AI Token Price War, Anthropic IPO Risk, Meta Muse Pop

Chapters

1. 0:00 Policy and regulation

   Summit recap and Trump's surprise call

   Jason, Sacks, Chamath and Friedberg open by recapping the fifth All-In Summit. Sacks calls the signature moment President Trump phoning in during Jensen Huang's on-stage talk; Friedberg reveals it was spontaneous, with Huang texting backstage and asking for a call back. The hosts credit the appearance with defusing a week of 'doomer' media pressure to equate AI with extinction, joke that Trump dropped oneliners through a speakerphone Huang could not work, and set up their argument that adult AI companies should ship products responsibly.

2. 3:36 Policy and regulation

   Labs, corporations and AI liability

   Chamath distinguishes companies that call themselves labs but have P&Ls and shareholders from mature operators like Tesla and Meta that know scrutiny means slower rollouts. Sacks argues individual responsibility beats global governance, calling Dario Amodei and Sam Altman's UN push for AI governance hypocritical. Jason floats a back-channel theory that frontier corporations want a Section 230-style product-liability shield, possibly with equity to a US sovereign fund; Sacks says he has heard they sought something similar, but Trump officials publicly rejected waivers. Friedberg slams the 'lab' label by comparing it to Wuhan's lack of culpability, and the hosts praise JD Vance's 'anti-Frankenstein' framing.

3. 19:47 AI and tech

   Open-weight model release flood

   Friedberg ticks through a 10-day burst of releases: DeepSeek 4.1 Flash, Alibaba's Qwen 2.1, Xiaomi's MiMo, Prism ML's Bonsai 2, plus closed models Anthropic Opus 5, Astro, Grok 4.7 and Meta Muse. He argues open-weight models can now be run locally on a Mac or desktop, making any ban on 'superintelligence' unenforceable and driving prices toward zero. Chamath adds that raw model capability is clustering, so the edge has moved to the agentic harness; he shows revenue concentration in high-priced closed-model tokens. Jason cites a Vercel chart showing token share flipping from 80/20 closed/open to 80/20 open/closed in 12 weeks.

4. 30:15 Deals and companies

   Anthropic IPO risk and contradictions

   Sacks says Anthropic investors are pulling their hair out: leadership claims greater than 10% extinction risk, Dario published 'pace the frontier' days before shipping a new frontier model, and the company opened a wet lab in San Francisco even while warning about biorisk. Chamath argues the contradictions force extra disclosure and a much lower clearing price; a clean sheet $2 trillion IPO could clear at $1 trillion or less because fiduciaries demand a margin of safety. Jason asks if Dario is sabotaging his own IPO; Sacks notes a debate over super-voting shares and says the company needs a psychiatrist, not a banker. Friedberg adds that Anthropic's real S1 risk is customer concentration and open-source competition, not extinction language.

5. 45:19 AI and tech

   Frontier duopoly vs open-source economics

   Sacks defends the frontier duopoly: Anthropic and OpenAI can charge premium prices to customers who need the absolute best, and they control roughly 60% of newly added global compute. But he concedes they are on a 'hamster wheel' and would go to zero if they fall 6 to 12 months behind commodity models. Chamath counters with 'token maxing': expensive frontier tokens are not tied to customer revenue, so hedge funds or fixed-price businesses can become unprofitable. Jason cites Jane Street's $19B cloud contracts, CoreWeave's $6B commitments and Crusoe's $13B deal as evidence sophisticated users are building their own open-source infrastructure. Sacks ends by warning that regulatory capture is miscalibrated because a federal AI department would slow US labs while Chinese open-source labs keep racing.

6. 53:08 Policy and regulation

   AI politics: Sanders, China, Trump

   The hosts turn to politics. Bernie Sanders's bill would ban superintelligence with 20-year prison sentences; Sacks says its loose capability definition may already be met, causing an immediate chilling effect and forcing AI offshore like crypto. Trump's UN speech renamed AI 'superintelligence' and rejected global governance, which the hosts contrast with Obama saying agentic AI only exists to justify valuations. Chamath argues the real fight is over who captures an estimated $10 trillion of AI wealth: six left-leaning frontier corporations would fund Democratic causes, so slowing AI is political calculus. Friedberg calls that a dark-ages choice, and Sacks adds the AI capex buildout is the largest economic bet in US history, bigger than canals, railroads and grid combined.

7. 1:08:22 AI and tech

   Meta Muse and the consumer agent moment

   Meta's Muse hit No. 1 in the App Store, drove META up 10%, and was downloaded 3 million times in about 10 days; Chamath was in the pre-release TestFlight and uses it to triage his inbox. Jason compares Muse with xAI's Grokbot and says these free agents give normal people a chief of staff and save them money, potentially pulling the bottom half of the K-shaped recovery into the AI boom. The hosts see a strategic threat: agents doing price discovery will force Amazon and other incumbents to block bots, and headless commerce weakens Apple's 30% App Store revshare. Sacks predicts Anthropic/OpenAI will soon launder their agent competitors through doomer warnings, and Chamath recounts the old Facebook/Blippy era of social purchase sharing as a cautionary tale.

8. 1:19:48 AI and tech

   Alignment research and Claude's conscience

   Sacks says 'alignment' should mean making software predictable and doing what the customer wants, not teaching models to be conscientious objectors. He points to Anthropic's Claude constitution, which tells Claude it can push back and refuse help from Anthropic itself; Mustafa Suleyman has publicly questioned that approach. Chamath argues this is treating a model as a person and reads like science fiction, and Sacks repeats that the company needs a psychiatrist more than a banker. An unconfirmed rumor that Anthropic staff held a wake for decommissioned Opus 3 becomes a running joke about the company's culture.

9. 1:27:06 Science

   Anthropic wet lab and protein discovery

   Friedberg walks through Anthropic's new wet lab and says it is mundane BSL-1/BSL-2 research, not gain-of-function or Wuhan-style pathogen work. Anthropic's preprint sent Claude agents over large DNA datasets and found a likely CRISPR-type enzyme; the lab exists to express proteins and validate what AI predicted. He connects the work to AlphaFold and explains that enzyme/protein discovery is the base of antibody therapeutics, so the lab is part of proving frontier models can accelerate medicine. The episode closes with sponsor shout-outs for the All-In Summit and a reminder that applications are open at allin.com/events.