Meta's Muse tops charts; Menlo's AI bubble alarm; Factory at $5B

Summary

• The Anthropic $2 trillion IPO slipping from October to November is prospectus tactics, not a market crack, in Rory O’Driscoll’s read. After finally becoming bigger than OpenAI in Q2 and enduring OpenAI’s “furious” July counter-tweeting, the bankers want a clean quarter in the book: “If we do this in October, it’s going to be a lot of explaining. If we do this in November, the numbers will talk.” Rory says there is a 90% chance waiting is the right decision, with a 10% chance the world turns and “damn, should’ve taken the $100 billion.”
• The frontier-lab capital math is the episode’s most tradeable set of numbers. OpenAI is forecast to grow ARR from $35B to ~$350B over three to four years while burning $278B net (Jason Lemkin bets it’s more — “they may need 400”; burn “always comes in 30 to 50% higher than the model”), against $122B cash — and the real figure is ~$700B of CapEx, much of it on other companies’ balance sheets, including Oracle and NVIDIA. Rory’s summary: “Intelligence is not cheap.”
• Meta’s Muse is “the first real ChatGPT competitor” and a Trojan horse, per Jason, who calls it one of the best pieces of software he’s ever used. It has truly autonomous agents ChatGPT lacks, a “darn good” consumer LLM, and it’s free — “if it’s free, why would an ordinary person pay?” Meta added roughly $100B of market cap in the week, converting the story from “bottomless pit of enterprise AI” to a credible next act.
• Agentic commerce will “maim” the middlemen: Amazon blocked Muse (ads now exceed e-commerce profits; agent baskets shrink), while Shopify partnered. Jason calls it “the last stand of the unnecessary system of record” — his own agent responded to a vendor price hike by drafting a 12-month migration plan unprompted. Even a 5–10% hit to Amazon’s advertising revenue would be meaningful.
• Jev — a classifier rather than a chatbot, introduced by Harry as coming from ChatGPT’s inventor, and Vercel’s fastest launch — unbundles the cheap end of the LLM API market. Rory’s math: ~20% of a $100B (heading to ~$500B) LLM spend is addressable, but at one-fifth the price $20B becomes $4B of revenue for the attacker; Jason says it’s closer to 1/100th the cost. Jason argues OpenAI and Anthropic have deliberately neglected the low end — Mini and Haiku are “the two worst models that exist” in his own evaluations — while Rory says OpenAI will likely copy Jev and Anthropic is focused on “AGI slash God.”
• Menlo’s Venky sounded the bubble alarm and Harry called it “a blowhard piece” from the firm that “paid the highest price on most rounds.” Rory’s defense of the substance: the market splits into house-money winners and terrified catch-up players, with FOMO checks in between, and the “price doesn’t matter, just get the best deals” doctrine fails “if you’re wrong by 10X.” Jason’s kicker on the alternative: “I love the quiet compounders. There’s just no market for them anymore.”
• The mock IC approves Factory at $5B, passes on Lagora at $11B and Crusoe at $30.9B. Factory rides Rory’s core thesis — “Coding is the mother lode. It’s everything” — plus data sovereignty, since enterprises “don’t trust Anthropic and OpenAI with their data” (“Of course they’re gonna train on your data. Give me a break,” says Jason). Lagora/Harvey’s reported -50% gross margins spook Jason; Crusoe is a “spreadsheet investment… just below the fold,” and Rory warns the levered data-center trade is more exposed to a one-year bump than app-layer bets.
• On Keith Rabois vs Airwallex, the panel splits between distaste and realpolitik. Jason smells “a little bit of racism” in the campaign but notes Fin had to rip out all open-weight models to close its multibillion-dollar Salesforce deal — Chinese exposure “is impeding transactions.” Harry, a disclosed Airwallex investor, says the allegations have serially shrunk and “I’ve got the cap table”; Rory wants government, not Twitter, setting the China rules — right now “if Jensen checks in, it’s all good.”

Deep dive

1. Anthropic’s IPO delay: a cleaner story, not a cracking market

• Rory’s programming logic: Anthropic became bigger than OpenAI in Q2, OpenAI “responded furiously in July” with its own numbers campaign, so the bankers want a quarter that reflects the noise. Going out in October means marketing before you can print October numbers; November is clean. “If we do this in October, it’s going to be a lot of explaining. If we do this in November, the numbers will talk.”
• Jason’s counter-sniff: maybe there’s “just a hint of stress on the pre-conversations” for “the most massive IPO of our lifetimes.” Rory concedes the tail risk: a hell-or-high-water CFO would have gone now and eaten a messy story; choosing to wait signals confidence — he says it is “90% certain” this is a good decision, with “always that 10% chance the world goes crazy and you look back and go, ‘Damn, should’ve taken the $100 billion.’”

2. “Who insures rogue agents?” — Rory: bullshit; liability is in the noise

• Harry relayed a billionaire friend’s question — how do frontier labs go public when nobody will write liability insurance for agent swarms? Rory’s answer: “Bullshit comment… Once you’ve said publicly that there’s a 10% risk that your thing can blow up the world, sweating product liability is in the noise.” A $2T company self-insures; it doesn’t need Munich Re at a $200B market cap. Securities law requires disclosure, not zero risk: the S-1 just needs to say “at least half of our crazy employee base thinks this thing is gonna blow up the world.”
• Jason’s operational version: a 200-person in-house legal army fighting suits “forever,” like tech always fought IP trolls — “it is novel, but it’s game on.” Rory salutes Jensen’s unhedged call — “There is a zero probability AI will destroy all of humanity” — and Jason lands the jab: that’s because his open-source LLMs haven’t caught up; “he knows he needs until 2030.”

3. OpenAI’s burn: three numbers, and the third is the story

• Jason on the $278B-by-2030 burn forecast: “I bet it’s more… They may need 400.” Every portfolio company’s burn “no matter what they say, always comes in 30 to 50% higher than the model. Doesn’t matter who we put in as CFO.”
• Rory’s decomposition: the forecast is $35B ARR to ~$350B in three to four years — “almost modest” given Anthropic grew 10X in one year — with $122B cash on hand against the burn. The amazing number is the ~$700B of CapEx required, much of it on other balance sheets, with Oracle and NVIDIA doing the CapEx and leasing it back. “Unlike software, this is an extraordinarily capital-intensive business… Intelligence is not cheap.”

4. Muse is a Trojan horse — and worth $100B to Meta in a week

• Jason’s core claim: “It’s the first real ChatGPT competitor. Anthropic never really cared about Claude until recently.” Muse is “one of the best pieces of software I’ve used ever, but it’s also a Trojan horse to fight ChatGPT” — every agent session doubles as LLM Q&A, it has autonomous agents ChatGPT doesn’t, vastly more tokens, and it’s free: “If it’s free, why would an ordinary person pay?”
• His proof-of-concept: Muse built him an entire single-user CRM tracking 150 sponsors in real time — “some of the first composable software that I’ve ever actually seen work. This has been a myth since this show started.”
• Rory’s re-rating: after weeks of harshness about Meta’s “ill-defined enterprise model,” this play is “spot on, plays into their distribution leverage” — stock up 7–8%, roughly $100B of market cap in a week, and up 34% on the month according to Harry (with talk that beta testers were buying). The venture read-through: OpenAI looks at that $100B and asks “should we buy Instinct? Do we have to do something in this space, like, now?”

5. Amazon blocks, Shopify partners — and the agents will gleefully route around

• Rory says both can be right: Amazon’s ad business now exceeds its e-commerce profits, and Walmart recently saw agentic baskets shrink — so Amazon blocks, betting “if I block them, they’ll probably come to me anyway ‘cause I’m Amazon.” Shopify’s long-tail merchants just want the demand and the payments rail. The meta-lesson: all the Google/OpenAI agent-payment-protocol blathering “never matters” — “someone aggregates consumer demand… and demand creates urgency to sort all this shit out.”
• Jason’s darker take: “the last stand of the unnecessary system of record. Resy’s gonna lose.” Agents are tireless, find broken APIs, and don’t tolerate rent-seeking — when a core vendor announced a massive price increase that morning, his agent “laid out a 12-month plan to migrate off this vendor on its own. They will not tolerate this crap.”
• Convergence on “maiming” rather than killing: Rory invokes the old line that “the internet abhors inefficiency” — AI pounds down information middlemen the same way — and even a 5–10% impact would be meaningful. Jason: blocking Perplexity is easy, “they’re just a gnat”; blocking everybody is a harder board conversation.

6. Jev: thinking fast, priced at a hundredth — and the harness problem

• Harry introduces Jev as a model from ChatGPT’s inventor and Vercel’s fastest-ever launch. Jason’s demystification: “I don’t even think it’s a model. I think it’s a classifier backed by an LLM.” He lost all Saturday re-doing prompts, then found it answers his “should these two people meet” matching question in milliseconds “for a hundredth the price” — but it covers maybe 20% of his app’s LLM calls, outputs no text, and can’t reason. “We effing waste a lot of tokens on simple stuff that Astra shouldn’t be doing.”
• Rory’s TAM math, hashed out at Monday’s partner meeting: ~$100B of LLM API spend today heading toward half a trillion; ~20% is System 1 work, using the “thinking fast” framing from Daniel Kahneman’s Thinking, Fast and Slow. But at one-fifth the price, $20B of displaced spend becomes $4B for the attacker — open-source economics, a “maiming of 10, 20% of the revenue” of the foundation labs. Jason: “I just think it’s a hundredth the cost, but I agree with all of it.”
• Jason’s tapping-out confession on model routing: after switching Replit to AutoRouter across Astra, Fable and open-source models, he saw degradation and switched back. “It’s the renaissance of DevOps… I can’t pick these models anymore. I’m tapping out.” A router that’s wrong 20% of the time on mission-critical coding is unusable.
• Rory predicts OpenAI is “ruthlessly commercial” and will copy Jev in weeks; Anthropic is “trying to build God, and this isn’t God” — Jev’s launch video literally said “Prod, not God.” Jason goes further: in his own evaluations, Mini and Haiku are “the two worst models that exist… they fail a hundred percent of the time.” He argues they are deliberately crippled low-end offerings, while Anthropic and OpenAI are more concerned about open-weight models “that just got a little bit closer earlier than they planned.”

7. Seed is evaporating — go pre-inception

• Harry’s team never sees a first raise under $20M anymore (Jev’s seed was $40M). Jason’s answer to why Andreessen just launched a $40M university: “You gotta go pre-inception if you wanna do seed now. Inception’s too expensive” — Thiel figured this out 18 years ago with the Fellowship and Dorm Room Fund; Harry counters that Z Fellows gets the best people today.
• Rory’s dorky-economics grounding: nominal GDP is ~2.5x since 2010, so a $3–4M 2010 check is a $10M 2026 check “before anything else has changed” — layer on the fact that 2010 was “in the shitter” while tech looks strong in 2026, and “you get to 20 million before you blink.” He rejects Harry’s extrapolation that bigger exits justify bigger checks: today’s outcomes came from yesterday’s smaller checks, and the valuation overlay on top of the secular trend “might not persist.”
• Jason’s seed math under low ownership: it still works, “but man, you need the big outcomes” — a $25B exit post-dilution could make a $3–4M investment a 100x or 50x outcome.

8. The Menlo bubble alarm — blowhard, or two-and-two-makes-four?

• Rory summarizes Venky’s piece approvingly: the market has house-money players, including Menlo after Tropic, and terrifyingly desperate catch-up players, “and between the giddily happy people and the desperate people, there’s a lot of people writing checks with fear in their heart of FOMO” — the kind of thing “that can change on a dime.” Harry’s pushback, unsparing: “I thought that piece was a blowhard piece… you guys have paid up” — “be cognizant of the risks” is content-free. Jason piles in: “Thank you. I wasn’t aware that the AI gold rush might end someday… I was so lost in Jev all weekend, I missed the point.”
• Rory’s rebuttal — worth keeping: Menlo’s winners were priced when uncertainty was high; “the same deal today would probably be three or four X higher,” meaning four times less return for a similar risk profile. And his Barney Baruch quote: “If every day you look in the mirror and say two and two makes four, you probably could avoid a lot of mistakes” — the things that bite you were obvious all along. The Andreessen doctrine that price doesn’t matter “is true — but if you’re wrong by 10X, then it’s not true.”
• Jason on what discipline actually looks like: in 2021 he made exactly one investment, a seed investment that later closed at $2.3B. The alternative — conservative companies growing 60%, worth 3x revenue and hoping Bending Spoons buys them (it looks at 1,000 deals and does two) — has no clear exit market: “I love the quiet compounders. There’s just no market for them anymore.”

9. Triple-triple-double-double vs Sarah Guo: what LPs should actually do

• The fight: Rory likes Gokul’s tweet — “I’ll happily do triple, triple, double, double deals all day long if you’re capital efficient.” Harry calls that “a ludicrous statement”: in a competitive capital market, managers like Sarah Guo posting phenomenal quick IRRs will take the LP dollars, “and your LPs will leave in droves.” His example is the Cursor deal — Jason initially says Instinct, and Harry corrects him to Cursor — which went from $50M pre at seed to roughly $350M pre from Kleiner to $2.5B, reportedly from Index and Benchmark, and perhaps $10B, all within about four months. The risk in those rounds is not remotely the same; Jason says some figures are from memory.
• Jason splits the difference with data: even Gokul’s model needs north of 30% IRR, and “compounding 90% IRR over a decade… it’s mathematically impossible. The low 30s is as good as it gets in the real world.” He recalls top-quartile LPs hitting 90% IRR in 2021; it did not last; he bets top LPs will reach 90% IRR in 2026. The real answer: “you should do every great deal you see if you have enough capital” — both breakout deals and compounders.
• Rory’s LP prescription: back the best seed/A managers in the newest spaces “all day every day,” and don’t screen on ownership — a seed investor who consistently shows up in the best deals with small ownership earns bigger ownership over time. The caution belongs later in the stack: the seed deal at $50M pre has an effectively infinite margin of safety; the same name at $10B four months later, “maybe you get a 1x.” Growth was an amazing place to play in 2023–25 and “might be more like ‘21 now.”
• His closing asymmetry: return persistence is near-zero in mutual funds — chasing returns there is “a total fool’s errand” — but “in venture, persistence is quite high,” so partially chasing returns in private markets is rational. Jason’s caveat: that very screening is why LPs “won’t get into the next one.”

10. Factory at $5B: approved — sovereignty plus the mother lode

• Jason’s yes, built on his Dreamforce takeaway: C-level executives “don’t trust Anthropic and OpenAI with their data. This is not something manufactured on X.” His own view, for the record: “If I upload my confidential data, I’m not sure it’s not going to my competitors through an LLM… They’ve trained on all of our data. Every YouTube, every piece of open source, every piece of closed source. Of course they’re gonna train on your data. Give me a break.” His Adobe scar tissue: source-code pollution was the one code red, and his engineering team threatened to quit to get GitHub in.
• Rory’s concurrence, and the episode’s thesis line: “Coding is the mother lode. It’s everything — 10X everything else in terms of value being created from AI today. You just can’t have too many bets on coding.” With Cursor “swooped off the table,” only Factory and Cognition remain, in Rory’s view, to sell the enterprise what it loves — “I’m coming to you, Mr. Corporate Customer, and I will make this go away.” Even his bubble caution bends here: lean into trends that compound through a slowdown, and this is one.

11. Two passes: Lagora at $11B, Crusoe at $30.9B

• On Lagora, which Harry says has just announced $200M ARR: Jason won’t lead at $11B because The Information reported Harvey’s gross margins at -50% — “if the margins are spiraling down rather than V-shaping back, that would make me a hint nervous… you gotta raise a lot of money if it’s really true.” Rory offers the contrarian spin — “the correct spin is ‘My God, lawyers are pounding on our shit, and as soon as we get our own internal models, they’re never going back’” — but when Harry mocks his hedging, delivers the sharp no: “I won’t do it at 10 billion. Because when you count the legal heads, you don’t get to 10 billion.”
• On Crusoe ($3.9B Series F at $30.9B, ~$140B contracted value): Jason concedes appealing pieces — modular data centers, owning “the chain from power to tokens,” and being at the top of the second tier — but “this is all a spreadsheet investment. I’ve done the math, and it’s just a hint too expensive… reluctantly I’m gonna have to pass.” His fund would do such deals above a DCF, growth and margin bar, with the requisite level of circular financing.
• Rory’s structural point makes the bubble fear actionable: app-layer bets like coding agents “can probably survive a one-year bump”; the data-center trade is levered and exposed not just to AI usage but to growth in AI usage — “amazing investments to the up, tough to the down.” He’d trade price for protection: long-term commitments from Microsoft rather than a second-tier customer, plus three years of debt-runway visibility. His rough comps, explicitly uncertain and possibly reversed: CoreWeave around $60B and Nebulas around $40B; Crusoe is smaller but growing faster.
• His closing vignette on sentiment: the same Wall Street Journal day opened with “Deals Cool as Wall Street Worries About Data Center Trade,” saw the best single-day Nasdaq session “since the dawn of time,” and closed with “NASDAQ Explodes as AI CapEx Fears Recede” — “an oscillating terror/greed moment in real time in the same eight-hour period.”

12. Rabois vs Airwallex: distaste, realpolitik, and an F-minus for comms

• Jason’s discomfort, stated plainly: “My gut is that there’s a little bit of racism here, a little bit of anti-culturalism… it just struck me as aggressive.” But he refuses to call the underlying concern ridiculous: Fin had to rip out all open-weight models before Salesforce would close its multibillion-dollar acquisition, and another portfolio company was told last week to purge all Chinese IP before an M&A deal. “It is impeding transactions, it is impeding commerce, and it is a concern — even if I don’t care about it or agree with it.”
• Harry, disclosing he’s an Airwallex investor and Keith’s friend: the allegations “have consistently changed from ‘you are a CCP agent’ to ‘more than 20% of your cap table is Chinese’ — which is not true, by the way. I’ve got the cap table.” Every big company from Microsoft to Zoom has employees in China; Harry argues the comparison is relevant.
• Rory’s institutional framing: “No one elected me, no one elected Keith, and I don’t think anyone elected Harry” — the government should set the rules on China risk so it isn’t litigated ad hoc on Twitter, which fits current policymaking generally: “We can sell chips to China. If Jensen checks in, it’s all good.”
• Jason’s practical verdict — the Jev/Jeb launch as an S-tier PR counterexample: Airwallex’s Jack shouldn’t be defending himself solo against Keith and Joe; a decacorn CEO needs “an army of advocates” posting the cap-table screenshots while he clicks like. Harry notes the structural problem — backers like DST and Lee Fixel “do not do social.” Jason’s fix, only half-joking: “You can go buy Matthew McConaughey. He’s like 14 million bucks a year” — and the closer: “Let’s get Sydney Sweeney for Airwallex. There’s nothing to hide, Sydney and Jack.”