Anthropic's Fable Banned by US Government | Wix & Adobe Hit All-Time Lows | Mistral Raising at $20BN

Summary

• SpaceX’s $2.7TRN debut was executed to perfection — zero price discovery, Elon “just told everyone the price,” then a 19% day-one “designer pop” — yet both guests take the under on the stock six months out: only ~4% of shares float, options just listed, and gamma squeezes plus index inclusion are engineering the price higher near-term. “Nothing matters until the lockup’s gone… it’s almost like a private mark — you can’t take it to the bank.” Even the bear trade is prohibitive: 6-month at-the-money puts cost ~20% of the share price, so a $200 stock must hit $120 just to double your money.
• The durable Elon thesis, from someone who lost his college savings shorting Tesla: he “sells the market on these incredible long-dated call options” — FSD, Optimus, Starship, now orbital data centers — and usually delivers, while a “small village” of blindly loyal funds that “all got stupidly rich” hands him a cost of capital no one else has. Crusoe exercised at $60BN plus roughly $2BN/month of Anthropic and Google compute contracts means SpaceX’s AI run-rate exceeds the entire rockets-plus-Starlink business from September; prediction markets put ~70% odds on a Tesla merger within two years.
• The Fable ban is “a Rubicon moment” — the first time the US has ostensibly regulated an AI model on capabilities, via an Export Restriction Act with no judicial review. The tell arrives in 3-6 months when OpenAI and Google hit Mythos/Fable-quality: ban them too, or Anthropic has a due-process claim. Rory’s verdict on Dario: “You’ve made yourself part of the political process… you better get good at politics really fucking quick” — and “good intentions bite you in the ass more than evil deeds.”
• The ban may also be incoherent — and accidentally bullish for tokens: with the right harness and enough test-time compute, open models can find the same vulnerabilities, and “they haven’t found where the wall is.” A guest expects another kink in the token-consumption curve, great for inference platforms; Anthropic IPO odds stay “way more than 50%,” with one founder predicting $150-200BN ARR this year, not $100BN.
• Sovereign AI finally gets a real bid — Mistral raising $3BN at $20BN — on the logic “would you prefer the second-best model that you have access to, or the best model that gets cut off once every 6 months?” But “outside of China, there are basically no good open source models,” Mistral itself has fallen far behind on models, and the endgame is a 2-3 player US oligopoly; Europe will “care enough to fund it, but not enough to find the $100 billion.”
• Fin’s $3.6BN sale to Salesforce is the golden path for zombie SaaS: burning the boats from seats to outcomes at 99 cents per resolution turned equity that was “essentially worthless” into cash — because “any liquidity for pre-AI SaaS companies is top decile performance. Any liquidity at all.” The caveat: “burn your boats only after you’ve checked where the boats are” — not every category has Fin’s obvious AI wedge.
• Public software now has a filtering mechanism: token-linked usage models (Datadog, Snowflake) trade back above 15x NTM revenue while Wix trades at 1x. At that price Wix should go private (“price has its wicked way — I’d be lining up the debt providers”); Adobe at 8x LTM free cash flow has every bad attribute and no buttons left to press; and the trade that beat everyone remains short SaaS / long semis — the BVP cloud index is -44% over five years against +325% for semiconductors.
• Robotics is a prime candidate for the next trillion-dollar companies, but the field reality is sobering: ~3 million robots worldwide against a billion people doing manual work after 30 years, and “reality has a surprising amount of detail” — one large order is blocked by wrinkled poly bags defeating a barcode reader. What changes now: edge LLMs make robots flexible instead of brittle, and a guest’s 2060 sketch is 10,000 digital agents per human and roughly 1:1 useful robots.

Deep dive

SpaceX Has a Designer Pop

• Rory’s read on the largest IPO in history: Elon “didn’t do any price discovery — he just told everyone the price he’s going to take, went out and got it,” then landed a 19% day-one pop — “the perfect pop, the designer pop,” at “the high end of perfect” after intraday touching 30%, where “Bill’s going to be mad” about money left on the table. It was up 30-40% around the debut and has traded up nicely since; Musk added roughly Warren Buffett’s lifetime net worth in 24 hours and sits about a trillion dollars richer than the next person.
• The self-aware absurdity, verbatim: “It’s really hard to say I made 1.2 trillion dollars but I might have left 50 billion on the table. How do I feel? I think he feels okay.”
• Harry’s wished-for rule: a moratorium on net-worth math while shares are locked up — “it’s almost like a private mark. It looks good, you put it in your little spreadsheet, but you can’t take it to the bank.” The real barometer is the price when the lockup releases in six months; plenty of IPOs saw the day-one victory lap and then a 60% drawdown.

SpaceX Floats Thinly

• Options on SpaceX began trading the day of recording, and Ev walks through the gamma squeeze: call buyers force market makers to hedge by buying stock, lifting the price, drawing more call buyers — “a self-reinforcing loop, where the more call options people are buying, the more forced buying there is” — which turns violent when only ~4% of the share count trades.
• Asked over/under in six months, both go under. Ev: “I would personally probably take the under” — despite “engineered things” pushing near-term price (index-inclusion forced buying, retail call-option mania, no shares available). Rory agrees on 6-12 months but flags that strong news plus strong technicals on a low float make “the short-term bet much harder to call.” He also notes the room to fall: $1.8 trillion of value between this and the last private round at $400BN.
• Rory actually priced the bear trade, since “options are the coward’s way of shorting the stock… and I’m a baby”: 6-month at-the-money puts cost roughly 20% of the share price — $40 on a $200 stock — so breakeven is $160 and a 2x needs $120. “Do I have the nuts to put a million bucks on the line and say I believe it’s going down? Decided I didn’t have that just yet.”

Elon Sells Long-Dated Calls

• A guest’s coming-of-age story, as told: a college value-investing-club kid (“you read Benjamin Graham and Howard Marks and think you’re smarter than everyone else”) who shorted Tesla after graduation — “I lost all my college savings shorting Tesla. That was such an unbelievably valuable lesson” — and rectified it by leading Kleiner Perkins’ SpaceX investment in 2022 at around $120BN.
• The framework: Elon “sells the market on these incredible long-dated call options” — full self-driving, then Optimus; rocket reusability, then Starship, now orbital data centers, Mars, the moon. “He honestly usually accomplishes what he’s going to lay out,” so markets give full credit for the next trillion before it exists — on the P&L alone “it’d be very hard to even get a $2 trillion valuation as a fair market value.”
• The paradox worth keeping: option markets price the stock’s volatility as extreme while the valuation treats the long-dated calls as “highly predictable.” If time-to-deliver ever stretches, “there’s a big gap between fundamental value and anything you got here” — but today “he gets the benefit of the doubt like no one else.”

Loyalty Lowers Capital Costs

• A guest on the shareholder base: “anyone that has been blindly loyal to Elon… they’ve all got stupidly rich” — “a small village of these people,” funds you’ve never heard of with ~95% of cumulative invested capital in SpaceX and other Elon companies. The surplus goodwill runs so deep that “SpaceX could go nowhere operationally for 5 years and people would still be a believer.”
• That loyalty is the moat: walk into a bank asking for $24BN to build Colossus with no contracts and “you’d have got laughed out of there.” Rory: “he’s earned the right to play” — low cost of capital enables bets nobody else can make, which produce returns, which lower the cost of capital further.
• Rory’s discomfort, unsmoothed: “the rational in me says that’s not the way capitalism should work” — and it’s a single point of failure: “if it goes wrong, it will go wrong for everything.” On a Tesla-SpaceX merger, prediction-market odds (he cites them loosely, “not that indicative”) sit around 70% within two years; with one board easier than two, “there ain’t going to be a ton of votes saying no. Especially after you’ve made him 2.2 trillion.”

SpaceX Becomes an AI Business

• Exercising the right to buy Crusoe at $60BN now looks prescient — roughly $4BN of revenue today, $6BN potential by year-end, the team retained via lockup. Rory: it cost “a price equal to a third of the variability of the stock between yesterday and today… it was a great deal when he did it, it’s a better deal since then.”
• He solved the what-do-I-do-with-this-compute problem twice: Crusoe filled the gap in Colossus, then contracts with Anthropic and Google — roughly $2BN a month, $24BN a year of “CoreWeave-like revenue.” When those kick in in September, “the revenue run rate of the quote-unquote AI business… is larger than the revenue run rate across the entire SpaceX and Starlink business.”
• The meta-point is velocity: “Everyone else has spent two years thinking, yeah, we should probably do something in AI. Maybe we should build a data center. He’s like, no — I’ve built two data centers… so I did two huge contracts. Moving right along, people.”

Fable Triggers a Communications Clash

• The facts as Rory assembles them: Fable is “effectively a front end” to Mythos, which Dario had pre-labeled too dangerous for general release. Amazon discovered you could interrogate the model for cybersecurity input and called the government; after a 90-minute call with an assistant chief of staff (reporting to “Susan Walsh”), the administration invoked the Export Restriction Act. “They just said after 90 minutes: fuck this — you’re not taking us seriously.”
• Both sides were right on their own terms. Dario’s defense was coherent: the jailbreak “didn’t represent a meaningful threat” because Mythos’s danger is scale, not any single bug — “the Russian army: quantity has a quality all its own”; a thousand bugs in two hours is terrifying, a thousand in a thousand hours is not, and this fault enabled neither. But politically: “if you run around telling everyone you’ve got the scariest thing in the last 20 years, and it shouldn’t do any cyber, and then it does cyber, and they don’t like you anyway — they’re going to pull the pin on you.”
• Rory’s Machiavelli lesson: “Good intentions bite you in the ass more than evil deeds.” And the deeper rule: “Private citizens don’t get to run around and say this could cause world damage — oh, but by the way, we’re the arbiters of the decision-making. You are ipso facto political, and you better get good at politics really fucking quick.”
• Unlike a prior fight he references, this is “pretty untrammeled” national-security power with no judicial review — Anthropic has little leverage in court.

Capability Regulation Reaches a Rubicon

• Rory’s core frame: “at the face of it, this is a Rubicon moment in the history of the AI industry — the first time the US has ostensibly regulated an AI model based on capabilities” rather than contract terms. The test comes in 3-6 months, when OpenAI and Google reach Mythos/Fable-quality models: does the government ban them too? Rory: “if they are comparable and they don’t ban all of them, then they have a problem” — Anthropic may even have a due-process claim.
• Extend the logic to ASI and it “gets pretty freaky”: “What happens when the government is gating access to superintelligence? Do NATO allies?… imagine the economic implications of the United States and China having access to superintelligence and Greece not having it.” That’s why sovereign AI is suddenly taken seriously — “the only issue is none of these sovereign AI plays have amounted to anything yet.”
• Practical fallout inside Anthropic, per a guest: non-US-citizen researchers “step away from the keyboard” or the organization breaches the Export Restriction Act — “a huge percentage of these teams are overseas.” Zooming out, a guest insists this outlives Anthropic: models can now “autonomously find and chain multiple vulnerabilities together and orchestrate an attack,” so the nation-state-adversary conversation “isn’t an Anthropic discussion anymore.”
• Everyone switched sides: Dario the regulation-funder must argue “it’s really dangerous and we should regulate it — but not regulate me,” while the deregulation camp’s administration just took “one of the strongest actions we’ve ever seen against the technology.” Both flag the epistemic caveat: “a lack of reliable narrators” — two parties that don’t like each other, contradictory reporting, “zero communication, zero trust.”

Test-Time Compute Expands Demand

• A guest’s nothingburger case: “if you talk to some smart people in AI, they will tell you that with the right harness and the right amount of test-time compute, you can use open models to find all of the vulnerabilities that Fable can find.” Rory accepts it and draws the conclusion: then you’d logically have to restrict any open model above X compute — “a very hard decision to implement with any degree of coherence across 12-24 months,” because “the original premise, that this is dangerous, is in and of itself incoherent.”
• The tradeable flip side: Noam Brown’s point that benchmark scorecards are the wrong frame — models “perform very, very differently if you just continue to throw more compute at it at test time,” and “they haven’t found where the wall is.” A guest expects another kink in the token-consumption curve, great for everyone “in the token path” — frontier labs and inference platforms like Fireworks.
• Should Anthropic still IPO this year? Rory the “simple Bayesian”: the prior should have been 90%+ after SpaceX’s reception — “you’d be pretty much of an idiot to see SpaceX trading at 2.4 trillion and say, let’s hold on for a better market.” The ban is “a significant wrinkle,” but odds stay “way more than 50%.” A guest’s kicker: a portfolio founder, after one day on Fable, said Anthropic won’t hit $100BN of ARR this year — “they’re going to hit 150 to 200.”

Sovereignty Drives Mistral’s Bid

• Mistral is raising $3BN at $20BN, having scaled past half a billion with some of Europe’s biggest enterprises. Rory’s sovereignty trade: “would you prefer the second-best model that you have access to, or the best model that gets cut off once every 6 months on a random basis?” The more the US exerts sovereignty, the more that logic bites — but Europe will do “the classic European thing: they’ll care enough to fund it, but we won’t be able to care enough to find the 100 billion dollars to compete.”
• A guest’s sobering inventory: “Outside of China, there are basically no good open source models… where are the US open source models?” Nvidia’s NeMoTron is “pretty solid”; Mistral itself “has fallen very far behind on the actual model side” while building a good inference platform. Frontier model-building — pre-training through post-training — is “an extremely scarce skill set,” not fungible; most countries will likely fine-tune a Chinese open-weights model. The guest saw a claim that the city of Rio post-trained a near-frontier “city model” — “might have been fake.”
• Rory’s endgame: given the cost to play, “a small oligopoly, like cloud, of two to three players in the US with some kind of affiliations in Europe… unless the US is ludicrously obnoxious in its sovereignty acts.” Nvidia wins either way — “everyone will feel the need to have their own chips.”

Benchmark Missed Model Providers

• Harry asks directly how one of the best firms processes not being early in a model provider. A guest, no spin: “No, it fucking sucks. It’s terrible… a complete and utter failure on our part.” If you claim to be one of the best firms in the Valley and pass on “a 30x on scaled capital in four or five years… that’s always a failure, no matter the way you cut it.”
• The sting is social too: “all your friends are sending you their implied look-through ownership of Anthropic and OpenAI and SpaceX… those are eye-watering numbers.” The fund still looks awesome — the Sierra, Firework, Lagoras, Recores, Langchains, and Hagens of the world — but the miss stands, and another top fund at dinner was in the same boat.
• Harry’s consolation, with teeth: “attribution is everything in venture. Those great deals were all you” — the number of Excel GPs he’s met who claim the Facebook deal “could fill a series of podcast episodes.”

Fin Shows SaaS’s Golden Path

• Salesforce buys Fin — formerly Intercom, founded 2012 — for $3.6BN. Rory: “these guys became the canonical example of an old-school SaaS company that made the transition and pulled it off… and whatever’s in the water, that’s what Salesforce needs to do” if Agentforce is to reignite growth. The mechanism was moving from seats to outcomes — 99 cents per resolution — which “was putting your ass on the line” and forces the vendor into the resolution path. His arc of software history: license key (“you didn’t even care if they deployed it”) → per-seat SaaS → pay-per-business-outcome, each step pushing vendors closer to the customer.
• The degree of difficulty: taking roughly $300M growing 7% to $400M growing 25% “is literally like pushing a rock uphill,” and they pushed for two or three years. The stakes-setting: a Benchmark partner says “any liquidity for pre-AI SaaS companies is top decile performance. Any liquidity at all” — equity that was “essentially worthless” (“no one’s going to buy 300 growing seven… that’s just a zombie company”) became $3.6BN. If this isn’t topic one at the next board meeting of every pre-AI SaaS company, “it’s a failure of the board.”
• Both immediately hedge the moral. Rory: “burn your boats only after you’ve checked where the boats are” (Alexander in Persia, as he tells it) — customer support was an obviously AI-addressable wedge; some categories should choose profitability and steady growth instead. Another guest: the most annoying board member will now demand “why can’t you do what Fin did?” where it’s unrealistic — “I would love to grow three times faster or increase our margins… but these things are very hard.” Still, the proof matters: “it turns out in AI you can teach old dogs new tricks” — incumbents will buy 14-year-old companies as AI plays.

Wix Faces a Replicability Problem

• The news: Wix slashes 2026 guidance, cuts 20% of staff (1,000 people), trims outlook by $50M and revenue by $25M; the stock trades around 1x revenue, and the buyback sits well underwater — “the acquisition was a great idea, the buyback was obviously a bad idea.”
• The T-chart for what markets now pay for. Good: a usage-based component that scales with tokens (Datadog, Snowflake), a clear AI tailwind for the use case (cybersecurity), and AI-accelerated share gains for challengers. Bad: perceived business-model exposure, a product easily replicable by coding agents (Wix; Intuit via TurboTax), being the incumbent “with only share to lose,” or “your product’s just lame” as IT budgets rotate toward AI. The premium names — Palo Alto, CrowdStrike, Cloudflare, Datadog, Palantir — are back above 15x NTM revenue: the apocalypse became “a filtering mechanism.” Wix scores massive replicability plus share-to-lose, and even Figma Make gets no credit “because people think the bad outweighs the good.”
• The contrarian case: Base 44 (the acquisition) at $150M ARR against Lovable and Replit at $400-500M with Replit priced at $10BN means “you’re essentially getting the core business for free.” The prescription: “price has its wicked way. I’d be lining up the debt providers… I’m going private” — the Michael Dell/Silver Lake move, “the sweetheart deal of all sweetheart deals.” Wix at least has capital, profits, and a real AI story: “they are definitely in the burn-the-boats because you’re stuck in Persia.”
• The trap for the levered: PE “bought them all in 2021 and 2022” and now holds companies worth ~30% of cost — “there’s no price at which you want to have six times revenue and death.”

Adobe Lacks Strategic Options

• Adobe beats and raises, the stock falls 6% — Rory reverses the order: “The CFO announces he’s leaving [for Marvel]… when the first sentence is the CFO is leaving, no one was on that call for the second sentence. They were pressing the sell button.” A guest checked the multiple — “this thing trades for eight times LTM free cash flow… oh my god” — but Adobe has “basically every single one of the bad attributes”: ~80% share with only share to lose, an increasingly replicable product, a seat-based model that’s now wrong, and no internal AI talent.
• The insidious 2026 trap, a guest’s “last comment”: with AI valuations going “thermonuclear up” while yours is cut 60%, you can’t buy your way out — a big AI acquisition “would make Adobe’s stock price go down another 30%,” because the only holders left own it for the 8x cash flow. “They should have robbed the cradle of all these AI companies two years ago… now every single good AI company they could acquire is too big for them.” A speaker’s vent: Adobe has “milked their users for so long it just feels like a piece of financial engineering… like being PE-owned without ever being PE-owned” — an institutional catalyst (“someone like Owen running it”) must precede any pricing catalyst, though he rules out Owen-to-Salesforce: “there’s no fucking way [Benioff is] leaving unless he’s 80 or 90, in a coffin.”
• The allocator’s shortcut, via a guest quoting Brad from Altimeter: why untangle Adobe’s problems at 8x FCF when Nvidia trades at 16x earnings as “the poster child of every single tailwind”? The SOX is up 45-50% this year against SaaS down 20-25%; over five years the Bessemer Nasdaq Emerging Cloud Index is -44% versus +325% for the iShares Semiconductor ETF. “Go short SaaS and go long semis and you’ve made better money than any other hedge fund manager in the world — besides Leopold.”
• Rory’s caveat: know each trend’s catalyst. Semis break on a capex flattening — “all bets are off and those things are going to go down so fast and hard.” SaaS has no single catalyst, just “the winnowing of the weak”; the practical move for incumbents is small acquisitions of founder-led AI teams — he’s watched acquired founders each running a $20M BU a year later — but “if they hire more financial engineering, you need to shoot it in the head.”

Robotics Faces Physical-World Friction

• A guest likes Standard Bots’ $200M raise and its Packy McCormick-published thesis: humanoids are “putting a whole bunch of money into legs” most industrial use cases don’t need — the play is a US-built, Apple-style integrated hardware-software arm between over-scoped humanoids/robotic foundation models and old-school arms (today’s incumbents: German, Japanese, Chinese). A guest steelmans both sides — pro-humanoid: “the world is human-shaped”; anti: actuators are expensive and legs “gratuitous or vain” for pick-and-pack. Benchmark’s own bet: Eric Vishria led Sunday Robotics’ Series A — a pseudo-humanoid for the home that “looks like Ness from Super Smash Bros” and calmly kept folding laundry as employees ripped jeans away mid-fold.
• Rory’s sobriety from the Locus Robotics board (15,000 robots in the field, ~$180M/year): there are only ~3 million robots in the world against a billion people doing real work — under 1% of manual labor replaced in 20-30 years of arms. Warehouse buyers pay minimum wage and “know to the penny how much labor costs — if the robot doesn’t cost half that, you’re not going to switch.” And per the essay he loves, “reality has a surprising amount of detail”: one robot company’s biggest order is blocked because wrinkled poly bags defeat the barcode reader — “I did not have that in my memo. The poly bag problem. Who knew?”
• Why now anyway: edge LLMs make robots flexible instead of brittle — move the object mid-demo and “the machine stops and it thinks, like the Raptors in Jurassic Park, and then it goes: I see it… that is the brain working, and that is the future.” A guest’s 2060 thought experiment: perhaps 10,000 digital agents per human and roughly a 1:1 ratio of useful robots to humans — one of “the prime candidates” for the next trillion-dollar companies. The one to watch per Rory: Unitree, going public in China at ~500M revenue and profitable — “you see that take off and I’m wrong.”