Google executives interpret Q2 financial report: The most important task is to solidly advance Gemini 4

Alphabet announced that its cloud business revenue for the second quarter exceeded Wall Street expectations, but sales related to its search engine business were slightly below expectations, which may intensify market concerns about its significant investment in artificial intelligence.

After the earnings report was released, Google CEO Sundar Pichai, Chief Business Officer Philipp Schindler, and CFO Anat Ashkenazi held an analyst conference call to answer questions related to the business.

Here is a summary of the analyst conference call responses:

Morgan Stanley analyst Brian Nowak: I have two questions. First, Sundar, over time, we have seen more generative artificial intelligence (GenAI) products and tools emerging in the market, and the company is continuing to increase its investment. Could you talk about how you view the return on invested capital (ROIC) in the entire generative AI space now compared to a year ago? Specifically, has your view changed regarding the scale of this opportunity and the timeline for realizing investment returns?

My second question is for Anat. In the previous briefing, you spent considerable time sharing with us the current issues the company faces regarding limited computing power supply. My question is related to this and concerns the company's future capital expenditures. As management looks ahead to 2027, if the company wants to address the current computing capacity bottleneck, what approach will you take in formulating future capital expenditure budgets? What new changes will there be in the company's planning and control principles for capital expenditures? How does management determine how much funding the company should invest in 2027 to effectively alleviate the current capacity constraints?

Sundar Pichai: In my view, we are still in a very early stage. I believe generative AI will drive structural changes with long-term impact across multiple fields.

In terms of our core information business, the current cutting-edge AI capabilities allow us to realize various possibilities, and I firmly believe we still have a lot of work to do to truly translate these advanced capabilities into consumer-facing product experiences. For example, one can imagine end-to-end intelligent agents that enable AI to accomplish more valuable tasks for users. I think these all represent tremendous growth opportunities. If we can seize these opportunities, I believe it will also bring us considerable investment returns.

The enterprise user market is similar; our product demand performance reflects this, and you can observe it from our growth rates and other performance indicators. Moreover, in my conversations with various CEOs and enterprise users, I find that most companies are actually just beginning to explore the potential of AI technology, and there is still a long way to go before realizing true value.

In the past, we often discussed the development of cloud computing. At that time, the workloads that had truly migrated to the cloud accounted for only a small portion of the total workload. Now, one can think of a similar question: what proportion of workloads are currently truly AI-native or AI-enabled? In my opinion, this ratio is still very low, and we can say that we are still in the early stages of AI technology development.

From the perspective of return on invested capital, we are adopting a full-stack strategy. Whether in the consumer market, enterprise market, or from the perspective of the developer ecosystem, we have seen a sustained momentum of growth.

In summary, if there has been any change in our thoughts and ideas over the past year, it is that we are more optimistic about future development opportunities than we were a year ago, and we are more firmly convinced of the long-term value inherent in these opportunities.

Anat Ashkenazi: Regarding your mention of future capital expenditures and how to address the constraints on computing power supply, the environment we are currently in has not fundamentally changed; supply remains the main factor limiting business development.

We have actually mentioned this repeatedly over several quarters. Whether it is the demand from external users of Google Cloud or the demand for computing resources from various internal businesses, both have maintained a very strong growth trend. Our principle of capital investment has always remained unchanged: as long as we believe that this investment can bring attractive returns to the company, we will continue to increase our investment. As Sundar just mentioned, we are confident in the current investment opportunities.

In terms of capital planning, we will still adhere to a long-term perspective. In other words, we will plan comprehensively from the perspective of business development needs over the next few years, while also focusing on the needs of the next year and shorter-term demands, continuously accelerating infrastructure construction to meet the growing market demand.

Over the past three years, we have significantly increased our infrastructure capacity, but market demand is still growing faster than our expansion rate. Therefore, like the entire industry, we are currently still in an environment of tight computing power supply and limited resources. At the same time, we are also actively enhancing our supply capabilities. Google adopts a full-stack layout strategy, which allows us to continuously improve the operational efficiency and technical efficiency of our technological infrastructure, thereby achieving higher computing power under existing resource conditions.

Therefore, as long as we can see attractive investment opportunities, we will continue to maintain our investment intensity.

JP Morgan Analyst Doug Anmuth: I also have two questions. First, Sundar, could you please talk about your confidence in maintaining a global leading position with the Gemini model?

This week, Google just released the new Gemini Flash series models. However, compared to some other industry-leading large model laboratories, Google does not seem to release products as frequently and does not have as high a market presence. Therefore, I would like to take this opportunity to hear your views on external concerns: can Google continue to develop and maintain industry-leading large model capabilities? Additionally, could management discuss Google's layout planning in the AI programming field? And how does management plan to further narrow the gap with competitors in the enterprise market segment? My second question is for Anath. Recently, Google has conducted both equity and debt financing. Could you please introduce how the company currently views its optimal capital structure? At the same time, how does the management weigh the "cost of debt financing" against the "cost of equity financing" in the process of advancing financing decisions?

Sundar Pichai: Currently, the competition for cutting-edge global models is in a rapidly evolving stage, with a very fast pace of technological iteration. At any given point in time, the industry landscape may change.

Google has always maintained a globally leading level of models, and we still hold an industry-leading position in many key capabilities. Of course, there are some areas where we believe further improvement is needed, such as AI programming and agent coding, and our team is currently focusing resources on continuously optimizing these capabilities.

Take the Gemini 3.6 Flash released this week as an example. The Flash series is our core and most widely used model product, and it is currently one of the most in-demand models. It achieves a very good balance between performance, cost, reliability, and response latency, which has been widely welcomed by users. We have now widely applied Gemini Flash across our entire product system, including various solutions such as cybersecurity and data analysis.

In user service scenarios, enterprises not only need high-quality voice capabilities and real-time interaction capabilities but also require models to have real-time reasoning capabilities; for professional service institutions, they rely even more on high-quality content summarization and text generation capabilities. The Gemini Flash model has demonstrated excellent performance in these application scenarios.

In terms of agent coding, we are also continuously iterating rapidly and will continue to release new versions in the future. For example, compared to Gemini 3.5 Flash, the newly released Gemini 3.6 Flash has improved its performance in the DeepSuite benchmark test by over 10 percentage points; at the same time, during the process of achieving performance improvements, the efficiency of Token usage has also been further enhanced.

Currently, we have fully applied this model internally at Google and are also testing its application effects in programming scenarios with several enterprise users. In terms of time, it has taken us only about six weeks to achieve significant progress from the previous generation product to the latest version. In the future, we will continue to maintain this rapid iteration pace.

As for the competition for cutting-edge global models mentioned in your question, we are both confident and will always remain committed. For the next generation of globally leading models, we believe that larger foundational models are needed as support. We have already initiated pre-training for Gemini 4 and set very high goals. I am very excited about the current progress of Gemini 4's research and development and believe that when it is officially released, it will surprise users.

To maintain our global leading position in the next phase, we need to have foundational models that are larger and more capable, like Gemini 4. Therefore, I believe our most important task right now is to solidly advance this work and do it to the best of our ability Anat Ashkenazi: Regarding the capital structure mentioned in your question and the recent equity and debt financing issues. When planning future investments, we look at both the next year and assess the funding needs for the coming years from a longer-term perspective.

First, we will evaluate how much investment can be supported by the cash flow generated from operating activities. From the performance announced today, you can see that Google continues to maintain strong and healthy operating cash flow, which has always been the company's primary source of funding.

On this basis, we will comprehensively consider debt financing. Over the past 12 months, we have significantly expanded the scale of debt financing. About a year ago, the company's debt balance was approximately $16 billion; it has now increased to about $100 billion, forming a debt financing system that covers multiple currencies and regional markets, with more diversified debt financing channels. In addition, while supporting the company's continued growth, we also hope to maintain a robust and resilient balance sheet. This is also an important consideration for our recent entry into the equity financing market.

At present, except for special arrangements, we have no plans for further equity financing. As you know, our previous equity financing plan included a portion of the ATM (At-the-Market) issuance plan. In the coming period, the company will continue to use this mechanism to offset the share dilution caused by stock compensation and to pay the tax costs associated with stock compensation.

Overall, in terms of capital allocation, we will mainly balance around three dimensions, including operating cash flow, the scale of debt financing, and equity financing arrangements. Our goal is to maintain a healthy and robust balance sheet while meeting ongoing investment needs.

Goldman Sachs analyst Eric Sheridan: I have two questions about TPU (Tensor Processing Unit). Sundar, could you please talk about the important experiences and insights Google has gained in the process of continuously expanding TPU deployment? Specifically, on one hand, as the application scope of TPU continues to expand, how does management view the future market demand for TPU? On the other hand, in the coming years, facing the simultaneous growth of external customer demand for TPU and Google's internal business demand for self-developed chips, how will the company allocate and balance resources between the two?

I have a follow-up question. Anat, in your briefing, you mentioned that TPU has had a positive impact on Google's cloud business. Could you further introduce us to more specific situations? For example, what percentage of the currently contracted but unrecognized revenue orders in Google Cloud is related to TPU? Additionally, how will TPU-related business gradually convert into revenue in the coming years? What impact will it have on the company's profit margins?

Sundar Pichai: First of all, we are very satisfied with the progress of the TPU product roadmap. Whether in terms of performance or the competitive advantages it brings, it has met our expectations, so we are also using TPU extensively internally As for the resource allocation of TPU, our primary principle has always remained unchanged: first, we must ensure that there are enough TPUs to support Google in maintaining a global lead in the field of artificial general intelligence (AGI). This is the foundation of all our work. At the same time, the market demand for computing power is very strong, so we also need to balance internal demand and external customer demand.

For most Google Cloud customers, we currently mainly use TPUs and GPUs to run and provide Google's own AI model services, such as Vertex AI, Gemini Enterprise, and various rapidly growing AI intelligent applications, all of which rely on these computing resources. As for those customers who wish to use TPUs directly as infrastructure, they want to rent TPU computing power to run their own models. We are meeting this demand by expanding our infrastructure deployment. For example, we are evaluating and advancing more plans to deploy TPUs to customer data centers or to the data centers of other partners. The project we are currently working on in collaboration with Blackstone is an example of this.

In this way, we hope to better balance the allocation of computing resources: on one hand, ensuring the research and development needs of the most advanced AI models; on the other hand, we can continue to support Google's own AI model services provided to consumer and enterprise customers.

Anat Ashkenazi: Regarding the revenue recognition method for TPU system sales, you can understand it this way: when we sign the relevant agreements mentioned earlier, these contracts will be included in the backlog of Google Cloud. Currently, the $514 billion backlog of Google Cloud mainly comes from customer contracts of Google Cloud Platform (GCP), but it also includes orders related to TPU system sales. After signing the agreement, we will build inventory in advance to ensure that we can deliver these TPU systems in the future. Since we need to produce and prepare in advance during the business expansion and scaling process, this will be reflected in the cash flow from operating activities. When we start delivering these TPU systems to customers, we typically begin to recognize revenue from that stage.

In this quarter, the revenue we recognized only accounts for a small portion of the total contract amount. In the future, we will continue to increase the delivery scale in 2026, and most of the revenue from this agreement is expected to be recognized in 2027.

Barclays Analyst Ross Sandler: I want to return to the "big model war" in the current industry. Sandler, I would like to further ask about the speed of model releases. We noticed that Google recently reached a third-party computing power cooperation agreement with SpaceX. I would like to understand, besides this initiative, what other measures Google has taken to accelerate the iteration and release pace of the Gemini model, thereby further enhancing the update speed? You just mentioned the Gemini Flash series models, which have performed very successfully, aligning with your introduction. However, considering the current fierce competition in the low-cost model market with numerous participants, do you believe that the market positioning of the Flash models is still the right direction? What are the management's views on the future development trends of the AI model market and how Google will participate in it?

Sundar Pichai: I want to address two points. First, we have been thinking about how to ensure our products cover the entire competitive landscape of AI models. For Google's users, we want to provide the best model options across different price ranges. In other words, on one hand, we place great importance on creating industry-leading top models; on the other hand, we also need to offer models that are powerful, cost-effective, and suitable for large-scale applications.

Therefore, you can see that we have launched different versions of large models such as Gemini Flash-Lite and Flash Pro. Our goal is to cover the entire AI model market, from the highest performance cutting-edge models to high-cost performance models, fully meeting the needs of different users.

Regarding the speed of model releases, we will continue to accelerate the iteration speed. For example, we released Gemini 3.5 Flash at the Google Developer Conference, followed by Gemini 3.6 Flash. In the future, you will continue to see this series of models being updated, and we will also continuously enhance capabilities in areas such as agent reasoning.

At the same time, we are investing significant resources to advance the development of Gemini 4. This is a very important project. Our goal is that with the release of Gemini 4, it can maintain a leading position in the competitive landscape of AI cutting-edge models at that time. Therefore, we are concentrating a lot of computing power and R&D efforts to achieve this goal.

In the process of building Gemini 4, we will also establish a stronger foundational platform. Based on this, we will be able to launch subsequent versions more quickly in the future, continuously iterating and upgrading. Improving the speed of model updates and achieving a model release cadence of approximately once a month is one of our important plans in the development of the Gemini 4 large model.

MoffettNathanson analyst Michael Nathanson: My first question is for Sundar, still regarding the model competition discussed throughout today's conference call. I would like you to talk about what you believe Google's true long-term competitive advantage is in this new AI model competition. In other words, even if all companies can achieve similar large model capabilities in the future, what strategic advantages do you think Google has that can help the company continue to maintain its growth momentum? What advantages do you think will keep Google ahead in the competition?

Anat, I would also like to follow up on Eric's earlier question about TPU. I would like to further understand what impact TPU might have on the company's profit margins in the future. Specifically, will the profit margins from the TPU business be higher than the overall profit margins of Google Cloud, thereby enhancing the overall profit margins? Or is it said that due to factors such as hardware investment and infrastructure costs, the TPU business will lead to profit margins lower than the current level of Google Cloud business?

Sundar Pichai: First of all, the level of solutions we provide is very rich. One of the values of our "full-stack layout" is that users choose Google not just to purchase a model, but to obtain a complete solution. For example, in the fields of cybersecurity or data analysis, customers deploy complete business solutions, of which the model is just a part. In the case of cybersecurity, for instance, customers can use products like Chronicle, Wiz, and our upcoming CodeMender to leverage AI tools to discover security vulnerabilities and fix them.

Similarly, in the field of data analysis. In the past, enterprise data was often scattered across different systems and isolated from each other. Now, users can integrate this dispersed data and build an intelligent analysis layer on top of the data through Gemini Enterprise.

Overall, in these applications, the model is just one component of the entire solution. I think this is very important.

Even in scenarios that seem to only use models, AI models themselves are gradually evolving into end-to-end intelligent systems. They are no longer just simple models but a complete system that includes workflows and intelligent agent workflows. To build such a system, you need not only the computing power required to train and run models but also high-quality data, suitable development and operating environments, the ability to continuously optimize models, and the capability to ensure data security and privacy protection that reassures users. Users need to be confident that their data and the behavioral trajectories generated from using the data are known only to themselves and that this information will not be fed back into the model training process in any way.

At the same time, enterprises also need to be able to configure systems, deploy services, manage resources, and allocate computing power in a secure manner. These are all complex end-to-end capabilities, and this is precisely the direction our entire cloud business is focusing on. Currently, we see very strong customer demand across these different links.

Of course, having our own models also allows us to further optimize these solutions and provide users with a more integrated product experience. At the same time, we will also offer model options from other companies within these solutions. In addition, we will provide underlying infrastructure capabilities, etc. From this perspective, we are adopting a complete full-stack strategy.

I believe Google has a very good competitive advantage and capability in this field.

Anat Ashkenazi: Regarding the profit margins of TPU, we will not disclose the profit margins of any specific product or type of infrastructure component separately.

Of course, designing and producing chips independently does bring us certain advantages. You can understand our business this way: TPU actually expands our overall Total Addressable Market (TAM). In other words, by providing solutions to customers who wish to deploy such systems in their own data centers, we have expanded new business opportunities From the perspective of overall profit margin for Google Cloud, the profit margin for cloud business this quarter has improved to 35.6%, which is an outstanding performance. This reflects the strong operational management capabilities of the entire business and also benefits from the efficiency improvements brought about by revenue scale growth.

Looking ahead to the third quarter and the remainder of this year, as I mentioned in the previous briefing, since the company is still in an environment where computing power supply is constrained, we plan to increase the use of third-party computing resources in the third quarter. This is a transitional arrangement; in other words, while we continue to build our own infrastructure capabilities, we will also utilize external resources to meet current demands.

However, due to the high costs of third-party computing resources, this investment will put some pressure on Google Cloud's operating profit margin in the short term.

Additionally, we have previously mentioned the issue of the Wiz acquisition and integration. This acquisition will also exert some pressure on the company's profit margin in 2026 in the short term.

Bernstein Research analyst Mark Shmulik: Anant, regarding the third-party computing procurement you just mentioned and the capacity supplement as a transitional solution, I would like to follow up on this issue. Is the most severely constrained business currently concentrated in a specific area? Or, from an overall perspective, are all business lines facing a relatively common supply pressure for computing power?

Moreover, from a macro perspective, or from the company's overall view, has there been any change in management's thinking when allocating computing resources among different businesses? For example, in the face of limited computing resources, how do you weigh the allocation? How much will be used for Google Search? How much for AI model training? How much for Google Cloud? How does management assess the priority of computing power investment among these directions?

Sundar Pichai: I can answer this question.

Regarding computing power allocation, I want to emphasize again that our fundamental principle is to ensure that Google can continue to maintain its leading research and development capabilities in the field of AGI (Artificial General Intelligence). Of course, the specific investment priorities will change as the cutting edge of AI model technology develops, because how much computing power future model competition requires depends on the stage of industry development.

Therefore, we will first prioritize meeting the research and development needs of next-generation AI technologies.

On this basis, we will prioritize supporting the company's core product areas, such as search business, YouTube, etc., while also supporting Google Cloud. Within the Google Cloud business, we will prioritize ensuring computing power for running and providing our core AI model services, specifically including Vertex AI, Gemini Enterprise, data analytics solutions, cybersecurity solutions, and so on.

Overall, our current computing power investment is mainly focused on two directions: on one hand, supporting consumer-facing core products; on the other hand, supporting core AI services for enterprise customers. This is our basic approach to computing power allocation Regarding the third-party computing power transition plan you just mentioned, I want to emphasize one point: in the short term, we are indeed helping some very important large Google Cloud customers get through this special phase. These customers have presented us with a significant amount of new demand. To meet these demands, we may need to incur higher costs in the short term, such as an increase in computing power procurement costs over the next few months. However, from the perspective of the entire cooperation cycle, as we gradually increase our own computing power supply, these long-term cooperation opportunities will bring very good investment returns. This is also a key factor we consider when evaluating these opportunities.

In other words, we ask ourselves: Is it worth incurring higher initial costs for a few months to seize an important customer cooperation opportunity that lasts for several years? If this customer can bring us very attractive profits and returns in the coming years, we believe such an investment is worthwhile.

I hope my explanations can help everyone understand the considerations the company takes into account when making these decisions.

Citi Analyst Ron Josey: I would like to switch topics to discuss Google's search business. Philip, could you please elaborate on the commercialization of the search business and YouTube? The management now has more observations on the changes in user behavior and commercial value brought by AI search, and you also mentioned the strong performance of the YouTube business in your earlier briefing. Could you further introduce how advertisers are leveraging Google's stronger personalization and precise targeting capabilities to enhance advertising effectiveness and return on investment?

Additionally, investors often ask us a question: Although Google already has such a large business scale, the advertising revenue from the search business has still maintained a year-on-year growth of 17%. Given this scale, what factors do you think have driven the search business to achieve record-high growth performance?

Philip Schindler: In the second quarter, revenue from search and other businesses grew by 17% year-on-year, which is actually the result of multiple business segments working together, and it also benefits from the deep integration of Gemini into the advertising system.

From a more macro perspective, the growth of our search business comes from the contributions of multiple industry sectors. Among them, the retail sector contributed the most to growth, followed by finance, technology, media, and entertainment, which also contributed significantly to growth.

I think it is very important to note that Gemini has significantly enhanced our ability to understand user needs and match the right ads to users. This directly addresses your question. We are applying the Gemini model across the entire advertising technology system, including improving ad quality, optimizing the tools used by advertisers, and supporting ad displays in new AI search experiences. We are deeply integrating Gemini into various tools used by advertisers to help them create and optimize campaigns more efficiently. And this is just part of the overall capability enhancement In addition, we have launched AI-driven advertising products, such as Google Ads AI Max. It helps advertisers break through the limitations of traditional keyword advertising and automatically discover more potential opportunities. In other words, it allows us to gain a deeper understanding of user intent, match ads more accurately, and find potential demands that were previously uncovered. As we mentioned earlier, AI Max is helping us tap into billions of new search opportunities that could not be monetized before.

Regarding YouTube, overall, the growth of YouTube's advertising business mainly comes from two directions: performance advertising and brand advertising. I previously mentioned that we are seeing rapid development in the "living room scene" (TV). This means that more and more users are watching YouTube on large-screen devices like TVs, which is creating new advertising opportunities for us.

In the future, whether in brand advertising or performance advertising, we have very promising advertising product development paths. Among them, Demand Gen advertising and YouTube Shorts remain very promising growth directions. Of course, these businesses also rely on precise targeting capabilities, which means finding the right users better.

Additionally, we are continuously innovating in performance advertising. For example, we have launched shoppable ads in the TV large-screen scenario, helping users complete product purchases directly. These will further drive our growth in the retail advertising sector.

Wells Fargo Analyst Ken Gawrelski: My first question is about computing power investment. Considering the supply chain constraints and rising supply chain costs currently seen by management, how does management view the returns on the company's investment in computing power infrastructure by 2027? How does management believe the returns on computing power investments in 2027 and beyond will differ from those in the past few years, especially in 2025 and 2026?

Additionally, I would like to understand how management assesses the return levels of future computing power investments in the current supply-demand environment, and what changes there are compared to infrastructure investments in the past few years.

My second question is about Waymo's autonomous vehicles.

If management were to adjust Waymo's corporate structure or organizational form, what key factors would you primarily focus on? I know management has been cautious about this issue in the past and has been reluctant to comment too much. However, now that Waymo's business is continuously expanding, it is clear that the business has reached a certain scale, and the leadership of the team is strong, with good business momentum. Under what conditions does management believe that allowing Waymo to operate independently from Alphabet would be a more appropriate choice?

Sundar Pichai: Regarding your first question, if I understand correctly, you are actually asking how we view the returns on future computing power investments Our investments are always based on a strict return on invested capital management framework. If input costs rise, we will take these cost changes into account and ensure that we can achieve corresponding returns by adjusting our pricing capabilities. All of these factors will be incorporated into our investment planning.

Regarding the company's computing power investment in 2027, as I mentioned in my previous answer, we are currently seeing very strong demand signals, such as long-term cooperation agreements, existing customers continuously renewing contracts, and sustained growth in future demand, etc. We will formulate investment plans and allocate resources reasonably based on these demand situations.

If there is any change, I believe the current market environment is actually healthier than it was a year ago. It is precisely due to these changes that we are more confident in continuing these investments.

Regarding the Waymo business you mentioned in your question, I want to say that our primary focus right now is on driving Waymo to scale and achieve commercial growth. We have supported the Waymo business through internal arrangements within Alphabet and provided the team with ample room for development.

One of Alphabet's long-standing important advantages is our ability to think and plan long-term, continuously invest in business development, and provide the team with a clear long-term development roadmap. For a business like Waymo that requires long-term investment, it is very important to advance according to a long-term plan and gradually scale up.

In summary, our current real focus is on continuously expanding Waymo's business scale and transforming its enormous development potential into actual results. This is also our current business priority.

(Continuously updating...)