AI Costs Pressure Margins as Google Search Risks Sending Less Traffic

Apple and Palantir both beat earnings expectations, but their shares were down around 2% on the morning discussed. John Coogan John Coogan said that positive results were not enough to quiet investor concerns about future costs and margin compression.

Apple reported record September-quarter revenue. Coogan pointed to a Financial Times report that the company warned rising memory-chip prices and the cost of running its AI tools could cut profit margins. Apple has historically held prices steady for extended periods, he said, preferring to set a clean price and stick with it for a year or two, or even longer. If memory prices rise in the middle of that cycle, Apple may have to absorb more of the increase rather than pass it promptly to customers. The pressure, as Coogan described it, is that component and AI costs can change while the price customers see tends to remain fixed.

Palantir’s third-quarter revenue came in at $2.3 billion, which Coogan described as a $9.25 billion annualized run rate. He called it a $400 billion company. US commercial revenue reached $984 million, exceeding US government revenue of $953 million for the first time. Coogan treated the shift as evidence of commercial growth and as a point in a wider debate about where AI’s business value will accrue: to foundation-model companies such as OpenAI and Anthropic, or to companies that help organizations deploy AI.

He compared Palantir’s pitch with Microsoft’s: work with customers on their data, help implement AI, and support organizational change. Jordi Hays Jordi Hays agreed that Alex Karp and Satya Nadella have very different personalities, while their business pitches have a similar shape. Coogan said Palantir’s commercial growth suggested customers had been receptive to that implementation-focused message.

The cost side of AI surfaced more directly in the discussion of Grok Bot. Elon Musk had posted that the cost of serving it was “unsustainably high relative to revenue.” A post by Sheel Mohnot, shown alongside Musk’s, made the broader point that computer use is expensive today even as AI companies race to lower costs and offer the resulting capabilities before they have worked out how to profit from them.

Coogan said Musk’s admission surprised him. Hays said the current read was that Grok Bot had been using Anthropic’s Opus 55 under the hood, but he was still unsure what to make of the cost announcement. He questioned why a company with substantial resources would not absorb costs for some time, particularly when Grok Bot was still a relatively small product. Hays put its weekly active users at around one million.

The uncertainty was sharpened by a separate statement from Musk, as Coogan recalled it, that Grok Bot could eventually consume most of the company’s compute. Coogan offered a possible reconciliation: perhaps a future version would be valuable enough that people would pay more for it. Hays pushed back, asking why that possibility would not coexist with falling costs to serve users. The two did not resolve the tension.

They were also unsure what had changed for users. Coogan initially wondered if the issue was a denied reset. Hays instead recalled complaints that usage had climbed to a limit and then stopped. Coogan’s tentative description was that the company had “right-sized the tokenomics” internally. He also observed that Grok Bot had not simply been turned off.

For Coogan, the limits raised a product and distribution question beyond the immediate expense. Grok fits naturally as an answer bot inside X, including for questions about posts or summaries of news articles. But he found the move from a social platform used largely for entertainment and discussion to workplace tools such as Slack less obvious. Compared with Microsoft or Palantir, he said, Grok appeared to have a less established enterprise go-to-market path.

A post by BuccoCapital Bloke about Reddit’s earnings call framed Google’s AI Mode as a threat to sites that depend on search referrals. Coogan read through the post’s claims: AI Mode was becoming a default Google experience; it supplied richer information inline, reducing the incentive to leave search; and its conversational format could keep users engaged over more turns, creating more opportunities to show ads.

The post argued that Google would extract as much revenue as it could from search while transitioning toward cloud services and chips. If human traffic became scarcer, it said, Google could become increasingly adversarial to the rest of the internet. The post characterized media companies that are not destinations as exposed, and described content-driven customer acquisition as a headwind for smaller businesses. It named HubSpot and Monday.com as companies it said had been “crushed” by a collapsing funnel.

Coogan connected the referral concern to remarks attributed to Reddit CEO Steve Huffman: people still want Reddit, but Google is not sending them there as it used to. He saw the complaint as part of a longer-running concern sometimes called “Google Zero”—the possibility that changes to Google search reduce the traffic other sites receive. Coogan recalled that the idea had been discussed for years and noted that a Condé Nast figure had spoken on the show about the company embracing the reality that search was no longer a growth engine. Hearing the concern applied to Reddit suggested to him that it was spreading.

The shift leaves Google with a product dilemma. Coogan said the company has to meet users where they want to go, even if that means changing search. But he also said people are increasingly voicing dislike of AI Mode and asking to return to an older experience. He described the new flow as potentially jarring: users can land in AI Mode, then have to work out how to navigate from its answer to the destination they want. The tension is between adapting search to the product Google is building and users’ preference for the familiar version.

Hays argued that it remains difficult to make a coherent bearish case on Google if AI is likely to become a large business. Even if search revenue declines, he said, Google has other assets and can use high-margin search revenue to fund infrastructure and acquisitions. Coogan raised a countervailing signal from a post by Tae Kim, citing Menlo Ventures: OpenAI and Anthropic together accounted for 86% of enterprise LLM API spending, up from 67% the previous year, while Google’s share among builders had fallen by more than half. The figures concerned enterprise use by builders, not consumer AI usage.

Hays’s response was that some of this demand still benefits Google through Google Cloud and its TPU business. The company might lose direct model usage while collecting revenue from the infrastructure that supports it. He pointed to Google’s exposure across cloud and chips, as well as its investments in companies including SpaceX and Anthropic, and called it one of the greatest businesses in history.

That broad asset base did not settle a separate question about how Google is run. Coogan asked whether the company was in “founder mode,” pointing to reports of Larry Page and Sergey Brin becoming more involved alongside the presence of an operating CEO. Hays said Google did not come across as being in founder mode in the way Meta does. He also named the recent talent exodus as a headwind, while arguing that there was still significant upside if Google could remain competitive in cloud and chips.

The schools dispute made the product transition more concrete. A New York Times headline shared in a post by Andrew Curran said Google would turn off search for schools that reject AI. Curran argued that the decision was being framed incorrectly: even when users do not turn on AI Mode, Gemini is involved in the search box, so “the old Google search is gone.” Hays summarized the apparent position as an argument that AI now runs throughout Google’s products; rejecting AI altogether would therefore mean giving up access to those products.

Coogan accepted the distinction between turning on AI Mode and using a search product that incorporates Gemini, but noted the practical consequence for schools that want a more traditional experience. They may need to use another search engine, he said. The exchange left the central tension in view: Google says AI is embedded in search, while some schools may want a version that excludes it.

John Coogan Coogan read CrowdStrike’s planned acquisitions of Fiddler AI and Dreadnode as a sign that companies were taking AI security seriously. A post shown during the discussion put the combined price at $495 million, describing Fiddler as providing agent observability and Dreadnode as testing AI systems for vulnerabilities.

The post characterized Fiddler’s work as tracking agents’ decisions and tool use, explaining why they took certain actions, and identifying hallucinations, behavioral drift, and unintended actions. It described Dreadnode as testing models, applications, and agents before deployment, using more than 100 attack strategies, including jailbreaks, prompt injection, and attacks specific to agents.

Coogan recalled CrowdStrike CEO George Kurtz discussing an AI “kill switch,” connecting that language to broader calls for a way to shut down AI systems. The hosts briefly joked about names for the infrastructure behind AI—“compute campuses,” “data bases,” and “supercomputers”—before turning to other security developments.

OpenAI had appointed former US Army Secretary Dan Driscoll to a newly created role as senior counselor for national security modernization, according to a post sharing a Financial Times headline. Coogan recalled Driscoll’s earlier appearance on the show and the change in tone that came with interviewing a former Army secretary rather than a founder or investor.

OpenAI had also partnered with Cathedral Systems on internal security testing. Coogan said Cathedral had been involved in breaking into Anthropic’s Slack and described its penetration-testing expertise as relevant to the effort to secure AI systems. He argued that agents could become a “master key” as users give them access to passwords and services. If an agent can reach email and other accounts, he said, a compromise could quickly put more than the model itself at risk. He described cybersecurity as important not only for protecting model weights but also for the user experience of AI products.

Jordi Hays Hays described SpaceX’s Starfall as a service for affordable, routine access to microgravity. The SpaceX material shown specified payloads returning from low Earth orbit of up to 1,200 kilograms. It named three areas of use: biomedical research, including biomaterials, medical implants, stem-cell organoids, and replacement tissues; pharmaceutical discovery through growing more uniform protein and drug crystals; and materials science, including manufacturing next-generation materials and semiconductors.

Hays said the market had been hard to assess because it sounded experimental and distant from the familiar startup proposition of making a product, selling a known quantity, and projecting growth. He had viewed the category as complementary to Varda, which had already been developing work in space. SpaceX’s entry meant Varda would face a major competitor, but Hays also saw it as validation: if SpaceX was pursuing the category, he reasoned, it must see a substantial opportunity and expect meaningful demand.

Coogan separated two questions that could otherwise be blurred together. SpaceX might be providing launch and return capacity, while Varda had developed its own capsule and built partnerships with biotechnology companies. Those activities could leave room for collaboration even as the businesses compete. He also pointed to a broader pattern in critical infrastructure: customers may not want to rely on a single provider. In his examples, companies seek alternatives in launch capacity as well as other strategically important services, partly to have negotiating leverage and avoid single-sourcing.

For both hosts, SpaceX’s move made the emerging market more consequential without establishing exactly how the division of work would develop. Hays emphasized the competitive threat to Varda and the signal of market confidence; Coogan emphasized the possibility of complementary roles and the appeal of having multiple providers.

John Coogan Coogan raised a privacy concern prompted by a post explaining how the sound of nearby airplanes might reveal where a video was recorded. The technique described in the post uses Doppler curves—the changes in pitch as aircraft move relative to a microphone—to help triangulate a location. Coogan said the exercise for viewers was to try to work out where he was from his audio, and wondered how much clean audio the method required.

Hays noted that Coogan was on a Zoom call, where noise cancellation might affect the sound. Coogan agreed that the call was not a continuous, clean recording and joked that soundboard effects might help confuse anyone trying to locate him. Beneath the jokes, he described the method as an example of AI-enabled surveillance.

He said the risk could apply to people who have recorded long, unedited vlogs outside their homes, with continuous audio and no cuts. Coogan compared the technique with geolocation by visual clues, which he said Rainbolt had long demonstrated through photographs and Google Maps. His point was that location privacy was already a layered problem, and the audio method added another possible source of clues. He called it a striking mathematical development, while leaving the exact amount and quality of audio needed unresolved.

A screenshot shared by Thomas Chavez showed an Instagram-story-style ad with the caption “Hey meta waste my money,” attached to a promotion for Ouloul Capital. Coogan said platforms make it easy to promote content, including by paying with a credit card in the app. Hays countered that X makes the process even easier.

The hosts joked about X’s prominent promotion button. Coogan also described what he experienced as pressure to enable notifications. When he declined, he said, the interface moved into a state that looked broken, as if the user had to correct an error. He called the pattern effective at encouraging people to switch notifications on and tied it to efforts to increase retention.

The hosts briefly discussed Rowdy, a new snack brand that Coogan described as Medici’s first savory consumer product. He praised its tangy-ranch flavor as close to Doritos Cool Ranch and predicted it could be a hit. Hays teased him about his enthusiasm for Doritos, and Coogan described his own life as “Doritos coated.”

A clip about the Philadelphia 76ers turned on confusion between Bloom Energy and an AI company with the same name. In the clip, its speaker said the AI company had paid about $30 million to sponsor the team. The speaker also said people had messaged him thinking he had signed the 76ers. The hosts found the mix-up funny.

They also noted Shopify’s purchase of shop.com for $99 million. Hays called it an impressive domain and said he thought Shopify had gotten a good deal, with Coogan agreeing. The exchange did not spell out Shopify’s plans for the domain.