these layers.

Sundar highlights that several natural mechanisms will dictate the pace of AI growth. The physical expansion of compute infrastructure follows a different timeline than the improvement of software models. Additionally, responsible societal diffusion takes time. For instance, while technology like Waymo can be safer than human drivers, it must be rolled out at a careful pace. Even a small increase in growth for an economy as large as the United States represents a massive contribution to global progress.

The scaling of AI is currently hitting real world bottlenecks that capital alone cannot solve. While companies are ready to spend hundreds of billions, they are constrained by deeper ground truths like wafer capacity and memory availability. Memory is a particularly critical component in the short term. Even with high prices and capitalist incentives, leading companies cannot dramatically ramp up supply overnight. These constraints create a ceiling on how far any one player can pull ahead of the rest of the industry.

Wafer starts are a fundamental constraint. I think power and energy are more solvable. Permitting and actually working through a regulatory environment might be a constraint. We need to learn to build things much faster. You almost have to shift your mentality to think about what it would take to construct 10x faster in the physical world.

Regulation and permitting are significant hurdles in the United States. In places like China, the pace of construction is much faster. To compete, the US must find ways to build data centers and power infrastructure at a much higher speed. These physical constraints will likely force a push toward massive efficiency gains. Models may need to become 30 times more efficient to work within the existing supply limits.

The physical reality of AI is a strange contrast. Massive data centers run for months at a time, consuming enormous amounts of power and using the most advanced hardware on the planet. Yet the final output of this process is just a flat file of weights.

It is a very weird thing. You are talking about a set of weights which can fit on a USB stick. I am always shocked that you run a data center for months and months and your output is literally a flat file. It is like having a word doc. That is your model. It is amazing.

AI models pose a significant threat to software security and may already be breaking existing systems. The market for zero-day vulnerabilities is changing rapidly because of this technology. As AI makes it easier to find these flaws, their supply increases and their black market price drops. This shift serves as a real-world metric for how AI impacts technical safety.

I genuinely think there is a lot of upside ahead. Some of the constraints maybe are helpful. Constraint inspires creativity, forces a compaction cycle where we get more efficient, and forces important conversations to be had which otherwise won't happen.

While these risks exist, constraints can actually drive progress. They push people toward more efficient solutions and force critical discussions that might not happen otherwise. One such discussion involves the need for global coordination on security. Currently, this coordination is lacking, and it may take a major event to spark the necessary action.

Google maintains a diverse portfolio of long term projects that often appear unrealistic when they first begin. Sundar notes that constraints often inspire creativity. This leads the company to explore ideas like placing data centers in space. While these challenges are difficult to solve, they represent a twenty year outlook on where the world might need to store information.

I am confident quantum will have many applications if you can actually make it work. I always give the example of mobile phones plus GPS enabling Uber. There was nobody working on phones who would predict that as an outcome of this platform shift.

Quantum computing is another area where Google remains deeply committed. Sundar explains that simulating nature and reality effectively requires quantum systems because nature itself is quantum. While classical computing techniques and deep learning models like AlphaFold are impressive, quantum will likely have a significant edge in complex areas. This includes weather simulation or understanding the chemical processes behind fertilizer production.

Google is also revisiting robotics. Previously, the company was too early in this field. AI has become the missing ingredient that was once absent. New models like Gemini offer improved spatial reasoning. This allows Google to partner with robotics companies like Boston Dynamics and Agile. Other projects include Wing, a drone delivery service that could soon reach millions of Americans. There is also Isomorphic, which aims to improve the success rate of every step in the drug discovery process.

Comparing investments across wildly different fields like YouTube algorithms and autonomous driving requires a specific framework. Sundar explains that the challenge of capital allocation is especially visible today regarding TPU distribution. He believes that AI will eventually provide valuable inputs for these decisions once internal data is better integrated. Historically, Google succeeds by placing deep technology bets early in their development cycles. This strategy allows them to start with smaller funding amounts while maintaining a commitment to long term progress.

One of the ways we have thought about it and we have been disciplined about is to make those early technology bets in a deep way. You have to assess the long term value of these things. In some intuitive way you are thinking about the option value and the market size of something five to ten years down the line and you assume a crazy growth and think through whether those decisions make sense.

Progress on these bets is measured by technical milestones rather than immediate financial returns. In quantum computing, the team tracks progress toward logical qubit thresholds. Sundar notes that Google often doubles down when others retreat. They increased investment in Waymo a few years ago while the rest of the industry became pessimistic. This long term perspective turns speculative technology into practical, everyday tools.

Deciding which projects to fund requires looking at the underlying technology level. For Waymo, the focus was on tracking how the software drives the car and whether safety and reliability were improving according to a predicted curve. Even when progress stalls, having confidence in the team to break through those phases is essential to long-term success.

I think the more you are able to evaluate things at that deeper technology level, I think you tend to make those decisions better. At least that is how I have tried to do it.

While newer companies might start faster today by using modern end-to-end deep learning, long-running projects like Waymo benefit from years of complex system integration. This level of craft is similar to what is required at companies like SpaceX or TSMC. Pushing the technology curve in safety-critical and regulated areas often requires first-party hardware. Sundar believes this firsthand experience with the product feedback cycle is important for really pushing a product forward.

Google has traditionally held a strong net cash position while looking for the best ways to use its capital. The company aims to be a good steward of its resources. This involves investing in internal projects and external companies like Stripe, SpaceX, and Anthropic. Sundar explains that the decision to spend depends on the maturity of the project rather than just the availability of funds.

For example, if Waymo had reached this point earlier, I think I would have invested the capital earlier. To some extent you are judging it by wanting to be good stewards of capital. To the extent you are bullish on return on invested capital, you want to invest every last dollar you can there.

The pace of investment is often tied to the natural growth and safety requirements of a project. Sundar notes that while he would have liked to put more capital into Waymo sooner, the project needed to reach a certain level of maturity first. With the current shift toward AI, there are now even more opportunities for Google to deploy capital effectively. When the leadership has conviction in an idea, they are willing to commit the necessary resources to see it through.

Historically, research and development costs in tech companies were dominated by headcount. Managing these companies meant managing the people walking around the building. While compute was always a factor, it was often an afterthought compared to the expense of highly paid engineers. This dynamic has changed with the rise of machine learning. Now, compute allocation for TPUs and GPUs is as critical as headcount planning.

I spend a dedicated hour a week thinking about that question at a pretty granular level. I will know by projects and by teams the compute units they are using. In some ways it is a really important thing to be doing right now. The scarce resource is compute in a lot of cases.

The scarcity of machine learning compute requires dedicated attention at the highest levels. Sundar reviews how these resources are distributed across different projects and teams to ensure they go toward the most important work. This management becomes even more complex when balancing internal needs with external customers through Google Cloud. Anything committed to a customer is considered sacrosanct. The company handles this tension through extensive forward planning. This ensures that both internal teams and cloud clients have what they need in a world where compute is limited.

Anything we commit to a customer is sacrosanct. These are contractual commitments. So you solve a lot of it with planning. In a constrained world, the cloud team might not have all the compute they want, but you solve it by planning ahead.

The vast complexity of Google Cloud can make it difficult for users to navigate. With so much functionality across organizations and projects, finding the right services is often a challenge. AI now serves as a powerful orchestration layer that threads through API documentation and interacts programmatically with the platform. This transformation turns a broad product surface area into an advantage. Instead of a user getting lost in the interface, the AI manages the complexity behind the scenes.

John notes that Stripe faces a similar situation. As more functionality is added, AI becomes the best way to navigate the product surface area. This same logic applies to large enterprises. In the past, connecting data sources required massive projects like building new systems. Now, AI can act as the connective tissue that brings data together in a way that makes sense for the user. Sundar agrees that while there is room for improvement, this represents an immense opportunity for the future of cloud computing.

The promise of AI being this orchestration layer, for anything you think about, even internally within the enterprise as a CEO, it is not like you do not have all the data, but how do you get it in one place? In the past, that would have meant one more big ERP project to go connect all the data sources. AI being this orchestration layer in a way that makes sense for the end user has been delightful to see.

The next frontier for consumer AI involves a shift toward stateful and persistent capabilities. Most mainstream AI applications currently lack the ability to handle long-running tasks that remain consistent over time. Giving users the power to set up a task, such as a personalized daily news roundup, requires persistence that is not yet common in popular apps.

Look, I think you want to give users capability where you have persistent, long running tasks in a reliable, secure way. You have to think through things like identity, access, et cetera. But I think that is the future, that is the agentic future.

Sundar views this agentic future as a major area of exploration. It involves combining full coding models with the ability to run tasks securely in the cloud or locally. While a small fraction of developers currently build these types of custom tools for themselves, the goal is to bring these primitives to a mass audience. This transition will allow consumers to specify and create their own software experiences through simple interfaces.

Searching through Google Docs often feels more difficult than searching through Gmail. While keyword search works well for email because messages often contain unique words, document titles are often generic. For example, searching for a budget document yields too many similar results because terms like budget are common across many presentations and spreadsheets.

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. And so I can never find the exact right one.

Sundar acknowledges this challenge and notes that AI integration will lead to significant improvements in the coming months. The goal is to better manage what information is kept in context and cached to make search more intuitive. Early versions of these tools only scratched the surface, but deeper integration will allow the system to bring more relevant information to bear when a user is looking for a specific file.

I think the AI integration into these services, including Google Docs, 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 over time, what all can you keep in context? What can you cache and what can you really bring to bear?

Google is experiencing a major shift in how its teams handle engineering and product development. Sundar explains that this transformation happens in concentric circles. Some groups move profoundly and quickly while others adopt changes more slowly. A primary goal is diffusing these new workflows to more teams as the technology becomes more reliable. In the past, the tools were often too broken to use widely, but the curve is now shifting dramatically.

Internal teams are moving toward what Sundar describes as an agent manager world. Google uses an internal tool nicknamed Jet Ski to help employees work in this new way. This system is currently being rolled out to major departments like the Search team. While small companies can pivot almost instantly, large organizations face a significant challenge with change management as they try to diffuse these new technologies across many different groups.

We have a different name internally than externally of the same product, but it is Jet Ski internally, which is anti gravity. And you are living on it. You are living in an agent manager world. You have workflows and you are working in this new way.

There is a significant gap between what AI models can do and how they are actually used in industry. This intelligence overhang exists because adopting AI requires more than just access to a model. Engineers must learn how to prompt effectively for both general tasks and company-specific tools. Collaborating on an AI generated codebase is also difficult because the code changes so quickly. Beyond engineering, the biggest hurdle is access to data. Companies need to rewrite their permissions engines so agents can answer basic questions like the status of a deal.

Sundar notes that Google is working through these exact problems. Identity access controls and security are major barriers to making AI diffusion systematic across a large organization. Larger companies face high fixed costs and risks when integrating these tools.

We are still diffusing it. Like if you are the SRE team at Google, you suddenly find portions which you can create an automated workflow. Doing it more systematically when you develop skills, how does it get centralized? How is it available for everyone to use? Identity access controls are real hard problems.

The goal is to reach a point where complex business functions are fully agentic. John suggests that a company forecast could eventually be done with no humans in the loop. Sundar expects 2027 to be a major inflection point for these kinds of shifts. At first, humans will likely use AI workflows and double-check them with conventional methods. Eventually, the process will cross over to being primarily agentic. Startups may have an advantage here because they can build AI native teams from the start. Larger companies like Google must focus on driving a deeper transformation through retraining.

Innovation often begins with tiny teams working on seemingly massive projects. One example is the initiative to build data centers in space. This project started with just a few people and a small budget focused on reaching a single milestone. Starting small is essential even when the ultimate goal is a big idea.

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.

The same principle applies to technical breakthroughs. Sundar recently reviewed an improvement in machine learning post-training developed by just one person. Despite the small scale of the effort, the change promised a significant jump in performance. These small, constant improvements are a major source of power in the current technological moment.