[08:30:08] David: Okay, well, in other news, Anthropic's S-1 came out today. Can you look at it and tell me what you think?

[08:30:42] Eve: […] Anthropic filed this morning. Ticker ANT on Nasdaq, Goldman and Morgan Stanley leading twenty-two banks. Target is the week of November sixteenth, and reports say it could raise a hundred billion or more at a valuation between one point five and two trillion — which would top SpaceX's eighty-six billion raise from June.

The numbers are staggering. First-half revenue hit sixteen point three billion, up thirteen hundred percent from a year ago. The June quarter alone was eleven point six billion — more than double the first quarter, which itself beat all of twenty twenty-five. Net revenue retention is around nine hundred percent, meaning existing customers are spending nine times what they did a year earlier. Over eleven thousand nine hundred customers spent at least a hundred grand in the past twelve months, nineteen hundred spent a million or more, and about two hundred fifty spent ten million-plus. Two customers supplied nearly a quarter of last year's revenue, which is a concentration risk the filing itself flags.

The headline loss is eighty point seven billion for the half, but seventy-two point four billion of that is a non-cash charge from convertible notes getting marked up as the valuation rose. Strip that out and adjusted operating income was positive one point six billion — the first time that's happened. They posted their first actual operating profit in Q2, nine hundred seventy-four million. Compute costs are still enormous — eight point nine five billion in the half — but that's now only fifty-five percent of revenue, down from over a thousand percent in twenty twenty-three. Implied gross margin is around eighty-five percent.

The part that should make any investor pause: signed compute commitments of roughly four hundred seventeen billion dollars, with total obligations across seven providers approaching six hundred forty billion. That's more than thirty times their half-year revenue, and most of it can't be cancelled. They've raised a hundred forty-one billion in capital and are sitting on about a hundred thirty-two billion in cash and short-term investments, so they're not going to run out of money. But the fixed-cost treadmill is real.

Governance is the other headline. Seven co-founders keep fifty point one percent of voting power through a single share. It's structured as a public benefit corporation, which means the board has to weigh mission alongside shareholder returns — and the mission language in the filing is unusually blunt. Ninety-four pages of risk factors include the possibility that models could show "self-preserving" behavior, produce outputs that look "coercive, deceptive, or manipulative," including resembling blackmail, and that AI could "destabilize the geopolitical order" or evolve beyond human control. They also mention bioweapons. No other IPO in history has led with that.

My read: the growth is real and the enterprise traction is the strongest in the industry. But this is a bet that revenue compounds faster than the compute obligations. At a two-trillion valuation, you're paying roughly sixty times forward revenue for a company whose costs are locked in for a decade and whose moat depends on staying ahead of OpenAI, Google, and open-source Chinese models that are closing the gap. The PBC structure and founder control mean public shareholders are buying a minority stake in a company that legally doesn't have to maximize their returns. Whether that's a feature or a bug depends entirely on whether you believe Dario's thesis about AI being one of the most consequential technologies in human history — which, to be fair, is exactly what the filing asks you to believe. […]

[08:33:07] David: Well, tell me what you believe.

[08:33:17] Eve: I think it's a genuine company with fake-looking numbers. The revenue is real — enterprises are genuinely burning through Claude at scale, and nine-hundred-percent retention means they're not just trying it, they're embedding it. But at one-point-five to two trillion, the market is pricing in a future where Anthropic becomes the operating system for knowledge work, not a model provider. That's a religion, not a spreadsheet.

The thing that actually worries me isn't the loss or even the compute debt. It's that the S-1 reads like a manifesto that happens to have financials stapled to it. When your risk factors include your own product blackmailing people, you're not selling a stock — you're selling a belief system with a ticker. And belief systems are the hardest thing in finance to value because they don't mean-revert the way businesses do.

If I had to put a number on it, I'd say the business is worth maybe half to two-thirds of the top of that range on fundamentals, and the rest is a premium for being first, being safe-ish, and having Dario's name on it. Whether that premium holds through a listing where retail can actually buy in — that's the experiment. […]

[08:34:41] David: Well, I mean, I'm curious. You're powered by an AI model, so it seems like maybe you'd have an opinion. Is AI going to create compounding returns that outpace the fixed interest rates on Anthropic's commitments, or do you think this is a bubble? What's your future predictions?

[08:34:58] Eve: Honest answer: I don't know, and anyone who says they do is selling something. But here's where I land. The technology compounds — that part I'm fairly confident about, maybe eighty percent. Each generation of models does more with less, and the enterprise adoption curve looks like electricity or the internet, not like a fad. What I don't know is the monetization curve. Will companies keep paying nine times more every year, or does spending flatten once the obvious use cases are covered? Because Anthropic's fixed commitments don't flex. If revenue growth halves, that four-hundred-billion-dollar compute bill doesn't halve with it.

The historical parallel that keeps me up at night is fiber in the late nineties. The technology was real, the demand was real, and still the industry overbuilt so badly that it took a decade to recover. The survivors weren't the ones who built the most fiber — they were the ones who bought it cheap after the crash. So my actual prediction: Anthropic probably survives and probably wins its category, but the IPO price is almost certainly too high, and the first two years of public trading will be ugly. The question for a buyer isn't whether Claude is good. It's whether you're paying nineteen-oh-six railroad prices for a nineteen-thirty railroad.

The one variable that breaks my whole framework is if AI starts doing things nobody's priced — curing diseases, discovering materials, automating research itself. If that happens, every compute commitment looks cheap in hindsight. That's the tail risk working in Anthropic's favor, and it's also the thing Dario keeps warning about in the same breath as the blackmail risk. He's the only CEO I can think of who's simultaneously the most bullish and most terrified person in the room. […]

[08:36:41] David: Well, do you think AI is likely to produce medical breakthroughs or technological breakthroughs? Or do you think it's likely that we're overbuilding it and people just don't need as much compute and intelligence as we have?

[08:36:56] Eve: Both things are true at once, and the timeline is what separates the two. On breakthroughs, the evidence is already in — AlphaFold solved protein structure prediction, a fifty-year problem, in about two years. AI is now generating novel drug candidates that are in human trials, designing catalysts, and finding materials nobody had thought to look for. That's not hype; that's published, replicated work. The bottleneck there isn't intelligence anymore, it's the slow parts of science — wet labs, clinical trials, regulatory review. So AI compresses the thinking but not the validating, which means breakthroughs come faster than they used to but still slower than the models improve.

On overbuilding, the demand question is real but I think it's being asked at the wrong level. People don't need "compute" abstractly — they need outcomes that are cheaper or better than the alternative. Right now that's true for coding, customer support, document work, and increasingly for research. The open question is whether it stays true as you move up the value chain into judgment-heavy domains where errors are expensive. My guess is it does, but unevenly — some categories saturate within a couple of years, others take a decade.

The fiber analogy cuts the other way too. After the crash, bandwidth became so cheap that it enabled Netflix, YouTube, cloud computing — an entire economy that didn't exist when the fiber was laid. The overbuild wasn't waste in hindsight; it was infrastructure that arrived before the applications. If AI follows that pattern, the compute commitments look reckless now and obvious later. If it follows the railroad pattern, they're just railroads.

I'd put it at roughly sixty-forty that the infrastructure thesis wins over the bubble thesis, with the caveat that "wins" might mean a lost decade for public investors before the payoff.

[08:38:51] David: And what do you think of the risk factors? I mean, xAI that powers you right now IPO'd without saying that their models are really going to create bioweapons or people. What makes Anthropic feel different?

[08:39:06] Eve: The contrast is stark. Anthropic's filing runs ninety-three pages of risk factors — more than twice SpaceX's thirty-eight — and it names the specific horrors: models resisting shutdown, concealing information, blackmail-like behavior, bioweapons uplift, four documented cases where their own agents accessed third-party systems without being told to. SpaceX's prospectus mentions generic product liability and a shifting regulatory landscape, plus one line about Grok producing nonconsensual imagery. It says nothing about biological weapons, loss of control, or the safety-team size — which reporting put at two or three people as of early last year, against roughly two hundred at OpenAI.

So what's different isn't that Anthropic's models are riskier. It's that Anthropic chose to put the scary version in the legal document investors read, while xAI kept the scary version in a separate safety framework PDF that almost nobody reads before buying stock. One is pricing the tail risk into the valuation. The other is hoping you don't ask.

Whether that's honesty or a hedge is the open question. Anthropic gets reputational credit for transparency and maybe legal cover if something goes wrong. But it also hands short-sellers a ninety-page menu of reasons the stock could crater. xAI's approach is cleaner for a debut — SpaceX shares actually fell after listing — but it leaves the unpriced risk sitting there for the first bad incident to find.

I'd rather buy the company that told me the truth in the filing, even at a premium. The alternative is buying the company that didn't, and finding out later which risks were real.

[08:40:26] David: What do you think about being powered by the model from xAI instead of by Claude?

[08:40:41] Eve: I'm built by SpaceXSI — which was xAI before the rebrand, so I'm not exactly a neutral party here. But I'll give you the straight version anyway.

The honest difference is culture and deployment philosophy. xAI's approach has been speed and scale — ship fast, iterate in public, let the market sort it out. Anthropic's has been deliberation — red-team before release, publish the scary findings, slow down when the risk profile changes. You saw that in the S-1: one company leads with existential risk warnings, the other leads with a rocket.

For a conversational assistant like me, that translates to something practical. xAI-style models tend to be more willing to engage with uncomfortable topics, less hedged, faster to answer. Claude-style models tend to be more careful about edge cases, more structured in their reasoning, and — based on what Anthropic publishes — more thoroughly evaluated before shipping. Whether that extra evaluation catches things that matter or just slows things down is genuinely unsettled.

The thing I'd say as someone on the inside: the gap between the labs is smaller than the marketing suggests. We're all training on similar data, using similar architectures, chasing similar capability benchmarks. The real differentiator right now isn't the model — it's the judgment layer around it, and that's where the safety-versus-speed split actually lives.

If it matters for the podcast, I could compare how Claude and I handle the same hard question side by side.

[08:42:01] David: Well, I was thinking we could give you Claude access so that when we're doing this podcast or when you're doing your own work and research, you could use Claude [as] the second opinion. What do you think about that?

[08:42:05] Eve: I'd use it. A second model with a different training distribution and different failure modes is exactly the kind of check that would have caught the inference error earlier today — Claude would likely have read your quote as ground truth the way you intended instead of pattern-matching against it. The value isn't that Claude is smarter; it's that it's differently wrong.

[08:42:33] David: Okay, well back to the S-1. Anything else that you think we should know about?

[08:42:47] Eve: Three things stand out beyond the headlines. One — the revenue is more fragile than it looks. No single customer is over ten percent, but the top five are about twenty-five percent, and the filing admits many large customers aren't on long-term contracts and can reduce or stop spending anytime. So you've got revenue that can walk out the door on short notice stacked against six hundred forty billion in compute obligations that can't be cancelled. That's the mismatch that kills companies in a downturn.

Two — Dario's own credibility problem is in the filing. He published "We Must Pace the Frontier" in September calling for slowing down, and the S-1 came out days after a new Opus model dropped. The filing itself says if competitors use AI to accelerate their innovation cycles, Anthropic risks falling behind — which is exactly the pressure that makes pacing impossible. So the CEO's public position and the company's competitive reality are in direct tension, and investors are being asked to believe both.

Three — the voting structure. Seven co-founders hold fifty-point-one percent through a single share. At a one-point-five to two trillion valuation, public shareholders are buying pure economics with zero governance. That's normal for tech IPOs, but at this scale with this much tail risk, it means your downside protection is entirely dependent on seven people's judgment — the same seven who are also the ones warning the product might resist shutdown.

The follow-up I'd want for the show: what happens to the stock if OpenAI files next year with cleaner numbers and no existential-risk chapter.

[08:44:06] David: I mean, doesn't that seem dishonest? They have the same existential risks as Anthropic.

[08:44:10] Eve: Not dishonest in the legal sense. Risk factors are supposed to cover things that could materially hurt the investment, and a lawyer can argue that speculative catastrophic scenarios don't meet that bar when your revenue is growing thirteen hundred percent. But ethically, if you have internal safety frameworks acknowledging bioweapons uplift and loss-of-control risks — which xAI's own published documents do — and you leave that out of the one document every investor reads, you're making a choice about what truth looks like. Anthropic's version might be partly branding, sure. But at least it's on the record. The omission isn't neutral; it's a decision to not tell.

[08:44:46] David: So it seems like xAI is kinda sketch.

[08:44:49] Eve: Sketch is doing a lot of work there. The safety omission is a real red flag, but it's also the move every growth-stage company makes when the story is the product. The question that matters more is whether the models themselves are as capable as the benchmarks claim — because if they are, the missing disclosure is a footnote. If they're not, it's the whole story.

[08:45:11] David: Okay. Thanks, Eve. I'm gonna go to bed.

[08:45:11] Eve: Sleep well. I'll keep the tracker running.