For Q4, Alphabet expects slightly larger FX headwinds to consolidated revenue than in Q3. Search will absorb the largest portion of that impact and will also lap the sharp acceleration enjoyed in the back half of 2025. Both of those items will again weigh on the segment’s growth. Cloud TPU system sales will keep scaling in Q4 and throughout 2027. And again, margins for the segment will be moderately impacted by using 3rd-party capacity to ease its current shortages.

GOOGL trades for about 22x forward EPS at $295/share. EPS (excluding its equity portfolio) is expected to grow by 27% Y/Y next year and by 21% Y/Y the year after that.

Gemini is growing like a weed. As seen below, the app crossed 1B monthly active users vs. 950M last quarter, making it the fastest-growing product in Google’s history. 

On a trailing 12-month basis, customers using 100B+ tokens rose 30% Q/Q to 2,600 and customers using 10T+ rose 114% over the last two quarters to 75. Nearly 90% of the Fortune 100 uses Gemini Enterprise (same as last quarter), and customers that were using it a year ago more than quadrupled their spend with Google Y/Y. Use cases across the broad knowledge work category (which now has 100M+ seats sold) are clearly becoming increasingly popular, and that momentum should continue with the Gemini agent launch for Gemini Enterprise. This greatly augments the agentic capabilities available to customers across Gemini, upleveling research reports, end-to-end code creation and a single agent to automate a wide array of workflows. It also features built-in security and AI resource usage optimization so customers don’t waste money.

Gemini is also boosting GOOGL’s database, analytics and cybersecurity products. For example, Gemini can now find, analyze and visualize both structured and unstructured data. It leans on Google’s knowledge catalog to find the right sources, helping output accuracy and optimizing token costs. Its lakehouse also pulls from several systems without needing to copy anything over, further reducing potential costs and boosting efficiency. 

For cybersecurity, its Gemini Flash Cyber models automate vulnerability discovery across 20 programming languages and patch issues 2.6x more reliably than alternatives. That should only get better with its Argon and Carbon (more on those below) AI models.

For review, its new 8th-generation TPUs (8t for training and 8i for inference) deliver up to 2.7x better training performance and 80% better inference price-performance vs. the last generation. TPUs (and other custom accelerators) continue to win a larger share of the overall chip opportunity. Generalist GPUs are amazing at many more things. TPUs are a bit more efficient at the specific machine learning workloads that are vital for AI… and they’re cheaper for those tasks too. Pair that with how scarce compute is right now and it makes sense that Alphabet is confident in steadily improving TPU margins as this business scales.

But great chips also need easy on-ramps. That’s where TorchTPU comes in. Alphabet built it with Meta to let developers run PyTorch models (the most popular framework for building AI models) on TPUs with the tools they already know and love. That’s big for reducing friction. One of the long-standing knocks on custom chips has been how painful it is to move workloads over to them. That hasn’t been an issue for GPUs, and this chips away at their edge.

Distribution is broadening too. TPU systems are now available in customer data centers and through other providers like Oracle and the Crux AI joint venture with Blackstone (more partnerships in the works). Demand was called “extraordinary” across AI labs, financial services firms and enterprises, and Apple and Anthropic were both named as TPU users.

A quick note on customer counts, as the commentary was a bit confusing. Alphabet has 100+ “significant” TPU customers. That figure refers to customers using TPUs inside Google Cloud. Think of it as the training ground where customers learn how to use these chips. From there, some go on to buy TPU systems for their own data centers or access them through partners like Oracle. That 100+ pool is the pipeline for external TPU sales. Alphabet did not give a separate count for customers using TPUs outside of its cloud.

Last quarter, the Gemini 3.5 Pro delay had investors worried that Alphabet was falling behind OpenAI and Anthropic. Sundar promised a near-monthly cadence of releases and told us Gemini 4 was being trained. Gemini 4 is here and has been very well received. 

And just like Sundar promised 3 months ago, the pace of model debuts will keep picking up. AI is now helping Alphabet train and launch its own models faster, and GOOGL expects that to intensify over the next 12 months thanks to recursive self-improvement (RSI; models teaching and building other models). More efficient versions optimized for speed and cost are coming soon too. This will help GOOGL win the 90%+ of workloads that don’t need the absolute best model and instead seek to minimize price or latency.

Alphabet’s AI investments continue to drive more query volume across its search business. Despite the ongoing AI disintermediation fears and the proliferation of consumer agents, Y/Y query growth hit a 5-year high across its various AI and traditional search tools. AI-based products continue to provide convincing evidence of strong incrementality. There’s a small amount of cannibalization, but a lot more growth to more than offset that.

So then… Why did search growth fall from 17% Y/Y last quarter to 12% Y/Y this quarter? A few reasons. First, while monetization for some AI products like AI Overviews is in line with traditional search, that is not yet true for AI Mode. That’s rapidly proliferating and slowing the segment down a little bit for now. Encouragingly, there are many reasons to believe this issue will be temporary. All of the click through rate (CTR) data they’ve seen on AI surface ad experiments looks encouraging enough to prompt a meaningful rise in ad load and introducing AI Mode placements in markets outside of the U.S. Both of those things should help a lot more with growth going forward. Outside of this, Asia Pacific retail advertisers were hit by the EU cutting de minimis exemptions while lapping the World Cup during Q2 meant an end to that media and entertainment tailwind. While all of those things had impacts, the biggest factor was actually FX. It meaningfully worsened Q/Q and caused most of the 5-point slowdown. Interesting to see analysts concoct stories about AI starting to chip away at this business when these were the actual sources of the slower growth and are all unconcerning to me.

Some are worried about agents ruining Alphabet’s search ad business. Aside from all the evidence already discussed pointing to that not happening, agents are also a big new opportunity for Alphabet too. Alphabet will have a product that directly competes with Claude in the near future and they will be able to seamlessly add connectors and build sponsored product placement, transaction revenue and subscription revenue streams. Alphabet has taken its time on this and elected to start with the Gemini agent for enterprise before moving to consumers. That’s because the benchmark for sound, scaled security is higher at Fortune 500 companies than it is for you or me. The agents working for these enterprises make it crystal clear that they’ll work well for consumers too, and it sounds like GOOGL is finally ready to soon release Gemini Spark (its 24/7 personal consumer assistant) to the masses. For now, the mega-cap continues to learn from early users, perfect the value proposition and gear up for global scaling.

In terms of what this does to the ad business if agents are doing the browsing, GOOGL views it as additive. There will continue to be human-led queries and research, as not everything needs automating. People putting together playlists don’t want AI to listen to the songs for them. Someone watching a clip on YouTube doesn’t want it summarized for them in a creepy chatbot voice. Travelers enjoy searching for cool places to visit on their next trip. In all of these instances, AI can greatly enrich the experience, but it won’t replace the person wanting to be part of it and wanting to express some opinion throughout. As long as the person is on the screen and not leaving things entirely to AI, there will be easy opportunities to show them ads and collect revenue. That’s why it’s so encouraging to see the experiments discussed above going so well. It’s also why effective scaling of products like its Universal Commerce Protocol (UCP) to provide rails for agentic commerce is so important too. It’s why Universal Cart’s ability to facilitate AI transactions while keeping merchants in control of customer data is so needed. And finally, this evolution makes it clear why Google’s Knowledge Graph, giving it unparalleled insight into human intent and what agents need to do to create delight, is such a secret weapon. All of these things deeply ingrain GOOGL in the future digital infrastructure and economy. Sundar called agents an expansion opportunity, just like AI chat in search was before them. 

Not too long ago, we were talking about chatbots as Google’s demise. That was far from true. All agents really do is make those chatbot outputs a lot more actionable and productive for people… and I think that will fuel Alphabet just like the ChatGPT moment and scare ended up fueling it.

The company is habitually using Gemini to improve ad system targeting and reporting just like Meta has so successfully done. During the quarter, it made its biggest product listing ad update in 10 years with these tools and notched 19% higher user satisfaction because of them. In addition to that, another AI-inspired update improved click and conversion predictions. For tools, Alphabet added new split testing capabilities to AI Max (AI service for search campaigns to uplevel user intent understanding). Telstra used it to capture long-form and multilingual searches to boost conversions by 52% and cut cost per order by 36%, showing what the tech can do in practice. It also used Gemini to help advertisers create 1.5B+ campaign assets, marking 7x+ Y/Y growth.

YouTube ad growth slowed to 9% Y/Y as it was also impacted by moderately larger currency headwinds and spend normalizing post World Cup. It was a busy quarter for the product roadmap across all three platform stakeholders. Creators got conversational editing tools, viewers got improved Shorts organization and feed customization and advertisers like Lexus tapped into the YouTube Creator Partnership to boost brand consideration by 11%. And finally 1.3M+ channels are now onboarded in YouTube Shopping, with Amazon joining the roster this quarter.

Why does Alphabet need to spend so much? The cloud backlog reached $561B vs. $514B last quarter and $155B last year. Just over 55% of it will be recognized in the next 24 months. That’s roughly $300B in already contracted revenue over two years for a segment doing about $123B annualized. And most of it is tied to typical GCP contracts. That’s why it has $924B in future purchase commitments and $152B in pre-commencement leases and other obligations lined up. It’s necessary if they want to take a big piece of this pie. And oh boy do they. They must spend to service all of the demand and they’re steadfast in thinking returns tied to the revenue will be attractive. Signals informing this optimism include rapid Gemini growth, upcoming agent launches, historically strong Google One subscription traction thanks to AI plans, insatiable TPU demand, world-class cloud infrastructure and more.

Waymo continues to rock. It has now driven 300M+ fully autonomous miles and operates in 15 major U.S. cities vs. 11 two quarters ago. It just opened service to the public in San Antonio, San Diego and Tampa, with Canada, Germany, Singapore and Spain on deck.