How the Latest AI Leap Is Disrupting Business Models

The episode opens by arguing that recent model improvements may represent a new capability jump rather than an incremental advance. The host then turns to Anthropic’s pre-IPO investor day, where the company reportedly told institutional investors that it generated $16.8 billion in Q3 revenue, 45% above Q2. The host interprets this as actual quarterly revenue rather than an annualized run rate, while noting that Anthropic likely counts gross revenue before payments to Amazon and other compute partners. Anthropic also said it was profitable on an adjusted basis, generating positive free cash flow, with 85% adjusted gross margins and 11% adjusted operating margins. Its adjusted figures have typically excluded stock compensation, training costs, and revenue shares. The market reaction illustrated how closely investors are already watching a private AI company. Some analysts called the result a flop despite roughly 50% quarter-over-quarter growth, while others questioned how $17 billion in quarterly revenue could be considered disappointing. Reuters separately reported that Anthropic expects $110 billion in annualized revenue by year-end; its Q3 figure annualizes to about $70 billion, implying another 57% growth surge in Q4. The host says this likely depends on a major upcoming model release, referred to as “Fable 5.5” or “the Wizard.” Anthropic reportedly plans a $2 trillion IPO, and accounting rules may allow it to disclose only older Q1 and Q2 numbers, making the unaudited Q3 figure especially important to the market.

The financial headlines then move to CoreWeave, whose unsecured bonds maturing in 2032 fell to about 80 cents on the dollar. The bonds now imply roughly a 15% interest rate, suggesting that investors see a serious default risk. CoreWeave’s stock fell 15% in one day, compared with a 1.6% decline in the Nasdaq. DA Davidson analyst Gil Luria again called CoreWeave “the subprime tranche of the AI industry” and said the stock could go to zero if data-center financing stalls. His central argument was that borrowing at 9% to earn a 5% return destroys value regardless of how large the company becomes, particularly as data centers outside the hyperscalers become increasingly commoditized. CEO Mike Intrater defended the company in a staff memo, saying markets sometimes misunderstand new technology companies and that CoreWeave remained confident in its business model, customers, execution, and financial discipline. The host notes, however, that Intrater has sold CoreWeave shares every week since April according to SEC filings, while other insiders have also cashed out. CoreWeave characterized the founders’ sales as responsible personal financial planning, but the selling has not helped investor confidence. At the same time, demand for AI infrastructure appears undiminished. ASML, the monopoly supplier of the extreme-ultraviolet lithography tools needed for leading-edge chips, said demand is running far ahead of its ability to increase supply. Customers have reservations extending into 2029 and even 2039, and ASML warned that achieving major additional capacity by 2028 would be extremely difficult even if construction began immediately.

NVIDIA CEO Jensen Huang was repeatedly asked in the United Kingdom about AI safety. In a Bloomberg interview, he was asked whether he would stop supplying chips to an irresponsible AI developer. Huang answered that he would, after first having a conversation, while emphasizing that NVIDIA’s technology is general-purpose and that the company cannot fully govern how other companies use it. During a later press exchange, a journalist summarized Huang’s supposed position as a belief that there was a zero-percent chance of AI extinction and that safety was merely an engineering problem already solved. Huang rejected that characterization. He said safety is “in our control,” but that between the present and the end of humanity there would be a great deal of work to do. In his view, every increase in capability must be accompanied by advances in safety technology; capability and safety are not in conflict but are “one.” The host describes this as an “AI-realist” position: Huang takes risks seriously but believes they are tractable through engineering rather than primarily through political camps or abstract declarations. NVIDIA also announced $250 million in compute over five years for AI-safety work in the United Kingdom. Huang said the commitment would put accelerated computing in the hands of UK researchers to develop, test, and strengthen safeguards for increasingly powerful systems.

The host reports that Bank of America has hired 1,000 contractors to support 3,000 employees in a large effort to repair bugs revealed by “Mythos.” The software patching effort, combined with increased anti-money-laundering compliance, added hundreds of millions of dollars in costs during the third quarter, though the bank expects those expenses to decline over time. CEO Brian Moynihan’s response was that the company had to “get to work to fix it.” A follow-up to the previous episode’s story about an OpenAI employee accused of planting a listening device says the suspect had booked a one-way flight to Saudi Arabia for October 12. Prosecutors had sought a sealed indictment and arrest warrant on October 8 because they believed public disclosure could cause the suspect to flee. He was arrested the next day, narrowly before the planned departure. The host connects the case to broader concerns about espionage in the AI race. A Wall Street Journal report described alleged attempts to obtain information from AI researchers through attractive strangers in bars and people posing as families at playgrounds; OpenAI reportedly devotes part of employee onboarding to recognizing insider threats and intelligence agents. The episode ends the headline section with a lighter example: a fake image of a bear in a Russian village spread through local group chats and prompted real police reports and an animal-control hunt. After the village was found to be bear-free, the woman who made the image was fined 15,000 rubles, approximately $178.

The main segment begins with a developer named Brandon Thomas, who posted that “Software is over” after using Claude Opus to decompile Photoshop and create a clean-room reconstruction. Traditionally, clean-room rebuilding required one engineering group to study and document a product without passing along the original code, while a separate group built an independent implementation. Because that process was expensive and slow, it was rarely practical. Thomas argued that AI has made a version achievable by an individual. His project produced AI-generated bootlegs of much of Adobe’s software, although the result still contained bugs and was not feature-complete. He offered a free version of an enterprise suite that Adobe sells for $100 per month, and Adobe’s stock opened 3% lower after the story spread. Chamath Palihapitiya took the most extreme view on the All-In podcast, saying software and intellectual property had effectively become worthless because headless services, agents, and MCP connections let users combine capabilities without caring who owns the underlying software. He said his own AI SaaS business had stopped negotiating to retain IP for client systems because “the IP is irrelevant.” The host pushes back on this as potentially “podcast hyperbolic.” Software has been pirated for decades without destroying Adobe, and SaaS companies generally sell support, maintenance, reliability, and procurement convenience rather than code alone. He compares the claim to the earlier vibe-coding prediction that companies would replace Salesforce and accounting systems with internally prompted tools; after nine months, many experimenters had returned to commercial contracts. His conclusion is that enterprise SaaS is likely to change but is not disappearing quickly.

The host argues that the video-game industry is more exposed because it is already suffering a severe downturn. Console sales are at a 13-year low, major studios have laid off tens of thousands of workers, and many new releases are selling poorly. AI-assisted software decompilation and reconstruction have therefore arrived at an especially vulnerable moment. Developers have produced mashups such as Super Mario with Elden Ring or Spider-Man with Arkham Asylum, while traditional crackers used AI to accelerate a PS5 jailbreak that could let it run games from other systems. Bloomberg Intelligence believes the combination of streamed games and AI could eventually eliminate consoles as gaming’s main gateway. AI-generated AAA games may be better served by data-center hardware, while semiconductor competition has pushed the base PS5 from $500 at launch to $650. Bloomberg’s view is that consoles will remain for enthusiasts and collectors but become one relatively uneconomic access point among many. More radically, commentators including George Kal and Andrew Curran say models such as Opus 5.5 can now help with code, shaders, system design, level logic, sound, tooling, and debugging. Curran said stable, playable game fusions can be produced in a day and suggested the industry may be close to a “1% game studio”—a solo developer capable of what once required a large team. The proposed end state is not simply cheaper competition on Steam. If a player can ask a model for a game tailored to a preferred genre, aesthetic, characters, and style, studios may lose their role as the default source of games. The remaining moats could be taste and audience. The host says this personalized-game future, associated in part with DeepMind’s Genie work, may be closer than many traditional game companies realize.

Epic Games has made one of the most aggressive moves, banning handwritten code and shifting developers toward an AI-first production model. The company also delayed a Disney shooter featuring Marvel and Star Wars characters so that its art and other production processes could use full AI workflows. CEO Tim Sweeney said the goal was not merely to make the same products with fewer people, but to let teams build more and better work. When an indie developer accused him of failing to read the room, Sweeney replied that much of the room wanted 2016 back, but Epic’s responsibility was to design for the present and near future. Epic’s new vibe-coded Games Launcher did launch faster than Steam, but users reported visual glitches that made its quality control look poor. Xbox took a less immediate but still substantial approach by creating Xbox Labs, an internal group to experiment with AI and other emerging technology. CEO Asha Sharma specifically cited decompiled games running on hardware Microsoft does not sell and said those experiments challenge the link between software and hardware and between creation and play. She argued that making games more accessible, reviving old games, and enabling players to become creators could be good outcomes even when they disrupt existing business models. The episode then gives examples of companies trying to become more valuable in an AI-heavy economy. Ithaca Holdings agreed to take Bumble private for roughly $1 billion, arguing that as intelligence becomes abundant, meaningful relationships may become more valuable and technology can help Bumble reinvent itself. Separately, Berkshire Hathaway invested $330 million in Bending Spoons after already investing $600 million in its July IPO, giving Bending Spoons more acquisition capital for brands such as AOL, Vimeo, Meetup, and Evernote. Finally, David’s Bridal is pivoting from selling dresses to AI-powered wedding planning. Its “Aisle to Algorithm” strategy connects an AI planner to a vendor marketplace using customer data, and the company says a project that once would have required more than 100 people was built by a team of only tens. The host’s final conclusion is cautious: rapid-change predictions can be overstated, but legacy companies increasingly believe that evolving is necessary to survive.