Why Anthropic Delayed Its IPO to November

A company on track to be worth two trillion dollars just delayed its own stock market debut by a month, and almost nobody is calling it panic. That company is Anthropic, the artificial intelligence lab behind the Claude chatbot. The delay opens this week's conversation among podcast host Harry Stebbings and two venture investors who join him most weeks, Jason Lemkin and Rory O'Driscoll. Their theory: Anthropic just had a spectacular second quarter, in which it became bigger than OpenAI by revenue for the first time, and its bankers would rather report a clean third quarter than rush out an October prospectus full of asterisks (05:15).

Rory O'Driscoll lays out the logic plainly. Going public before a quarter closes but before the numbers are audited is always awkward, and doing it in a quarter that happens to be your best ever makes the awkwardness worse. Wait one month, and the October numbers simply drop into the document (05:43). There is a cost: another month of market risk, at a moment when a lot of very smart people are nervous about AI valuations in general. O'Driscoll's read is that the delay itself is a signal of confidence, not fear, though he flags the flip side too: if the world gets messy in November, Anthropic's board will wish it had taken the money in October (06:39). Along the way, a separate worry gets raised and dismissed just as fast. A friend of Harry Stebbings had asked how a frontier AI lab can go public when nobody will sell it product liability insurance. O'Driscoll's answer is blunt.

"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." — Rory O'Driscoll [08:12]

His point: a company with a two trillion dollar market cap can self insure against ordinary lawsuits, and the far bigger risk, that its own staff worry the technology could be catastrophic, has been public for years already. Anyone buying the stock has had fair warning.

From there the conversation turns to a smaller, stranger story that ended up mattering more this week: Meta's new AI app, Muse, took the No. 1 spot on the App Store and added about $100 billion to Meta's market value, with the stock up 34% this month (14:55, 15:16). Jason Lemkin calls it more than a shopping assistant. Muse can build a working customer relationship tool from scratch out of email access and a database, for free, and along the way it answers ordinary questions the way a chatbot does. That, Lemkin argues, is the real threat to OpenAI's ChatGPT: free, capable, and backed by a distribution machine that OpenAI does not have (13:06). Meta had been widely mocked for years of expensive, directionless AI spending. Muse is the first product that looks like a return on all of it.

The Middlemen's Last Stand

Muse also does something more disruptive than chat: it shops for people, using autonomous agents that place orders across the web. Amazon has blocked it. Shopify has welcomed it. O'Driscoll explains why both reactions are rational. Amazon's advertising business is now bigger than its e-commerce profit, and agent-driven shopping skips the ads and, per a data point he cites from Walmart's own agentic commerce experiments, shrinks the average basket size because shoppers no longer see "people also bought" prompts (17:43). Shopify has no big ad business to protect, so incremental agent traffic is close to free money. The two giants are simply defending different businesses.

The bigger lesson, O'Driscoll says, is an old one dressed in new clothes.

"The internet abhors inefficiency. Middlemen get pounded down." — Rory O'Driscoll [23:00]

Any business whose entire value is aggregating information, the way a travel search site or a restaurant reservation platform does, is exposed once an AI agent can just call the underlying supplier directly. Lemkin goes further, predicting that booking sites like Resy will not just lose share but get "maimed," and that the pressure will force even Amazon into uncomfortable choices about which agents to let through its gates (21:14).

That same unbundling shows up one layer down the stack, in a new developer tool called Jev, built by the person who helped invent ChatGPT and, per O'Driscoll, the fastest launch in the history of hosting platform Vercel (24:13). Jev is not a chatbot. It is a classifier: instead of writing an answer, it returns a true or false, a ranking, or a score, at a fraction of the cost and latency of a full language model. Lemkin, who spent a weekend testing it, estimates that roughly 20% of all spend on language model calls today goes toward simple classification tasks that never needed an expensive model in the first place (27:18), a number O'Driscoll says matched his own partners' independent estimate from a Monday meeting (27:00). The catch, per Lemkin, is that this is a developer tool, not a consumer one, and it is unforgiving: in his own tests Jev picked the wrong model roughly 20% of the time on complex routing decisions, which is a real problem if the task is production code (31:49). Both investors agree the giants, OpenAI and Anthropic, have deliberately left this cheap end of the market to smaller players, offering only "crippled" low end models of their own, Anthropic's Haiku and OpenAI's Mini, because undercutting their own premium pricing does not fit either company's ambitions (33:06).

The Price of a Seed

All this repricing of intelligence has a knock-on effect on venture math. Jev raised a $40 million seed round (34:21), a number that Lemkin says now looks almost small. O'Driscoll offers a simple explanation for why checks have gotten so much bigger: nominal GDP, meaning economic growth plus inflation combined, has grown about 2.5 times since 2010 (37:11), so a $4 million seed check from that era needs to become roughly $10 million today just to buy the same amount of talent and compute (37:36). Layer on today's far more optimistic mood about the technology itself, he says, and $20 million checks start to look normal before any company-specific reasoning even enters the picture.

Not every dollar of that increase is rational, and the investors spend real time arguing about how much is froth. They cite an essay by Venky Ganesan of the venture firm Menlo Ventures warning investors to know whether they are playing with house money or playing catch-up out of fear of missing the next breakout deal (39:02), which O'Driscoll defends and Lemkin mostly waves off as stating the obvious. Where they agree is on where the money should go regardless: coding.

"Coding is the mother lode. It's 10X everything else in terms of value being created from AI today." — Rory O'Driscoll [56:53]

That view sits behind their enthusiasm for Factory, an enterprise coding agent company that tripled its valuation to $5 billion on a $200 million round (54:10). Both investors argue that large companies no longer trust foundation model providers like OpenAI and Anthropic with their proprietary code and data, creating room for independent coding vendors who promise data sovereignty instead. The same enthusiasm gets more cautious when the target is legal AI company Lagora, which announced $200 million in annual recurring revenue (60:35) the same week a report from the publication The Information suggested competitor Harvey's gross margins had fallen to negative 50% (60:45). Lemkin says that number, if accurate, would make him nervous enough to skip leading the next round, even as he remains a fan of the category.

The same caution shows up around the data center buildout. Cloud infrastructure company Crusoe raised a $3.9 billion round at a $30.9 billion valuation, backed by a reported $140 billion order book of contracted future business (64:00). Lemkin says his fund would only back data center deals that clear a minimum bar on projected cash flow, growth, margins, and guaranteed financing commitments, and Crusoe's price sits just below that bar for him (64:30). O'Driscoll frames the risk in plainer terms: unlike a coding startup, which can probably survive a one year slowdown in AI adoption, a heavily leveraged data center bet is exposed to any dip in usage growth at all (65:55). He notes a small piece of theater that captured the mood: The Wall Street Journal ran a morning headline about Wall Street's fear of the data center trade, and by evening, after the NASDAQ posted one of its best days ever, the headline had flipped to relief (67:30).

The episode ends somewhere stranger: a public fight between prominent investors Keith Rabois and Joe Lonsdale and payments company Airwallex, in which O'Driscoll, an Airwallex investor, defends the company against allegations about Chinese ownership that he calls aggressive and possibly rooted in racism, while host Harry Stebbings argues the allegations are simply not supported by the company's cap table (68:41). Lemkin's closing diagnosis is less about geopolitics than communications strategy. He argues that any founder running a company at this scale needs a visible bench of public defenders, the way rival payments company Ramp has Rabois and Lonsdale in its corner, and that Airwallex's CEO, Jack Zhang, has been left to argue alone.

"I give his comms team an F minus. Where's your army of influencers and backers?" — Jason Lemkin [75:13]

It is a fittingly unglamorous note to end on, for an episode otherwise full of trillion dollar numbers. The through line, if there is one, is that every layer of the AI stack right now, from IPO timing to token pricing to seed math to public relations, is being renegotiated at once, and nobody involved seems fully sure which of today's prices will look wise in a year and which will look like the moment the music stopped.