The Zhitong Finance App learned that the core line presented by Google's parent company Alphabet (GOOGL.US) during the 2026 third quarter results conference is that AI is driving the accelerated expansion of search usage, cloud revenue, and TPU businesses, but there is still a time lag between the conversion of search traffic to revenue. Significant investment in computing power construction and TPU inventory occurred before the cash flow recovery associated with AI occurred. The company's third-quarter revenue increased 25% year-on-year to US$127.6 billion; the year-on-year growth rate of search queries was the highest in five years, but the growth rate of search advertising revenue slowed to 12%. Management clearly acknowledged that in addition to changes in exchange rates and advertising spending in some industries, the rapid growth of AI model queries and the current low level of monetization associated with them are also one of the reasons for the slowdown in revenue growth in Google's traditional flagship business, the search engine business.
Management said at the performance conference that the company is expanding advertising to more AI overviews and AI model inquiries, promoting multiple rounds of dialogue advertising in the international market and the US, and exploring the commercialization of smart devices through general commercial agreements, smart payment agreements, and general shopping carts. At the same time, the Gemini first-party model API processes nearly 32 billion tokens per minute, an increase of 40% over the same period last year, more than four times that of the same period last year; Gemini application users reached 1 billion. The scale of use continues to expand, providing a foundation for subsequent monetization, and the improvement of advertising coverage and business models determine how this portion of traffic is converted into revenue.
Cloud computing and TPU formed the strongest growth end, but they also brought about a new cash flow structure. Google's strongest business in the third quarter earnings season — Google Cloud's revenue for Google's cloud computing business — increased 103% year-on-year to US$30.7 billion, with operating profit reaching US$11.1 billion, more than three times that of the same period last year. The operating profit margin rose from 23.7% to 36.2%; backlog orders reached US$561 billion, and it is expected that slightly more than 55% of this revenue will be confirmed over the next 24 months. Management emphasized that Google Cloud Platform contributes the vast majority of cloud revenue, and even if TPU system sales are excluded, growth is still extremely strong; enterprises using Gemini are also further driving computing, storage, database, and network security requirements.
TPU, on the other hand, extends from its own cloud services to customer data centers and partner channels, and system sales revenue is expected to increase significantly in 2027. However, since this business requires purchasing components, establishing inventory, and then recovering cash through delivery and invoicing, cloud business profits are growing rapidly and can occur at the same time as the group's free cash flow is under pressure. Operating cash flow for the third quarter was US$42.3 billion, capital expenditure was US$51.4 billion, and free cash flow was negative US$9.1 billion; the increase in TPU inventory and the payment of US$5.2 billion in European Commission fines also further dragged down cash performance. Curian specifically stated that TPU related inventories all correspond to signed contracts, and the financial impact is reflected in the use of working capital before contract delivery, and not all of the company's own infrastructure capital expenses.
2027 is expected by management to be an AI-driven “golden year of productivity,” and it will also be a year where capital expenditure continues to grow significantly. The “golden year” emphasized by Google's parent company CEO Pichay mainly refers to models such as Argon and Carbon unleash productivity in programming, long-term tasks, and product innovation: internal teams have achieved a 35% increase in code delivery and a 30% reduction in errors. The company is also developing distilled versions of models, safe operating environments, and tools to connect the ecosystem to expand enterprise intelligence capabilities to a wider range of consumer products. Therefore, “gains” are more accurately reflected in productivity growth, product iteration, and commercial expansion. It is not management's promise that free cash flow will be fully corrected or substantially improved next year.
Alphabet maintains the 2026 capital expenditure guideline of $1950 billion to $205 billion, and is expected to increase significantly in 2027; infrastructure depreciation, energy and operating costs will continue to affect profit and loss, and the expanded use of third-party computing power in the fourth quarter will also put pressure on recent cloud profit margins. Management still anticipates that the return on invested capital throughout the investment cycle will be significantly higher than the cost of capital, but the phone call will reveal that the key transaction lines are already very clear — AI demand and cloud revenue are being realized. The next thing the market needs to evaluate is whether search and monetization efficiency, TPU delivery and repayment, and continued expansion of investment can work together to form stronger cash generation capabilities.
The following is the full text of Google's parent company Alphabet's performance conference call (artificial intelligence tool assisted translation):
Operator:
Welcome everyone. Thank you for waiting for Alphabet's Q3 2026 results conference call. Currently, all participants are in listen-only mode. After the management has finished speaking, we will move on to the Q&A session. (Operator instructions)
I now entrust the conference to today's host, Jim Friedland, Head of Investor Relations. Please get started.
Jim Friedland:
Thank you. Good afternoon everyone, and welcome to Alphabet's 2026 third quarter results conference call. Together with us today are Sundar Pichai, Philip Schindler, Thomas Curian, and Anat Ashkenazi.
Let me now briefly introduce the safe harbor clause of the forward-looking statement. Some of the statements we make today about the company's business, operations, and financial performance may be forward-looking statements. Such statements are based on current expectations and assumptions, and are subject to various risks and uncertainties, and actual results may vary materially. Please refer to our 10-K and 10-Q reports, including risk factors. We are under no obligation to update any forward-looking statements.
In this conference call, we will introduce financial indicators prepared in accordance with US GAAP and non-US GAAP. A reconciliation table between non-US GAAP metrics and US GAAP metrics has been included in the results press release issued today. The press release has been made available and accessible to the public through our investor relations website abc.xyz/investor. Unless otherwise stated, our reviews are compared year over year.
Now I'll leave the conference call to Sundar.
Sundar Pichai:
Thanks Jim, hello everyone. The third quarter was a strong quarter for Alphabet. Our deep full-stack layout is driving the growth of all businesses, and there are many opportunities ahead.
Let me first give you a brief overview of what happened this quarter. Alphabet's revenue increased 25% year over year, and the current annualized revenue scale has reached 500 billion US dollars. This is excellent growth. The search business continues to be strong. Our investment in AI is driving growth in usage and overall enquiry volume, including commercial inquiries. In fact, the year-on-year growth rate of our enquiry volume hit the highest level in the past five years. Google Cloud revenue increased 103%. AI infrastructure and solutions are the main drivers of cloud business growth, and customers see us as a leading supplier in this field, so demand is strong. We are working to bring more production capacity online to meet this long-term demand.
At the same time, demand for TPU systems from a broad range of customer groups is strong, and we expect this demand to continue as the scale of the business expands. I am delighted to welcome Thomas to this conference call, where he will further explain the momentum of the cloud business. In terms of AI models, customers are using our models at an unprecedented level. Our first-party model API currently processes nearly 32 billion tokens per minute, a 40% increase over the previous quarter, and more than four times the processing volume a year ago. With the launch of Gemini 4 Argon and Carbon announced today, we're excited to be at the cutting edge of technology. The Gemini app currently has 1 billion users and is the fastest growing product in Google's history. Finally, including trial users, we have over 150 million YouTube Music and Premium subscribers worldwide.
Today, I'll take a closer look at our AI advancements, global consumer platforms represented by search and YouTube, and our long-term layout. Start by introducing AI, the most notable of which is the Gemini 4 series cutting-edge model. Argon, which we launched two weeks ago, is receiving strong demand and usage. In fact, even with limited computing power, Argon's usage climbed 3.5 times faster than Gemini 3 in the first week of launch on our developer platform. Argon has excelled in finance, law, cybersecurity, and science. Developers and cloud customers are already using it to sort through legacy codebases, accelerate scientific research, robotics, discover product improvement opportunities, and specialized field tasks that require deep reasoning.
Within Google, in the past few weeks alone, Chrome, advertising, and search teams have used Argon to dramatically speed up code writing and optimization, delivering 35% more code and 30% fewer errors. Argon is just the beginning of the Gemini 4 series. Today, we're announcing the launch of Carbon. This is our strongest cutting-edge model, with the most advanced cybersecurity and long-range mission capabilities. In the recently updated Artificial Analysis Intelligence Index, it ranked first with a score of 59. Carbon surpasses Argon in almost every field, including full-stack web development, multi-modal understanding, complex knowledge work, and scientific reasoning. We're first rolling out Carbon to some trusted cyber defenders, then expanding the scope of use; the US government is also testing it.
In addition to the Gemini 4 series, yesterday we also announced the launch of Genie 4, the most advanced world model to date, capable of simulating a safe, realistic, interactive, and natively audio-enabled world in real time. This is an important basic module for training robots and intelligent bodies. This rapid progression highlights the strong momentum in the development of our model. We've reached a stage where AI can help us speed up training and release. I expect this process to continue to accelerate over the next 12 months, driven by recursive self-improvement. We'll soon launch more models, including more efficient versions, to help developers optimize speed and cost, and deploy to billions of users.
Next, let's talk about searching. Search is bringing the benefits of AI to more people than any other product in the world. Just like when we transitioned to mobile over a decade ago, we first focused on creating the best search user experience before commercializing it. We're building a unified experience that seamlessly connects AI overviews, AI models, and traditional search. People are increasingly using our AI capabilities and searching in new ways. Their queries are longer, more personalized, and more multi-modal. The average query length in AI mode is three times that of traditional searches. Globally, voice, image, or document is used once in every five AI mode searches. We currently have over 50 billion monthly visual searches. One area I'm excited about is helping people actually get things done, whether it's booking, filling out forms, scheduling appointments, or shopping. As AI becomes more personal and begins to act on behalf of individuals and businesses, trust and security are more important than ever. People already trust Google and rely on us to complete everyday tasks, and our leading position in this field remains an important differentiator.
We first focus on improving the experience of these agents through Gemini agents for enterprises, because enterprises have extremely high requirements for security, protection, and governance. Next, we'll bring these features to search, and of course the Gemini app's 1 billion users. People are using the Gemini app not just to ask questions, but to get things done, including complex multi-step agent queries. The Gemini 4 series has made significant progress in this regard and is already able to handle these complex tasks very well. Smart devices will be a unified layer across all of our products, so you'll see us further expand these smart experiences from search and Gemini apps to Chrome, Android, and Workspace products.
Next, let's talk about YouTube. YouTube is still the number one streaming platform in the US in terms of length of time spent watching, and continues to lead the living room viewing scene. As mentioned, its subscribers have also maintained good growth. At last month's Made on YouTube event, we announced more than 30 products and features for creators and viewers. One great example is a custom stream of information: you only need to choose a topic you like and get customized video homepage recommendations around that topic.
Philip will further introduce YouTube. Next, let's talk about our long-term layout. Waymo's development trajectory is very good. We have accumulated over 300 million miles of fully autonomous driving. Waymo currently operates in 15 major US cities, recently opened its services to the public in San Antonio, San Diego, and Tampa, and plans to expand to Canada, Germany, Singapore, and Spain. Our other long-term projects have also made strong progress. Isomorphic is using Argon to support drug design research. For the first time, we demonstrated magic state cultivation technology in quantum computing and launched TPU into space through the “Sun Catcher Program.” So it's been a very busy quarter, with more progress ahead.
Thank you to the employees for their hard work this quarter. I leave it to you, Philip.
PHILIP SCHINDLER:
Thanks Sundar, hello everyone. I'll focus on the momentum of search, YouTube, and partnerships. We continue to accelerate the deployment of the Gemini model across our advertising infrastructure to improve performance in the three areas I previously highlighted: ad quality, advertiser tools, and AI user experience.
The first is ad quality. Integrating the Gemini model into advertising infrastructure is improving our ability to understand user intent and increase ad relevance. For example, in the third quarter, we made the biggest update to product listing ads in ten years, adding new AI features to help users make more confident purchasing decisions. We've seen a 19% increase in satisfaction and a drop in dissatisfaction by more than half. We've also seen that AI models can better predict clicks and conversions, leading to stronger overall performance and deeper ad personalization. For example, through multiple innovations in personalized modeling, we brought 18% incremental advertiser value to Discover ads in 2026 and increased user sentiment by 8%.
Overall, AI is speeding up our ad development process and driving the launch of most of our products. The second is advertiser tools. We're making sure every advertiser, from global brands to small businesses, can seamlessly participate in our AI product interface. AI Max and Performance Max continue to gain widespread adoption. To give marketers more control, we're now introducing a new way for them to test different budgets and return on investment goals for multiple search campaigns in an AI Max A/B test.
To further improve performance, we've integrated AI creative tools directly into AI Max. In the third quarter alone, the Gemini model helped advertisers generate more than $1.5 billion worth of campaign material, an increase of more than seven times over the previous year. For example, after using AI Max to capture long queries and multi-language queries, Telstra increased conversions by 52% and reduced the cost per order by 36%, enabling customers to reach them at critical moments no matter how they search.
The third is AI user experience. Sandal mentioned the obvious behavioral changes we're observing: users are starting to ask a whole new category of questions. We look at search from a holistic perspective and see it as a unified product that encompasses traditional results and AI search. In the AI model, many queries are at the top of the marketing funnel, enabling us to capture new, higher-value commercial intentions. Furthermore, most queries in the AI model are incremental queries, which will create more monetization opportunities over time.
Additionally, our user research shows that people are able to accept ads in these interfaces and consider them highly relevant. In the AI model, the click rate of paid promotion of product list items and natural display of product list items is already at a similar level.
Based on these comprehensive results, we are now promoting the AI overview and high-quality advertisements in the AI model to more users and queries. This also includes expanding our monetization coverage globally for the first time. Today, we are announcing the expansion of AI-mode advertising to the international market, and the launch of multi-round dialogue advertising in AI mode in the US. As always, we take a prudent approach, prioritizing user experience and interaction, then expanding ad formats that have been proven to generate return on investment for advertisers. It is this discipline that provides a solid and lasting foundation for our future.
As Sandar said, the smart experience in our services is still in its early stages, and we're excited about future business opportunities. The Universal Commercial Agreement, or UCP, is being rapidly adopted by retailers. Currently, UCP supports direct checkout for hundreds of thousands of brands on Google platforms, including Chewy, Nike, Tapestry, and many others. We're building on the future of commerce by combining it with smart payment agreements. We are focusing on two directions: one is to unleash the ability of intelligent bodies to act on behalf of consumers, and the other is to design practical tools to give shoppers confidence and control. To support the second direction, we launched a generic shopping cart this quarter, which fits the way people actually shop on Google platforms today. This enables retailers to maintain relationships with customers while helping drive sales and reduce cart abandonment.
Next, let's talk about YouTube. We think of YouTube as a flywheel made up of creators, viewers, and advertisers. The first is the creator. Over 20 million videos are uploaded every day. We've made content creation easier with new generative AI tools in YouTube Studio and conversational editing assistants in Shorts and YouTube Create.
Let's take another look at the viewing situation. In the Made on YouTube campaign, we launched a short video series feature where creators can organize their short videos into different seasons and episodes. Continuous playback allows viewers to seamlessly discover new short videos and keep watching. The third part is monetization. First, let's talk about advertising. Our focus remains on creating value throughout the marketing funnel.
In the third quarter, the living-room viewing scene contributed more than one-third of total YouTube ad revenue. Small to medium advertisers continue to use Demand Gen to drive direct response results. On YouTube, Wayfair's Demand Gen campaign achieved a return on ad spend of 2.3 times that of traditional video coverage campaigns. Our YouTube Creator Partnership Program connects brands with over 3 million creators to help run effective advertising campaigns. Brands such as Lexus continued to collaborate with creators like Colin and Samir, increasing brand consideration by 11%. Over 1.3 million channels worldwide have joined YouTube Shopping.
Building on this momentum, we recently announced the addition of Amazon to YouTube Shopping. US creators can now directly tag Amazon products in videos, short videos, and live streams to drive direct conversions.
Importantly, our subscription ecosystem continues to complement the advertising business. Driven by YouTube Music Premium, the subscription business continues to grow faster than advertising. YouTube Premium has over 150 million subscribers. With Premium and Premium Lite, viewers have different options. As more people choose the Premium experience, viewers, creators, and our business all benefit. Subscription revenue helps support the creator ecosystem and move the platform forward.
Finally, let's talk about strategic partnerships. Businesses choose Google because we combine cloud, AI, advertising, and consumer platforms to accelerate their transformation process. We have expanded our partnership with American Express to support the continued modernization of its contact center and smart business partnerships around search and wallets. Unity has also signed an extended multi-year cloud agreement to support the next generation of engines using our full-stack AI capabilities. Our business priorities are clear: increasing advertisers' return on investment, systemically improving AI monetization efficiency, and expanding ecosystem partnerships to support the next generation of discovery experiences.
Finally, thanks to Google employees around the world for their outstanding execution, and for the continued trust of customers and partners. Thomas, leave it to you.
Thomas Curian:
Our strong momentum comes from our integrated AI product portfolio, spanning chips, models, data, security, and agent platforms, all designed to work together. Our growth has been highly diversified in terms of product lines, industries, and customers. Google Cloud currently has 18 product lines, and each product line has an annual revenue of more than $1 billion. This quarter's growth was driven by strong growth in Google Cloud Platform, and both AI and core GCP businesses performed well. Sales of TPU systems have also made an important contribution. Over the past 12 months, more than 2,600 customers have used at least 100 billion tokens on Gemini Enterprise, and nearly 75 customers have used at least 10 trillion tokens.
Nearly 90% of Fortune 100 companies use Gemini Enterprise. Customers using Gemini Enterprise in the third quarter of 2025 more than quadrupled their spending on us year over year. One of the key ways customers use Gemini is to do intellectual work. We've sold over 100 million seats for knowledge work across all of our AI services. Gemini Enterprise now includes the Gemini Agent, which is deeply integrated into our entire cloud product portfolio. Through an agent, an API, and a prompt window, it can conduct research, answer questions, write and run code, and automate processes. It provides businesses with strong security and cost control, as well as the widest selection of cutting-edge models.
Customers include BNP Paribas, which is deploying Gemini Enterprise for more than 65,000 employees to streamline workflows while maintaining strict enterprise-grade security and data governance. Tata Steel deployed more than 300 professional AI agents in just nine months to improve manufacturing safety, optimize the supply chain, and reduce customer service time by 50%. Ulta Beauty has built the Ulta AI shopping assistant, and its conversion rate is three times that of traditional browsing methods on the website.
Gemini is also driving the accelerated growth of our database, analytics, and cybersecurity products. Gemini now helps users find, analyze, and visualize information from structured and unstructured data. Gemini uses our knowledge catalog to find the right information sources and definitions to improve accuracy, reduce token costs, and make responses more reliable. Gemini uses our Lakehouse integrated platform to access information from multiple systems in a zero-copy manner.
To support agent analysis, BigQuery's query speed increased by more than 40% year over year, and query processing costs were reduced by more than 50% year over year. Customers include Deutsche Telekom, which uses BigQuery, Cloud Storage, and a knowledge catalog to integrate more than 40 legacy local systems into a unified lakehouse platform. Etsy, on the other hand, migrated 3Pb of data to a unified lake warehouse integrated platform. By using BigQuery and Iceberg, they are currently up to 60% faster in critical data correlation operations. Growing cybersecurity threats are also driving strong customer demand for our AI security platform. Our Gemini Flash cybersecurity model is a cutting-edge model that can automatically detect vulnerabilities in 20 programming languages, and its patching reliability is 2.6 times that of other models. With the launch of Argon and Carbon, there will also be better performance in the future.
Our security platform incorporates threat intelligence, prioritization of network responses, automated scans, code fixes, and monitoring. Gemini and Google Threat Intelligence independently analyze millions of external events every day, with an accuracy rate of 98%. Over 95% of US customers are adopting AI capabilities, and growth continues. These security tools protect critical infrastructure, including financial institutions such as Banorte and Jack Henry, and telecom service providers such as AT&T. All of this is supported by our leading AI infrastructure, which includes TPU systems, Axion CPUs, the latest Nvidia GPUs, and Agent Compute Engine designed specifically for AI agents. The TPU system is a dedicated system built for AI. It combines hardware, computing, storage and network with system components and software optimized for training and inference. Compared with the previous generation TPU, the maximum training performance of the TPU 8t reached 2.7 times, and the inference cost ratio of the TPU 8i increased by 80%.
To enable developers to easily migrate workloads to TPU, we recently worked closely with Meta to launch TorchTPU, which enables users to train, provide services, and deploy PyTorch models on TPU using the tools they already know and rely on. Currently, nearly 100 customers use TPU, and we're seeing strong and continuing growth in demand from many other customers. In addition to working directly and through Broadcom with Anthropic, our public customers include Apple, which is expanding the pre-training scale of its base models on the latest generation of TPU. We've also partnered with new cutting-edge model companies, including Mirendil and Physical Intelligence; with capital markets agencies, including Citadel Securities and Jump Trading; with universities such as Caltech; and public sector service providers such as Cassava and Cerescale.
As demand continues to be strong and continues to accelerate, we have recently established TPU system delivery partnerships with Crux AI and Oracle respectively, and more partnerships are underway.
Anat, leave it to you.
Anat Ashkenazi:
Thanks Thomas. Unless otherwise stated, my review will focus on year-over-year changes in the third quarter. I will first introduce Alphabet's overall performance, then introduce the performance of each division, and finally explain the outlook for the fourth quarter and the full year of 2026. We performed strongly in the third quarter, with our 13th consecutive quarter of double-digit revenue growth.
Consolidated revenue was US$127.6 billion, an increase of 25% in terms of reporting and fixed exchange rates. Total cost of revenue was $51.4 billion, an increase of 24%. Traffic acquisition costs were $16.4 billion, an increase of 10%. Other revenue costs were US$35.1 billion, an increase of 32%, mainly driven by hardware costs, increased depreciation, content acquisition costs, and technical infrastructure operating costs associated with TPU system sales. Total operating expenses increased 20% to US$35.6 billion. R&D expenses increased 36%, mainly driven by remuneration, mainly investment in AI talent, as well as an increase in depreciation. Sales and marketing expenses increased 22% due to increased investment and compensation in Gemini. General and administrative expenses fell 17%, mainly due to higher legal fees in the third quarter of last year, which had a comparative base impact, but this was partially offset by current year's legal expenses.
Operating profit increased 30% to US$40.6 billion, and operating margin was 31.8%. Other net revenue and expenses were negative $8.2 billion, mainly due to unrealized losses from our investment in SpaceX due to the fall in its stock price in the third quarter. Net profit and earnings per share declined due to a decrease in net other income and expenses. In the third quarter, we generated strong operating cash flow of $42.3 billion, with operating cash flow of $179 billion over the past 12 months. The increase in inventory required to sell the TPU system and the $5.2 billion cash payment of the 2018 European Commission fines after the appeal process ended had an adverse impact on operating cash flow.
Capital expenditure for the third quarter was $51.4 billion, the vast majority of which was spent on technology infrastructure to support our AI investments. About 65% of technology infrastructure investments in the third quarter were related to servers and network equipment, and these assets typically have a lifespan of six years. The remaining 35% is used in data centers and has a useful life span of seven to 40 years.
In the third quarter, our free cash flow was negative $9.1 billion, mainly driven by capital expenditure. The increase in inventory mentioned above to support sales of TPU systems and the payment of European Commission fines also affected free cash flow in the third quarter.
Free cash flow for the past 12 months was $19.7 billion. At the end of the quarter, we held $265.1 billion in cash and marketable securities, including $90.4 billion in tradable equity securities.
Long-term debt was $125.1 billion. We have also signed multi-year agreements to ensure technical infrastructure capacity, inventory, and leasing and energy supply. At the end of the third quarter, our future procurement commitments were $924 billion, and unleased leases and other obligations were $152 billion. These amounts reflect the maximum amount of contractual obligation; however, some of these agreements have variable pricing and purchase volume flexibility, which may reduce actual expenses as a result. As stated in the press release, the Board announced a quarterly cash dividend of $0.22 per share on common stock, which will be paid in December.
Next, I will introduce the segment results. Google Services' revenue grew 11% to $96.3 billion, reflecting strong growth in the search and subscription business. Total advertising revenue increased 11%, with growth across all major vertical industries. As stated in our outlook for the previous quarter, the exchange rate had a slight negative impact on the year-over-year growth rate, and the impact on the month-on-month performance was even more significant. Google search and other ad revenue grew 12% to $63.3 billion, with retail contributing the most, followed by finance. YouTube advertising revenue increased 9% to $11.2 billion, with brand advertising contributing the most, and the living room scene was particularly strong.
Online advertising revenue was $7.5 billion, up 1%. Subscription, platform and device revenue increased 11% to $14.3 billion this quarter, mainly driven by strong growth in YouTube subscriptions, particularly YouTube Music and Premium, and strong demand for Google One's AI package. Google Services' operating profit increased 17% to US$39.2 billion, with an operating margin of 40.7%. The Google Cloud division achieved record growth in the third quarter. Cloud revenue grew 103% to US$30.7 billion, mainly driven by strong GCP growth. Among them, AI businesses, including AI infrastructure and AI solutions, and core GCP businesses all performed well. Sales of TPU systems have also made an important contribution. GCP contributed the vast majority of revenue in the third quarter. After excluding the impact of the TPU system, the growth rate of cloud revenue is still extremely strong.
The operating profit of the cloud business was 11.1 billion US dollars, which more than tripled from the previous year. The operating margin increased to 36.2% from 23.7% in the third quarter of last year. The third-party transition computing power we discussed on our last conference call wasn't available until the end of the third quarter, so the negative impact on profit margins for the quarter was limited. Google Cloud's backlog of orders reached $561 billion in the third quarter, most of which were related to regular GCP contracts. We expect to recognize just over 55% of the total backlog of orders over the next 24 months as revenue.
Other Bets had revenue of $372 million and operating loss of $2 billion because we continued to expand Waymo's business and invest in important Other Bets projects. The operating loss from Alphabet-level business activities was US$7.7 billion, mainly driven by shared AI R&D expenses.
Now I'll present my outlook for the rest of 2026. The first is revenue. Based on the current spot exchange rate, we expect the negative impact of the fourth quarter exchange rate on consolidated revenue will be slightly larger than the previous quarter. This impact will mainly be reflected in search and YouTube ads. On the Google Services side, we will continue to face the high base formed by the acceleration of search business performance that began in the third quarter of last year, which will put pressure on revenue growth in the fourth quarter.
On the Google Cloud side, we expect TPU system sales to continue to expand in the fourth quarter and achieve significant revenue growth in 2027. In terms of profit margins in the cloud business, due to the limited supply environment, we began using third-party computing power at the end of the third quarter, and will expand usage in the fourth quarter. As we use this computing power, we expect profit margins to come under pressure in the short term.
Next, let's talk about investing. Our guidance range for capital expenditure of $195 billion to $205 billion for the full year of 2026 remains unchanged. We still expect a significant increase in capital expenditure in 2027 and will provide more details during the fourth quarter conference call. On the cost side, the sharp increase in technology infrastructure investment will continue to affect the income statement through higher depreciation costs and associated data center operating costs, such as energy costs.
Before entering the Q&A session, I'd like to briefly explain how we view return on invested capital and free cash flow. We are in a unique position to monetize AI investments through a variety of products for billions of users and businesses. Our investment in AI in the search field is driving the highest growth rate in five years in search volume. Google Cloud has generated tens of billions of dollars in AI revenue so far this year; driven by demand for AI packages, Google One is our fastest-growing subscription product.
Throughout the investment cycle, we expect to achieve an attractive return on invested capital that is still significantly above the cost of capital, and expect to generate strong operating cash flow. As the TPU system business expands, our large-scale investment in technical infrastructure and TPU inventory construction will put pressure on free cash flow. The third quarter was a strong quarter, highlighting our innovation and rigorous execution across all of our businesses, including the launch of Gemini 4. I would like to take this opportunity to thank our employees for their contributions to the company's performance.
The following is the performance conference Q&A session (AI-assisted summary and translation):
operator
(Q&A session)
Thank you. (Operator's explanation) The first question came from Doug Ammus of J.P. Morgan Chase. Your line is connected.
Q — Doug Ammuth
Thank you for taking the question. I have one question for Sundar and one for Philip. Sundar, you recently stated in a podcast that 2027 could be Google's golden age for productivity due to a new class of models and their integration with other products. I hope you can expand further and share what you think of future opportunities. Then, Philip, search revenue growth is the highest in five years, but search revenue growth is clearly slowing down. Can you help us understand the difference between the two? Is it just because you show fewer ads in the AI overview and AI model, and this part of the query now accounts for an increasing proportion of total queries? Thank you.
Answer - Sundar Pichai
Thanks Doug. I'd like to say that as the model's ability to perform long-term tasks, intelligent programming, etc. has increased, we have already begun to see incremental changes when using 3.8 Flash internally. But now I do see a very significant inflection point. Since we began deploying Argon internally, it has become one of the most important internal productivity levers we've seen in software engineering. We directly track the submission of high quality code, the speed of submission, and the reduction in the number of rollbacks. More importantly, people are starting to solve large-scale problems; at the same time, we are also beginning to see innovations in product and consumer experience after deploying these models. But all this only happened about four weeks ago. We are now starting to deploy Carbon internally, and Carbon has made a significant improvement over Argon.
So as we roll out these models and provide more computing power for Google employees to use widely, I anticipate — and that's what I meant before — and I'm even more excited now after that comment on the podcast. I think the next year will be a golden year for us to increase our productivity. So we're excited for the innovations that are coming. Philip?
Answer - Philip Schindler
OK, Doug, to answer the second part of your question. Search revenue increased 12% year over year to $63 billion in the third quarter. If we look at the change in growth rate from the second quarter to the third quarter, there are several driving factors behind it. The first is the exchange rate, which is the single biggest factor in the slowdown in search growth compared to the second quarter. Second, there are changes in specific vertical industries. On the retail side, some of our advertisers in the Asia-Pacific region have been affected by the EU's cancellation of the small import tax exemption policy. In terms of media and entertainment, the second quarter was unusually strong, driven by the FIFA World Cup, before spending returned to normal. The World Cup had more games in June, so the advertising budget allocated to June was actually higher than in July.
Next, to answer your next part of the question, it's worth noting that the strong increase in AI model queries also had an impact because of the lower level of monetization we saw in the AI model at the time. What needs to be clear is that we are very confident about the development trajectory of this field. The ad trial has shown encouraging results, and we are now promoting the AI overview and advertising in AI mode to more users and more queries, including the international market I mentioned earlier, as well as multiple rounds of conversation ads in the US.
Q — Doug Ammuth
Thank you.
operator
The next question comes from Morgan Stanley's Brian Novak. Your line is connected.
Q — Brian Novak
Thank you for taking my questions. I have two questions. Let's start with smart devices. You launched Spark to some users during the year. I know, Sundar, you've talked a lot about plans to launch more smart products to a wider user base. Can you help us understand what key technical barriers you need to overcome in order to launch smart products to all users? Is it model efficiency? Is it computing power? Or is the technology not ready yet? From a technical point of view, what more do we need to do?
Second question, next Doug's question about search, Anat, can you help us better understand the factors you mentioned that affect the performance comparison of the search business? How should we look at these comparison bases as we head into the fourth and first quarters? Thank you.
Answer - Sundar Pichai
Thanks, thanks Brian. First, intelligence is an exciting and dynamic field. As I look at our new models and see that they can automatically take on complex tasks and perform at extremely high quality, it's clear that the intelligent capabilities we've always imagined will become reality. So I'm really looking forward to bringing the next generation smart body experience to users. In response to your question, I would like to make a few comments. The Gemini 4 model series is very important. We are developing a distilled version of the Carbon model so that it can be efficiently deployed to 1 billion users in terms of computation, latency, and quality.
As such, it is an important investment area. Achieving large-scale deployment clearly requires not only a powerful model, but also a robust intelligent execution framework and a highly secure operating environment, and this is where GCP comes into play. We need a secure operating environment in the cloud that is available 24/7 for every user, as well as the right ecosystem of skills and connectors. All of these are foundational. We've been working hard to build and decided to launch it first, the Gemini smart device we announced for enterprises, because the company has the highest requirements for security governance and control. It has been well received, and we are expanding our deployment.
We'll bring the same experience to consumers later this year, apparently starting with the 1 billion users in the Gemini app. We've seen the way users use apps is shifting from chat to tasks and intelligent asana experiences. The trust users have placed in our products, and all of their contextual information is already there, I think, has created favorable conditions for us. But as a company, our real focus is on providing this unified experience across products, and you should see a lot of progress later this year.
Answer -- Anat Ashkenazi
Thank you. Regarding the second question, which is my comment, I think you're referring to the search business outlook. I mentioned a few factors. One of them is the negative impact of the month-on-month exchange rate. This was also the case this quarter compared to the previous quarter. Therefore, based on the current spot exchange rate, we expect this to have a negative impact on the fourth quarter results.
Then look at the year-on-year comparison. This quarter, we began to face the high base formed by the acceleration of search growth in 2025, that is, the situation in the second half of 2025. The comparative base figure associated with the third quarter includes an acceleration of growth of about 3 percentage points; we expect this situation to continue in the fourth quarter, facing the comparative base impact of about a few percentage points formed by accelerated growth at the end of 2025.
Q — Brian Novak
Great, thank you both.
operator
The next question comes from Goldman Sachs's Eric Sheridan. Your line is connected.
Q — Eric Sheridan
Thank you so much for taking the questions. Sundar, based on some of your recent comments on computing power limits, can you update what are your key priorities for computing power applications when the new computing power becomes available company-wide?
In the second part, I might ask you or Thomas to answer, specifically Google Cloud, how can you continuously adjust your views on computing power applications, that is, how to consider meeting the needs of cloud customers and providing TPU to the outside world? Thank you very much.
Answer - Sundar Pichai
Thanks, thanks Eric. I answered first, then Thomas answered the second part. The allocation of computing power is indeed an area where we invest a lot of time, because it requires advance planning. First, we feel very lucky to have an extraordinary array of opportunities ahead of us. We operate large-scale consumer businesses covering billions of people, hyperscale cloud businesses that are growing strongly, and of course other highly important areas under Google and Alphabet. We always plan for the long term, so every business area submits long-term plans under a very strict return on invested capital framework. So, overall, I would view all of this as a long-term planning process rather than a short-term allocation of work. Obviously, our AI model is the foundation for everything from search and YouTube to cloud, Waymo, and more. So it's the foundation, and we think it's a highly leveraged way. The full-stack approach brings significant efficiency improvements, and also allows us to have many choices in a wide range of opportunities. So this is clearly the foundation, and we will also be planning for the long term. In addition to that, as I said, we will actually review each business and adopt a rigorous long-term planning approach.
It's clear to us that we're seeing very strong demand signals in a wide range of areas. So we think this is an extraordinary series of opportunities, and we are investing in this moment. Thomas?
Answer - Thomas Curian
Eric, when it comes to cloud requirements, we always manage our business portfolio to ensure a broad range of customers. We have a diverse range of clients, including digital natives, AI labs, traditional enterprises, and financial and capital markets clients. So this is a broad customer base. Many of these customers initially started with TPU systems. For example, we already have over 100 important TPU customers in our cloud, and they first started using TPU in our cloud to learn how to use TPU. Keep in mind that a TPU system is a combination of hardware, system software, and compilers, all of which are assembled into a complete system, which also includes other components such as optical networks.
We have taken a very careful approach in providing TPU to our customers. The first is to build the right chip, so we have 8t for training and 8i for inference. Compared to previous generations, both have brought significant improvements in cost performance and performance. Second, we've optimized the software. Software optimization is designed not only to make the system run better and provide excellent performance, but also to enable a seamless migration. For example, we recently introduced PyTorch, which opened up a much larger market for us by optimizing PyTorch for TPU. Third, we are expanding our distribution channels.
In the past, customers could use TPU in our cloud. Now, for customers who want to deploy TPU systems in their own data centers, we will also provide TPU systems to their data centers. We are also expanding our range of TPU offerings to other suppliers. For example, we announced our collaboration with Oracle. Additionally, we have jointly launched an agreement with Kuroishi where the joint venture Crux AI will provide customers with TPU systems.
Demand from AI laboratories, capital markets institutions, financial services institutions and enterprises is extremely strong, and growth is strong. If we imagine that more than 100 customers will grow further into more customers and expand their use to TPU systems, we can see strong demand in the future.
operator
The next question comes from Barclays Ross Sandler. Your line is connected.
Q — Ross Sandler
Thank you. I'd like to continue discussing the TPU resale business you just mentioned. Anat, last quarter, you seem to have said that revenue will be recognized when the system is sold. Therefore, if the computing equipment and various surrounding components of a 1 gigawatt complete TPU system cost about 35 billion US dollars, then how much of this will actually be recognized as revenue by us? Compared to the current profit margin level of the Google Cloud division, what kind of profit transfer will this generate?
Then I think the second part of the question is, what does this mean for Alphabet's capital expenditure intensity when it comes to helping companies like Anthropic bring more of their own infrastructure online? As you hand over some of your future capital expenses to the AI lab, does that mean your own data center no longer needs that high capital expenditure intensity? I'd love to hear your thoughts. Thank you.
Answer -- Anat Ashkenazi
OK, I answered these two questions together. First, obviously, we don't provide product-specific financial data, but I can explain how the financial performance associated with TPU system sales should be understood. As a reminder, Thomas can also further explain that the TPU system is not just a chip, but a special system that combines various elements such as hardware, computing, storage, and networking. As a result, we sell far more than just one chip. Software, training, and reasoning all provide significant added value for customers purchasing TPU systems. The way this business works is also related to the second part of your question: when we sign an agreement with the customer to sell the TPU system, you first see it reflected in our backlog of orders.
Then, we started building up our inventory. So think about this in terms of capital expenditure versus cash. We build inventory before generating revenue. I think I mentioned this on my last conference call. You'll see your inventory balance increase. Previously, the inventory balance was mainly for some of our hardware products; now, those TPU systems that have not yet been sold will also be reflected in it. As a result, you'll see cash flowing out, and we're building an inventory base. Then, when we start delivering these TPU systems and bill our customers, you'll see the cash flow in. So it's not the traditional capital expenditure you usually understand — I think, that's the core of your problem — but a different cycle of cash flow generation.
Currently, as we are in the process of establishing and expanding this business, you'll see these cash requirements in the cash flow statement and an increase in inventory. So far, we have recognized some system sales as revenue, but we expect more revenue to be reflected in 2027.
operator
The next question comes from Bank of America's Justin Post. Your line is connected.
Q — Justin Post
OK, thank you. I have a few questions about transitioning to smart bodies. You guys did launch a generic shopping cart. It's clear that people love and are using the smart device use case. As this situation becomes more common, how do you plan to manage the shift to how search is used as an agent? Do you guys think this might be a disadvantage? Second, in the long run, do you think smart entities can actually improve traffic monetization capabilities?
Then I'll ask again about TPU profit margins. We wanted to know if you think this is a very good business for the company and maybe explain why? If you can't provide profit margins, why do you think this is a good way to use this computing power? Thank you.
Answer - Philip Schindler
Justin, we're still in the early stages of smart body development. But maybe directly to the second part of your question, we think smart body experiences can clearly be incremental. Our focus is to create experiences that users love, are trustworthy, secure, and can operate on a large scale. Frankly speaking, these experiences are supported by smart business infrastructure modules, such as general commercial agreements, shopping carts, smart payment agreements, etc., as you mentioned. When users shop, there's a part of the whole process they don't like, that is, the tedious work, and frankly, automating it with AI is perfect. But we also want to enhance the product discovery experience rather than simply automating and eliminating that experience, because in many cases, people really love the process. Users love choices and want to consider many factors other than price, such as brand loyalty, customer service, etc.
Therefore, when we design these intelligent workflows, we will consciously leave time and space for users to explore these options. As a simple example, the smart body can respond: this is the insurance product we found, and there is also a very competitive option, one with a higher premium, maybe a lower deductible, and is currently offering a very special discount. So, given that there's almost never just one perfect option when making a decision, we're actually very confident that we can provide a very interesting and engaging advertising model for smart use cases. Obviously, you can also imagine subscriptions and other payment methods other than ads. But I'm really excited about the advertising opportunities in the smart world.
Answer - Sundar Pichai
Overall, I'd like to add that I would view this as a process of expanding opportunities. Just like a few years ago, incorporating AI chat use cases into search enabled people to ask more questions and get more done in search, and today's usage is reflecting that. I think smart bodies will once again be an opportunity to expand the types of problems we can solve for our users. In my opinion, people are always underestimating how many components are needed to provide users with these agents — I mean these complex experiences. The search may seem simple on the surface, but behind it are extremely deep investments and our long-term building capabilities, such as knowledge maps and our understanding of the world. So I'm excited about the smart body and will see it as a new frontier, bringing a new set of use cases to search. So we're really looking forward to continuing.
Answer - Thomas Curian
Regarding TPU systems and their profit margins, we do not separately disclose the profit margins for each product or infrastructure component. However, we believe that designing and manufacturing TPU systems can provide multiple benefits. First, let me remind you that the TPU system is not just a chip, but a complete system. Considering the cost performance advantages of TPU systems over competitive products, accelerated growth in computing demand, and accelerated scarcity of computing resources, we are very focused on continuously improving profit margins as the scale of our business expands.
There are two other points. First, Anat mentioned that we buy components to build inventory, but we only do this after signing a contract. Therefore, all inventory related to the TPU system corresponds to the promised contract.
Furthermore, deploying TPU systems in third-party facilities and customer data centers can first expand our serviceable market. Data for some industries is stored locally and cannot be moved to the cloud, and this approach allows us to bring accelerated computing to where the data is located. Additionally, it reduces the need for us to build facilities ahead of time to accommodate these TPU systems.
operator
The next question comes from Bernstein's Mark Schmullick. Your line is connected.
Q — Mark Schmullick
OK, thanks for taking the question. I'm still focusing on the same few topics. Philip, with regard to smart devices, if you looked at some use cases with more intelligent characteristics in Gemini and AI models, and also observed the integration of Google products with some third parties, did anything surprise you or change your opinion? I know we've talked a lot about business, but maybe it also includes areas such as consumer productivity.
Then, ask Thomas or Anat to answer. Regarding cloud growth, can you share the growth rate of Google Cloud after excluding sales of TPU systems, and how much of it was limited by computing power? With some third-party computing power going live in the fourth quarter, is it reasonable to assume that the Google Cloud business will accelerate growth? Thank you.
Answer - Sundar Pichai
As for our observations on these early smart asana use cases, maybe it's up to me to answer the first part. We're really seeing this trend. Recently, we added a “Task” tab to the Gemini app. People are actually entering agent queries, so we can observe these behaviors. In all of these experiences, users really want to be surprised, so they actually have high standards. What encourages users to reuse is whether you can better handle these time-consuming tasks and ultimately deliver truly satisfying results. Like what we've seen in other fields like programming, when it's only 70% complete, it doesn't seem like it's working at all; once it crosses a certain threshold, it actually takes off. So we're closely studying these situations, identifying gaps, and building on all the foundations I've mentioned before. I believe that later this year and in 2027, our various product interfaces will make a lot of progress in this area.
Next, talking about TPU, I would like to add one more point to the TPU system: in the semiconductor industry, economies of scale are very important. Therefore, I regard our ability to both deploy TPU in our own data center and provide TPU systems to other data centers as the ability to expand the scale of the product, which will directly translate into better economic benefits.
Answer - Thomas Curian
In response to questions about the rate of cloud growth, our Google Cloud division achieved record growth in the third quarter, with cloud revenue rising 103% to $30.7 billion. The biggest contributor to growth came from Google Cloud Platform, where AI and core GCP businesses all showed strong performance. As a result, GCP contributed the vast majority of third-quarter revenue. After excluding the impact of TPU system sales, cloud revenue growth is still very, very strong. Demand for all product lines is also strong.
We saw three aspects. First, when people use Gemini, we see Gemini's own revenue increase; but we also see that using Gemini also drives growth in other parts of the product portfolio. For example, when customers build agents, they need to use computing resources to perform tasks, so they buy computing resources. They use storage to preserve the agent's memory, and use databases to provide so-called context for the agent to reason. Obviously, given the various cybersecurity threats, our cybersecurity product portfolio is also growing strongly. Therefore, we are confident in growth and the diversity of sources of growth. We're also seeing new customers starting to use GCP, and new customer adoption and associated growth are accelerating. Therefore, we also see it as a direction of growth. Obviously, if we have more computing power, we can meet more demand at any given time, but we're carefully managing all of these factors and growth.
operator
The last question came from MoffettNathanson's Michael Nathanson. Your line is connected.
Q — Michael Nathanson
Thank you. I have two questions for Sundar. In an earlier response to Eric's question about TPU and cloud, we heard that you would be willing to work with some of your competitors, or companies we might think are competing with you in the model and cloud space. So I'd like to know what do you think about the idea of working with companies that may be competitive and why is this strategy an opportunity rather than a risk in the long run?
Second, with regard to consumer smart behavior, one of the debates relates to safety and trust. Can you guys share what you've learned so far about safety and trust? Are these variables important to consumers? If they matter, then what are you doing to protect consumer trust and privacy when building consumer intelligence tools? Thank you.
Answer - Sundar Pichai
Thanks Michael, great question. Regarding the first question, we have long divided the things we build into two categories: those built as products, and those built as platforms. We have to make a clear distinction. When you build something as a platform, you need to work very hard to get the platform right and ensure that it serves everyone.
We have a long history of operating many platforms, whether it's a browser, an operating system, or a very successful cloud platform. In all of these areas, we have been working hard to provide a platform to all participants and be excellent customers. As you've already seen, we work on models with other companies, including Apple, and we compete with Apple in many, many fields.
Similarly, we see the TPU system as an important platform under construction. We want to achieve world-class standards in construction methods. That means putting it under market forces, being tested, and competing with customers with the most demanding workloads. In my opinion, it is this approach that ensures that the relevant teams remain at the cutting edge of technology. That's the way we think. Over time, I think this approach will eventually always expand opportunities. We've been doing it this way for a long time, and we've been doing it very well.
Second, with regard to intelligent behavior, going back to the previous question, I think I should have emphasized this more: when people are exposed to these intelligent experiences, we clearly see that they care about safety and trust. This is really a very, very important factor affecting how they use these products. Over the decades, we've built trust by providing experiences in a secure way in many consumer products for billions of users; at the same time, we've implemented these requirements end-to-end when building models and agents. We conduct extensive testing before deployment, and monitor in real time during large-scale deployment. I think all of this will ultimately be a huge differentiator. We've heard users express these views, and I think they'll respond with real choices as this cutting-edge field continues to evolve.
Q — Michael Nathanson
Thank you.
operator
Thank you. This concludes today's Q&A session. I will now return the meeting to Jim Friedland to ask him to make a further statement.
Answer — Jim Friedland
Thank you all for attending the conference today. We look forward to speaking with you again during the 2026 fourth quarter results conference call. Thank you and have a nice evening everyone.
operator
Thank you all. That concludes today's conference call. Thanks for participating, you can hang up now.