20VC with Harry StebbingsMeta's Muse Hits #1 | Menlo Sounds the AI Bubble Alarm | Keith Rabois vs Airwallex: Who is Right?
20VC with Harry Stebbings
Anthropic Delays Its IPO 1:01
Anthropic pushed its expected two trillion dollar IPO from October to November. The hosts agree this is not a sign of market cracking but a deliberate timing choice. After a strong second quarter where Anthropic briefly surpassed OpenAI, banking advisors likely suggested waiting so the company could include clean October numbers rather than going public with a messy, unaudited quarter. One host notes this could reflect quiet confidence rather than stress, since a company willing to wait an extra month for a valuation pop is signaling it has time and a strong position, even though there remains a small chance market conditions could worsen before then. The conversation also touches on OpenAI's projected 278 billion dollar burn through 2030, with the company potentially running out of cash by 2028 and possibly needing another round near 1.5 trillion in valuation. One host points out three key numbers, revenue growing from 35 billion to 350 billion in three to four years, a net burn of 278 billion against 122 billion in cash on hand, and a staggering 700 billion dollar capital expenditure requirement, much of it carried on partners' balance sheets rather than OpenAI's own, showing how capital intensive frontier AI has become.
Meta's Muse Takes Off 10:01
The hosts introduce a new recurring segment, story of the week, starting with Meta's Muse assistant hitting the number one spot on the app store for several days and helping push Meta's stock up seven to eight percent, adding about a hundred billion dollars in market cap in a single week. They describe Muse as the first real competitor to ChatGPT, built on a genuinely good language model that also functions as autonomous agents, unlike ChatGPT's more limited agent capabilities. One host calls it one of the best pieces of software he has used, noting it can hold real conversations, offer opinions on movies or the news, and still complete tasks, all for free with generous token limits. The other host agrees it was a real win for Meta, crediting the company's distribution advantage through Instagram and Facebook and suggesting the launch might push OpenAI to consider building or buying something similar, perhaks even prompting deal-making with companies like Instacart.
Building A Custom CRM 13:01
One host describes personally using Muse to build a full customer relationship management tool for tracking around 150 sponsors, with real time email tracking and updates, entirely for free. He calls it some of the first truly composable software he has seen actually work, something the show has talked about as a possibility for years. He acknowledges it remains limited, since it cannot support collaboration or a full sales team, and that he is not a typical user, but the underlying pieces, database access, intelligence, and email integration, all worked together seamlessly. The other host adds that Meta appears to be giving users more standalone compute and sandbox space than Instacart currently offers, reinforcing the sense that Meta has made unexpected progress in consumer facing AI.
Amazon Blocks Muse, Shopify Embraces It 14:30
Amazon has blocked Meta's Muse shopping agent while Shopify has chosen to partner with it, and the reasoning behind each move comes down to what each company stands to lose or gain. Amazon's ad business is now larger than its e-commerce profit, so letting agents shop on customers' behalf means losing that ad revenue, and Walmart has already seen basket sizes shrink when shoppers use agents instead of browsing and picking up suggested items along the way. Amazon also believes it has leverage, since customers will likely come back to it anyway given its scale, so blocking Muse now buys time to negotiate a more favorable arrangement later. Shopify, by contrast, represents many small merchants who welcome the extra demand and don't have a competing ad business to protect, so its only real interest is preserving its payments rail. The broader point made is that months of talk from Google and OpenAI about agent payment standards mattered less than expected, because real change comes when a platform with actual consumer demand, like Instagram and Meta, starts hitting APIs directly, forcing everyone else to react. Amazon, Resy, and OpenTable are all expected to be scrambling to define agent policies as a result.
Agents Will Maim, Not Kill 17:31
The discussion turns to whether agents will destroy existing platforms like Amazon, Resy, or OpenTable, or simply weaken them. The view offered is that these systems of record won't be killed outright but will be "maimed," since agents are tireless, only cost compute, and will happily route around blocked or broken APIs to complete a task, whether that means calling a restaurant directly or finding a backdoor into a delivery service. A concrete example is given where an internal agent, faced with a vendor's price increase, unprompted proposed a twelve month plan to migrate away from that vendor entirely rather than tolerate it. Even a five to ten percent hit to Amazon's advertising revenue would be significant enough to move its stock meaningfully, echoing how a modest percentage shift boosted Meta's stock by over twenty percent. The comparison is made to how the early internet hollowed out inefficient middlemen, and AI is expected to do the same to anyone whose value is mainly reducing search costs rather than delivering something agents can't replicate themselves.
Jev Launches, Fast Cheap Classifier 22:01
Attention shifts to Jev, a new release from ChatGPT and Thinking Machines described not quite as a model but as a classifier built on top of an LLM, and reportedly the fastest launch the company has had. After a full weekend of testing, including a frustrating first day where nothing worked until prompts were restructured, it proved able to answer narrow, structured questions, such as matching two people for a meeting, in milliseconds at a small fraction of the cost of a full model like Anthropic's, sometimes a hundredth the price. Unlike a large language model, which returns computationally expensive text, Jev returns only true or false, a ranking, or a score, which is why it is described internally as "system one," referencing Daniel Kahneman's fast thinking versus slow thinking distinction. Estimates suggest roughly a hundred billion dollars is currently spent on LLM calls across OpenAI, Anthropic, and open source models, a figure expected to grow toward half a trillion dollars within five years, and about twenty percent of that spend is seen as better suited to a cheap classifier like Jev rather than a full foundation model. Because such tools can be five times cheaper, that twenty billion dollar slice could shrink to four billion in spend while still representing a real prize for whoever captures it. This is framed as a narrow but real erosion of foundation model revenue rather than a threat to advanced reasoning work, and it is expected to unfold mainly among software developers rather than everyday consumers, since ordinary users have no direct reason to touch Jev themselves.
The Exhausting Work Of Model Routing 29:00
Using a harness to pick which AI model to use for a task turns out to be exhausting. Every use case needs its own testing, because a model that gets one kind of question right, like whether two people should grab coffee, will fail badly on something more nuanced, like judging who between two candidates would make a better partner. Switching Replit over to an auto-router that shuffled between models like Astra and open-weight options led to performance degradation, forcing a switch back. The broader point is that picking models is becoming a full-time discipline of constant testing and quality assurance, almost a new kind of DevOps function, and if a router gets the model choice wrong even 20 percent of the time, a mission-critical coding feature ends up riddled with bugs.
Why Anthropic And OpenAI Ignore The Low End 31:31
OpenAI and Anthropic have deliberately left their cheapest models, OpenAI's Mini and Anthropic's Haiku, weak and unpolished, treating the low-cost end of the market as beneath their ambitions. OpenAI is described as ruthlessly commercial and willing to compete if forced, while Anthropic sees itself as building toward AGI and considers cheap commercial products a distraction from that mission. Both companies are said to have intentionally neglected the bottom of the market, viewing the recent narrowing of the gap with open-weight models as the real threat rather than losing budget customers.
Seed Rounds Are Getting Bigger 33:00
Seed rounds are no longer small, with $20 million or more now common for a first raise, and this has pushed investors like Andreessen Horowitz to launch something like a university for pre-inception investing with $40 million, echoing Peter Thiel's earlier Thiel Fellowship model. The reasoning offered is partly mathematical: nominal GDP growth of about two and a half times since 2010 means a $3 to 4 million check then should reasonably be a $10 million check now, and today's tech environment looks far stronger than 2010's downturn. But rising check sizes are also driven by fear of missing the next big winner, and a piece from Menlo circulating in the venture community warned about playing at the top of a cycle, a point one speaker calls somewhat obvious and condescending rather than genuinely useful.
A conservative fund manager's dilemma 43:01
One panelist describes a single seed investment made in 2021 that has since closed at a 2.3 billion valuation, a strong outcome from doing only one deal that year. He raises the harder question of what a cautious investor should do now: investing in a company worth 50 million and growing 60 percent a year, hoping it reaccelerates enough to be acquired, looks shaky because buyers like Bending Spoons only complete two to four deals out of a thousand reviewed each year. He worries there is simply no market anymore for slower, steady compounding companies, since capital seems to have shifted entirely toward AI-forward deals.
Chasing IRR versus disciplined pricing 46:31
The group debates whether it makes sense to chase managers who post spectacular IRR, using Sarah Guo's fund as an example, where the same company was priced at 50 million, then 350 million, then 2.5 billion, then reportedly 10 billion within four months. One side argues betting aggressively at the later, far more expensive rounds is intrinsically riskier than the same bet made early, even if the manager is excellent. Others counter that LPs will flock to managers with high IRR regardless, and that a model of doing many deals while staying capital efficient, similar to one described by Goku, can still work if the picker is skilled enough. They agree that sustained IRR above 30 percent is rare, that a 90 percent IRR year like 2021 cannot persist, and that being an LP today is harder because chasing recent returns risks missing the next great fund entirely, though they note private market returns show far more persistence than public market returns do.
Factory raises at five billion 54:31
Attention turns to Factory, an enterprise coding agent company known for its Droid product, which has just tripled its valuation to five billion dollars on a 200 million round. One panelist backs joining the round, citing rising demand for coding inference that current models cannot fully meet, and pointing to sovereignty and data trust as decisive themes from conversations at the Dreamforce conference, where enterprise executives said they do not trust Anthropic or OpenAI with their confidential data. He argues this distrust is genuine rather than manufactured online, and that wanting a choice of model and control over where data sits will keep pushing enterprises toward providers like Factory, making the round worth taking despite the froth elsewhere in AI valuations.
Coding as the biggest AI opportunity 57:30
The group agrees that coding is the standout value driver in AI today, calling it "the mother lode" because it produces roughly ten times the value of other AI applications. Corporate customers no longer trust OpenAI or Anthropic to be neutral custodians of their code, worrying less about the companies "blowing up the world" and more about their data retention policies. This has created room for standalone players like Cognition to pitch themselves to big enterprises as a safe, walled-off "software factory," separate from the model makers themselves. Even if AI adoption overall slows in the next year or two, the panel sees coding tools as one of the trends likely to keep compounding regardless.
Legal AI valuations and shaky margins 1:01:30
Discussion turns to Legora, which just hit 200 million in annual recurring revenue with a next round valued at 11 billion, and to Harvey, whose margins were reportedly reported by The Information at negative 50 percent. One partner says that number alone would make him nervous enough not to lead the next round, though he'd still back Harvey given its category strength, while voting against backing Legora due to undisclosed margin problems that surfaced abruptly before the investment committee meeting. Another partner argues the negative margin could actually be spun positively, as evidence that lawyers are using the product heavily before firms build their own internal models. The consensus is that AI for law remains a strong category long-term, even if it won't match software engineering in per-head spend, and that the real question now is valuation rather than the category's validity.
Data center bets and market swings 1:05:31
The panel debates Crusoe's 3.9 billion dollar raise at a 30.9 billion valuation, a company building data centers, GPUs, and managed inference with a 140 billion dollar contracted order book. One partner calls data center investing a spreadsheet exercise dependent on backlog, infrastructure access, and circular financing, ultimately passing on Crusoe as slightly overpriced despite liking the team. Another argues data center trades are highly leveraged and exposed to any AI usage slowdown, unlike coding or legal AI, making them riskier bets during uncertainty, though appetite would rise with stronger long-term commitments from major partners like Microsoft and longer debt runway. They close by noting a Wall Street Journal whiplash, from a morning headline warning about data center deal fears to an evening headline celebrating the NASDAQ's best day ever, all within the same eight-hour trading day.
Debating the China ownership attacks 1:11:31
The speakers debate whether Keith Rabois's attacks on Airwallex over alleged Chinese ownership are fair, noting the accusations shifted over time from claiming a CCP link to citing a cap table with more than 20 percent Chinese ownership, something one speaker insists is not true. They point out that many major companies, including Microsoft and Zoom, have employees in China, but also mention that Finn had to strip out all its open weight models before its acquisition by Salesforce closed, and that a portfolio company recently faced a similar demand to remove Chinese IP before an M&A deal could proceed, showing the issue has real commercial consequences even if the specific claims against Airwallex are wrong.
Calling for clearer government rules 1:15:00
One speaker argues this should not be settled ad hoc on Twitter by unelected people like Keith, Harry, or himself, but decided by the US government, which should set clear rules on trading high tech goods and handling risk with China, especially since both countries keep doing large amounts of commerce with each other, including leaders meeting in DC that same weekend to discuss expanding trade.
Jack needs his own advocates 1:15:31
Turning to Airwallex's Jack, the speakers argue his launch PR was excellent but he lacks a network of prominent backers, unlike founders with vocal supporters such as Keith or Joe Lonsdale, since his investors DST and Lee Fixel avoid social media. They say Jack should not be defending himself directly against Keith and instead needs an army of advocates handling the pushback, joking his communications team deserves an F and even suggesting he could hire Matthew McConaughey, before ending the show.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
