AMZNAmazonNeutralAmazon[scuttleslops] RDDT, AI and cybersecurity, AMZN, MU, META, BLDR, RYAN
Google is reading Reddit and sending it less traffic; cybersecurity is entering a dangerous new era; AWS is on its way to $1 trillion; Micron is trying to differentiate its memory; Meta’s advantage in personal AI agents; Builders FirstSource faces a brutal housing market and intensifying price competition; Ryan Specialty faces a tough E&S market
Google is reading Reddit and sending it less traffic; cybersecurity is entering a dangerous new era; AWS is on its way to $1 trillion; Micron is trying to differentiate its memory; Meta’s advantage in personal AI agents; Builders FirstSource faces a brutal housing market and intensifying price competition; Ryan Specialty faces a tough E&S market
View original →Neutral
AMZNAmazonAmazon
APPAppLovinAppLovin
GOOGLAlphabetAlphabet
SPGIS&P GlobalS&P Global
WMTWalmartWalmart [scuttleslops] GOOGL, AMZN, BRO, KNSL, ODFL, MCO, SPGI
AI could improve Google’s ad economics ; AI agents need Amazon ; more specialty insurance risks are becoming conventional ; Kinsale’s competitors are getting reckless; Old Dominion is sitting on excess capacity and taking price ; AI investment is boosting the credit ratings business; S&P Global wants to be infrastructure, not tools Upcoming posts (in no particular order): Domino’s Pizza, Constellation Software, Otis, Gartner, Badger Meter, Tractor Supply, Fortive, and Veralto. What are scuttleslops? The Synopsis podcast; Interview with Eric Seufert on whether AI will destroy Google’s search monopoly ; Oct. 22, 2026 Eric Seufert, the independent advertising analyst behind Mobile Dev Memo, believes investors have been far too quick to conclude that conversational artificial intelligence will undermine Google’s search business. Traditional Google Search was designed to send users elsewhere as quickly as possible, ideally through a relevant sponsored link, whereas AI Overviews and AI Mode can answer questions directly and encourage users to continue asking follow-up questions without leaving Google’s properties. Seufert argues that this transformation could actually improve Google’s economics because the company retains more of the user’s attention, learns more about what the user wants, and gets additional opportunities to show advertisements. He points to Google’s reported 2.5 billion monthly users of AI Overviews and one billion users of AI Mode, alongside approximately $63 billion in quarterly search revenue growing 17% year over year, as evidence that the company is successfully adapting despite ChatGPT’s popularity. Seufert is also unusually optimistic about advertising in conversational AI, rejecting the argument that introducing sponsored recommendations will necessarily undermine users’ trust in their assistants. He argues that consumers generally appreciate relevant advertising, and that modern advertising systems are sophisticated enough to distinguish situations where ads would be useful from sensitive conversations where they would be inappropriate. Conversational interfaces may ultimately offer advertisers better targeting than traditional search because users reveal their preferences gradually, refining a broad request into increasingly specific requirements. Someone initially asking about televisions, for example, might eventually specify a budget, screen size and intended use, giving the platform enough information to show a highly relevant advertisement near the end of the conversation rather than several less useful ones at the beginning. Seufert envisions a world of “better ads and fewer of them.” Regarding concerns about how shopping agents will eventually make money, with Seufert predicting that competition will push them toward advertising rather than commissions on purchases. A shopping agent charging retailers a fixed 5% commission would exclude merchants unable to afford that fee while collecting less than other merchants would willingly pay for access to a valuable customer, making a fixed commission inherently inefficient. An advertising auction solves both problems by allowing merchants to bid according to what each customer interaction is worth to them, with the platform selecting ads based on a combination of relevance and willingness to pay. Seufert doesn’t independent shopping agents to displace Amazon, because retailers have powerful incentives to prevent intermediaries from controlling their relationships with customers. Amazon’s advertising business is enormously profitable, and allowing agents to purchase products without bringing shoppers onto Amazon’s website could eliminate opportunities to advertise related products, recommend upgrades and encourage future purchases. Seufert contrasts Amazon with Walmart, which has been more receptive to working with outside agents partly because it has less advertising profit to protect and potentially more to gain from attracting additional online shoppers. He is considerably more optimistic about agents disrupting online travel agencies, because hotels and airlines already maintain infrastructure that allows intermediaries to search inventory and complete bookings, while customers must ultimately interact directly with the hotel or airline to consume the service. Booking.com ‘s ability to aggregate thousands of unfamiliar hotels and present them to travelers is valuable today, but an agent capable of finding a hotel within a 20-minute walk of a particular attraction, with a gym and a courtyard-facing room, could potentially perform that discovery function without requiring hotels to pay Booking’s substantial commissions. Turning to Apple, Seufert believes a more capable Siri could become an important interface through which consumers interact with other companies’ applications, particularly for simple tasks that don’t require users to spend time inside an app. He sees several possible ways Apple could profit from this change, including subscriptions for advanced Siri functionality, collecting fees from competing model providers seeking access to Apple’s users, and charging for processing tasks through Apple’s privacy-focused cloud infrastructure. Apple knows which applications people install, which subscriptions they maintain and where they spend money, potentially allowing it to establish a formidable advertising network if it decides to compete more aggressively with companies such as AppLovin and Unity, particularly because Apple controls the operating system through which competing advertising technologies operate. Subscribe now Read more
View original →
AMZNAmazonNeutralAmazon[scuttleslops] AMZN, MTCH, SPOT, Open vs. closed models
Agents are less disruptive to commerce than you think; how Match is trying to find its spark; the podcast economy; open models have a place Agents are less disruptive to commerce than you think; how Match is trying to find its spark; the podcast economy; open models have a place
Upcoming posts (in no particular order): Constellation Software, Otis, Gartner, Badger Meter, Tractor Supply, Intuitive Surgical, Fortive, and Veralto
View original →Neutral
AMZNAmazonAmazon
AMD
APPAppLovinAppLovin
AVGOBroadcomBroadcom
MRVLMarvellMarvell
NVDANVIDIANVIDIA [scuttleslops] MRVL, APP, AMZN, Chips
what are scuttleslops? The Circuit; Unpacking Marvell’s Analyst Day, OpenAI Revenue Rumors, and the Heterogeneous Rack ; Oct. 12, 2026 In a discussion of Marvell’s recent Investor Day, semiconductor analysts Ben Bajarin and Jay Goldberg argue that the company has finally articulated a more convincing strategy for competing in the rapidly expanding AI infrastructure market. For years, investors were preoccupied with how much custom accelerator business Marvell was losing at Amazon to competitors such as Alchip, but management is increasingly positioning Marvell as a supplier of the chips needed to connect and operate AI servers rather than primarily as a designer of the accelerators themselves. This includes networking switches, optical connectivity, and other components that sit alongside GPUs and custom processors, many of which Marvell has been producing for years but which are becoming much larger markets as data center spending explodes. Marvell’s new five-year revenue and earnings targets were much higher than investors expected, reflecting the growing demand for chips beyond the AI processors themselves. As AI data centers become more complicated, they increasingly need networking chips and other components to connect GPUs, CPUs, memory, and entire server racks. Bajarin believes Marvell is particularly well positioned because data centers are likely to use a mix of NVIDIA GPUs, custom AI accelerators, and processors from other suppliers, all of which need to communicate efficiently. He also speculates that Marvell could capture opportunities Broadcom chooses not to pursue, such as supporting NVIDIA’s NVLink technology to connect different types of computing hardware. Rather than competing directly with NVIDIA or Broadcom to design the main AI processors, Marvell can benefit by supplying the chips that connect those processors and allow them to work together. The speakers believe AI computing infrastructure is becoming increasingly specialized, potentially creating opportunities for semiconductor suppliers beyond NVIDIA. Goldberg argues that inference workloads are likely to require different types of processors for different tasks, particularly prefill, where a model processes the user’s input and context, and decode, where it generates the response token by token. These operations place different demands on computing power and memory bandwidth, making it potentially more economical to use different chips rather than a single general-purpose GPU for everything. NVIDIA is already moving in this direction with specialized products, while Google is developing different versions of its Tensor Processing Units for particular workloads. Goldberg believes this increasing specialization could open the door for AMD, Broadcom, Marvell, and other competitors, although he does not expect the market to support dozens of independent accelerator startups. Bajarin takes the argument further, suggesting that hyperscalers ultimately want to connect multiple racks containing different types of processors into a single computing environment capable of sharing memory, accessing the same context, and running enormous models across the entire system. Today, even when thousands of AI processors are connected together, large AI models are typically divided into pieces that run on different groups of processors, rather than having all those processors operate as one large computing system with shared access to memory and data. Bajarin believes the industry will eventually move toward connecting entire server racks into a single computing system, potentially combining NVIDIA GPUs, custom AI accelerators, and processors from other vendors. This would allow data center operators to run workloads on whichever hardware can handle them most cheaply and efficiently, rather than maintaining separate systems for each type of processor. Making different chips work together this way is an enormous technical challenge, but it could create substantial demand for the networking and connectivity technologies supplied by companies like Marvell. Marvell is also developing a new type of chip to help manage increasingly complex AI servers. Traditional servers use baseboard management controllers (BMCs) to monitor things like temperature, power consumption, and fan speeds. But modern AI servers contain far more processors, consume much more electricity, and have hundreds of sensors that need to be monitored and controlled. Existing management chips weren’t designed to handle this complexity, so engineers have been adding field-programmable gate arrays (FPGAs)—flexible chips that can be programmed to perform specific tasks—to fill the gaps. Marvell may see an opportunity to develop specialized chips that handle these functions more efficiently. Marvell has suggested that it is developing silicon specifically for these increasingly complicated systems, although the company provided few details about the actual product. Goldberg points out that even a 2% improvement in electricity consumption becomes economically meaningful when applied across a gigawatt-scale data center, illustrating why seemingly minor components can become valuable in AI infrastructure. The larger point is that conventional server designs were relatively mature, whereas today’s rapidly increasing power densities and networking requirements are creating engineering problems that existing chips were never designed to solve. Subscribe now Read more
View original →
AMZNAmazonNeutralAmazon[scuttleslops] Chatbot advertising, AMZN, ROP, China EV’s and Humanoid Robots, APH
Upcoming posts (in no particular order): CoStar, Gartner, Tractor Supply, Intuitive Surgical, Fortive, and Veralto What are scuttleslops? Mobile Dev Memo Podcast; Season 7, Ep. 31: Understanding the privileged position of chatbot advertising ; September 2, 2026 Chatbot advertising could become unusually powerful because a chatbot can combine the strongest feature of search advertising—explicit commercial intent—with the strongest feature of social advertising—behavioral targeting based on the user. Search usually sees a short query, while a chatbot may see a long conversation in which the user gradually explains what they are trying to accomplish, what constraints they have, and what trade-offs they care about. The author calls this a “privileged position” because the ad system can target either the conversational task itself or the user more broadly, depending on whether the current session has obvious commercial intent. In other words, if somebody is actively planning a Disney trip, the system can use that context; if somebody is discussing a noncommercial topic, it can instead rely on first-party behavior, customer lists, conversion data, or other off-platform signals much like Meta does. That makes chatbot advertising structurally different from traditional search. A search engine historically tried to answer a query well enough to send the user somewhere else, so the successful interaction often ended with one click. A chatbot tries to keep the user inside the product until the problem is resolved, meaning the economically relevant unit can become an entire multi-turn session rather than a single query. The author’s example is useful: “Hotels on Monorail Disney World August 2025” contains obvious commercial intent, while “How can I get back and forth to Disney World without a car?” initially looks informational but may develop into hotel, transportation, restaurant, and itinerary decisions after a few follow-up questions. That means an initially unmonetizable prompt can become commercially valuable as the chatbot learns more about what the user actually wants. The author thinks Google and OpenAI are converging on the same end-state from opposite directions. Google is turning search from what he calls a “distribution mechanism” into an “engagement sink” through AI Overviews and AI Mode, keeping users inside Google longer and creating more opportunities to monetize the subsequent steps of a search journey. OpenAI started with the engagement sink and is now layering on the infrastructure of a performance-advertising business. This is already visible in Google’s rising query volumes and strong search revenue growth, even as AI-generated answers were supposed to cannibalize clicks and therefore hurt monetization. The simple “AI destroys search ads” thesis is too crude because conversational search can actually create more queries, more ad exposures per session, and more opportunities to pull commerce directly into the interface. Subscribe now Read more
View original →