Earlier this month, I published More Of The Latest Thoughts From American Technology Companies On AI (2026 Q3). In it, I shared commentary in earnings conference calls for the third quarter of 2026, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. 

A few more technology companies I’m watching hosted earnings conference calls for 2026’s third quarter after I prepared the article. The leaders of these companies also had insights on AI that I think would be useful to share. This is an ongoing series. For the older commentary:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here
• 2024 Q2 – here and here
• 2024 Q3 – here and here
• 2024 Q4 – here, here, and here
• 2025 Q1 – here and here
• 2025 Q2 – here and here
• 2025 Q3 – here, here, and here
• 2025 Q4 – here, here, here, and here
• 2026 Q1 – here, here, here, and here
• 2026 Q2 – here, here, here, and here

With that, here are the latest commentary, in no particular order:

Airbnb (NASDAQ: ABNB)

Airbnb’s management thinks AI is helping the company build and improve products faster than it has ever done before; Airbnb shipped more than 50 products and features in 2026 Q3, up 70% from a year ago

Then AI changed the speed of everything. We had spent years laying the groundwork, but now AI is helping us build, test, and improve products faster than we have ever done before. In Q3 alone, we shipped more than 50 products and features, nearly 70% more than a year ago because of AI.

Airbnb’s management recently introduced AI search for guests to improve the search experience; guests can find the right experience by typing in entire sentences; AI search is especially popular with Gen Z and younger travellers

In Q3, we brought AI into more of the guest and host experience. For guests, we introduced AI search so they can describe what they’re looking for in their own words and find the right home service experiences for the trip. If you want to see AI search, just go to our app or our website, and you’ll see the very top of the page, there’s a toggle. If you open the feature and toggle on, you can basically type entire sentences, like natural language, and then it will basically deselect all the filters. It’s working really well. It’s especially popular with Gen Z and younger travelers who are very AI native. AI can also suggest relevant search filters, it can summarize important details of homes, and help guests compare places they’re saving.

Airbnb’s management recently introduced new AI-powered features for hosts, such as making it easier to create 3D floor plans, price homes competitively, and understand a property’s performance; Airbnb has improved hosts’ customer support with an AI assistant; the AI assistant is available in more than 50 languages, and around 40% of all customer support issues were resolved entirely by the AI assistant in 2026 Q3, up a third from 2026 Q2; management will expand the AI assistant to provide voice support in the US; the AI assistant helped lower Airbnb’s customer support cost per booking by 17% year-on-year in 2026 Q3; Airbnb’s customer service NPS (net promoter score) remains as high; Airbnb is using AI to help hosts with dynamic pricing of their properties; dynamic pricing is still early, but the adoption is great and has real potential to help hosts earn more over time while lowering prices across the market place

In Q3, we brought AI into more of the guest and host experience…

…For hosts, we’re using AI to make it easier to create 3D floor plans. These are really cool. They can help you price your homes competitively and understand how the listings are performing. We’re also using AI to transform customer support. Our AI assistant is now available in more than 50 languages, and this next stat is really key. In Q3, approximately 40% of all customer support issues were resolved without a human agent, and this is up from a third in Q2. This has been huge for us, and we’re expanding our assistant to voice support in the U.S. so guests and hosts can help get faster help over the phone. What this means is that customer support cost per booking has declined more than 17% year over year, driven in large part by improvements to our AI assistant. And if you’re wondering, customer service NPS is as high as it’s ever been…

…We do not have any data to share today about dynamic pricing, but the adoption is really great. I think one of the things we have noticed, this is a really key point, is that there is this win-win where dynamic pricing can generally lower prices across the marketplace and still allow hosts to make more money. So some nights they make more money, other nights they make less. But we are really able to find these sensitivities where if you just charge like $10 less, you will make enough more bookings that you will actually make a lot more in the long run because your occupancy is much higher. And that is not an intuitive thing, and it is good for the guest because it makes it a little more affordable. And it is a really huge lever. It is extremely early. This is something that AI is really good at tuning, and the thing about AI is you need to tune it and get more data. So, I think in the coming quarters, we should be able to have some more results.

Airbnb’s management thinks people will not being using one AI agent for everything; management thinks the popular AI agents of today are good for routine tasks, but planning a trip is a different order of difficulty; 80% of travellers enjoy spending time researching trips; Airbnb is building its own AI agent; management thinks Airbnb has 3 advantages when building its own agent, (1) a thorough understanding of the Airbnb marketplace, (2) Airbnb is part of the reservation, and (3) Airbnb is building a new interface specifically for travel; Airbnb’s agent will not be relying only traditional chat, but will instead also include maps, photos, comparisons etc; over time, management wants Airbnb’s agent to help guests plan their entire trip on Airbnb; management will roll out the agent in 2027 H1; management envisions Airbnb’s agent to be able to go outside the walls of Airbnb for travel bookings and this will be done via MCPs (model context protocols); management thinks general agents must be 2x-5x better than vertical agents; management is developing sponsored listings for Airbnb but it’s in the context of the company’s AI agent

There’s a lot of discussion about whether people will eventually use one AI agent to do everything. And if they do, what does that mean for Airbnb? Are we going to be cut out? Well, I don’t think people are going to use one AI agent for everything. You see, today’s AI agents, like Muse and Instinct, are good at simple routine tasks like making a restaurant reservation or scheduling an appointment. But planning a trip is different. Travel is aspirational. Discovering where to go and where to stay is part of the experience. In fact, 80% of travelers say they enjoy spending time researching trips. People want to look at photos, they want to compare homes, they want to make decisions with the people they’re traveling with. And I don’t think the general-purpose agents are the best to do that.

That is why we are building our own AI agent. And we have three distinct advantages when we build our own agent. First, we understand the Airbnb marketplace better than anyone. For nearly 20 years, we’ve helped guests and hosts find places to stay with millions of hosts. Every home is different. Guests care about the home, the location, the reviews, and the person hosting them. And we spend years learning how guests make these decisions and how hosts operate. The understanding gives us a significant advantage in building an AI agent that helps people find and book the right place to stay. Second, we are part of the reservation itself. This is important because we know that travel doesn’t end with a booking. In fact, that’s when it starts. Because Airbnb handles reservations, payments, and customer support, our agent is able to help guests throughout their trip. This means we can manage reservations, we can help you make changes to your plans, and accessing Airbnb support and protections when they need them. These are things an outside agent cannot do easily on their own. Third, and perhaps most importantly, we’re building a new interface specifically for travel.

We don’t think the traditional chatbot is the right way to plan travel. It’s the wrong interface. So what we are doing is developing a custom-designed, AI-native experience that responds to what guests are looking for. Instead of relying on traditional chat alone, which is very text-heavy, which again, is definitely not the way you should plan travel, our agent draws from a library of hundreds of travel-specific interface components, maps, photos, comparisons, calendars, itineraries, and compose them in real time to match what they’re looking for. This is going to feel a little bit like generative UI in the sense that it’s going to offer up the right interface at the right time for your specific travel needs…

…One problem with a general agent is you have to create like a general interface, because the agent has to be able to do everything. The general interface generally becomes almost entirely text-based, right? Because you can ask about travel, medical information, deliveries. You can’t know what they’re going to ask. As you get more specialized, you can develop a really custom interface that is perfectly bespoke for that type of use case. With travel, we’re able to design these really custom components. If you want a car rental, the agent generates this little UI of a car, and it’s like beautifully in a custom-generated interface. If you want to invite co-travelers, the agent helps you. There’s going to be, I think, an incredible user interface experience that what I would describe is a kind of post-chat-based interface. It’s going to have a chat element to it. That’s what we’re going to be doing…

…Over time, we want our agent to help guests plan their entire trip on Airbnb, all in one place. We’re going to start rolling it out to guests in the first half of next year…

…With regard to agents going outside the walls of Airbnb, yes, of course. That will be through MCPs. We’re going to be integrating a number of MCPs with our trip agent in the first half next year…

…There is not one part of travel that we are not eventually going to do, and every part of travel will have MCPs, will have APIs, and eventually, we do anticipate all bookings going through the app for every vertical in travel…

…I think basically the way to think about it is, in order for specialization to win, it must not be 20% better. It must be like two, three, 5x better than a general agent at its task, because you do have the inertia and the switching cost…

…The thing I will say about sponsored listings in particular is, we are developing our AI agent, and we are developing sponsored listings, but it will be developed in the context of the AI agent. We first need to launch the AI agent, we need to refine it, then we need to figure out how ads work in that new world.

Airbnb has seen its operating costs improve as a result of AI-driven efficiencies in customer support

Margin expansion in Q3 was driven by strong revenue growth and cost discipline, including operating leverage from AI-driven efficiencies in customer support, as well as the shifting of certain marketing expenses to Q4. 

Airbnb’s management thinks every company is still figuring out how advertising will work in AI agents

We first need to launch the AI agent, we need to refine it, then we need to figure out how ads work in that new world. By the way, everyone is figuring this out. ChatGPT is figuring this out, so I do not think it is clear. It is not clear what the implementation is, but I think if anyone can figure it out, I think our design team could.

Airbnb’s management is seeing very little traffic and bookings coming from popular 3rd party AI agent products, and the trend has not increased over the last 4 weeks; Airbnb has data showing that people find researching travel to be highly aspirational, so they’re unlike to offload the task to agents; management thinks 3rd party AI agents will be great handoffs to Airbnb

[Question] I was wondering if you could just respond to some of the data that’s out there about early adoption of these agents that’s showing agents like Muse and Instinct, that’s showing that a lot of commercial activity happening there is travel-related. Is that something you think strengthens the mix over time, or how does that gel with your overall view of how people use these horizontal tools?

[Answer] We’re not seeing a lot of traffic, and even fewer bookings from Muse and Instinct yet. We do have nearly half a million bookings a day, and a very non-material number of bookings happen on Muse and Instinct. So, we’re talking very low numbers. We’re not seeing a lot of traffic today, and it’s not really increased over the last four weeks. It kind of peaked in early October. It’s kind of been stable in the last four weeks…

…Most of our traffic happens during 9:00 A.M. to 5:00 P.M. hours. These people are supposed to be working, and they’re not. We have evidence. People are actually secretly using Airbnb. Maybe they’re trying to escape their job. But the idea is that it’s highly aspirational, so I don’t think a lot of people want to offload all that to an agent. I think they want to offload the things they don’t enjoy, not the things they do enjoy…

…But I also just think they’ll be really helpful handoffs to us. But I think that to do. I’m not saying agents can’t do travel. I just think it’s a focus on travel, and that’s what we’re going to do. We’re going to focus on travel.

Apple (NASDAQ: AAPL)

Apple has an all-new Siri AI; Siri AI has strong privacy protections; Siri AI was launched in beta in English in the US in September and has since expanded to a number of Asian and European languages; the new Siri AI is able to understand users and hold natural conversations; management is pleased with the reviews and feedback of the new Siri AI; usage limits will apply to Siri AI features that rely on server-side models; Siri AI is primarily request-based today, but it is built on an orchestrator-driven architecture that gives Apple plenty of room to expand Siri AI’s capabilities over time; management thinks the privacy features of Siri, and the data that the iPhone has on users, are core differentiators of Apple’s AI experience; management expects Siri AI to handle queries both on-device and in the Private Cloud Compute environment, but they are unclear about the eventual mix; Siri AI has no real cost associated with handling queries on-device, unlike when routing to Private Cloud Compute 

We’re unlocking more powerful AI experiences across our lineup, including with the all-new Siri AI, a profoundly capable and deeply personal assistant seamlessly integrated across our platforms. Like everything we make, Siri AI is built with essential privacy protections, so no one can access users’ data, not even Apple. We launched Siri AI in beta for U.S. English users in September and expanded to French, Japanese, Korean, Portuguese, and Spanish in October…

…Siri AI marks a new chapter in AI at Apple. It is a completely reimagined assistant that is powerful, personal, private, and deeply relevant to our users’ daily lives. Siri AI picks up what’s relevant to our users, understands what’s on their screen, and holds a natural conversation, all to help them with the things they do every day. It’s the Siri we’ve always wanted. We are very pleased with the reviews and feedback we’ve received, and we’re so excited that more and more users are getting to experience it…

…There’s going to be some usage limits on features that rely on server-side models, which include Siri AI and intelligent photo editing and some other things like that…

…With Siri AI, we really have the Siri that we always wanted. It is primarily request-based today, but as we talked about at WWDC, it is built on a modern, orchestrator-driven architecture that gives us real room to expand its capabilities over time…

…We really do think of the iPhone as the intelligent personal hub. I mean, if you think about it, so much of your personal data is on your iPhone, whether it’s your contacts, your communications, health data, financial data, all of that. Siri can tap into that and help you access it in a way that’s very secure and private and that nobody else has access to. Even Apple doesn’t have access to it. We think that is a core kind of differentiator that we’re excited about…

…We do envision Siri AI to kind of have queries that kind of are managed on-device and also in the Private Cloud Compute environment that we have outlined. The mix of that is still very early days, so it is hard to predict exactly what the mix of those costs are going to look like. Obviously, on-device is very, very efficient, no real cost associated with that. But on the server side, we would have some inference costs, and that is really the connection of the Siri AI-related costs that I mentioned that would be recognized with services cost of sales.

Apple’s management sees the company’s line up of products as the best platforms for AI

From Mac to iPhone to iPad, we are proud that our products are the best platforms in the industry for AI, for both first-party and third-party applications, and we are excited about the opportunities that creates.

Apple’s management sees iCloud+ becoming a monetisation route for AI, not just storage

We have these iCloud+ plans that now include Apple TV and Apple Arcade in more than 100 countries and regions where we think we can really reach new potential subscribers. Then, as you mentioned, we think there’s good opportunities on Apple Intelligence. There’s going to be some usage limits on features that rely on server-side models, which include Siri AI and intelligent photo editing and some other things like that. Anyone who needs greater capacity will be able to purchase expanded access in the future. So, kind of to your point, in the beginning, iCloud tiers were largely about storage, but we’ve been adding more capabilities and features to it.

Apple’s management thinks of the iPhone as the intelligent personal hub for users

We really do think of the iPhone as the intelligent personal hub. I mean, if you think about it, so much of your personal data is on your iPhone, whether it’s your contacts, your communications, health data, financial data, all of that.

The MacBook Pro and iMac now has the M6 chip, which includes the world’s fastest CPU core and a dual 16-core Neural Engine for private on-device AI tasks; every Mac in Apple’s lineup right now is a powerful platform for AI; the Mac Mini and Mac Studio allow users to work with local, private AI, in ways that were previously not possible; Gap is running AI entirely on Mac Minis; Christie’s is running local AI models on Mac Studio 

We also brought the power of M6 to the new 14-inch MacBook Pro and the new iMac. This latest generation of Apple silicon includes the world’s fastest CPU core to speed up everyday tasks, in addition to a dual 16-core Neural Engine to drive private on-device AI tasks…

…More than 40 years after the first Mac, the category is having its biggest moment yet, with every Mac in our latest lineup a powerful platform for AI. Mac Mini and Mac Studio in particular are pushing the boundaries of local, private AI, giving users the ability to do things once considered impossible…

…Gap Inc. is running AI entirely on Mac Mini for loss prevention across hundreds of stores, eliminating cloud token costs. Christie’s, a world-leading auction house, is running local models on Mac Studio to improve core productivity workflows such as pricing estimates.

The Apple Watch will soon have an Audio Intelligence future, which will help users remember moments they might otherwise miss; Apple has introduced the AirPods 5, which has industry-leading active noise cancellation; AirPods can now allow users to experience live translation when speaking with other people in a different language; the Apple Watch can send alerts to deaf users on what’s happening around them

We are excited for Audio Intelligence, which will be coming in a future software update and will help you remember moments you might otherwise miss…

…Meanwhile, AirPods 5 offers best-in-class open-ear active noise cancellation at an incredible price, while improving sound quality, battery life, and durability. With iPhone and Apple Intelligence, including Siri AI, AirPods allow users to do remarkable things, like connect with someone who’s speaking another language with live translation, or get answers with broad world knowledge entirely hands-free.

…It means developing intelligent capabilities like sound recognition on Apple Watch to alert users who are deaf or hard of hearing to what’s happening around them.

Apple’s management thinks the current memory pricing situation is unprecedented; management expects even higher memory costs for the 2026 Q4 quarter; Apple has increased its manufacturing purchase obligations to $33 billion, driven by purchases of memory chips

What we’re seeing with memory pricing right now is unprecedented. As we said earlier in the September quarter, we did see higher memory costs as compared to the June quarter, which were partly offset by the benefit of carry-in inventory and lower costs on certain non-memory components. For the December quarter, we expect even higher memory costs, partly offset by a few factors. As you know, we changed the price on many of our products, including Mac and iPad at the end of June and iPhone in September, and these price increases provided some offset. We also expect to see some benefit from carry-in inventory in the December quarter as well, as well as lower costs on certain non-memory components. Looking beyond the December quarter, we see the pricing for memory purchases continuing to increase. As we said before, we do believe we will see a decreasing benefit from the carry-in inventory on memory over time. There is one more thing I wanted to add. As we work to improve supply, our manufacturing purchase obligations have increased by $33 billion from June to the September quarter, which you will see in our annual filing when it comes out tomorrow.

Arista Networks (NYSE: ANET)

Arista Networks’ management is seeing customers increasingly coalescing the backend and frontend of AI data centers

Our customers have increasingly coalesced the backend and the frontend of AI and data centers.

Arista Networks’ management thinks standards-based ethernet can modernise networking for AI data centers with a lower total cost of ownership; Arista Networks’ customers can build an open network platform that can handle different models and AI accelerators at scale; Arista Networks’ Etherlink portfolio is designed to improve XPU utilisation, job completion time, and token latency for both training and inference workloads; Arista Networks recently demonstrated its rack scale-up designs, developed together with many different AI accelerators; management expects scale-up trials to begin in 2027, and production to begin in 2028; management thinks the AI industry is migrating from highly centralized training compute to disaggregated and distributed inference, pre-fill, and decode, which requires intelligent networking to provide optimal token efficiency; Arista Networks’ Etherlink products are designed to understand latency and bandwidth profiles deeply to avoid idling of precious compute cycles; Arista Networks was initially working with 4-5 AI customers with 50,000 to 100,000 NVIDIA GPUs, and NVIDIA’’s Infiniband was the networking solution; by end-2027, management expects more than 10 GW of deployment across 10 different AI accelerators; much of the 10 GW of deployment are new deployments; Arista Networks is new to scale-up, and management thinks the addition of scale-up will enable the company to add another 5 AI accelerator customers for deployment

Arista has long championed standards-based Ethernet. Customers have welcomed our open approach to modernizing cloud, campus, data centers, WAN centers, and now AI centers with a lower total cost of ownership…

…Arista customers can now construct an open network platform capable of handling heterogeneous models and diverse accelerators at scale. Our Etherlink portfolio of at least 25 switches for scale-up, scale-out, and scale-across is optimally designed to improve XPU utilization, job completion time, and token latency for these intense training and inference workloads.

In October of 2026, at the Open Compute Project conference, Arista showcased our rack scale-up designs, partnering closely with several AI accelerators infused with our Arista network diagnostics layer. Based on ESUN, Ethernet for Scale-Up, rack options, these include orthogonal designs, cabled backplanes, and heterogeneous cross-rack designs co-designed with many of our partners, including Arm, Microsoft Maia, AMD’s Helios, Qualcomm, Meta’s MTIA, and many more ahead. We expect scale-up trials to begin in 2027, driving Ethernet to larger radix of 126 to 144 within a rack and across 1,000 across racks, with production expected for scale-up in 2028.

We are witnessing the migration from highly centralized training compute to disaggregated and distributed inference, pre-fill, and decode, demanding much more intelligent networking and co-location of compute, storage, and KV cache resources, rather than simply focusing on raw feeds and speeds and flops of compute. A well-designed network is becoming a foundational cornerstone for optimal token efficiency. Arista’s decades of EOS engineering in both multi-planar and distributed model routing at scale, be it ARP, address resolution protocols, or FIB tables, forward information base, or access lists with rapid reconvergence comes to the rescue. Our Etherlink products are inherently designed to actively understand parameters at a deep level, including latency and bandwidth profiles of multi-tenancy to intelligently steer workloads for optimal placement between local and remote domains. The need for visibility into ports, packet queues, flows, buffer utilization, and port latency across the entire scale-out and scale-across fabric validates real AI system-level performance to avoid idling of precious compute cycles. It assures immediate and network-wide insight to the root cause, as well as rapid recovery of the AI system with smart system upgrades across millions of accelerators.

When we initially began this transformational AI journey, we spoke about four or five AI customers with approximately 50,000 to 100,000 NVIDIA GPUs. In fact, we were largely staring at a proprietary network cycle such as InfiniBand. In 2026, we are in a super cycle of AI, having already surpassed 100 Etherlink customers. By the end of 2027, we are forecasting greater than 10 GW of deployment across 10 or more different types of compute accelerators. What a stark inflection in just three years…

…Our guess is by the end of 2027, we will deploy largely in 2026 and 2027 is where the exponential ramp happened across or greater than 10 GW. Much of that gigawatt may be new, but remember, in scale-across, we can also have existing gigawatts being deployed. We are talking about new distributed 10 GW, but they may already be multiple gigawatts deployed across. The rationale here is, while we are new to scale-up, we are touching a variety of XPUs in scale-out and scale-across, across the rack and across data centers. We already know with our scale-across, we are touching multiple XPUs, four or five of them. With the advent of scale-up, we will add another five in 2027.

Arista Networks’ management thinks the AI industry is shifting from a compute-centric vertical stack to a holistic network-centric approach, driven by the rise of agentic AI models; as the shift to the network-centric approach happens, management sees Arista Networks being involved earlier in designing the systems, and also the systems having higher networking intensity

As agentic AI models increase in parallelism, compute diversity, and density, we have the choice of many more training and inference accelerators. The industry, we believe, is shifting from compute-centric vertical stack to a holistic network-centric approach…

…[Question] You talked initially about the industry moving towards a more network-centric architecture. As customers deploy different type of accelerators, does Arista become involved earlier in designing the overall systems rather than connecting them afterward? As you have to work with a more distributed, disaggregated compute network, does that change your networking content per deployment? I know you talked about 10 GW, and I’m just sort of wondering, does the network intensity start to change as well?

[Answer] Yeah…

…I think until now, we’ve enforced our architecture mostly on scale-out and scale-across, and a lot of that had to do with distributed power, location, moving to regional edges, a very software-defined type of network with the intensity, like I described, of package, flows, performance, predictable latency, et cetera. As we start participating with new accelerators, we get a greater chance early in the cycle to see the behavior of these accelerators. Two or three examples I’d like to highlight is our close work with Cerebras, our work with Microsoft and Maia. These have been going on the last couple of years, and has allowed us to tune our buffers, tune our software performance, and maximize the IO throughput, if you will. The recognition of those scale-up designs are still in the works, but they’re already manifesting itself in scale-out and scale-across.

Arista Networks recently demonstrated its rack scale-up designs, developed together with many different AI accelerators; management expects scale-up trials to begin in 2027, and production to begin in 2028; scale-up had historically been a tough market for Arista Networks because it does not have the software advantages that are the company’s strengths

In October of 2026, at the Open Compute Project conference, Arista showcased our rack scale-up designs, partnering closely with several AI accelerators infused with our Arista network diagnostics layer. Based on ESUN, Ethernet for Scale-Up, rack options, these include orthogonal designs, cabled backplanes, and heterogeneous cross-rack designs co-designed with many of our partners, including Arm, Microsoft Maia, AMD’s Helios, Qualcomm, Meta’s MTIA, and many more ahead. We expect scale-up trials to begin in 2027, driving Ethernet to larger radix of 126 to 144 within a rack and across 1,000 across racks, with production expected for scale-up in 2028…

…I think scale-up has been a tough market for us because it doesn’t have all the software advantages that we bring, but it absolutely has the hardware skills and all the non-NVIDIA accelerators that Arista is in the best position to connect to. This movement from standalone basic switches to rack scale is something that Andy, Alex, Hugh Holbrook, the entire team has been working on and co-developing for a couple of years. It’s hard work. Just as we were staring at InfiniBand and now we can go, “Of course it’s going to be Ethernet,” we’re looking forward to a time where we were staring at proprietary UALink or NVLink, and we’re going to go, “Of course it’s going to be ESUN.” But these type of migrations are multi-year, and so we’re very confident not only in our 2027 guide, but a promising 2028 ahead as well. The reason we are is because even though we have two quarters of visibility, we have two years of work behind us.

Arista Networks’ management recently expanded Arista AVA, the company’s autonomous virtual assistant; AVA allows AI agents to interact with Arista Networks’ network stack; AVA helps improve the security of networking; AI is an increasing threat for network security, and Arista has been collaborating with key players in the AI industry to integrate AI-enabled security capabilities to the company’s software security pipeline; Arista Networks is an early key infrastructure provider in important AI-related security projects such as Project Glasswing; management thinks Arista Networks’ EOS (extensible operating system) is the ideal architectural foundation for networking system for 4 reasons, namely (1) having a single high-quality operating system, (2) the ability to limit vulnerability impact through process isolation, (3) it provides a seamless upgrade process, and (4) there’s a compliance dashboard to stay current; Arista Networks provides Zero Trust Networking technology for customers to build secure infrastructure; Arista Networks offers protection of workload traffic that leaves the confines of the local data center through wire-speed ubiquitous link and tunnel encryption; Arista Networks has maintained one of the lowest CVE (common vulnerabilities and exposures) counts in the industry over the last 20 years, and it is often 10x-50x better than alternatives; management has just turned on AVA as an alpha trial and is working on the pricing model; management feels good about monetising AVA in the next few years

Central to this AI super cycle is also the expansion of Arista AVA, our autonomous virtual assist, to support efficient AI in the enterprise. By providing that open programmable interface within EOS and utilizing the model context protocol, MCP, AVA allows emerging AI agents to reason, to interact with the network stack, and perform advanced multi-domain event correlation across the entire enterprise…

…AVA, which is Arista’s autonomous virtual assistant, fueled by real-time data, AVA is incredibly powerful. It seamlessly navigates networking and security infrastructure from campus to WAN to data center to cloud, continuously monitoring and troubleshooting, isolating root causes without manual labor-intensive tasks…

…The bad news, of course, is the massive challenge AI poses for us and our customers by discovering and exploiting security vulnerabilities. This threat landscape demands secure and robust networking. Over the past few months, we’ve been collaborating with Anthropic, Google, OpenAI, and many others to integrate new AI-enabled security capabilities from frontier models into our existing software security pipeline.

We are being invited early as a key infrastructure provider into partnerships like Project Glasswing, building AI-driven vulnerability discovery and assessment into our established security vulnerability management process. True network security has become a must-have, not a nice-to-have, and Arista’s EOS operating system is the ideal architectural foundation based on four key principles. First, a single high-quality operating system. Arista has maintained one of the lowest CVE counts, less than 100, in the industry over the last two decades, while others are rising exponentially, due in part to the simple fact that multiple networking operating systems tend to have more vulnerabilities than just one. Second, limiting vulnerability impact through process isolation. EOS keeps each software agent, management plane or control plane, architecturally separate from each other and also from hardware forwarding, so traffic keeps moving even if there is a defect in one component. EOS’s state-sharing architecture also resists lateral movement, so a defect in one component typically cannot be exploited to attack other agents within the switch. Third, we provide a seamless upgrade process. Process isolation also means that we can ship a targeted fix to the affected component and install the fix without a maintenance window or network disruption. Fourth, delivering a compliance dashboard to stay current. When a security advisory lands, Arista CloudVision brings real-time visibility that turns a stressful advisory day into a manageable one.

In addition to providing secure switches, we also provide Zero Trust Networking, or ZTN technology, to help our customers build secure infrastructure. I’d especially like to highlight Macro-Segmentation Services, or MSS, a ZTN feature allowing for dynamic security enforcement, preventing undesired interactions among users, devices, and applications. Second, we provide Network Detection and Response, or NDR services, that profile, analyze, and proactively detect sophisticated security threats. Finally, once workload traffic leaves the confines of the local data center, it becomes even more vulnerable to malicious actors. So, we offer protection with wire-speed ubiquitous link and tunnel encryption, native in every port with zero performance degradation…

…Our CVE count, our vulnerability count, is often 10x- 50x better than alternatives…

…We have just turned the AVA on as an alpha feature for customer trial within our CloudVision service, and we are currently working on pricing models. I do not think I have any forecast or data to share with you at this time on that…

…We feel in a very good position to monetize AVA in the next few years.

Arista Networks won a deal with an Asian web services company as they wanted to prioritise speed, scale, and resiliency in the AI area, and they recognised Arista Networks’ reliability; Arista Networks’ 100 gig and 400 gig platforms will be the front-end of the Asian web services company’s AI deployments going forward

Arista won a strategic deal at an Asian web services company as they shifted their focus strategically onto speed, scale, and resilience in the new AI era. Reliability was the key to the engagement as their teams were frustrated by longstanding quality issues with their existing vendor. Arista was able to demonstrate reliability and scalability while leveraging standards-based technology and a powerful EVPN VXLAN architecture on our 100 gig and 400 gig platforms. This deployment will operate as the front end for their AI deployments moving forward.

Arista Networks’ management has raised its revenue guidance for 2026 to growth of 48% (previous guidance was for growth of 40%); management is targeting 38%-40% revenue growth in 2027, with the AI fabrics part of the business to be the fastest growing; management expects the AI fabrics business to be up by more than 100% to $7 billion in 2027; management expects scale-up networking products to be only be in early trials in 2027, so scale-up will not have a material revenue contribution in that year; management expects scale-out and scale-across networking products to be 70:30 of its business in 2026, and for the ratio to shift closer to 50:50 in 2027; management has confidence in Arista Networks’ supply chain in helping the company meet its revenue goals 

We are now increasing our annual guidance for the fourth time in 2026 to $13.3 billion, which represents a 48% annual growth…

…Our first look at 2027 guidance range puts us in the 38%-40% range for growth annually to approximately $18.5 billion. Within this envelope, we expect our AI fabrics target to be the fastest-growing segment and to double to approximately $7 billion, while our cognitive campus represents a healthy growth of 60%, much faster than the market, to approximately $2 billion in 2027…

…The scale-up will still be in early trials, so I don’t think there’s a material number associated with scale-up in 2027. I do think it could contribute in 2028. If scale-up is virtually zero, then we’re basically left with most of the use case being scale-out and scale-across. As you’ve heard us say in the past, scale-across, we think will be at least a third. It’ll probably go higher in 2027. The split this year is roughly 70/30. The split next year, I can’t venture to guess, but I think the 70 may come down and the 30 would go up when it comes to scale-across, because it’s a very strong use case for us. Arista’s highly differentiated with our 7800 scale-across platform…

…[Question] As you think about this 38%-40% guide for 2027, which is extremely impressive, can you just touch on your confidence around the supply chain execution to meet this demand?

[Answer] I am confident in the guide, and I am confident in the supply chain. What makes me confident in the supply chain is internally, Todd and Eugenia and the team doing a fantastic job every day, being scrappy to make sure we can do what we need to do to deliver to our customers. I also have confidence, a lot of great conversations and agreements with some of our key suppliers, and that’s why you see our purchase commitments increasing, because we have that confidence…

…We are proud of our strategic partnerships, and we have been investing heavily in long-term strategic agreements in order to secure supply, not just in silicon, but PCB capacity memory, which Chantelle mentioned earlier. So we are incredibly. We have very high confidence in the ability to deliver supply to the forecast.

Arista Networks’ management is investing more into liquid cooling technologies for its networking products; management is spending a lot on R&D because the opportunity is huge in AI; management sees a lot of liquid cooling being used for networking switches

Arista continues to demonstrate its commitment and focus on network innovation with additional investment in liquid cooling, high-radix switching, and leading AI software…

…[Question] I was just curious about the big sequential jump in R&D. It was really, really significant. I’m just wondering if that included some one-time items. Also, is this the new run rate in R&D going forward?

[Answer] We’re building so much new stuff. The opportunity is humongous in AI. These AI accelerating systems are. Our success with the hyperscaler AI companies depends on our ability to partner. Our ability to partner is based on our understanding of each of their unique requirements and constraints. Also, there’s a new technology introduction around liquid cooling that requires major infrastructure upgrades and technology upgrades on our part. I think that’s what’s really driving the expansion spending there…

…What’s very clear is there’s a lot of switches swimming in water. A lot of it’s liquid cooled.

Arista Networks’ purchase commitments has increased sequentially to $11.4 billion (was $9.7 billion in 2026 Q2 and $8.9 billion in 2026 Q1); Arista Networks’ purchase commitments are mostly for chips for new products and AI deployments; the chips in the purchase commitments now include memory chips; management expects the purchase commitments to increase in 2026 Q4

Our purchase commitments at the end of the quarter were $11.4 billion, up from $9.7 billion at the end of Q2. As mentioned in prior quarters, this expected activity mostly represents purchases for chips related to new products and AI deployments…

…Purchase commitments, from that perspective, there are some things additionally we added. Now we have memory in there, a memory agreement which has done that. So that’s a multi-year agreement that you’ll see as the Q comes out. I think on the purchase commitment side, you can probably expect that to go up a little bit more as we exit Q4, as we get ready for this 2027 guide we have.

Arista Networks’ ramp of new products and new use cases in AI is contributing to growth in deferred revenue

Our total deferred revenue balance was approximately $6.8 billion, down from $6.9 billion in the prior quarter. The majority of the deferred revenue balance is product-related. Our product deferred revenue decreased $86 million sequentially versus last quarter. We remain in a period of ramping our new products, winning new customers, and expanding new use cases, including AI. These trends have resulted in increased customer-specific acceptance clauses and an increase in the volatility of our product deferred revenue balances.

Arista Networks’ management expects more large customers to become 10%-customers for the company, and join the ranks of  Meta Platforms and Microsoft

We want to have diverse 10% customers too. As we said before, our dearest ones are M&M. We fully expect our top prior 10% customers to continue to be with us in 2026, and we hope beyond. We expect to add at least one, maybe two, and a lot of that depends on whether Todd and Eugenia can ship enough product. We will let them decide, and we will give them all the way into December 31st. We are feeling pretty good about at least three and maybe four.

Arista Networks’ management believes in copper whenever it’s possible, and only optics if necessary; management does not want to build proprietary co-packaged optics (CPO); management has been pioneering an open CPO approach; management thinks the open CPO approach will come into market sometime in 2028-2030

Optics is very integral to Arista’s strategy. Andy Bechtolsheim, our Chief Architect, lives and breathes this. It does not mean we build everything, but we actually systematically put it together. We, first of all, strongly believe in copper when you can and optics if you must, because the minute you go beyond a one or two-meter radius, the flexibility of pluggable optics and lasers is probably the most widely deployed and accepted. Andy was a pioneer of OSFP that gave us a very wide density of thermal limits and widely used. Now our new coherent optics, LPO modules, is something that goes long distance without requiring DSP. Then to meet higher front panel density, we are very excited about XPO, that is being supported by over 150 vendors. So we think the world is going to remain in a very majority fashion on pluggable optics. Having said that, I think there is always room for complementing the pluggables with a strong adoption at higher speeds to what we call near-packaged optics or open CPO. We do not want to build proprietary co-packaged optics. To solve these constraints, Arista has been pioneering this open CPO, that instead of soldering components, we can use socketed optical engines and pigtail fibers to allow these modules to be pre-tested, et cetera. So we are big fans of these. We think they will come onto the market in the 2028 to 2030 timeframe. It is not imminent. But meanwhile, long live pluggable optics.

Arista Networks is seeing its scale-across networking products used by both hyperscalers for training, as well as smaller customers

[Question] How concentrated is the customer base for Scale Across? Should we think about this as really being driven by the cloud titans for synchronous training, or are we also seeing Scale Across deployments even among smaller customers or for inference use cases?

[Answer] I think there are both small use cases and colos for Scale Across, which would be a smaller footprint switching and routing platforms, like the 7280. But the concentration of our large customers also is high, with our larger customers opting for the very large AI spines across data centers and requiring that kind of 576 core density and terabit scale and multi-gigawatt deployments for security and performance. We do see both, but I think the revenue concentration favors the 7800. The location concentration of multiple sites favors the 7280.

The lead times for Arista Networks’ AI-related networking products are longer than for the non-AI networking products

On the lead times, and it goes both for the component lead times and the finished good lead times. In the enterprise side of our business, we tend to have shorter lead times, somewhat shorter, and that’s driven by the fact that those switches are not always in the absolute cutting edge. They’re not leading brand new technologies requiring the highest process silicon, et cetera. We’re able to They also tend to be at a lower price point. We’re able to hold a little bit more buffered inventory there. So, we tend to have lower component lead times there and lower lead times to our customers. On the AI side, on the ultra-advanced hyperscale data center, lead times tend to be a little bit longer.

Cloudflare (NYSE: NET)

A fast-growing AI company expanded its relationship with Cloudflare for the Workers Developer Platform; the AI company’s platform has supported the creation of tens of millions of AI-generated apps, and legacy hyperscalers would charge the AI company a huge sum for idle compute; the AI company would not have been able to build what it wanted to without Cloudflare’s Dynamic Workers

A rapidly growing AI company expanded their relationship with Cloudflare, signing a one-year, $11 million pool of funds contract for our Workers Developer Platform. Tens of millions of AI-generated applications have been created on its platform, and legacy hyperscalers would charge this customer for massive amounts of idle compute. With Cloudflare’s Dynamic Workers, apps start in milliseconds, and the customer pays only for the compute they actually use. One of their lead engineers said, quote, “Without Cloudflare’s Dynamic Workers, we wouldn’t be able to build what we want to build.”

A Fortune 100 cloud infrastructure company expanded its relationship with Cloudflare for Magic Transit; the cloud infrastructure company’s legacy DDoS hardware apps could not keep up with the company’s global scaling for AI; Cloudflare stopped attacks that other vendors failed at doing

A Fortune 100 cloud infrastructure provider expanded their relationship with us, signing a 13-month, $15.5 million contract for Magic Transit. As the customer aggressively scales its global footprint for AI, their legacy DDoS hardware applications simply could not keep up, failing to scale cost-effectively and missing low-level attacks that were quietly draining capacity. During the evaluation, our global network stopped attacks that their appliances simply could not, and we fully displaced the incumbent. As one of their technical leaders noted, “Cloudflare’s threat mitigation is one of our superpowers.”

A leading European media company expanded its relationship with Cloudflare for application services and the Workers Developer Platform; the European media company chose Cloudflare because reliably identifies modern AI crawlers; management thinks the deal with the European media company is a blueprint for how media companies will monetise their content in the agentic AI era

A leading European media company expanded their relationship with Cloudflare, signing a five-year, $8.1 million contract for application services and our Workers Developer Platform. Replacing a 17-year deployment with a legacy CDN vendor, they chose Cloudflare because our best-in-class Bot Management reliably identifies modern AI crawlers, giving this customer the control they need to protect their intellectual property, but also creating a direct on-ramp for Pay Per Crawl and pay-per-use monetization. This win is a clear blueprint for how the world’s largest media companies will protect and monetize their content and innovate in the agentic AI era.

Cloudflare’s management thinks the company has seen an inflection in growth in Oct 2025 and Jul 2026 because (1) the company has built the development platform for the future of the AI internet, and (2) growing threats in the cybersecurity landscape because of AI; there was a recent meeting at the White House with US government officials and leaders of AI companies, and the leaders of the AI companies said the best company to protect organisations against vulnerabilities is Cloudflare

We’ve been growing throughout our entire history, but recently something changed. The rate of growth inflected in October of 2025, then accelerated again in July of 2026, and then again at the beginning of this October. So what’s going on? I think it’s the convergence of two things. First, we have built the development platform for the future of the AI-powered internet. Second, the security landscape is getting so scary that the best viable strategy, regardless of who you are, is to ensure you have Cloudflare protecting you…

…We’re seeing AI growth, but by no means would I say that this is really concentrated around AI in particular. We’re seeing strong growth around financial services, government, manufacturing, healthcare, e-commerce, across the board, that’s coming on board, and it is new customers. As I talked about in the prepared remarks, I think one of the biggest drivers is the fact that it’s a secondary effect of AI, which is AI is finding more and more security flaws that are out there…

…At some level, I guess that’s AI driving this because AI is helping find these bugs…

…It was discussed literally in the White House the other day, where a bunch of AI leaders got together with the White House officials and the White House was like, “How are we going to make sure that everyone’s protected against these vulnerabilities?” The answer came back, Cloudflare, because we sit in front of so much of the internet, and we are such a responsible player protecting the internet. I think that’s what’s driving a huge amount of our growth.

Legacy hyperscalers built their cloud computing services around container technology; Cloudflare’s management thinks container technology is not suitable in the agentic AI era because it is not cost effective; management thinks Cloudflare’s isolates technology works better; management thinks Cloudflare is the world’s first agentic cloud

Hyperscalers and even neoclouds were built around container technology. It was a significant improvement over bare metal or even virtual machines. Containers allowed these last-generation clouds to scale during the explosion of mobile devices that defined their heyday. Now we’re in the era of AI and agents…

…The container-based architecture of the hyperscalers and neoclouds just can’t keep up in a cost-effective way with the new agentic paradigms of today. Cloudflare’s approach is different. We have based it on what we call isolates, not containers, which initially let us scale geographically to span the globe, but increasingly is allowing us to scale densely with a level of efficiency no hyperscaler or neocloud can match. The smartest developers, whether they’re at frontier model companies or legacy enterprises, see that the world is changing and that they need a new type of cloud in order to keep up. In other words, Cloudflare isn’t a neocloud. We’re also not, as some of you have called us, the fourth public cloud. We’re something different. We’re the first agentic cloud, and developers are flocking to us to build the agentic AI future.

Cloudflare’s management now expects agentic traffic to be 1000x higher than human traffic within 5 years; Meta Platforms is a customer of Cloudflare; management thinks Cloudflare will benefit as Meta Platforms’ personal agent product, Muse, scales; management is seeing companies building agentic products on Cloudflare, and the tools that agents use are also on Cloudflare

I predicted last quarter that within five years, agents would outnumber humans online 1,000 to 1. A quarter later, I’ll take the under on the timing of that bet based on everything we’re seeing today…

…As for Muse, I mean, first of all, if you haven’t played with it yet, it’s super cool, as are a number of the other agents that are out there. In the case of Muse, Meta’s a customer, and it’s been incredibly fun to watch them and help them scale that product and just see the explosive growth that we’re seeing. So we’re benefiting as that product grows directly…

…We’re in a really interesting position because the people who are building the agents are often building them on Cloudflare, and then the things that the agents want to get to are also on Cloudflare.

Cloudflare’s management thinks even bigger threats to cybersecurity will come because of AI; Cloudflare is among the first companies in the world to receive quiet disclosure of new vulnerabilities; Cloudflare is partnered with every major frontier AI lab to find vulnerabilities in existing software; Cloudflare has often patched discovered-vulnerabilities before they are publicly revealed by researchers; the AI-driven threats to cybersecurity have led to even legacy organisations signing up for Cloudflare at an accelerating rate

The second driver of our growth inflection has even broader applicability across every business. It may seem like the sky is falling with the number of new security vulnerabilities that have been reported over the last few weeks. I wish I could tell you we were through the worst, but we’re not…

…Just last week and the week before, we had six of the worst internet security bugs of all time announced over the course of fewer than 14 days. We’re not done. We have already gotten disclosures of many, many, many more extremely scary bugs, which researchers have found, have quietly disclosed to us, and we have protected our customers from…

…Because of the vast percentage of the internet that already sits behind Cloudflare, we are among the first to receive quiet disclosure of new vulnerabilities. We’ve partnered with nearly every major frontier AI lab and security company searching for flaws in existing software. When they find something, they notify us through a trusted reporter process, and we quietly put protection in place for all our customers, whether they pay us or not…

…The bugs you read about over the last few weeks, and the ones I suspect you will read about over the next few years, were less impactful because Cloudflare has the internet’s back…

…Companies increasingly know that if they are behind us, they are more likely to be protected, whatever the future brings. Even legacy organizations that aren’t pushing the bleeding edge of AI are signing up for Cloudflare at an accelerating rate.

Cloudflare recorded its first revenue for its Act 4 products in 2026 Q3, which are the products that will help shape the new business model of the internet; management is finding Cloudflare to be increasingly the mediator between publishers and companies building agentic products; management thinks Cloudflare is starting to set the rules for what the agentic internet is going to look like; as agentic traffic on the internet explodes, someone has to pay for the extra traffic, and the entire ecosystem is at the cusp of what the next business model is going to look like; Cloudflare had partnered with OpenAI earlier this year to help OpenAI crawl the internet in a more efficient manner; management thinks some form of micropayments will emerge whenever an agent visits a website; management also thinks some form of advertising can emerge from agentic traffic; management thinks Cloudflare is in a position to connect agents with the places that might have what the agent is looking for; management does not know when orchestrating agentic traffic on the internet will become a meaningful revenue source of revenue, but they think it would become the largest source of revenue in the next 10 years; management is aware that some websites will be happy to have agents visit, whereas some websites want agents to stay away

I’m proud to announce that we recorded in Q3 our first revenue for what we’ve called Act 4 products, those where we’re helping invent the new business model of the internet, and I expect we’ll talk a lot more about that in the quarters to come…

…We’ve had examples over the last several months, in the case of Meta, in the case of Google, where at times there’s been places where what they have been doing hasn’t aligned exactly with what the content creators, publishers, application developers, financial services organizations, e-commerce companies, what they want to do. Cloudflare is finding ourselves over and over and over again in the center of those questions and acting almost as a mediator between the two sides. In the process of doing that, I think what’s really powerful is we’re getting to set the rules of the road for what the agentic internet is going to look like.

The reality is that if agents are going to be 1,000 times human traffic online, that’s going to require more bandwidth, that’s going to require more servers, that’s going to require more things to actually power it, and someone has to pay for it. It can’t be a free ride for all the agents that are out there. What I’m encouraged by, and what we’re already seeing in terms of Act 4 revenue, is that the agentic companies get that they have to compensate the creators and the banks and everyone else who’s providing this information in order to be able to continue to access it. So I think we are at the cusp of figuring out what that next generation business model of the internet’s going to look like. Because of the fact that we’re helping the companies that are building the agents, and because of the fact that so much of the rest of the internet sits behind us, Cloudflare is the company that’s going to define what that looks like…

…The way of thinking of it is, I am sitting in Manhattan, and if I ask my agent, “Where am I going to go for dinner tonight after I am done with earnings calls and callbacks?” It might go and scour the internet and look at 1,000 different restaurants’ websites, pull down 1,000 different menus, and then I might choose to only go to one of those. The problem is that 999 of those restaurants have now had a cost that is just been then applied to them, but they are not getting any benefit. That was always true of the internet, but the scale was different. It was, a human might look at two or three things, but in this case, the scale is so high that we are going to have to figure out some additional way to pay for it.

I think the ways of doing that are twofold. So one, where we are already recognizing revenue, and recognizing significant amounts of revenue, is how do we make that more efficient? We talked last quarter about how we partnered with OpenAI to make sure that as they searched the internet, they did not do it in the kind of dumb way of just crawling everything all the time over and over again. Instead, Cloudflare could actually give them hints when a page had changed, when a menu price had updated, when anything had happened, and then they could get the freshest data, and they could actually crawl that. That helps OpenAI save money, and that actually helps the Cloudflare customers get the freshest data to them and make sure that it does not waste a ton of resources. I think that is a clear way where we are already seeing that, and that has helped. OpenAI is paying for that, and a lot of other companies in the space, I think you will see do that as well.

Beyond that, though, I think that we are going to see some version of micropayments that exist where small payments get made when an agent comes. I think it will be in exchange for, the more that an agent is willing to tell you about why it is there, the more that you may be willing to charge it less to access the page. If you imagine that an agent comes to a travel site and says, “Hey, I am looking for a vacation for my user and his or her family. It is a family of four. They want to go to Hawaii. What do you have for me?” That actually then might be something where you let the agent on for free, but it opens up all kinds of opportunities for advertising. Again, I think Cloudflare is in the perfect position to help bring both of those things together, both the agent and the places that might have whatever the agent is looking for. That is coming. I don’t know exactly how long it will be before it turns into major sources of revenue, but if you look out in the more distant horizon, if you look 10 years out, I would imagine that this becomes the largest part of Cloudflare’s business…

…Some customers see agents as wonderful, and they want to have them, as many of them come to them as possible. Others see them as a bane to their business and want to keep them away. We need to be able to serve both of those needs. For those customers that want more agentic traffic, we’re helping them with modern There are different terms for it, but sort of the new version of SEO, sort of how do you optimize to get as much agentic traffic as possible? How do you make sure that it’s being served in the most efficient way possible? We’re making that as easy as possible for them. On the other hand, customers that want to keep those agents away, that see them as a threat, we’re blocking that, and they’re not getting charged for that traffic.

Cloudflare’s management is seeing a lot of dumb deals happening in the neocloud space, and management has no interest in being a neocloud 

We have gotten approached over and over again to say, would we do something like what the neoclouds are doing? I think that what our answer has always been is, we are an engineering company. We are not a financial engineering company. We are focused on how we can actually deliver real operational excellence and get real advantages for every CapEx dollar that we spend in order to deliver more benefit. We have watched a bunch of really dumb deals get done. Obviously, the market is digesting those right now. Those just are not the sorts of deals that we would ever do. I think as the market gets more rational and as we actually figure out, okay, you cannot sell a dollar for $0.50 and be in business over the long term, what will happen is people will focus instead on where can we get more value, where can we drive more efficiency? It turns out that we are the platform that has built that.

Cloudflare’s management thinks the company is getting chosen by customers over legacy Zero Trust vendors because the legacy vendors do not have a strategy for dealing with agents, whereas Cloudflare does

[Question] Can you just talk about what you’re seeing in terms of win rates in relation to Cloudflare, one, maybe against some of the first-generation Zero Trust vendors, and why are your customers choosing overall Cloudflare products over some of the legacy vendors there?

[Answer] We were a second-generation entrant into the Zero Trust and SASE space. As such, a lot of what we had to do for a while was just get recognized, get in the conversation. I think what’s really helped us there is actually the developer platform that we’ve had, being out there talking about, “Here’s how you build agents, here’s how you think about agents.” Because it turns out those agents need identity. They need to have some way of being protected. If you’re going to turn agents loose within your organization, having the bulkheads that a Zero Trust system provides is more critical than it ever has before.

As we were talking about the developer platform, what we found is more and more of the time, the Chief Security Officer was sitting in the room taking notes. The number of meetings that I’ve been in where people have said, “I’m way down the RFP process with a traditional kind of gen-one Zero Trust vendor, and I’m going to tear it up because they really aren’t thinking about agents and the agentic space and how that is different than human users, whereas Cloudflare is.” I think that’s given us a significant leg up, and it’s improved our win rates. I think even more importantly, it’s improved the number of deals that we’re in.

Coupang (NYSE: CPNG)

With the help of AI, Coupang’s fulfillment and logistics network are producing higher customer service levels while lowering the cost to serve

Process improvements and deeper automation across our fulfillment and logistics network are producing higher customer service levels and lower cost to serve at the same time. This quarter, for example, we began deploying AI cameras in our packing stations to remove the need for manual scanning, paired with auto baggers that automatically size and customize packaging for each order, cutting both time and packaging waste. Each is small on its own, but they compound with the constant innovation, technology, and process excellence across our operations. 

Datadog (NASDAQ: DDOG)

Datadog’s AI cohort has continued to grow and diversify; management thinks startups to the largest enterprises are building with AI, and this is expanding adoption of the cloud and Datadog; Datadog’s largest customer (OpenAI) reduced usage in 2026 Q3, but the usage was above the minimum commitment; Datadog’s AI cohort growth is significantly faster than the rest of the business; new customers accounted for 30% of Datadog’s revenue growth in 2026 Q3, up from 25% a year ago

Our AI customer cohort remains a strong growth driver. This cohort continued to grow and diversify, both in the number of customers we serve and in the scale of those customers. Overall, we continue to see healthy demand as customers from the most tech-forward startups to the largest enterprises in the world are building with AI and with modern technologies, which we believe is fueling expanded adoption of the cloud and of the Datadog platform…

…As we communicated in August, our largest customer reduced their usage in the third quarter, while their usage remained above their minimum commitment. Meanwhile, our AI customers’ growth continues to significantly outpace the rest of the business…

Datadog’s AI cohort has continued to grow and diversify; all 15 of the largest AI companies are Datadog customers (was 10 of the largest in 2026 Q2); Datadog has more than 850 AI customers in 2026 Q3 (was 750 in 2026 Q2); management is seeing AI activity growing across its non-AI customer base; around 33% of Datadog’s customers now use 1 or more of the company’s AI products, including 75% of the $100,000-plus customers; the number of MCP tool calls is up 50x in 2026 9M; Datadog’s AI cohort growth is significantly faster than the rest of the business; 39 customers in the AI native cohort now spend more than $1 million annually (was 31 in 2026 Q2)

Our AI customer cohort remains a strong growth driver. This cohort continued to grow and diversify, both in the number of customers we serve and in the scale of those customers…

…All 15 of the top 15 AI leaders are our customers. More broadly, we now serve over 850 AI customers, up from about 750 last quarter. Across our broader customer base, we also see AI activity growing substantially. About a third of our customers now use one or more of our AI products, including three-quarters of our $100,000 plus customers, with many of these customers still just getting started with these products, leaving a large opportunity for future expansion. We continue to see rapid growth in agentic activity, with the number of MCP tool calls up by a factor of 50 so far this year…

…Our AI customers’ growth continues to significantly outpace the rest of the business. This over 850-strong group continues to diversify and grow, including 39 customers spending more than $1 million annually with us.

Reminder that Datadog’s management sees the company’s AI initiatives as being split into 2 buckets, namely (1) AI for Datadog, and (2) Datadog for AI; AI for Datadog is about making Datadog’s platform better with AI products and capabilities while Datadog for AI is about Datadog’s end-to-end observability and security capabilities across the AI stack

As AI changes how our customers build and run software, we have been investing and executing on two fronts: AI for Datadog and Datadog for AI. First, AI for Datadog. These are AI products and capabilities that make the Datadog platform more useful and increasingly automated for our customers…

…Datadog for AI. These are the products that deliver end-to-end observability and security across the new AI stack.

Under AI for Datadog, Datadog’s Bits AI product (first launched in 2023) is the fastest-adopted product in the company’s history; management made Bits Testing generally available in October 2026; with Bits Testing, users can discover and test critical user flows in an application; Datadog recently introduced Scout, the company’s AI model optimised for SRE (site reliability engineering) tasks; Scout is on par or even better than frontier models on alert investigations at a much lower cost; Scout is an important demonstration of what Datadog can achieve when it trains AI models with its own production data, critical user workflows, and internal tools; Datadog is busy training more models that will move the company closer to its goal of autonomous operations; Bits AI is monetised both directly, and through driving adoption; Bits AI also has a model where it charges for AI credits specifically for what Bits does, and Datadog only charges a little 

Our flagship agent, Bits AI, is off to a very strong start. Since we launched Bits, it has become the fastest adopted product in our history. In October, Bits Testing became generally available. With a single click, it discovers the critical user flows in an application and generates tests to cover them…

…Last month, we introduced Scout, our own reasoning model specifically optimized for SRE tasks. Scout was built using our in-house distributed reinforcement learning infrastructure and now powers our Bits Investigation. This is exciting for two reasons. First, it performs on par or even better than frontier models on alert investigations at a fraction of the inference cost. This is valuable as it allows Bits to investigate many more alerts, start earlier, and go deeper on incidents as an always-on first responder for our customers. Second, this is also exciting as a demonstration of what we can achieve when we train models with our own production data, critical user workflows, and internal tooling. Scout is only the beginning. With the Adaptive ML team, which joined us in June, we are broadly accelerating our work on reinforcement learning and model post-training. Over time, we expect our research efforts to drive a step change in capabilities across our platform and move us closer to our goal of autonomous operations…

…[Question] Could you help numerate that a little bit and maybe remind us how the pricing works there? Is it kind of directly monetized, or does it really drive the adoption or use of Bits AI overall?

[Answer] It’s actually a bit of both. Look, it largely sits on all of the data we acquire and the way we can process it and all the various products we have. That’s the core. But we also have a model where we charge for AI credit specifically for what Bits does on top of that. And the model we’ve chosen is a model where we stay fairly close to inference costs there, so that we can pass on the savings as the models become more efficient to our customers, and so we can run many more investigations. The reason we’ve done that is the field is changing so fast, and we’re improving the efficiency of those models so fast that it’s not very productive to have a fixed price on everything you do as you expect the volume of things you do to increase by orders of magnitude while the price goes down by orders of magnitude as well. We decided on that model for that, which works really well.

Under Datadog for AI, management thinks customers want important answers to their AI operations on the same platform as the rest of their technology stack; Datadog is able to monitor the performance of a company’s entire AI stack, from compute to agents; Datadog’s GPU Monitoring product helps customers improve GPU utilisation, training throughput, and inference costs; GPU Monitoring’s number of GPU devices monitored is up 8x in 2026 9M; Datadog’s frontier lab customers are collaborating on GPU Monitoring as design partners; GPU Monitoring is not monetised widely yet; the number of spans sent to Datadog’s Agent Observability product is up by 9x in 2026 9M; Agent Observability can be bought directly through frontier AI labs; management is seeing non-AI natives consume more and more of Agent Observability because they are getting deeper into AI; Datadog’s AI Guard product discovers custom agents in an organisation and protects them; Datadog’s Cloud Cost Management product now includes AI costs and lets customers see granular details on their AI spending; more than 33% of Cloud Cost Management use it to track AI costs, and the number of customers doing so is up 11x; management is seeing an explosion in the adoption of AI, which is driving a strong need for tooling, and products such as Cloud Cost Management; Datadog’s Agent Console product shows customers exactly how and which agents are used

As AI moves from experimentation into production, customers today need to answer three key questions. How is it performing? Is it secure? What’s the ROI, or is it worth the spend? Because AI doesn’t run in isolation, they want these answers on the same platform as the rest of their tech stack.

On performance, we cover the whole AI stack, from compute all the way to agents. GPU Monitoring now goes beyond hardware reliability, with new capabilities that help customers maximize fleet utilization, increase training throughput, and serve inference at a lower cost per request. The number of GPU devices monitored is up by a factor of eight so far this year, and some of our Frontier Lab customers are working with us as design partners on these capabilities…

…The GPUs are still largely on, we see them for our customers on training. We have products there, but we do not monetize those products widely yet…

…Agent Observability helps teams monitor and evaluate their agents in production, including the quality of their responses, latency, and token usage. The number of spans sent to Agent Observability is now up by a factor of nine this year. We’ve also made Agent Observability available in the Anthropic Marketplace, and we’re a launch partner to OpenAI’s new enterprise marketplace as well. So now, customers can buy our product directly through the Frontier Labs they already work with…

…We see growing demand for Agent Observability. There was always interesting demand from the AI natives, et cetera, but what is most interesting here is the non-AI natives are consuming more and more, and because they are getting deeper into AI, we see real demand, real volumes, and a really exciting pipeline for those products, actually. So we think that this is crossing the chasm, so to speak, and that is exciting for this product, even though we did not mention any hard numbers today on that.…

…On the security side, our AI Guard product discovers the custom agents running across an organization and protects them at runtime against prompt injection and data exfiltration, drawing on the same real-time data we collect to observe them.

On the cost side, companies everywhere are grappling with a very fast growth in token and AI spend. Beyond the three major clouds, AI providers are now the largest category of spend our customers monitor in Datadog Cloud Cost Management, or CCM. Now, with AI costs in CCM, customers can see which teams, models, and agents drive their AI spend, can govern that spend with budgets, forecasts, and anomaly alerts. More than a third of our CCM customers now track AI costs. The number of customers doing so is up more than 11 x…

…I think what we have seen is there has been an explosion of adoption of AI, in particular coding agents, and that was not planned largely. A year ago, nobody had the amount they spend in their budgets on AI. So, everybody had to try and figure out what that works and what part of that they actually need and what they get from that. So there is a strong need for tooling, which is something we provide. It is a great opportunity for us, and we see great growth and adoption of our Cloud Cost Management products from that…

Moving up the stack again, our Agent Console adds to our CCM data and shows exactly which agents are actually used, by whom, for what task, at what cost, and most importantly, with what outcomes.

Example of a 7-figure expansion deal with a leading software company; the software company has doubled its commitment to Datadog; the software company is moving AI agents into production and will use Datadog Agent Observability to monitor the agents; the software company is consolidating 3 tools into Datadog and will grow to 20 Datadog products 

We signed a seven-figure annualized expansion with a leading software company, more than doubling their annual commitment. This customer is moving AI agents into production and will replace a standalone commercial tool with Datadog Agent Observability to monitor their costs and quality. By using Flex Logs, they can gather and correlate larger log volumes at a lower cost. In total, they will consolidate three tools onto Datadog and grow to 20 Datadog products, including On-Call and Bits AI.

Example of a 6-figure land deal with a top European AI lab (possibly Mistral?); the AI lab’s developers would save time with Datadog’s CI Visibility and Test Optimization product; the AI lab will also be using Bits AI 

We landed a six-figure annualized deal with a top European AI lab. This customer’s developers were losing valuable time to slow builds and flaky tests, which delayed releases in a highly competitive market. They chose Datadog’s CI Visibility and Test Optimization, replacing a commercial tool and freeing their own engineers to focus on shipping product. Their engineers also use Bits AI to quickly build monitors and custom analytics.

Example of a 7-figure expansion deal with a leading AI coding company; the AI coding company is one of the largest users of Datadog’s MCP (model context protocol) servers; the AI coding company will grow to 13 Datadog products

We signed a seven-figure annualized expansion for an eight-figure annualized deal with a leading AI coding company. This customer is growing rapidly and renewed early to keep pace with their expanding footprint. With this expansion, they will adopt Flex Logs and our new Infinite Cardinality Metrics to scale costs effectively with volume. They are also one of the largest users of our MCP Server. This customer will grow to 13 Datadog products.

Example of an 8-figure expansion deal with the AI division of a large technology company; the technology company has tripled its annual commitment with Datadog just 2 quarters after starting as a customer; the technology company is scaling Datadog across its AI training workloads, and is adopting APM (application performance monitoring) and Bits AI

We signed an eight-figure annualized expansion with the AI division of one of the world’s largest technology companies, which nearly tripled their annual commitment. Just two quarters after we landed them, they are scaling Datadog across a large number of AI training workloads. They are also adopting APM and Bits AI to further reduce engineering friction.

Datadog’s management continues to believe that digital transformation, cloud migration, and AI adoption are long-term growth drivers of the company’s business; management is seeing AI workloads move into production; management thinks that as AI writes more code, more complexity is generated and at a faster pace than what engineering teams can track on their own, and this is a problem Datadog can solve; management is observing that AI training could become a sizeable opportunity for Datadog as more enterprises start to do their own model-training

There is no change to our overall view that digital transformation and cloud migration are long-term growth drivers for our business. AI is another large growth driver for us today, as AI workloads move into production to drive more cloud consumption and more usage of the Datadog platform…

…As AI writes and runs more of our customers’ software, they are shipping more code faster across a growing mix of models, agents, and infrastructure. Of course, this generates complexity that is very quickly outpacing what even the best engineering teams can track on their own. That is exactly the problem we are investing to solve. AI for Datadog takes on more of that work for them. Datadog for AI helps them run their AI with confidence…

…If you look a year or two years ago, basically nobody was doing training. It was only a few frontier labs were doing their own training, and nobody else was. Now we see an order of magnitude more large-ish labs or neo labs that are doing training at a large scale. Then there is another whole category of companies below that that are starting to do their own, including us. We started doing training at a pretty decent scale, and we showed some of the results. We discussed some of the results on the call as well. So, I think there is some interesting potential development there where training becomes a part of everybody’s business and everybody is customizing model and building their own. So we are investing. I would say it is still early to call it as a mass market.

Datadog recently did renewals with some large AI customers where the commitments were increased, but the customers got better pricing; these renewals are good for Datadog in the long-term, but create short-term revenue headwinds; the headwinds go away after a few quarters; Datadog’s AI customers are growing at a rate never seen before, so it’s creating new renewal dynamics

Sort of Q3, we did some renewals extensions with some large and large AI customers, where as normal, we essentially extend and increase the commit, but do better pricing for that. And that happened in a number of large customers….

…I could bring up a few of those large customers where we saw renewals way earlier than we thought we would. There’s another one that happened in October that we’re not discussing. These are all good news, but they do create headwind into the short term into the revenue…

…[Question] If I think about the customers renewed early, so they get better pricing, but obviously, the volume is going kind of going crazy there, but it kind of creates a short-term headwind because the same volume is now on a lower price, but then the volume grows, and from then on, in a couple of quarters, we should see the normal progression. Is that kind of the right thing to think about?

[Answer] Yeah, that’s broadly what happens.

…The point here is, there’s one part that is very usual, which is that we have all these customers that are spending and they’re growing and they’re renewing, and then when they renew, they get a better deal, and they also adjust their traffic over time with an underpinning of growth, which is what we’ve seen since the beginning of the company. There’s something that is new that we have some of these companies that are scaling incredibly fast to a level that we’ve never seen in the industry or any industry really before, which creates new dynamics. This is why we can be a bit surprised when we see a customer renew early just a few quarters into their starting to adopt the product.

Datadog’s management is seeing that both the company and its customers are using a mix of AI models, from frontier models, to open-source models, to 1st-party models; management thinks frontier models can do things other types of models cannot, but the open-source models are very cost-efficient for certain workloads

What we see internally and what we see with customers is there is a mix of models that are being used. I think I can say we and include both customers and us into that collective we because the behaviors are fairly similar. We are big consumers of the top labs, top models, but we also consume open-weight models, a growing amount of open-weight models, and we train our own and post-train our own. I think it is probably representative of where the industry as a whole is going. We do that because there are certain things you can only get from the top-of-the-line models that are being generally trained by others, and there is a number of things that are good enough and very cost-efficient with open-weight models you can just pick up and run. And there’s finally a number of other things that you can make way more efficient with your own data, your own workflow, and your own post-training, which is what we’ve done with some of the models we’ve announced. I expect that mix to remain about the same.

MercadoLibre (NASDAQ: MELI)

MercadoLibre has doubled the number of robots in its logistics network in the past year; robotics have been deployed in Mexico and Brazil; robotics have delivered productivity gains of 70%; management is confident that robotics will lower shipping costs as a percentage of GMV and revenue over time

Automation is moving faster still: we have more than doubled the number of sorters and robots in our network over the same period, with robotics now live in Mexico as well as Brazil, delivering productivity gains of 70% against manual processes in picking operations where robotics has been deployed. Progress will not be linear, because new capacity carries costs before the volume that fills it, but over time, as scale grows and productivity rises, lower cost per shipment should flow through into lower shipping costs as a percentage of GMV and revenue. 

AI is now central to how MercadoLibre sells ads, with AI improving relevance, retrieval, click-throughs, and advertising loads; MercadoLibre’s AI advisor now manages campaigns for 56,000 sellers, up 4x sequentially; MercadoLibre has launched Engage+, which uses AI to predict purchase intent and directs spend towards users who engage with the ads; in tests, Engage+ reached significantly more engaged users at a fraction of the cost per engaged user; Engage+ is live in 6 sites

AI is now central to how we sell ads. Better relevance, retrieval and click-through models lifted revenue across the quarter, and ad load is moving from fixed slots to dynamic decisions, so the number of ads on a page adjusts to the quality of the match for each search. Our AI advisor now manages campaigns for 56,000 sellers, almost 4x the number three months ago. And in branding, we launched Engage+, which uses AI to predict from our purchase data which users are most likely to buy in a brand’s category, and directs spend toward the ones who go on to engage with it. In testing against conventional reach campaigns, it reached significantly more engaged users at a fraction of the cost per engaged user. It is now live across six sites.

MercadoLibre’s usage of AI is growing much faster than its spend because the company is becoming more efficient at choosing models for tasks, and lowering the cost of each model used; management expects open-source models to be a large part of MercadoLibre’s model-mix in the medium to long-term; MercadoLibre’s cost per token in 2026 Q3 is much lower than a year ago; MercadoLibre’s returns on AI come in a few ways, namely, (1) higher revenue, (2) better user experience, (3) removal of costs, and (4) productivity; the revenue uplift from deploying AI in search more than offsets the incremental costs of the models used; management thinks MercadoLibre would need multiples of the 7,000 customer support representatives it currently needs if AI was removed; MercadoLibre’s fintech AI assistant, Mago, had 19 million conversations in September 2026; MercadoLibre’s technology teams have doubled their output in the past year because of AI; AI has freed 20% of the finance team’s hours

Our investment in AI grew by almost $30mn QoQ. Usage is growing much faster than our spend, because we are becoming increasingly efficient in two ways: we are getting better at choosing which model to use for each task, and we are lowering the cost of each model we use. This includes significantly expanding our use of open-source models, which we expect to be a large part of our mix over the medium to long term. The result is a cost per token that is much lower than a year ago, and fell again this quarter…

We have built many tools, but the bigger picture is that we are organizing our work, across the entire company, around AI. The returns come through several channels: revenue uplift, better user experience, cost avoidance and productivity, each with its own yardstick. Some of these returns we can measure precisely, and we do. The revenue uplift from deploying AI in search, for example, more than offsets the incremental cost of the models behind it. In customer support, we handle today’s volume with around 7,000 representatives; without AI in those processes, we estimate we would need a multiple of that, so the ROI is clearly positive. And in user experience, Mago, our Fintech AI assistant, held 19mn conversations in September alone. In productivity, the output of our technology teams has doubled over the past year, and the gains extend across the company: in Finance, AI has freed up almost 20% of hours worked. 

MercadoLibre’s management sees agentic shopping as an opportunity to accelerate the shift from offline retail to online retail; management currently does not want to provide MercadoLibre’s inventory to any external AI platform; management’s priority is to build the best agentic commerce experience inside MercadoLibre; management thinks that with an agentic experience within MercadoLibre, it can sustain the high levels of direct traffic; MercadoLibre’s own agent is already available to a small number of users; early results of the agent are promising

We clearly see agentic shopping as an opportunity to accelerate the structural shift from offline retail into online retail. Specifically, the way we’re thinking about this is not necessarily by integrating ourselves to third parties. We continue to look at what’s globally happening, trying to learn, and we are always ready to get new things and adapt. For now, we have decided not to provide our inventory to any external platform. Our priority is still to build the best agentic experience inside MercadoLibre, taking advantage of the 27 years of first-party data that we have, our unique combination of selection, pricing, logistics, financing, loyalty, and so on. E-commerce is much more than just search and find. Trust is built further down the funnel through all the key variables that I was just mentioning, and that is exactly what we excel at. We think that we can combine our strengths with an agentic experience within MercadoLibre, and with that, we believe that we will be able to sustain the high levels of direct traffic to our app or our properties through that asset that we are building. Today, our own agent is already available to a small portion of our user base in our three largest markets. Results are early, but promising.

Shopify (NASDAQ: SHOP)

Shopify’s management thinks personal agents have hit the mainstream; consumers can use Shop Pay as their wallet with personal agents; Shopify is powering product discovery across all major agentic channels;  agentic commerce volume is still small for Shopify, but it’s growing rapidly and management sees the potential for it to be a very large surface area for commerce; AI-driven traffic and orders to Shopify stores is up 2.6x year-on-year in 2026 Q2, and orders  are up 3.3x (AI-driven traffic to Shopify stores was up 3x year-on-year in 2026 Q2 and orders was up 3x); agentic channels are sending shoppers to stores at 3x the rate of organic search, and these shoppers are converting 64% more often; the products surfaced by agentic channels are highly specific; 3/4 of AI-referred sales in 2026 Q3 came from product categories outside the top 100; management thinks personal agents must earn the trust of consumers in order for delegated purchases to scale, and the trust component can be solved by Shopify’s Shop; Shop has all the details on a consumer that agents require, but the agent cannot access these details; Shopify has allowed consumers to use Shop Pay as their chosen wallet for any merchant for agentic purchases done through Claude Duet, Instinct, and Grok Bot; the activation of Shop Pay as the chosen wallet does not require any set up from consumers and merchants and developers; Shopify has connectors with popular agents that gives merchants complete optionality on which agents to sell on; management sees agents as becoming an on-ramp for new merchants to Shopify, although agents are still a small source of new merchants for Shopify today; Shopify’s support for agentic commerce comes in many different features, including Catalog, UCP (Universal Commerce Protocol), Personal Agent Protocol, Shop MCP, and more; there are 400 million shoppers using Shop and Shop Pay; if an agent leads a consumer to a merchant which is using Shopify, the economics to Shopify are exactly the same

Since our last earnings call, personal agents hit the mainstream. Muse, Instinct, Claude Duet, Grok Bot, all partner with Shopify. Our merchants’ checkouts are supported, and buyers can use Shop Pay as their wallet. We’re powering product discovery across all major AI channels, from ChatGPT to Microsoft to Muse…

…Let’s start by talking about the future first, specifically agentic shopping. While the volume is still small here, adoption is growing, and the early signals show that this has the potential to become a very powerful surface area. Here’s what we’re seeing. AI traffic is up 2.6x year-over-year. Orders are growing even faster at 3.3x. AI channels are sending shoppers straight to product pages at 3x the rate of organic search, and shoppers are converting 64% more often…

…AI channels are surfacing highly specific products that fit specific customers with a specific need…

…In Q3, three-quarters of AI-referred sales yet again came from product categories outside our top 100.

…Q3 was the first time we saw shoppers experimenting with delegated purchasing through personal agents, who can also act on behalf of their users with their permission. In order to do that, they must earn their trust. Does it know what I like? Will it get my size right? Can I trust it with my personal information? For Muse, Instinct, Claude Duet, and Grok Bot, the answer is yes, thanks to Shop. Through your Shop account, the agent knows your preferences and shopping history. Through the Shop app, you can track all your orders placed by your agent listed right next to your other orders. With Shop Pay as your wallet, your payment details are already saved. You stay in control of your purchases, and your agent never sees your card details.

With Claude Duet, Instinct, and Grok Bot, that reach even extends beyond stores built on Shopify. Shoppers can now use Shop Pay as their chosen wallet for any merchant, no extra setup required. It just works. For merchants, this means reach, a huge new pool of customers with zero lift for them. For buyers, this means trust, a familiar way to buy in an unfamiliar channel. For developers and AI companies, this means access to world-class infrastructure. Now all the tools they need to build a shopping experience, including catalog, checkout, and the Shop network are in one place. We call it shop.dev, one stop, all commerce complexity handled…

…We have connectors with agents like Muse, ChatGPT, Claude, Google Gemini, giving merchants complete optionality over how and where they run their business…

…Imagine an aspiring entrepreneur talking to ChatGPT or Claude or Muse about starting a retail business. The agent suggests they start building with Shopify, and then it helps them build their store inside the chat. Once they’re ready to start selling, they sign up for a Shopify account. Agents can effectively become ambassadors for Shopify, connecting new entrepreneurs to the platform at precisely the right moment. Of course, this is still a small source of new merchants today relative to our scale, but as more people use more tools to turn ideas into businesses, Shopify is part of that journey right from the start…

…AI and the AI products that we build are a rising tide lifting all boats. We saw that wave coming a long time ago. It’s why we started building Catalog three years ago, before agentic was ever in the lexicon of commerce, to make our merchants’ product data reachable across all AI channels with high fidelity and high consistency. It’s why we co-led the creation of the Universal Commerce Protocol, because we knew that in order for our merchants to take advantage of the opportunities inside agentic commerce, agentic commerce needed a shared language across the internet. It’s also why we’re actively driving other standards, like the Personal Agent Protocol, alongside other technology and retail partners, and why we have embedded Shopify into personal agents like Instinct, Muse, and more. We also recently released shop.dev and the Shop MCP, giving developers one easy connection to add Shopify’s AI tooling, sign-ins, and checkout into their apps or agents…

…The fact that we have 400 million shoppers already trusting Shop and Shop Pay, now that trust goes with them into every single agent…

…If you start, for example, on Muse, and if you as a consumer then do a search which leads to a merchant which is using Shopify, then the economics to us are exactly the same. Essentially, this is another form of discovery. They’re sending you to the merchant’s website, and everything else that you would have done in any other form of discovery ends up to the same economics back to Shopify.

Agentic searches powered by Shopify’s Catalog are converting at 2x the rate of general AI search; Catalog turns over 1 billion product pages into structured data that agents can understand; Catalog is a live, cross-merchant map of every product available across all of Shopify; management thinks Catalog can help long-tail merchants succeed by matching them with buyers’ specific needs

Catalog-powered searches are converting at 2x the rate of general AI search. That is because Catalog is quite literally written for agents. It turns over 1 billion product pages into structured dataset, translating details like size, color, and material into numerical values that an agent can read and understand. It is not just a linear product feed. It is live, cross-merchant map for what is available across all of Shopify, with current pricing and inventory that helps AI channels make more accurate recommendations, and that makes it insanely powerful…

…With Shopify Catalog and how it is optimized for LLMs, more long-tail merchants succeed by matching buyers’ specific needs, however unique, with our merchants’ tailored solutions.

Shopify recently launched Canvas, its upgrade to its online store builder; Canvas allows merchants to build a fully bespoke online store without any coding or design expertise; merchants using Canvas are launching more personalised stores 20% faster

This quarter, we launched the biggest update to our online store builder in our 20-year history. It’s called Canvas. Canvas lets merchants build a fully bespoke online store simply by describing what they want to Sidekick. No coding or design expertise required. The store you can picture is the store you can build. It’s only been live for about a month, but here is what we are seeing. Merchants are launching stores that are more personalized, and they’re doing it 20% faster.

Daily active merchants using Sidekick was up 3.7x year-on-year in 2026 Q3 (was up 3.6x in 2026 Q2), with daily sessions up 4.2x (was up 4.8x in 2026 Q2); in 2026 Q3, Sidekick handled 2/3 of all Shopify Flows, which are automations merchants build; a large furniture retailer used Sidekick to build custom Flows for many aspects of their business; Sidekick is the only AI tool that allows merchants to connect Flows to 3rd-party apps without manual integrations; management has been improving the cost-efficiency of delivering Sidekick; merchant conversations on Sidekick increased 14% sequentially in 2026 Q3 even as cost per conversation declined 28% and total Sidekick costs fell 18%; Sidekick attracts more merchants to Shopify; in 2026 Q3, Sidekick’s dynamic conversion guide for merchants led to a 3% lift in merchants reaching their next 5 orders in 60 days after using the guide

I want to talk a bit more about Sidekick, which is still overwhelmingly the most widely used tool for our merchants to run their business, and usage continues to grow at an even faster clip in Q3. Daily active merchants are up 3.7x year over year, and daily sessions are up 4.2x…

…In Q3, Sidekick created around two-thirds of all Shopify Flow. Remember, flows are those automations merchants use to take repetitive tasks off of their plates. Here is how that shows up in practice. I recently met a large furniture retailer who has built custom Sidekick flows for everything from warehouse inventory snapshots to delivery fee analysis to regional pricing checks. This is what makes Sidekick so powerful. It can be completely tailored to power a merchant’s particular business. Now these flows extend beyond the admin into supported apps that merchants use to run their business. Sidekick is the only AI tool that can connect to them without the merchant having to set up manual integrations…

…Although Sidekick drove year-over-year pressure in margins, on a sequential basis, we have been able to offset that pressure with efficiencies we built into how we deliver Sidekick. We have made improvements in model selection, distillation, and caching to support more usage at a lower cost. From Q2 to Q3, merchant conversations increased 14%, while cost per conversation declined 28%, and total Sidekick cost fell 18%…

…Sidekick attracts more merchants to Shopify, which manifests in the financials and subscription solutions, and it propels more merchant success, which materializes in merchant solutions. Last quarter, we shared that using Sidekick’s onboarding guide results in an 8% lift in merchants reaching five sales within 15 days. This quarter, we can demonstrate benefits beyond that. After their first sales, Sidekick offers merchants a dynamic conversion guide to prompt additional conversion tactics in those critical early months as a business is launching. Examples include help with answering customers’ questions, reaching out to customers with abandoned carts, and more. This resulted in a 3% lift in merchants reaching their next five orders in 60 days after using the guide. Incrementally small, but off large cohorts. That benefit makes a difference and shows the value Sidekick can bring in targeted, proactive merchant communication.

Shopify’s management is seeing AI improve the speed at which the company can execute; management thinks the increasing use of AI is making taste and judgement even more valuable

AI is changing what our teams can build and how quickly they can get an idea into someone’s hands. Small teams can try more things, learn from what works, and build again. That makes taste and judgment even more valuable, seeing a problem we are solving, knowing what a great solution looks like, and then having the curiosity to go after it. That mindset extends well beyond engineering. All across Shopify, our team is rethinking how they work and what they can accomplish with these amazing tools.

Shopify’s management thinks agentic commerce will expand Shopify’s addressable market, because of 2 reasons, namely, (1) Shopify’s merchants are driving a disproportionate share of global e-commerce growth, and (2) Shopify creates new commerce, and AI makes it even easier for entrepreneurs to start; Shopify’s US merchants’ online sales had a CAGR of 23% in the last 3 years, compared to 8% for overall US e-commerce; Shopify’s Europe merchants’ online sales had a CAGR of 37% in the last 3 years, compared to 8% for overall Europe e-commerce; 3 years ago, Shopify merchants was only 20% of incremental e-commerce sales in the US, and 25% in Europe; since the start of 2025, Shopify merchants have been 50% of incremental e-commerce sales in the US, and more than 40% in Europe; through Shopify, niche markets are not only addressable, but they are no longer that small, and this helps create new commerce; the number of 1st time sellers on Shopify has increased 7x since 2018; in 2025, 43% more stores made their first sale within a day of signing up to Shopify compared to 2023

I believe that agentic commerce will expand our market. Commerce continues to grow globally, but in addition to that, there are two important things to keep in mind regarding Shopify’s role and our future opportunity. First, we continue to power a larger share of commerce overall, and second, we are expanding commerce itself by helping more merchants survive and thrive, helping them discover new buyers, and lowering barriers to entrepreneurship. To my first point, the quality and capabilities of our solutions help our merchants drive a disproportionate share of global commerce growth…

…To my second point, measuring our opportunity against the market potential that exists today misses a critical component. Our products do more than help merchants compete for existing commerce. We help create new commerce. AI is an accelerant for our market, simultaneously making entrepreneurship easier to try, but also raising the stakes for what success requires. In that context, Shopify is a merchant’s not-so-secret weapon. We make the complex simple so that our merchants can focus on what matters, building strong businesses…

….Over the last three years, our U.S. merchants’ online sales grew at a compound annual growth rate of 23%, compared to 8% for U.S. e-commerce overall. In Europe, we saw a similar pattern. Our merchants grew online GMV at a compound annual growth rate of 37%, compared to a market growth of 8%, doubling their share of e-commerce in just three years. These trends have accelerated. Three years ago, Shopify merchants drove nearly a fifth of incremental e-commerce sales in the U.S. and nearly a quarter in Europe. Since the start of 2025, they now capture nearly half in the U.S. and more than 40% in Europe…

…When your niche passion products can more easily find the customers they are meant for, you are more likely to build that business. Smaller niche markets are now not only addressable, they are not so small anymore. With Shopify Catalog and how it is optimized for LLMs, more long-tail merchants succeed by matching buyers’ specific needs, however unique, with our merchants’ tailored solutions. Extending the long tail of commerce is helping create new commerce. in Q3, three-quarters of AI-referred sales yet again came from product categories outside our top 100…

…Since 2018, the number of first-time sellers on Shopify has increased seven times. Moreover, we help new entrepreneurs reach that milestone faster than ever before. Last year, compared to just three years ago, 43% more stores made their first sale within a single day of signing up. We’re not just finding high-intent entrepreneurs. We’re bending the curve for how quickly they reach early success. AI and the AI products that we build are a rising tide lifting all boats.

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Apple, Coupang, Datadog, MercadoLibre, and Shopify. Holdings are subject to change at any time.