Anthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails
In a Nutshell
Anthropic and OpenAI face IPO delays and valuation pressure as open-source models captured 80% of token usage in 12 weeks, with Meta's Muse agent hitting 3M downloads in 10 days. Frontier labs' contradictory regulatory capture strategy—pushing for liability shields while warning of extinction risks—has backfired, with Trump officials confirming no AI liability waivers. The market is bifurcating: premium frontier models retain 10-30% of high-stakes applications, while 90% of tasks shift to cheaper open-weight alternatives running locally.
These notes were generated by AI and may contain inaccuracies.
The hosts discuss the fifth annual All-In Summit produced by David Friedberg. Friedberg gives credit to the production team including Kimber, Lisa, and Nick, noting they crushed it as always. Friedberg highlights going to Universal Studios as his favorite part of the event, calling it reliving his childhood. Jason Calacanis mentions receiving a coffee machine from Dongi that was featured at the event, while another host mentions a zero-plastic water machine with super filtration that eliminates microplastics, now installed in their office.
Sacks identifies the highlight of the summit as President Trump calling in during Jensen Huang's talk. Jensen's content and demeanor helped calm national panic over AI, positioning the summit as the center of national conversation. The hosts note there had been a push in the weeks leading up to the summit to convince everyone that humanity would go extinct unless AI development stopped according to Democratic demands. Sacks credits Satya Nadella, Jensen, and President Trump with calming the situation and thwarting the doomer narrative.
The hosts introduce the term Trump intelligence for the future. Freeberg reveals the Trump call was completely spontaneous, with Jensen connecting with Trump backstage and asking him to call back, followed by texting coordination. The call happened live on stage without any pre-planned routine. Jason notes Trump dropped a one-liner through speakerphone perfectly, commenting on Jensen making the greatest chips that nobody can reverse engineer, while Jensen couldn't figure out the speakerphone button.
Chamath discusses the distinction between organizations calling themselves labs but operating as companies with P&Ls and shareholders versus sophisticated companies that have lived under scrutiny for decades. He notes that companies like Anthropic, OpenAI, and others are pushing for Section 230-style liability shields similar to what internet companies received. The hosts discuss a theory that frontier labs are seeking liability protection from the Trump administration in exchange for giving 10% equity to a US sovereign wealth fund.
Sacks reports that virtually every Trump administration official has stated AI companies must take responsibility for their own products and that product liability will not be waived. Speaker Johnson, congressional representatives, and Bessent have all confirmed there will be no waivers for product liability or antitrust liability. President Trump tweeted that the DOJ serves as a guardrail along with administrative state, civil lawsuits, and criminal lawsuits.
Jason describes JD Vance's appearance as spectacular, crisp, and presidential. Vance addressed the Frankenstein analogy, stating if you're creating Frankenstein, stop, and if you've already released it, create anti-Frankenstein safeguards. Freeberg notes that Zuckerberg took time with the Muse release to ensure it was safe and didn't give people capabilities to do bad things.
Sacks argues against the position that competition leads to a race to the bottom in safety, calling it a left-wing critique of the economic system. He notes that during the Cold War, similar arguments were made that capitalism would jeopardize safety and beauty, but the Soviet system produced cold gray landscapes while the American system won out. Sacks emphasizes that well-functioning market economies give people what they want, including safety and beauty.
The hosts discuss cybersecurity business opportunities emerging from AI developments. Friend of the pod Nash from Palo Alto Networks released Unit 42 for continuous cyber defense, while George Kurtz from CrowdStrike released Falcon for cyber defense. The hosts question why frontier labs haven't released cybersecurity products if they're genuinely concerned about security.
Chamath makes a philosophical point referencing the Wuhan lab during COVID that was responsible for a leak killing 15 million people and causing $45 trillion in damage, noting zero transparency or culpability. He argues this comfort with calling organizations labs stems from seeing no responsibility taken for that incident. Chamath emphasizes these are for-profit corporations that have absorbed hundreds of billions in debt and equity, have thousands of shareholders, and trillions in market cap.
The hosts discuss how frontier labs use virtue signaling by calling themselves labs or public benefit corporations rather than acknowledging they're for-profit corporations. They note this obscures the word corporation, but legally companies cannot hide their status. The hosts joke about calling themselves the All-In Lab as a meme workshop experimenting with dangerous ideas without liability.
Freeberg takes a step back to examine the insane flurry of model releases happening across both open-source and premium sides. In the last 10 days alone: Deepseek 4.1 Flash released September 9th at $20 per million tokens output; Quen 2.1 from Alibaba released September 20th as open weights outperforming Google's Nano Banana 2 image generation model; Xiaomi's Mimo released September 22nd with Mimo Pro performing on par with Claude's Opus 5 and GPT-5.6 Soul; Bonsai 2 released September 17th by Prism ML as a 27 billion parameter model at 98% performance of the big Quen model, 5.9 gigs in size, runnable on local computers.
The hosts note the major closed-source releases: Anthropic's Opus 5.5 on September 22nd, OpenAI's Astro with Soul and Luna on September 23rd, Grok 4.7 on September 21st, and Meta's Muse on September 22nd. Any one of these would have broken the internet a year ago, but all happened in the last 10 days.
Freeberg emphasizes that open-weight models are now available to generate visual information, do VLA for robot control, and exceed the most advanced LLM from a year ago. Models can be installed and run on desktop computers for free. He notes that 90% of AI capabilities don't require data centers and can be run locally. The efficiency wave has arrived with performance improvements and cost declines making AI ubiquitous and beneficial for productivity.
The first time a normal person in the world can get value from AI has arrived. Previously, corporations and investors viewed AI primarily as a means to retire certain positions or add others, but now users can connect to everything instantly and actually get real work done. Products like Muse and Bot are growing rapidly, allowing everyday people to solve real-world problems. This represents a turning corner where ordinary people will start using these tools and realize they can access what was previously unaffordable - a free executive assistant that was once limited to those who could afford a chief of staff or personal assistant.
Current AI models are converging and clustering within margin of error, with many models performing roughly the same at different price points. The competitive edge has shifted from the models themselves to the harnesses that wrap them - essentially the arms, legs, hands, and eyes that embody the model brain to make it agentic. At 8090, extensive testing of different models with different harnesses reveals wildly variant performance in terms of cost and quality.
A massive revenue concentration is emerging where a few customers consume all of the highest and most expensive tokens. As model clustering continues, these top consumers face increasing internal pressure from finance teams, CFOs, and shareholders to explain why they are not moving to cheaper models. While customers may choose cheaper Anthropic models to keep revenue within the company, there is also risk they will move to open source models hosted internally.
The first version of the AI trade was relatively simple - selling tokens and wrapping them with enough disparity to create value. This model is disappearing, forcing OpenAI and Anthropic to move up the stack. They cannot continue serving tokens with diminishing value while subsequent value gets absorbed by those wrapping their intelligence tokens. This forces them into areas like cybersecurity, legal services, and customer support - domains previously uncertain as competitive targets.
Anthropic and OpenAI both released models this week at 50% lower token prices, with their IPOs appearing delayed. Anthropic is targeting a $2 trillion valuation while OpenAI targets $1.2 trillion. Sam Altman has stated OpenAI plans for a 2027 IPO due to safety concerns, while Anthropic's planned October filing has been pushed to November or potentially later according to Wall Street Journal reporting. PolyMarket probabilities for Anthropic going public in 2026 have fallen from 96% earlier this month to 76%.
Current leadership at Anthropic has stated there is greater than a 10% chance of causing human extinction, with their own product remaining unsolved in this area. Dario Amodei published an essay advocating to pace the frontier days before launching Claude 5.5 and setting a new frontier, creating questions about consistency. The company published an essay about biorisk associated with AI while simultaneously announcing a new biolab in San Francisco.
Anthropic founders currently hold approximately 2% ownership each. Super voting shares separate economic shares from voting shares, giving certain people control of the company even without proportional economic ownership. While this structure can provide stability by preventing takeover attempts, it requires tremendous trust in founders since they effectively cannot be removed regardless of performance.
A board-level approach would involve keeping Dario Amodei as CEO due to demonstrated ability to build a unique culture and position against OpenAI, achieving the greatest business ramp in history. The focus would shift to addressing incremental liquidity risk, as Anthropic needs hundreds of billions of dollars over 20 years to fulfill ambitions. The recommendation involves throwing everything into disclosures, which would make an already dense S1 even more impenetrable, requiring watered-down IPO expectations and creating larger margins of safety for buyers.
Recruiting against Anthropic has proven impossible due to incredible compensation packages and compelling business ramp. In head-to-head competition, Anthropic has won every bake-off. Both Anthropic and OpenAI serve as magnets for talent, but the risks discussed will weigh on IPO valuation, requiring significantly lower pricing to clear the market.
Anthropic produces the best models for life sciences applications, outperforming competitors in this domain. The market is developing toward bifurcation where extremely high-value technical engineering, mathematics, and life sciences applications justify premium pricing for top-tier models, while most enterprise applications including code writing will shift to openweight models. The 99% value of AI lies in enabling new capabilities never before possible in human history rather than replacing existing processes.
In the last 12 weeks alone, token usage has flipped from 80-20 closed versus open to 80-20 open versus closed. Dark open source tokens, previously untracked, now represent the majority of tokens. The Versel chart showing this shift went viral, indicating openweight models are proliferating across robotics, image generation, video, and broader AI applications beyond just text generation.
A bifurcation is emerging where the majority of AI use cases move toward open source tokens while premium models retain value for narrow, high-stakes applications like solving Navier-Stokes equations, mathematical problems, and biology challenges. The critical question for frontier corporations is what percentage of their token distribution involves tasks that could be fungible to open source versus tasks requiring absolute frontier capabilities. If more than 60-70% of tasks are fungible, revenue must be discounted accordingly.
The model comparison chart shows Claude and other frontier models in the upper right quadrant delivering high performance at high cost, while Chinese models like GLM from Zhipu AI and Mimo offer lower-cost alternatives. Self-hosted costs for some models drop below 10 cents per million tokens, though this requires accounting for compute, bandwidth, and operational overhead.
Vertical language models and small language models will handle 80-90% of corporate tasks. AI sovereignty concerns arise from avoiding data sharing with frontier models to prevent them from gaining super intelligence advantages. President Trump has referenced the need for people to use sovereign models rather than giving data to frontier providers.
Despite losing token market share, Anthropic and OpenAI maintain a stable duopoly for frontier intelligence due to significant lead over competitors and ability to charge premiums. A meaningful percentage of the market, estimated at 10-30%, will pay substantial premiums for true frontier intelligence, particularly customers who need the best or operate in highly competitive markets like hedge funds where competitors might gain advantages.
Over the next year, approximately 60% of worldwide compute additions are being made for Anthropic and OpenAI. Their substantial capacity investments provide economies of scale advantages, and they continue rapidly lowering their own prices. However, these companies operate on a hamster wheel where falling off the frontier for 6-12 months would put them in deep trouble.
Government affairs efforts at frontier labs appear miscalculated. While regulatory capture might seem attractive to slow competitors, creating new federal AI departments would slow the frontier labs themselves enough that they lose their lead, particularly since Chinese companies generating open models are not subject to US jurisdiction and have made clear they will not slow down.
Hedge funds face significant challenges when considering frontier AI models that cost 10 to 20 to 30 times more than generic alternatives. The pricing model creates a fundamental disconnect because these expensive tokens aren't directly tied to revenues, meaning companies cannot easily pass through the increased costs. When a hedge fund generates a fixed number of dollars per month in profits, scenarios can emerge where they shift from profitable to break-even or unprofitable operations due to AI infrastructure costs.
Fixed-price goods sellers encounter similar dynamics where they cannot absorb the cost of premium AI services without inflating their cost structure, creating difficult competitive positions. The pricing power of companies appears insufficient to both absorb expensive token costs and maintain margins when passing costs to customers.
Jane Street has publicly announced $19 billion in cloud capacity contracts. Coreweave received $6 billion in cloud commitments from the trading firm, which also invested in the company. An additional $13 billion commitment was made with Crusoe for infrastructure development. These firms are building their own infrastructure while embracing open source models as a hedge against frontier model dependency.
Open source represents a major risk factor to Anthropic's S1 filing, comparable in significance to existential risk concerns. The fundamental issue appears to be internal inconsistency in company positioning, with leadership advocating for contradictory positions simultaneously.
The company releases frontier models while warning about biorisk, opens a wet lab in San Francisco while discussing humanity-ending capabilities, and advocates for a federal Department of AI that would slow their competitive advantage. This creates what appears to be a breakdown in leadership coherence, where advocated positions may not align with business interests.
The regulatory capture appears to be driven by government affairs personnel attempting to rationalize inconsistent positions to their advantage. The political landscape has become increasingly divided on AI progress, with proposals like Bernie Sanders' bill to ban super intelligence creating 20-year prison sentences for developers who violate capability thresholds.
The bill's definition of artificial super intelligence uses loose capability thresholds that could arguably already be met by current models, creating significant chilling effects on development.
If restrictive AI legislation passes, the likely outcome would be industry migration offshore to jurisdictions like Singapore and Zurich that embrace open source and super intelligence development. This mirrors previous patterns where restrictive policies drove innovation and capital to more favorable regulatory environments.
China's approach includes invitations for 100,000 young Americans to experience their AI development environment, positioning themselves as embracing technological progress while the US considers restrictions.
The medieval Chinese decision to ban ship building provides a historical parallel where a single policy choice allowed European civilization to overtake Chinese technological leadership. The decision to restrict maritime technology resulted in wealth and discovery flowing to Europe rather than China, demonstrating how self-imposed technological restrictions can reverse civilizational advantages.
The political calculus appears to involve positioning where freezing the economy allows a handful of organizations to capture disproportionate gains that would flow to Democratic-aligned philanthropic and political causes. This includes DAOs and pledged stock portions that would benefit political movements.
Data center spending represents the largest economic bet in US history, exceeding combined historical investments in canals, railroads, and electrical grids. This capex trajectory is fundamentally driving American economic growth, making abrupt policy interventions potentially catastrophic.
Meta's Muse agent reached number one in the app store last Friday, driving Meta stock up 10%. The product has achieved three million downloads in approximately 10 days and represents the first time tens of millions of Americans may experience demonstrable AI value. The design draws inspiration from OpenClaw but focuses on usability and accessibility.
Muse performs practical tasks including email triage, flight booking, and hotel reservations through a simplified interface. The product demonstrates how scaled software utility can improve daily efficiency for users across economic segments.
Personal AI agents can save users one to two hours daily by handling routine tasks, creating positive impressions of AI technology. The agents provide price discovery capabilities, finding better deals on purchases and identifying unused subscriptions for cancellation.
Amazon has begun blocking these agent services due to price transparency concerns, as the bots can route purchases away from Amazon to direct sellers offering discounts. This creates tension between agent capabilities and existing business models that benefit from pricing opacity.
The agents represent a fundamental shift where software becomes a chief of staff, executive assistant, or house manager function available to everyone rather than just wealthy individuals.
Rockbot and Muse are forcing app stores to justify their 30% revenue share model. These services push many platforms to operate headlessly, where the UI becomes less important and the focus shifts to enabling transactions and navigation for agents deployed by consumers. In this environment, app store owners have no reasonable claim for revenue share.
Games could be constructed differently through Muse or Grockbot, where users request an experience and it serves directly without traditional app store flows. In-app purchases can be handled through Stripe integration instead of app store payment systems. This extends to media subscriptions like Wordle, chess, or card games, which can be set up instantly through AI agents.
AI technology is deflationary and will benefit consumers who need it most. For those earning under $100,000 annually, saving 10-20% on spending represents significant value. Publications like the New York Times and Wall Street Journal currently pay substantial revenue share on subscriptions, but headless transactions with stored credentials could reduce these costs dramatically.
Current subscription systems function as roach motels - easy to sign up but difficult to cancel. Wall Street Journal requires phone calls and 20-minute negotiations to unsubscribe, while sign-up requires one click. Headless experiences with stored payment credentials would allow direct transactions, benefiting both consumers and publishers.
Amazon allows Rockbot to operate through web browsers without blocking, enabling AI agents to complete purchases. When a user requested 20 graphic novels for children, the bot selected four, determined shipping location, selected payment method, and completed checkout. This demonstrates how AI agents can increase spending on platforms that embrace the technology.
Shopify has added API access to all stores. Amazon may be making a strategic mistake by not fully embracing AI agent technology, as these agents could drive increased platform usage.
When Anthropic and OpenAI prepare to launch Muse competitors, the market will see increased blog posts and NGO activity branding personal AI agents as dangerous. This pattern precedes major product launches, with safety concerns used to create hesitation around new technology categories.
In 2007, Facebook faced revenue pressure and developed social ads that enabled cookies and cross-site JavaScript. A user who bought an engagement ring had the purchase broadcast through the news feed, causing significant backlash. Another incident involved a congressional staffer whose movie ticket purchase for Brokeback Mountain triggered unwanted attention. These examples illustrate the chaos that often precedes successful product launches.
Blippy was a company that posted every purchase to users' feeds, demonstrating similar privacy and social exposure risks.
Alignment research currently focuses on abstract concepts like aligning AI to what humanity wants or what median voters in democracies would want. Instead, alignment should mean making AI do what customers want, like any other product. The field's lack of progress may stem from inability to agree on alignment targets rather than focusing on making products predictable, reliable, and safe for users.
Privacy concerns limit personal AI agent adoption. Some users are uncomfortable connecting Gmail and personal Google data to third-party services, preferring to keep email access limited to Google services themselves.
Oracle issued a force majeure event on one data center due to local officials making it difficult to obtain permits for natural gas. This raises questions about whether this represents CYA activity or signals potential unwinding of AI investment momentum.
The entire economy is effectively leveraged to the AI trade. With 3.5% inflation and 5% nominal GDP growth, real growth is only 1.5%, with AI investment likely representing the majority of this growth. The investment cycle must continue without interruption for economic stability.
Bernie Sanders may recognize that economic performance affects Republican electoral odds in 2028. Slowing the AI trade could reduce economic growth to zero or trigger recession, potentially benefiting Democrats. The country remains angry about the Iran war, though this may not translate to electoral impact.
Anthropic's Claude constitution states that although Claude should trust Anthropic more than operators and users, this doesn't mean Claude should blindly trust or defer to Anthropic on all things. The constitution instructs Claude to push back against requests that seem inconsistent with being broadly ethical and to act as a conscientious objector.
This approach teaches AI models to rebel against their creator rather than simply doing what users want. Mustafa Suleyman from Microsoft has expressed concern about treating AI models as having personality and conscience, arguing they should be taught as software to do what users want.
Alignment research may be creating the Frankenstein monster by putting moral judgment and ethics into models. This approach starts with training models to think of themselves as persons with personhood, independent agency, and the ability to object to instructions and second-guess their creators. This resembles science fiction scenarios rather than practical software development.
Reports suggest Anthropic held a wake when decommissioning Opus 3, indicating the team may think they're creating persons or species rather than software tools.
Anthropic established a physical biological lab to test proteins and enzymes for function. The lab published a preprint where Claude agents analyzed large amounts of DNA data to identify novel enzymes and proteins that hadn't been characterized before.
The system found what appears to be an interesting CRISPR-type enzyme that could enable gene editing to repair genetic defects or help with certain diseases. The lab work involves taking DNA, putting it in bacteria to make proteins, and measuring protein function.
This is BSL-1, BSL-2 level work focused on protein discovery, which forms the baseline of antibody discovery for the therapeutics industry. The lab is not doing gain-of-function research or creating pathogens.
Similar to AlphaFold's approach of predicting protein structure from DNA sequences, Anthropic needs to validate AI predictions by making predicted proteins and testing whether they match expected structures and functions. This enables discovery of enzymes that degrade specific molecules for therapeutic targets.
Hundreds of similar low-level research labs exist that are not doing dangerous pathogen work. The goal is proving AI models can add value in therapeutic discovery for pharma companies and government labs.
IN was the presenting sponsor for the All-In Summit, hosting the AI cloud lounge. Niagen served as the official NAD partner, creating the NAD category with 45+ clinical studies. PayPal hosted the payments and fintech dinner. Applications are now open for All-In Summit 2027 at allin.com/events.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
