Alphabet's second-quarter cloud revenue surpassed Wall Street expectations, though sales related to its search engine business were slightly below forecasts, potentially heightening market concerns over its significant investments in artificial intelligence.
Following the earnings release, Alphabet CEO Sundar Pichai, Chief Business Officer Philipp Schindler, and CFO Anat Ashkenazi held an analyst conference call to address questions about the business.
Key Takeaways from the Analyst Q&A Session
Morgan Stanley analyst Brian Nowak: I have two questions. First, Sundar, over time, we've seen more generative AI (GenAI) products and tools enter the market, and companies continue to ramp up their investments. Could you discuss how your view on the return on invested capital (ROIC) for the entire generative AI space has evolved compared to a year ago? Specifically, has your perspective changed regarding the scale of this opportunity and the timeline for realizing returns?
My second question is for Anat. In the prepared remarks, you spent considerable time detailing the compute capacity constraints the company is facing. My follow-up relates to future capital expenditures. Looking ahead to 2027, if the company aims to address the current compute capacity bottleneck, what is the thought process behind setting future CapEx budgets? How might the principles for planning and controlling capital expenditures evolve? How does management determine the appropriate level of investment needed by 2027 to effectively alleviate the current capacity constraints?
Sundar Pichai: In my view, we are still in the very early stages. I believe generative AI will drive structural changes with long-term impact across multiple domains.
For our core information business, the current frontier AI capabilities open up numerous possibilities. I remain convinced there is significant work ahead to translate these advanced capabilities into consumer product experiences. For instance, consider end-to-end agents where AI can perform more, and more valuable, work for users. I see these as substantial opportunities. Capturing these opportunities, I believe, will yield very attractive returns.
The same holds for the enterprise market, as reflected in our product demand and observable in metrics like our growth rates. In my conversations with numerous CEOs and enterprise users, I find most companies are just beginning to explore AI's potential, with a long journey ahead to unlock its full value.
We often discussed the evolution of cloud computing in the past. Initially, only a small fraction of enterprise workloads had migrated to the cloud. Now, consider a similar question: what percentage of workloads are truly AI-native or AI-powered today? In my view, that percentage remains very low, indicating we are still in the early phases of AI technology development.
From an ROIC perspective, we are pursuing a full-stack strategy. We see strengthening momentum across consumer markets, enterprise markets, and from a developer ecosystem standpoint.
In summary, if our thinking has changed over the past year, it is that we are more optimistic about the opportunities ahead and more confident in the long-term value these opportunities represent compared to a year ago.
Anat Ashkenazi: Regarding your question on future capital expenditures and addressing compute supply constraints, the fundamental environment has not changed; supply remains the primary constraint on business growth.
We have highlighted this for several consecutive quarters. Demand for compute resources, both from external Google Cloud customers and from our internal businesses, continues to grow very strongly. Our capital allocation principle remains unchanged: we will continue to invest as long as we see attractive returns. As Sundar mentioned, we are confident in these investment opportunities.
In capital planning, we maintain a long-term perspective. This means we plan for business needs over several years while also focusing on the next year and near-term requirements, accelerating infrastructure build-out to meet growing demand.
Over the past three years, we have significantly increased infrastructure capacity, yet demand growth has outpaced our expansion. Consequently, like the broader industry, we operate in a constrained environment with tight compute supply. Simultaneously, we are actively enhancing our own supply capabilities. Alphabet's full-stack approach allows us to continuously improve the operational and technical efficiency of our technical infrastructure, delivering more compute from our existing resources.
Therefore, as long as we see attractive investment opportunities, we will maintain our investment pace.
JPMorgan analyst Doug Anmuth: I also have two questions. First, Sundar, could you discuss your confidence in Alphabet's ability to keep its Gemini models at the global forefront?
This week, Google released new Gemini Flash models. However, compared to some other leading AI labs, Google seems less frequent in product releases and has lower market visibility. I'd like to hear your perspective on external concerns: Can Google continue to develop and maintain leading-edge model capabilities? Also, could management discuss Google's strategy in the AI programming space? And how does management plan to further close the gap with competitors in this specific enterprise segment?
My second question is for Anat. Recently, Google conducted both equity and debt financings. Could you outline how the company currently views its optimal capital structure? Also, in the financing decision process, how does management weigh the "cost of debt" versus the "cost of equity"?
Sundar Pichai: The competition for frontier global models is evolving rapidly, with a fast pace of technological iteration. The landscape can shift at any given point.
Google has consistently had models at the global forefront, and we maintain leadership in many key capabilities. Of course, there are areas we believe require further improvement, such as AI programming and agentic coding, where teams are focused on advancing these capabilities.
Take Gemini 3.6 Flash, released this week. The Flash series is our core, most widely used, and one of the most in-demand model products. It offers an excellent balance of performance, cost, reliability, and latency, making it highly popular. We have integrated Gemini Flash across our product suite, including solutions for cybersecurity, data analytics, and more.
For customer service scenarios, enterprises require not only high-quality speech and real-time interaction capabilities but also real-time reasoning. Professional services firms rely more on high-quality content summarization and text generation. The Gemini Flash model performs excellently in these applications.
In agentic coding, we are iterating rapidly and will continue to release new versions. For example, compared to Gemini 3.5 Flash, the newly released Gemini 3.6 Flash shows over a 10 percentage point improvement on the DeepSuite benchmark while also achieving higher token efficiency.
We are using this model extensively internally and are testing it with several enterprise users in programming scenarios. The progress from the previous version to the latest took about six weeks. We will maintain this rapid iteration pace.
Regarding the competition for frontier global models, we are both confident and committed. For the next generation of leading global models, we believe a larger-scale foundational model is necessary. We have initiated pre-training for Gemini 4 with very ambitious goals. I am excited by the progress seen so far and believe its eventual release will impress users.
To maintain a leading position in the next phase, we need a larger, more capable foundational model like Gemini 4. Therefore, I believe our most important task currently is to execute this work steadily and excellently.
Anat Ashkenazi: Regarding your question on capital structure and the recent equity and debt financings. We plan for future investments by assessing needs for the next year as well as over a longer-term, multi-year horizon.
First, we evaluate how much investment can be supported by cash flow from operations. As seen in today's results, Google generates strong, healthy operating cash flow, which remains our primary funding source.
On top of that, we consider debt financing. Over the past 12 months, we have significantly expanded our debt financing. Approximately a year ago, our debt balance was around $16 billion; it has now increased to about $100 billion, forming a diversified debt financing system across multiple currencies and regions. Additionally, while supporting continued growth, we aim to maintain a robust, resilient balance sheet. This was a key consideration in our recent entry into the equity financing market.
For now, barring special circumstances, we have no plans for further equity financing. As known, our previous equity financing plan included an At-the-Market (ATM) program. For the foreseeable future, the company will continue using this mechanism to offset dilution from stock-based compensation and cover related tax obligations.
Overall, in capital allocation, we balance three main dimensions: operating cash flow, debt financing scale, and equity financing arrangements. Our goal is to meet ongoing investment needs while maintaining a healthy, robust balance sheet.
Goldman Sachs analyst Eric Sheridan: I have two questions about TPUs (Tensor Processing Units). Sundar, could you discuss key learnings from Google's ongoing expansion of TPU deployment? Specifically, as TPU adoption broadens, how does management view future market demand for TPUs? Also, in the coming years, as demand from external customers for TPUs grows alongside internal business needs for custom chips, how will the company allocate and balance resources between these areas?
I have a follow-up. Anat, in the remarks, you mentioned TPUs positively impacted Google Cloud. Could you provide more detail? For instance, what percentage of Google Cloud's committed but unrecognized revenue backlog is related to TPU business? Furthermore, how will TPU-related business convert to revenue over the next few years, and what impact might it have on company margins?
Sundar Pichai: First, we are very pleased with the progress of our TPU product roadmap. Both performance and the competitive advantages it delivers meet expectations, which is why we use TPUs extensively internally.
Regarding TPU resource allocation, our primary principle remains unchanged: first, ensure sufficient TPUs to support Google's leadership in Artificial General Intelligence (AGI). This is foundational. Simultaneously, market demand for compute is very strong, so we must balance internal and external customer needs.
For most Google Cloud customers, we currently use TPUs and GPUs to run and provide Google's own AI model services, such as Vertex AI, Gemini Enterprise, and the rapidly growing array of AI agent applications. For customers who want direct access to TPUs as infrastructure to run their own models, we are meeting this demand by expanding infrastructure deployment. For example, we are evaluating and advancing more solutions to deploy TPUs into customer data centers or partner data centers, similar to our project with Blackstone.
This approach allows us to better balance compute resource allocation: on one hand, prioritizing the most advanced AI model development needs; on the other, continuously supporting our own AI model services for consumer and enterprise customers.
Anat Ashkenazi: Regarding revenue recognition for TPU system sales: after signing the aforementioned agreements, these contracts are included in Google Cloud's backlog. The majority of Google Cloud's $514 billion backlog comes from Google Cloud Platform (GCP) customer contracts, but it also includes orders related to TPU system sales. Post-agreement, we build inventory in advance for future delivery. As we scale and expand, this process is reflected in operating cash flow. Revenue recognition typically begins when we start delivering TPU systems to customers.
This quarter, we recognized only a small portion of the total contract value. We will ramp up deliveries through 2026, with the majority of revenue from this agreement expected to be recognized in 2027.
Barclays analyst Ross Sandler: I'd like to return to the "model wars" within the industry. Sundar, a follow-up on model release velocity. We noted Google's recent third-party compute agreement with SpaceX. Beyond this, what other measures is Google taking to accelerate the iteration and release pace of Gemini models to further increase update speed?
You also mentioned the Gemini Flash series models, which are performing successfully. However, given the already intense competition in the low-cost model market with many players, do you believe the Flash model's market positioning remains the right direction? What are management's views on future AI model market trends and how Google will participate?
Sundar Pichai: I'll mention two points. First, we constantly consider how our products can cover the full AI model competitive landscape. For Google's users, we aim to offer the best model choices across different price points. That means we focus heavily on building industry-leading top-tier models while also providing powerful, cost-effective models suitable for large-scale applications.
Thus, you see we have launched different model versions like Gemini Flash-Lite and Flash Pro. Our goal is to cover the entire AI model market, from highest-performance frontier models to high-value models, meeting diverse user needs.
Regarding model release velocity, we will continue to accelerate iteration. For example, we launched Gemini 3.5 Flash at Google I/O, followed by Gemini 3.6 Flash. You will see continued updates to this series, with improvements in areas like agentic reasoning.
Simultaneously, we are dedicating significant resources to advancing Gemini 4. This is a crucial project. Our goal is for Gemini 4, upon release, to be competitive at the frontier of AI models at that time. Therefore, we are concentrating substantial compute and R&D efforts to achieve this.
In building Gemini 4, we are also creating a stronger foundational platform. This will enable faster subsequent version releases and iterative upgrades. Increasing model update speed, aiming for roughly monthly release cadences, is a key part of our Gemini 4 development plan.
MoffettNathanson analyst Michael Nathanson: My first question is for Sundar, again on the model competition discussed throughout today's call. Could you discuss what you believe are Google's true long-term competitive advantages in this new AI model race? In other words, even if all companies eventually reach similar model capability levels, what strategic advantages do you think will help Google maintain growth momentum? What advantages can keep Google ahead?
Anat, I have a follow-up on Eric's earlier TPU question. Could you elaborate on the potential margin impact of TPUs? Specifically, would TPU business margins be higher than Google Cloud's current overall margins, thus lifting them? Or, due to hardware investment and infrastructure costs, would TPU business margins be lower than current Google Cloud business levels?
Sundar Pichai: First, we offer solutions at multiple layers. One value of our "full-stack" approach is that users choose Google not just for a model, but for complete solutions. In cybersecurity or data analytics, for instance, customers deploy full business solutions where the model is just one component. In cybersecurity, customers might use Chronicle, Wiz, or our upcoming CodeMender, leveraging AI tools to find and fix security vulnerabilities.
Take data analytics. Previously, enterprise data was often siloed across different systems. Now, users can unify this data and build an intelligent analytics layer on top using Gemini Enterprise.
In these applications, the model is just one part of the overall solution. This is very important.
Even in scenarios that appear to be just model usage, AI models themselves are evolving into end-to-end intelligent systems. They are no longer simple models but include workflows, agent workflows. Building such systems requires not just compute for training and running models, but also high-quality data, appropriate development and runtime environments, continuous model optimization capabilities, and trustworthy data security and privacy assurances. Users need confidence that their data and usage patterns are known only to them and are not fed back into model training.
Enterprises also need to configure systems, deploy services, manage resources, and allocate compute securely. These are complex end-to-end capabilities, which is precisely what our cloud business is building. We see strong customer demand across these different layers.
Of course, having our own models allows us to further optimize these solutions and provide a more integrated product experience. We also offer other companies' model choices within these solutions and provide underlying infrastructure capabilities. Viewed this way, we pursue a comprehensive full-stack strategy.
I believe Google has very good competitive advantages and capabilities in this area.
Anat Ashkenazi: Regarding TPU margins, we do not disclose profitability for individual products or infrastructure components.
Certainly, designing and producing our own chips provides advantages. Think of our business this way: TPUs effectively expand our total addressable market (TAM). By offering solutions to customers who want to deploy such systems in their own data centers, we unlock new business opportunities.
From a Google Cloud overall margin perspective, cloud operating margin improved to 35.6% this quarter, a strong performance reflecting robust operational management and benefits from revenue scale.
Looking ahead to Q3 and the rest of the year, as mentioned earlier, given the constrained compute supply environment, we plan to increase usage of third-party compute resources in Q3. This is a transitional arrangement; while we continue building our own infrastructure capacity, we utilize external resources to meet current demand.
However, due to the higher cost of third-party compute, this will pressure Google Cloud's operating margin in the near term.
Additionally, we previously mentioned the Wiz acquisition integration. This integration will also pressure margins in 2026 in the near term.
Bernstein Research analyst Mark Shmulik: Anat, a follow-up on the third-party compute procurement mentioned as a transitional capacity supplement. Is the constraint most severe in a specific area, or is it a broadly felt compute supply pressure across business lines?
Also, from a macro or company-wide perspective, has management's thinking changed regarding allocating compute resources across different businesses? With limited compute, how do you currently weigh allocation? How much goes to Google Search, AI model training, and Google Cloud? How does management assess the priority of compute investment among these areas?
Sundar Pichai: I can answer this.
On compute allocation, I reiterate our foundational principle: first, ensure Google can continue leading-edge R&D in AGI. Of course, specific investment priorities will evolve with the frontier of AI model technology, as future model competition's compute needs depend on the industry's development stage.
Thus, meeting next-generation AI technology R&D needs is our baseline.
Beyond that, we prioritize supporting our core product areas like Search, YouTube, and Google Cloud. Within Google Cloud, we prioritize compute for running and providing our core AI model services, including Vertex AI, Gemini Enterprise, data analytics solutions, cybersecurity solutions, etc.
In summary, our compute is primarily allocated in two main directions: supporting consumer-facing core products and supporting core AI services for enterprise customers. This is our basic framework for compute allocation.
Regarding the third-party compute transitional approach, I emphasize: in the short term, we are helping some very important, large Google Cloud customers through this unique period. These customers have presented us with significant new demand. To meet it, we may incur higher costs in the near term, such as increased compute procurement costs over the next few months. However, over the full partnership lifecycle, as we gradually increase our own compute supply, these long-term partnership opportunities offer very attractive returns. This is a key factor in evaluating these opportunities.
In other words, we ask: is it worth bearing higher initial costs for a few months to secure a multi-year, important customer partnership? If the customer can deliver very attractive profits and returns over several years, we believe such investment is justified.
I hope this helps explain our decision-making considerations.
Citi analyst Ron Josey: I'd like to shift topics to Google's search business. Philipp, could you elaborate further on Search and YouTube monetization? Management now has more observations on user behavior changes and commercial value from AI Search, and you mentioned strong YouTube performance in the remarks. Could you discuss how advertisers are leveraging Google's enhanced personalization and targeting capabilities to improve ad performance and ROI?
Also, investors often ask: despite Google's massive scale, Search advertising revenue still grew 17% year-over-year. At this size, what factors are driving Search to its highest historical growth rates?
Philipp Schindler: Search and other revenues grew 17% year-over-year in Q2, resulting from contributions across multiple business areas and the deep integration of Gemini into our advertising systems.
From a broader perspective, our Search growth came from multiple industry verticals. Retail contributed the most, followed by Financial Services, Technology, and Media & Entertainment, which also showed significant growth.
A crucial point: Gemini significantly enhances our ability to understand user needs and match relevant ads. This directly addresses your question. We are applying Gemini models across our advertising technology stack, improving ad quality, optimizing advertiser tools, and supporting ad experiences in new AI-powered search. We are deeply integrating Gemini into tools advertisers use to create and optimize campaigns more efficiently. This is part of the overall capability enhancement.
Additionally, we launched AI-powered ad products like Google Ads AI Max. It helps advertisers move beyond traditional keyword-based constraints, automatically discovering more potential opportunities. It allows deeper understanding of user intent and more precise ad matching, uncovering previously untapped demand. As mentioned, AI Max is helping us unlock billions of new search queries that were previously unmonetizable.
On YouTube. Overall, YouTube ad growth comes from two main areas: performance advertising and brand advertising. As mentioned, we see strong growth in the "living room" (TV) context, where more users watch YouTube on large screens, creating new ad opportunities.
Looking ahead, we have exciting ad product roadmaps for both brand and performance advertising. Demand Gen and YouTube Shorts remain high-potential growth areas. These also rely on precise targeting—finding the right users.
We are also heavily innovating in performance ads, such as introducing shoppable ads on TV screens, enabling direct product purchases. These will further drive growth in retail advertising.
Wells Fargo analyst Ken Gawrelski: My first question is about compute investment. Given the supply chain constraints and rising costs management sees, how does management view the returns on compute infrastructure investments planned for 2027? Compared to past years, especially 2025 and 2026 compute investments, how might returns in 2027 and beyond differ?
Also, in the current supply-demand environment, how does management assess the return profile of future compute investments compared to past infrastructure investments?
My second question is about Waymo.
If management were to adjust Waymo's corporate structure or organization, what key factors would you focus on? I know management has been cautious about commenting on this previously. But as Waymo scales, the business clearly has significant scale, strong leadership, and good momentum. Under what conditions would management consider it more appropriate for Waymo to operate independently outside of Alphabet?
Sundar Pichai: Regarding your first question, if I understand correctly, you're asking about our view on returns for future compute investments.
Our investments are always based on a disciplined ROIC framework. If input costs rise, we factor those changes into our considerations and adjust pricing power accordingly to ensure appropriate returns. All these factors are incorporated into our investment planning.
For 2027 compute investments, as I mentioned earlier, we see very strong demand signals—long-term agreements, existing customer renewals, continued future demand growth. We base investment plans and resource allocation on these demand dynamics.
If anything has changed, I believe the current market environment is actually healthier than a year ago. Thanks to these changes, we are more confident in continuing these investments.
Regarding your question on Waymo, I'd say our current focus is on scaling Waymo and driving commercial growth. We have supported Waymo through internal Alphabet structures, providing the team ample runway.
One of Alphabet's long-standing strengths is our ability to think and plan long-term, continuously investing in businesses and providing clear long-term roadmaps. For a business like Waymo requiring long-term investment, having the ability to execute against a long-term plan and scale gradually is crucial.
In summary, our real focus is on scaling Waymo's business and translating its tremendous potential into tangible outcomes. That is our current business priority.
Comments