Apple reported record revenue for its September quarter and beat earnings expectations. However, the Financial Times reported that the company warned rising memory-chip prices and the cost of running its AI tools could cut profit margins. The hosts noted that Apple typically holds prices steady for long periods, so higher component costs mid-cycle could put the company in a difficult position. At the time of the discussion, Apple shares were down about 2%; the hosts linked weakness in stocks in part to market concerns about possible future margin compression.

Palantir reported $2.3 billion in third-quarter revenue, equivalent to a $9.25 billion annualized run rate. For the first time, US commercial revenue exceeded revenue from US government customers: $984 million versus $953 million. The hosts said revenue had nearly doubled year over year and that shares were down around 2% at the time of the broadcast. They also described Palantir's commercial pitch as resembling Microsoft's: helping companies implement AI and transform their organizations. One participant stressed that the similarity was in the business message, not in the personalities of executives Alex Karp and Satya Nadella.

In a discussion of Reddit's results, the hosts relayed the argument that Google's AI mode is increasingly becoming the default search interface: it presents answers directly in search and supports dialogue, potentially reducing the incentive to visit third-party sites. In their view, this model lets Google keep users engaged and serve ads at each turn, while putting the company in greater conflict with the rest of the internet ecosystem. Citing Reddit CEO Steve Huffman, they said people still want to use Reddit, but Google is sending it fewer visitors than before.

The hosts described this as a serious headwind for media and other companies reliant on search traffic, especially content-led customer acquisition among small and midsize businesses; they cited HubSpot and monday.com as examples. One participant called the situation an “endgame” struggle for a dwindling volume of human internet traffic. In a discussion of Google's prospects, another argued that even if search revenue declines, the company retains substantial cash flows and assets in cloud and chips. The hosts also cited an estimate that OpenAI and Anthropic's share of enterprise spending on large-language-model APIs had risen from 67% to 86%, while Google's share among enterprise builders had fallen by more than half. This was an estimate presented during the broadcast, not data independently established by the participants.

The hosts discussed a New York Times headline saying Google would turn off search for schools that reject AI. One participant called the framing overly aggressive and relayed Google representative Andrew Kern's clarification: he said the decision was being incorrectly portrayed as punitive. Kern argues that Google search without Gemini no longer exists; even when users do not turn on AI mode, Gemini is still used in the search box. The hosts noted that schools seeking a more traditional search experience without large language models could choose another service, such as DuckDuckGo. They also expressed the view that users need a classic-search option.

On November 1, Elon Musk said serving Grok Bot was too costly relative to revenue. The hosts contrasted this with his earlier public prediction that the bot would eventually consume most of the company's compute. One participant speculated that the service might have been using Opus 5.5 under the hood; this was an unconfirmed guess. The hosts said Grok Bot had around 1 million weekly active users and that the company behind it was valued at about $2 trillion. They expressed surprise that the service was not allowed to operate at a loss for longer and said the situation was unclear. One participant noted that the episode may not have involved shutting down the bot, but rather denying a user's request to reset a limit after usage reached 100%. The hosts also offered a hypothetical calculation: 1 million paying users at $100 per month would generate $100 million in monthly revenue; this was not presented as the actual subscriber count or price.

The hosts discussed how xAI might right-size Grok’s economics. One host speculated that even if the service generated 10% of Anthropic’s revenue, its costs would be less than $1 billion; as a separate calculation, he said one million people paying $100 a month would generate $100 million. He said the business could still lose substantial sums under those conditions, but had significant capital, and suggested that Elon Musk might sharply cut costs and wait.

The host also said he was surprised by the level of Grok usage, while qualifying that his impression might be driven by discussion on X. He considered the move from an X-integrated bot for answers and summaries to an enterprise product and Slack integration less convincing than going to market through more mature enterprise channels, such as those of Palantir or Microsoft. These were the host’s assessments, not a verified financial forecast or company statement.

A host suggested that someone’s location could be identified from a long outdoor audio recording: distinctive Doppler curves from passing aircraft could be analyzed, matched to a location, and used to triangulate where the person is. He had heard that this might require about 15 minutes of clean audio, but questioned how clean the recording would need to be, particularly when sound is transmitted through a video call with noise cancellation.

Another host said analogous geolocation from photos has long been possible: GeoGuessr creator Rainbolt, he said, has identified places using images and Google Maps. The first host concluded that long, unedited vlogs recorded outside someone’s home could reveal their location, calling the scenario an example of AI-enabled surveillance. The participants did not establish that the method had been tested on the recording under discussion.

A Shopify representative said the company’s third-quarter GMV was $119 billion, up 29.5% year over year; revenue reached $3.7 billion, up 32%; and free-cash-flow margin was 22%. He said MAC, Clinique, Balmain, and Lululemon moved to Shopify during the quarter, and that the company was beginning to see more commerce through agentic services and personal agents embedded in them.

The representative described agentic shopping as a developing area that remains small relative to Shopify’s overall volume: Shopify’s total GMV last year was $400 billion, he said. Still, traffic from AI agents was up 2.6x year over year and orders were up 3.3x. Conversion using the Shopify catalog was 2x higher than with general AI search, which the representative characterized as effectively scraping. He said early AI beneficiaries include “long-tail” products, including those outside the top 100 categories.

A host said he had made his first agentic purchase: he asked Dot in ChatGPT to find and buy a raincoat, then paid with Shop Pay. The Shopify representative said connecting Shop to personal agents can give them access to purchase history for recommendations; Shop Pay also simplifies payment, and the Shop app can track packages. He characterized Shop Pay as a kind of identity system for agents connected to Shop Pay.

In the representative’s view, agentic commerce is moving from product discovery toward delegated purchases. He said agents may be especially useful when shoppers have not chosen a specific brand or need an item with particular attributes: an agent can compare many products with the shopper’s needs. He cited a specialized Roborock robot vacuum, WeThings smartwatches, and a car seat that fits three children and the back of a sedan as examples.

The representative advised merchants to check Shopify’s knowledge center to see what agents can learn about their products and to provide relevant attributes, such as whether a raincoat works in cold weather. In his view, complete product information can make a product more likely to surface in agent responses. A host added that sizing details and precise measurements—including conversions between metric and imperial units—could help an agent match an item to a user’s measurements. The representative also said merchants selling across borders should clearly state which products can be shipped to each country: restrictions on particular types of fish, for example, may depend on duties and tariffs.

A Shopify representative said transactions through shopping AI agents currently carry the same fees as online-store purchases; when a payment goes through Shopify Payments, the same economics apply. From a merchant’s perspective, he said, nothing changes for now, whether a customer arrives through Muse, Instinct, Google AI Mode, Microsoft Copilot or GPT.

The company does not assume agentic shopping will necessarily remain ad-free. Shopify has already integrated a ChatGPT ads app: announced in Q3, it lets all Shopify merchants create and manage campaigns. The representative suggested that a merchant might eventually pay an application a commission when a shopper chooses a product in a sponsored placement while searching with a personal agent. He stressed that this scenario has not happened yet, and compared it with Google search advertising, where, in his view, people often click organic results.

The participants also discussed whether agents might be paid for generating demand. One host argued that it could be fair for a merchant to share revenue with an agent if it researched products and created the buyer’s interest. The Shopify representative said such payments may happen eventually, but merchants’ current economics have not changed. In his assessment, AI applications are currently focused first on attracting users and giving them a good experience.

Northwood Space founder and CEO Bridget Mendler said the company has completed a $375 million Series C round. She said the fundraising followed validation of the capabilities of its two products: Portal, a C2 system, and Prism, a high-throughput product. Their performance was validated not only in low Earth orbit (LEO) but up to geostationary orbit; the company was able to field the demonstrated capability within days.

Mendler said Northwood has produced 29 hardware units across the products and completed two production runs. The company is in a new, expanded 185,000-square-foot facility designed to produce hundreds of antennas. She described production scaling as the next milestone.

According to the CEO, Northwood has signed new customers and is seeing substantial interest in space solutions from commercial and government customers. She cited more solidified government programs and funding. Portal and Prism are designed to serve a broad range of use cases. Key customer requirements include scale and reliability; Mendler said third-party ground networks historically have not always provided the required uptime. She also described rising throughput and the growth of data in space as promising areas, and called national security a key priority.

On supply chains, Mendler said the new capital will help the company scale production by deepening vendor relationships and bringing some work in-house. Northwood considers insourcing, among other reasons, when vendor timelines constrain development or when performance or cost makes it appropriate. The company is building in nine countries and, Mendler said, expects to increase both its geographic reach and network density substantially; she did not specify a target number.

Mendler added that Northwood is enthusiastic about new-graduate hires who are energetic and able to get things done quickly. In her assessment, junior employees already take on significant ownership; the optimal balance between new graduates and more experienced staff remains an open question.

Members of SemiAnalysis’s tokenomics team presented GDPx and described it, in their own view, as the best benchmark for measuring frontier-model capabilities. They criticize approaches based on an industry’s share of GDP or SOC and O*NET taxonomies: in their assessment, the first trillion dollars in AI revenue will be concentrated in fields that can adopt the technology quickly and earn high returns from frontier intelligence. They named chip design, biology, cybersecurity and RSI among GDPx’s focus areas.

A team representative said SemiAnalysis is working on a demand model to project OpenAI and Anthropic revenue in 2027 and 2028, and declined to detail the analysis at this stage. Another team member challenged the assumption that broad public adoption of AI will itself drive large AI spending: in his view, most people will struggle to use AI effectively across many tasks. He also said the cost of AI at a given capability level is falling by 100x to 1,000x a year, driven by progress in research, data and compute. He argued that the greatest demand will therefore come from users able to get rapid returns from large amounts of machine intelligence.

The team believes white-collar work will eventually be automated, but will not drive the first trillion dollars of AI spending: in their assessment, adoption in such industries will be slow because companies will need time to reorganize operating expenses.

A SemiAnalysis representative argued that replacing white-collar workers will not drive the first $1 trillion of AI spending: companies are slow to change job responsibilities, compensation and staffing, so adoption in those areas will take time. He estimated that the cost of a given level of AI capability falls by 100x to 1,000x a year, and said the fastest scaling will be in fields where additional intelligence can generate a quick return.

He named chip design, cybersecurity, financial services, life sciences, materials science and software engineering. In science and engineering, he said, much remains beyond the human frontier; in finance and cybersecurity, competitive rivalry sustains demand for intelligence. He predicted that OpenAI, Anthropic and other leading labs could approach $1 trillion in revenue within a few years. If fields requiring frontier models do not generate that demand, he said, the market could quickly saturate existing tasks while white-collar replacement diffuses too slowly—potentially making it harder for AI labs to meet their compute commitments.

In the interview, a SemiAnalysis representative pushed back on the idea that AI investment in materials science would quickly run out of ideas and be constrained by the capital available to test them. He said science involves both discovery and engineering: most spending goes toward the engineering needed to bring discoveries to market. As an example, he described Periodic Labs as focusing mainly on increasing the pace of materials engineering.

Small improvements can still carry substantial economic value, he argued: a 5% increase in battery capacity, a more efficient semiconductor deposition process or better manufacturing yields. He said yields for some CPU substrates are around 40–60%, and invited listeners to consider the value of raising them to 95%. He also said Intel generates about 100 terabytes of data for each wafer but does not use most of it. In his view, AI, simulation and data analysis could accelerate engineering loops and improve yields; this was a prospective opportunity, not a report of an achieved result.

As another example, he said Nvidia takes 2.5 years to go from design concept to shipping a chip in volume production, and suggested that faster simulations and engineering could potentially cut that to a year. That was a hypothetical scenario, not a claim that the production cycle has already changed.

Interview participants described an evaluation they conducted of models on materials-science tasks. One said Astra solved one subtask better than a published state-of-the-art reference. He characterized model performance as “spiky” across areas and said that in some cases models reach above-superhuman baselines. However, the best Astra materials-science model scored only about 28% on their benchmark.

The speaker said it is difficult to establish a human baseline for a test like this: tasks may require teams of researchers or hardware engineers, and it is hard to estimate a team’s work in hours or days. He specifically noted that the chip-design task would involve many people, not a single engineer. Participants also argued that buyers want models with superhuman capabilities in particular domains, rather than models merely comparable to ordinary people. The conversation did not provide details on the methodology, benchmark composition or independent validation.

Discussing Elon Musk’s comment that Grok Bot is too expensive to run, interview participants suggested that personal AI agents could become economical in about a year. One expected serving costs to fall by more than 100x as AI becomes cheaper. This was the speakers’ forecast, not confirmation that the product has already reached those economics.

A SemiAnalysis representative attributed a possible limit on subsidizing Grok Bot to Musk’s need to finance a buildout of compute. He estimated that Musk had about $80 billion in cash on his balance sheet, of which around $20 billion would go to rockets and $60 billion could be spent on compute; Musk was also raising another $40 billion, the speaker said. The stated target was to build 13 gigawatts of compute. The conversation put the cost of building all of it at more than $400 billion, and later at more than $500 billion. The analyst also cited a gap between selling GPU compute for about $15 an hour and a total cost of ownership of roughly $4 an hour.

In his account, selling available capacity, including to Anthropic and Google, could help fund further capital expenditure, even if that meant temporarily reducing compute available to the xAI team. Another participant contrasted five-year contracts at $12–15 billion per gigawatt with a short-term Anthropic deal priced above $40 billion per gigawatt. These explanations and figures were estimates voiced in the interview, not independently verified financial data.

The guests said that the on-premise and open-model market would continue to grow quickly in absolute terms. But if companies can earn high returns relatively quickly from frontier AI in domains where intelligence is a key bottleneck, they predicted that the frontier-model market would grow faster relative to that segment. One guest suggested that most tokens may be served by non-frontier models, while most spending may go to frontier models.

Discussing a report that Goldman Sachs might ban or restrict Claude Code, Codex and other coding agents internally, the participants did not confirm that a decision had been made. One guest described a possible restriction for junior analysts as potentially short-sighted, but predicted that Goldman Sachs would still need to spend heavily on penetration testing.

As context, a participant cited his team’s finance benchmark: according to him, none of the models tested could successfully complete an analyst task, and the closest model received about 50% of the credit. He argued that open models would perform worse than frontier models. These were the participants’ assessment and benchmark results, not an independently verified model comparison or an established fact about Goldman Sachs policy.

CrowdStrike is acquiring two AI security companies, Fiddler AI and Dreadnode, for a combined $495 million. The hosts also noted that CrowdStrike CEO George Kurtz had posted about creating an AI “emergency button.” One host called the idea effective positioning and branding amid public calls for an AI shutdown mechanism; this was the host’s characterization, not confirmation of an operational feature.

A host said OpenAI had hired Dan Driscoll, a former U.S. Army secretary and a past TBPN guest. The hosts recalled interviewing him and said his appearance was an unusual moment for a show previously known mainly for lighthearted conversations with founders and investors. The segment provided no further details about Driscoll’s role or responsibilities at OpenAI.

A host said SpaceX had introduced Starfall, a project intended to provide affordable, routine access to a microgravity environment. Potential applications listed included biomedical research; developing biomaterials, medical implants, stem-cell organoids and replacement tissues; pharmaceutical research and growing more uniform protein and drug crystals; and materials science and manufacturing advanced materials and semiconductors using space’s vacuum and microgravity. The hosts discussed possible competition with Varda, but also suggested the companies could complement each other: it is unclear whether SpaceX will provide only launch and return capacity, while Varda has developed a research capsule and partnerships with biotechnology companies. That was the hosts’ analysis and speculation, not a confirmed agreement.

A host said Shopify had spent $99 million on the shop.com domain, calling it a good deal and suggesting the company likely has big plans for the address. The latter was the host’s speculation; Shopify’s specific plans were not disclosed in the segment.