20VC with Harry StebbingsAMD Buys Fei-Fei Li's World Labs for $8.2B | Meta Poaches MongoDB's CEO | Bessemer Raises $5.75B
20VC with Harry Stebbings
Anthropic's leaked S1 draft 0:00
A draft S1 filing from Anthropic leaked, showing 4.6 billion dollars in 2025 revenue, an 8 billion dollar operating loss, and 518 billion dollars in compute commitments. The hosts argue the leak contained almost no useful information, since the 2025 numbers are already old news and the real decision point will be the Q3 revenue figures, which are 90 percent of what matters for pricing the company. One genuinely interesting detail stood out: two customers accounted for 25 percent of Anthropic's revenue, meaning someone spent roughly half a billion dollars with the company last year. The group also discusses how going public might make investors think more long term, in quarters or years rather than month to month, though one host worries that retail investors could fixate on the existential risk disclosures and massive losses, potentially souring sentiment even after a strong IPO, and that this disclosure of risk could even fuel political and legal backlash, pointing to a Florida case trying to use safety admissions against the company in court.
Instinct raises a billion at 10 billion 7:00
Instinct, led by 23 year old founder Noah Shin, closed a billion dollar Series C at a 10 billion dollar valuation, backed partly by Benchmark, whose partners Peter and EV led the deal. The conversation frames this as a new paradigm of consumer agents that act on your behalf across the internet rather than just answering questions, comparable in scale to the shifts chat and coding tools caused. The panel compares Instinct to Meta's Muse, noting Muse has impressed people and reinvigorated Meta's organization, but questions whether either product will become something people use for eight hours a day the way coding tools like Harvey or Logora already are for their users. They trace the agent concept back to OpenClaw, noting Muse's team has said they were directly inspired by it, and observe that these product cycles are shrinking fast, with labs now catching up to startups within weeks rather than the long head starts something like Cursor once had.
Benchmark's growth bet 13:00
The hosts note that Benchmark, a firm known since the 1990s for staying away from growth stage investing, broke that pattern decisively by writing a huge pre revenue check into Instinct at a 10 billion dollar valuation, after first backing the company at 2 and a half billion. One host jokes that Benchmark did not tiptoe into growth investing but went all in immediately, while a Benchmark partner clarifies that internally the firm actually viewed it as an early stage bet despite the size of the valuation.
Sizing a bet like Instinct 14:00
The conversation turns to how investors should size a position in a company like Instinct, treating it as a kind of Kelly betting problem where you weigh the odds of success against the potential payoff. One guest says that if a bet truly had fifty times upside with only one times downside protection, the math would say put thirty percent of the fund into it, but in practice no investor could stomach that, since limited partners would revolt. A more realistic answer offered is five percent, treating it as a full-size position within a twenty-deal fund, especially given that Instinct has only raised about a billion and a half dollars, has a strong team, and carries real downside protection since a large provider like Microsoft could plausibly acquire it.
Extreme dispersion in today's market 20:30
One speaker argues this is a strange moment where both valuations and real traction are simultaneously extraordinarily high, making it almost impossible to know if you are investing at the right pace. He admits that right now every venture investor is probably convinced the market is overheated while still wiring fifty million dollar checks to new AI labs. An anecdote is shared about economist Tyler Cowen being asked about venture at an annual meeting, who predicted that returns will become highly skewed, variance will rise with AI, and many investors will fail while a few get very rich. The group agrees this is why spreading bets across enough companies matters more than ever, and why even a pricey round like Instinct's does not yet call for a concentrated, all-in position the way OpenAI or Anthropic might.
AMD acquires Fei-Fei Li's World Labs 23:30
The discussion shifts to AMD buying Fei-Fei Li's World Labs for 8.2 billion dollars in stock, just two and a half years into the company's life, marking the first big exit among AI lab startups. The speakers praise Fei-Fei Li's personal story, noting she built the ImageNet project that helped spark the deep learning boom when a 2012 model using neural networks, created by Ilya Sutskever and others, dramatically outperformed rivals. They call her path from Stanford and Google researcher to startup founder a genuine immigrant success story. On AMD's side, the deal is framed as a modest price, about eight percent of its market value, to secure a world-class team and keep its stock's 279 percent yearly rally alive against Nvidia. The group predicts this will be the first of several such acquisitions, as major AI labs look to buy into robotics and world-model technology, even though skepticism remains about how many of the more than one hundred new AI labs that have raised over 70 billion dollars combined will ultimately find buyers.
A flood of billion-dollar acquirers 28:00
The speakers note something unusual about the current market, there are now roughly ten companies capable of and willing to do ten billion dollar acquisitions, which makes selling far easier and faster than going public. They point to World Labs and Fei-Fei Li as an example of how credentialism shapes these deals, since having a recognized AI pioneer and a proven, polished technical team matters enormously when a company becomes an acquisition target, even though scrappier teams can still thrive at the application layer.
Seed rounds have split in two 29:30
The conversation turns to fundraising sizes, with one guest noting that some teams now treat a hundred million dollars as a normal seed round, while another pushes back that two to three million dollars can still take a small team eighteen months down the road, especially with free compute credits from labs like OpenAI or Anthropic. They conclude there are really two separate worlds now, one where huge upfront capital is required to build a model or a chip, costing hundreds of millions to enter spaces like neoclouds or semiconductors, and another where lean teams ship products on ten million dollars or less because the expensive infrastructure already exists for them to build on.
Fast growth versus slow compounding 34:01
One guest argues that most of today's attention goes to fast, headline-grabbing growth stories playing out in one to three years, while a quieter, slower form of venture investing still exists, where a software company compounds steadily over ten to twenty years in markets like police departments or libraries. Others counter that this quiet compounding path is inherently unstable, pointing to a hundred-million-revenue board they sit on where competitors multiply overnight, and arguing that in a market reshaped by a 2022 discontinuity, slow stable growth without matching hypergrowth is a weak trade. The talent war is cited as a key reason compounding is hard to sustain, since top people are constantly pulled toward the most exciting, fastest-moving opportunities.
MongoDB's CEO leaves for Meta 39:00
The group discusses MongoDB's chief executive, who had taken the job less than nine months earlier, accepting an offer from Meta to run its enterprise division, a move that sent MongoDB's stock down twenty percent in a day before the previous long-time CEO stepped back in. The reported pay jump was roughly ten times his prior package, from about fifty two million to around five hundred million dollars, and the group frames this as proof that price and momentum in hot AI companies are now powerful enough to pull even sitting CEOs out of their jobs, sending a signal that pulls talent toward wherever the market's heat currently is.
Talent liquidity in AI 42:00
The hosts discuss why AI work is so alluring right now, pointing out that it is not just about money but about being at the center of attention, the product everyone is using and talking about. They note a historic level of talent mobility, with skilled people moving quickly to the most important opportunities instead of staying stuck in place, which they see as a broadly positive force for society, even if individual moves look confusing in the short term. California's culture of easy job switching, where someone can leave a company and start a new one the next day, is held up as the best use of that talent.
Mock investment pitch for Jev 44:02
The segment turns into a playful investment committee exercise around a company called Jev, said to be raising at a 10 billion dollar valuation just a week after a 200 million dollar seed round. One host argues for putting 20 to 30 percent of the fund into it, citing that Jev already handles 17 percent of traffic on Open Router and 20 percent on Versel's router, and that it is a 70th of the price and 100 times faster than rivals, framing it as a low-downside bet even if Jev itself does not end up the long-term winner. The group jokes about the risk of deploying too much capital too fast and about simply raising a new fund from limited partners if needed.
Inference boom and cost pressure 46:32
The conversation shifts to the broader inference market, noting big rounds for companies like Modal, at a 15 billion dollar valuation, and Base10, in talks at 26 billion. The view offered is that after a period where the best move was to invest in AI labs, the better trade became investing in inference providers such as Modal, Base10, Fireworks and Together, since unsustainable model costs are pushing demand toward cheaper, good-enough options. The idea of intelligence saturation is raised, using the example that a correctly filed tax return or a properly hammered nail does not improve by adding more intelligence, which supports continued growth in open source and inference usage even as frontier labs keep strong structural advantages in compute, users and monetization.
Peak open weight models 50:00
One host argues open-weight models have peaked in market share, for two reasons: Anthropic and OpenAI can freely cut prices on their non-frontier models to stay competitive, as shown by a new Sonnet model priced only 20 percent cheaper after one week with room to cut further, and enterprise customers, observed firsthand at Dreamforce, are increasingly uncomfortable running Chinese-origin open models due to data and security concerns. Another host pushes back that the real issue is specifically Chinese open models rather than open weights generally, and raises interest in a low-cost, US-based open alternative. The discussion closes by framing compute ownership, roughly a gigawatt-scale footprint split among major inference providers, as the dominant factor determining long-term token and revenue share, just as OpenAI reopens its 200 dollar plan at reduced capacity after running out of compute.
Hard to Predict Token Efficiency 56:01
The hosts admit that trying to calculate whether AI delivers more intelligence per gigawatt is nearly impossible, since it means multiplying several numbers that each carry huge error bars. One host ran his own test on the newly released Sonnet model for an app he is building called Saster Connect, finding input tokens were 42 percent higher, output tokens 44 percent higher, quality improved, yet cost only dropped about 10 percent instead of 30. Given how hard this is to measure even for a single task, they argue the only real signal is watching how much money big buyers are actually allocating, noting that some spent half a billion dollars last year on a single AI vendor, which suggests they have run the numbers and found it worthwhile.
Aura Pulls Its $16 Billion IPO 58:30
Consumer company Aura unexpectedly withdrew its planned IPO just before pricing, despite having banks Morgan Stanley, Goldman Sachs, and JP Morgan on board and being a profitable, well known consumer business. Early investor 4Runner had said it would sell its entire position, an unusually confident move for an IPO, which made the later pullback more puzzling. The hosts debate whether the deal collapsed because secondary sellers would not accept a lower price, comparing the process to a bobsled run where once you file the S1 there is little room to back out. They note the irony that this happens amid record market highs and heavy M&A activity, and they wonder which AI native company will be the first to test public markets, guessing many are waiting to see how the first mover fares, possibly pushing real IPO activity into 2027.
New Bank Eyes Monzo Acquisition 1:03:30
British digital bank Monzo is reportedly being eyed for acquisition by New Bank at a price between 8 and 12 billion dollars, a move the hosts call surprising since it pulls New Bank's focus away from winning the US market. They reason Monzo wants to sell because it is too small to matter in US public markets and Europe's public markets are weak, while New Bank's interest might reflect a strategy of targeting inefficient banking markets like Brazil and the UK rather than the hyper competitive US. They also mention New Bank's stock fell on the news and is down over 12 percent for the week, prompting one host to jokingly buy shares in real time. The segment closes by noting Monzo's own boardroom turmoil, where founder Tom stepped back, a new CEO was swapped in by the chairman, investors pushed back, and the chairman eventually retired.
Board Turmoil After Leadership Swap 1:10:00
The conversation opens on a boardroom drama where investors backed a CEO, not the chairman, and were blindsided when that CEO was swapped out. The company had been losing 100 million while doing 100 million in revenue, then a new leader identified only as TS turned it around, only to be pushed out himself a few months later. The speakers describe venture investors reacting with alarm, since they had committed capital based on trust in a specific operator, and that kind of instability is seen as exactly the setup where an outside buyout offer becomes an easy yes.
Fund Strategies at Opposite Extremes 1:11:02
Bessemer's new 5.75 billion dollar fund and NFX's decision to stop taking outside investor money are framed as two different bets. Bessemer's move is tied to its growth-stage win backing Anthropic, reinforcing a pattern of pouring more money into what already works, while the group jokes that a seed round now needs to be imagined at 30 million, meaning even a billion dollar fund starts to feel small. NFX going all-in on its own partners' capital is compared to Homebrew's earlier move, freeing a firm from explaining itself to outside limited partners, though the trade-off is that building a lasting firm with junior staff and career paths usually requires outside money to pay salaries and give people something to work toward.
Direct Investing and Founder Voting Control 1:17:00
One host admits he considered skipping funds to invest directly but says his comfort zone is five million dollar checks, not fast hundred thousand dollar ones, so a traditional fund fits him better. The talk closes on Anthropic's founders locking up 50.1 percent voting control, with the group puzzling over why it took so long, guessing it stems from pre-IPO voting structures resetting once shares convert to common stock after going public.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
