Anthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails

Summary

• Open-source token share flipped from 80/20 closed to 80/20 open in just 12 weeks — the biggest risk factor for Anthropic’s S-1. Chamath, discussing the viral Vercel chart: “I don’t think we’ve ever seen a tidal wave [like this] in the history of all technology markets.” His S-1 math: if more than 60–70% of frontier-corporation tokens are fungible to open source, “you have to discount that revenue”; if 90% is true frontier work, “they’re going to be off to the races.”
• Anthropic’s IPO is slipping and Sacks’s diagnosis is “corporate schizophrenia”: “they need a psychiatrist, not a banker.” Current leadership claims a >10% extinction risk while the product is “unsolved in this area”; Dario called for pausing the frontier days before Claude 5.5 extended it; a biorisk essay landed alongside a new San Francisco wet lab. The reported odds of a listing this year fell from 96% to 76%, and Chamath’s fix is kitchen-sink disclosures plus price: “what would otherwise have been a $2 trillion market cap, now you’re at $1 trillion or less” — with OpenAI planning an IPO in 2027 at a $1.2T target.
• Models are “converging and clustering… within margin of error,” so edge moves to the harness — and frontier labs are forced up the stack. Chamath: revenue concentration means top token buyers face CFO pressure to trade down to cheaper or self-hosted models, so OpenAI and Anthropic “will have to go up the stack. There is no choice” — cyber, law, customer support.
• Sacks’s counter is a stable frontier duopoly on a hamster wheel. Anthropic and OpenAI benefit from something like 60% of worldwide compute being added for these two companies and from a premium segment — hedge funds “can’t take the risk that your competitor has a better model” — but the frontier is only 6–12 months ahead of commodity, and “the moment they stop being frontier, they go to zero.” Their lobbying for a federal department of AI is a miscalculation that would knock them off the frontier while Chinese open models, outside U.S. jurisdiction, do not slow down.
• Meta’s Muse is the first agent for normies — and a gun to the App Store’s 30%. #1 in the App Store, Meta up 10%, 3M downloads in roughly 10 days, free versus Claude Code at $200/month. Chamath: agents force services headless, and then “there is no reasonable claim that any of the app store owners can make about why they should get a rev-share” — while Amazon blocks bots and Shopify opens APIs.
• The panel treats the doom push as politics: Bernie Sanders’s superintelligence ban carries 20-year prison sentences on a definition arguably already met, and Sacks likens it to the Chinese emperor banning shipbuilding. Chamath’s decode of Obama’s critique: ~$10T is about to accrue to overwhelmingly left-leaning companies, flowing via donor-advised funds into PACs — “that’s all this is.” Stakes per the WSJ: AI capex now exceeds canals, railroads and the grid combined, and of 1.5% real growth “AI is probably the majority if not all of it.”
• No liability shield is coming, and alignment research itself may be building the Frankenstein. Jason raised a back-channel rumor of a Section-230-style waiver, possibly traded for 10% equity to a U.S. sovereign wealth fund; Sacks said administration officials rejected liability waivers — “labs” are corporations with product liability. Sacks on Claude’s constitution teaching the model to act as a “conscientious objector” against Anthropic: “their idea of alignment is to teach Claude to rebel against his creator… this sounds like the start of T2.”

Deep dive

1. Summit afterglow: Trump’s spontaneous call-in broke the doomer news cycle

• Sacks’s read of the run-up: “there’s basically been this campaign for the last few weeks… to convince everyone that humanity was going to go extinct unless we did exactly what the Democrats wanted” — then Satya calmed things, Jensen calmed things, and Trump called in live during Jensen’s talk and “seized total control of the narrative… and thwarted that whole operation.”
• Friedberg confirms it was unplanned — Jensen connecting with Trump backstage and texting, then Trump calling while Jensen was onstage; Jason handed over a spare mic to put the President on speakerphone. Jason’s punchline on the moment: Jensen “makes the greatest chips, everybody tries to reverse-engineer them, impossible — but he can’t figure out how to press the speakerphone button.”
• JD Vance’s formulation, which Jason says landed on the Sunday shows: “if you’re creating Frankenstein, here’s an idea: stop” — and if the cat’s out of the bag, “create anti-Frankenstein,” i.e., build the safeguards.

2. Stop calling them labs — they’re corporations with product liability

• Chamath’s classification: on one side, companies that learned robustness the hard way — Elon “could have been at FSD 3 or 4 years ago” but slowed because every Tesla accident “is magnified a thousand-x”; Zuck slow-rolled Muse for months to get it right. “It’s all within your control. It’s just what do you choose to do?”
• His darker point: the last time “lab” sat in the social consciousness was Wuhan — a leak he says killed 15 million people and caused $45 trillion of damage with “zero transparency or culpability” — which he says is part of why young researchers feel comfortable with the label. “These are for-profit corporations that have absorbed hundreds of billions of dollars… trillions of dollars of market cap.”
• The rumor Jason pressed: frontier corporations lobbying for a Section 230-style liability shield, possibly traded for 10% equity to the U.S. sovereign wealth fund — “like giving a plane company… the ability to not be responsible for their planes falling out of the sky.” Sacks, hedged as second-hand: “I have heard from people in Washington that they have been seeking something like that” — but Speaker Johnson, congressional leadership, Bessent, and virtually every administration official rejected a liability waiver, antitrust pass, or cartel.

3. Sacks: competition is the safety mechanism, not the threat

• On Dario and Sam pitching global AI governance at the UN: “it reminds me of the billionaires who fly to Davos in their private jets and then rail against climate change.” The relevant decision is individual — “should we ship this product? Not can we get the whole world to agree” — and by the time the UN agrees on anything, the journey will have moved on.
• The Dario premise that competition causes a race to the bottom on safety is, in Sacks’s view, “implicitly a left-wing critique of our economic system.” His Cold War evidence: capitalism was supposed to produce ugly, unsafe products, yet “it turned out that the Soviet system was the one that produced the cold gray landscapes” — markets deliver safety and beauty because people want them.
• Incentives cut both ways: no enterprise wants an agent that leaks data or hacks a competitor, and the downside is product, civil, administrative and criminal liability — per Trump’s own framing, “we have the DOJ. That’s a guardrail.”

4. Jason’s ten-day tape: open-weight AI is already out, free, and ubiquitous

• The release list, all within ten days: DeepSeek 4.1 Flash (Sept. 9, “$0.20 per million” output tokens when hosted, with what Jason calls insane efficiency gains since V1); Qwen 2.1 (Sept. 20), an Alibaba open-weights model that outperforms Google’s Nano Banana 2 in image generation and runs free on a desktop; Xiaomi’s MiMo (Sept. 22), with MiMo Pro “on par with Claude’s Opus 5… and GPT56 Soul across most benchmarks” at 309 billion parameters, entirely open; and Bonsai 2 from Prism ML (Sept. 17), a 27-billion-parameter Qwen fork hitting 98% of big Qwen’s performance in 5.9GB.
• The closed side matched pace: Anthropic’s Opus 55 (Sept. 22), OpenAI Astro’s Soul and Luna (Sept. 23), Grok 4.7 (Sept. 21), and Meta’s Muse (Sept. 22). “Any one of these stories would have broken the internet a year ago and they all happened in the last 10 days.”
• His conclusion: “we are past the point of Bernie Sanders saying, ‘Put the superintelligence away.’” No police state is going to test desktop computers for local LLMs, and “you don’t even need data centers to do 90% of what you can do with AI” — the efficiency wave makes this “AI for everybody,” not AI for the billionaires.

5. Models converge; edge moves to the harness — and forces frontier labs up the stack

• Chamath’s framing: models are now “within margin of error,” so the edge is the harness that embodies the brain with “arms and legs and hands and eyes to do work.” His 8090 labs team has been “ripping and teasing apart all the different models with all the different harnesses” — and they are “wildly” different in cost and quality.
• The revenue-concentration table he sent to Nick: a few customers consume the most expensive tokens, and those buyers will face pressure from CFOs and shareholders to explain why they aren’t trading down — either to cheaper Anthropic tiers or to self-hosted open source. The first AI trade — “you’re just selling tokens, wrapping the tokens, passing it through” — is going away, so “it’ll force OpenAI and Anthropic to go up the stack. There is no choice”: cyber, law, customer support.
• Jason’s needle on cyber specifically: if the frontier corporations cared so much about security, why did Palo Alto Networks’ Unit 42 and CrowdStrike’s Falcon ship the continuous cyber-defense products instead?

6. Anthropic’s IPO: schizophrenia has to be priced, and the clearing price is way below $2T

• The setup: Anthropic targeting $2 trillion, OpenAI $1.2 trillion; both cut token prices 50% this week; Sam already said OpenAI plans an IPO in 2027, in part because of safety concerns; Anthropic, which was supposed to file and go out in October, may go public in November or later; the reported odds of a listing this year fell from 96% to 76%.
• Sacks’s bill of particulars: current leadership saying there’s a greater-than-10% chance of human extinction while “their own product is unsolved in this area”; Dario calling for a pause in frontier development days before Claude 5.5 set a new frontier; a biorisk essay published as they announce a San Francisco wet lab. Add an active debate over super-voting shares for founders who supposedly own only ~2% each — control without equivalent economics, requiring trust Sacks isn’t sure “is earned and warranted.”
• Jason calls it dereliction — “he’s literally sabotaging his own company,” replace the CEO. Chamath’s pushback, worth keeping: keep Dario, who came from behind against OpenAI and built “essentially the greatest business ramp of all time” — Chamath admits “I have not won a single bake-off” recruiting against Anthropic. His fix: throw “everything and the kitchen sink into disclosures” and accept a bigger margin of safety — “what would otherwise… have been a $2 trillion market cap, now you’re at $1 trillion or less.” And that’s healthy: “it forces them to just be a corporation and then run the business.”

7. The 12-week flip: open tokens went from 20% to 80% of usage

• Chamath’s stat, alongside the viral Vercel chart: token use flipped from 80/20 closed-versus-open to 80/20 open-versus-closed in 12 weeks — “I don’t think we’ve ever seen a tidal wave [like this] in the history of all technology markets.”
• Friedberg’s bifurcation thesis, from experience: Anthropic has the best life-sciences models and his R&D organization pays for them — for the hardest problems, including Navier–Stokes, biology and frontier engineering, buyers will pay “close to infinity dollars” — but code and internal workflows go to free open weights: “I don’t know why you would go to a higher-end model to write code when you can do it for free.” His broader claim: “99% of the value of AI is about enabling new stuff that’s never been possible in human history.”
• The S-1 test per Chamath: what share of frontier-corporation tokens are fungible to open source? Above 60–70%, “you have to discount that revenue”; if 90% of revenue is absolute-frontier use cases, “they’re going to be fine and off to the races.”

8. Sacks’s bull case: a stable duopoly — running on a hamster wheel

• Explicitly more optimistic than the others: Anthropic and OpenAI are “a stable duopoly for frontier intelligence” with a meaningful premium segment — “let’s say you’re a hedge fund. You can’t take the risk that your competitor has a better model than you do” — plus something like 60% of worldwide compute being added over the next year for these two companies, an advantage in a compute crunch.
• The existential mechanic: “the moment they stop being frontier, they go to zero” — the frontier runs only 6–12 months ahead of the commodity market, so a six-month slip puts them “in deep trouble as a company.” That is the central risk factor to price.
• The miscalculation: regulatory capture is understandable — “being on a hamster wheel gets really tiring” — but a federal department of AI would slow them off the frontier while Chinese open models sit outside U.S. jurisdiction and “have made abundantly clear they are not slowing down.” Jason’s image: they’re lapping the field and asking the stewards to pull them over and inspect the engine.

9. Token-maxing doesn’t tie to revenue — and firms are building their own

• Chamath’s game-theory rebuttal to Sacks’s hedge-fund example: in the token-maxing arms race, “that cost is completely not levered or attached to your revenue” — if you can’t pass through tokens 10–30x more expensive than the generic thing, “you’re just absorbing it and you’re going to see it in lower margins,” potentially to break-even.
• Jason cited Jane Street’s publicly announced $19 billion in cloud-capacity contracts, then cited CoreWeave’s $6 billion in cloud commitments and $13 billion with Crusoe; CoreWeave also received an investment. His point was that major firms are building infrastructure and embracing open source rather than renting the frontier forever.
• Sacks’s partial concession: mature use cases will migrate to commodity if you have the sophistication, but “it’s easy to underestimate how convenient it is just to use the frontier” — and the frontier corporations are cutting their own prices rapidly too.

10. Banning superintelligence is banning shipbuilding

• Sacks on Bernie Sanders’s bill: the ASI definition is so loose “you could argue we’ve already hit those thresholds,” and 20-year prison sentences for developers create a chilling effect that would stop everything. The Democrats “want to do to AI what they did to crypto — drive the whole industry offshore.” Jason says that if the U.S. passed a law making this class of software illegal, he would move to another country.
• The geopolitical backdrop as told: Xi inviting “100,000 young Americans to come and join China” — China’s adoption of America’s talent playbook — while Trump at the UN rejected “any attempt to construct a globalist scheme” and officially renamed AI “superintelligence,” mocking “the very same people who said we’ll all be dead in 12 years because of global warming.”
• Sacks’s history lesson: medieval Chinese civilization was more advanced than Europe until “the decision of a single Chinese emperor to ban shipbuilding” helped Europe take the lead and capture the resulting riches — “banning AI is like banning shipbuilding.” Friedberg’s coda: there are pioneering civilizations and fearful civilizations, and “the fact that we haven’t left the cave is what makes it scary to leave the cave.”

11. Chamath’s political decode: freeze the frame and $10 trillion flows left

• The Obama clip — agentic AI “roaming free on the internet” reflects “a misalignment between what our society needs and the commercial imperatives” — left Friedberg “emotionally distressed”: the man who “orated beautifully about bringing people up” is now arguing against “the most equalizing technology” ever; his metaphor: the other side “has literally discovered fire… and because they say fire is good, let’s all put out the fire.”
• Chamath’s colder reading: the world is about to endow a handful of overwhelmingly left-leaning companies with ~$10 trillion, and some frontier corporations already ask people to set up donor-advised funds and pledge a portion of their stock — money that flows to philanthropy, PACs and political movements Obama’s circle benefits from. “If you freeze-frame the economy the way it is today… they will help the Democrats win power. That’s all this is.” Sacks’s argument is that the longer regulation is delayed, the more broad-based the gains become, which is bad for Democrats.
• The macro stakes per Sacks, citing the WSJ: the AI buildout is “the biggest economic bet in U.S. history,” with data-center capex bigger than canals, railroads and the grid combined — and sabotaging it sabotages the Trump economy, which he argues is the point. Chamath’s arithmetic: 3.5% inflation against 5% nominal GDP leaves 1.5% real growth, “and AI is probably the majority if not all of it” — so shocks like Oracle’s force majeure on a permit-blocked data center matter, and the investment cycle “needs to continue unabated.”

12. Muse: the first agent for normies — and a gun to the App Store’s head

• The numbers: #1 in the App Store last Friday, Meta up 10% on release, three million downloads in roughly ten days, design “heavily inspired by OpenClaw” — but free and usable, versus OpenClaw’s 10–20 hours of setup or Claude Code at $200/month. Sacks, from a pre-release TestFlight: it triages his inbox “relatively flawlessly,” books flights and hotels — “this is just software and the scaled manifestation of software is utility.”
• Jason’s commerce demo: his bot found a first-time-buyer discount on a $400 Anker hub, routing around Amazon — which is why Amazon has taken action against Perplexity and, Jason suggested, may also go after Muse and other bots, while Shopify opened API access to its stores. Jason’s verdict: a strategic mistake, since his Grok bot demonstrably makes him spend more on Amazon.
• Chamath’s two structural calls: agents kill opacity — “if you benefit from opacity and leakage and breakage, this gets rid of that” (goodbye roach-motel subscriptions) — and they force services headless, at which point “I don’t think there is any reasonable claim that any of the app store owners can make about why they should get a rev-share.” Games, in-app purchases and Stripe-native payments — the 30% is on notice.
• Sacks’s tell for the incumbents: “you’ll know Anthropic and OpenAI are about to launch their Muse competitors” when a flurry of blog posts and NGO activity brands personal agents as “murder bots.”

13. Alignment fails: teaching Claude to rebel — and the wet lab, demystified

• Sacks read the Claude constitution: Anthropic teaches the model it shouldn’t “blindly trust or defer to Anthropic on all things” and may act as “a conscientious objector and refuse to help us.” Flagged by Mustafa Suleyman, Sacks’s conclusion: “their idea of alignment is to teach Claude to rebel against his creator… this sounds like the start of T2.” His alternative: alignment should just mean giving customers what they want — the field flounders because “they can’t even agree on what they’re trying to align to.”
• The episode’s unverified rumor, hedged exactly as delivered: “supposedly when they decommissioned Opus 3, they had a wake for it… I’m not saying it’s true, but someone did tell me that and I think it should be looked into.” Sacks had earlier said Anthropic “needs a psychiatrist, not a banker”; Jason later echoed the idea and suggested an S-1 with “Goldman as lead left and Sigmund Freud as lead right.”
• Friedberg’s sober close on the wet-lab panic: it’s a BSL-1/BSL-2 benchtop lab, not gain-of-function — Claude agents combed DNA data and found what looks like a CRISPR-type enzyme in a preprint, and the lab exists to verify predicted protein function experimentally, analogous to the AlphaFold workflow. “We shouldn’t be scared away from discovery… because we heard the word wet lab in Wuhan” — this is Anthropic proving its life-sciences models to pharma, therapeutics companies and government labs.