The Twenty Minute VC · 1 Oct 2026 · From the week of 28 September
20VC: Instinct Raises $1B at $10B Valuation | AMD Buys Fei-Fei Li's World Labs for $8.2B | Meta Poaches MongoDB's CEO | Bessemer Raises $5.75B | Oura Pulls IPO & Nubank Eyes $8–12B Monzo Takeover
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Harry Stebbings hosts his weekly news roundup with Rory O'Driscoll and Jason Lemkin, joined by Jack Altman of Benchmark. They cover Anthropic's leaked draft S-1, Benchmark's participation in Instinct's $1B raise at a $10B valuation, AMD's $8.2B purchase of World Labs, Meta hiring MongoDB's CEO, a startup seeking a $10B valuation a week after its seed, inference companies, open-source models, Oura pulling its IPO, Nubank's possible Monzo deal, new venture funds and Anthropic founders' voting control. The central argument is that AI has made venture a high-variance, high-dispersion market. Early-stage risk now needs late-stage capital, many acquirers can write $10B checks, talent moves to wherever price and attention point, and the traditional small seed round and quiet-compounder model look strained in many segments, though the speakers disagree on how much.
For founders
- Jack Altman reports that well-networked founders in hot categories now often skip the traditional '6 at 40' seed and raise very large first rounds, and he believes the classic $3-6M seed check for 8-15% ownership is broken for that cohort.
- Jason Lemkin argues that $2-3M can still take a team of three or four about 18 months, especially with free model credits, so the natural seed size for many software companies has not really changed.
- Jack Altman says roughly ten companies can now do $10B acquisitions and want to, which makes M&A a quicker and easier exit than an IPO for AI lab-type companies.
- Jason Lemkin warns that 'quiet compounder' software companies are less stable than before because they must ship far more software, far faster, against many more funded competitors.
- Rory O'Driscoll's account of Oura says that once an S-1 is public, pulling the IPO near pricing is the hardest move to make, and a large secondary component makes selling VCs very price-sensitive.
For revenue leaders
- Jason Lemkin reports from Dreamforce that none of the enterprise customers he spoke to were comfortable running Chinese-origin open models, and he expects C-level anxiety about this to rise as security concerns grow.
- Rory O'Driscoll argues that the only reliable signal of AI value is buyers' actual budget allocation, because token-cost and token-volume forecasts compound large error bars.
- Jason Lemkin's own test of a new model version found input tokens up 42% and output tokens up 44%, with quality higher but total cost only about 10% lower, so per-token price cuts may not translate into cheaper tasks.
- Jack Altman expects many tasks to reach 'intelligence saturation', where a cheaper model is good enough, and says this favours open-source models and inference providers for those workloads.
What was said 54, most useful first
Benchmark treated its $10B Instinct investment as an early-stage bet, because the company is days old even though the price looks like growth. Listen
Benchmark first invested at $2.5B and then at $10B pre-revenue. He said that in this cycle 'early stage risk requires super late stage capital'. Judged as growth, public companies with $2.5B ARR trade at $10B. Judged as early stage, the paradigm is too new to model, with several more evolutions expected before Christmas. He said Benchmark's thinking was closer to the early-stage view.
“in this cycle, we've had investments with early stage request. risk requires super late stage capital which is just definitely a strange time to be playing”
Bottom-up market sizing for consumer agent monetisation, such as travel and fine dining, looks small, so these valuations implicitly rest on a bet that 'if this matters, good shit will happen'. Listen
After listening to the Instinct founder's interview on monetisation, he checked how few Americans fly more than once or twice a year and noted that fine dining is about 10% or less of dining. He pointed to OpenTable as roughly a $1B business after 26 years. He concluded there is an implicit assumption of something bigger beyond the stated revenue lines.
“if you think of it as travel monetization, you could get pretty angsty about market size here.”
Rory O'Driscoll frames fund position sizing as Kelly betting: size by edge over odds, then decide whether to bet full Kelly or back off. Listen
For a fund doing 20 deals, about 5% each on average, he said capital needs might push one bet to 10% while risk might pull it down to 2.5%. He described a deal like Instinct as the hardest to size, with high upside pulling in, very high variance pulling out, and a need for meaningful dollars. He later said 5% in a 20-deal fund is already a 'full whack' position.
“you do some kind of Kelly betting and then you kind of, you know, how certain what's the upside and how certain are you it's going to happen edge over odds. And then you say, do I have the stones to do full Kelly or do I back off?”
AMD's $8.2B stock purchase of World Labs, about 8% of its roughly $1T market cap, is cheap if the team helps keep its run going. Listen
He noted AMD is up 279% this year and needs to stay relevant as number two to Nvidia. He said he did not know whether AMD will continue the 3D world models or use the team as its counterpart to Nvidia's model team. He argued that acquirers on big runs will spend a slice of market cap to maintain momentum, and that AMD would not do this if it were up only 3%.
“I'd certainly invest 8 % of my market cap to keep the good times going.”
A seed round should still be about $2-3M, enough for three or four people to get 18 months down the road. Listen
He accepted Rory's estimate of roughly 2.5x nominal venture inflation since 2010, but said the work a small team needs to do has not changed. He pointed to free model credits (up to $0.5-1M from labs) as a way to stretch the money. He said founders can do as much with $2-3M as ten years ago if they are frugal.
“the truth is, you can do as much, I think, for two to three million bucks as you could 10 years ago.”
Well-networked Silicon Valley founders now skip the traditional seed, which leaves the $3-6M check for 8-15% ownership 'fully broken' in that part of the market. Listen
He gave two reasons first rounds have become extreme: labs genuinely need $200M+, and well-connected founders can raise a $50M seed even when they don't need it. Those founders skip the old '6 at 40' round or raise about $300K for a month and then go big. He said traditional seed investing in that cohort 'just isn't there anymore'.
“there is a cohort of the market where traditional seed investing where you're going to write three to six million dollar checks by, you know, eight to 15%, where I just think that is fully broken slash just isn't there anymore.”
Jason Lemkin's own test of a new model version found it used far more tokens, so total cost fell only about 10% despite cheaper per-token pricing. Listen
Running his own informal evals for an app he is building (SaaStr Connect), the new Sonnet 5.5 used 42% more input tokens and 44% more output tokens. Quality improved and it passed more blind tests. He concluded that task-level cost is hard to predict.
“Input tokens, 42 % higher than before output tokens, 44 % higher.”
A large secondary component makes an IPO harder to price, because selling VCs care about the price far more than a company taking primary dilution. Listen
Discussing Oura pulling its IPO at a planned $16B, he noted that Forerunner, one of the largest investors, had announced it would sell its entire position, something he had never seen. A company taking 10% dilution barely notices leaving money on the table. A VC selling at $18 instead of an expected $22 sees its whole return fall by 20%. His 'positive version' was that investors liked the deal but would not pay up, and the sellers preferred not to transact at that price.
“it's always harder to get a deal done when there's secondary. And when the more secondary there is, the harder it is to get a deal done.”
One panellist argues that the UK-style split of chairman and CEO roles suits boring public companies but makes no sense for venture-backed companies. Listen
They cited Monzo's board turmoil: the chairman replaced a CEO whom investors had backed, investors objected, the decision was reversed and the chairman retired. In the discussion it was noted that the replaced CEO had turned the company around after the founder stepped back. The speaker suggested this board instability made selling more attractive.
“it's that UK chairman plus CEO role I've seen, which makes intuitive sense. for well -governed, public, boring companies. But in my view, it makes absolutely no sense for venture -backed deals.”
If a seed round is $30M, even a $1.75B seed fund like Bessemer's looks small once 1:1 reserves are set aside. Listen
Bessemer raised $5.75B, including a $1.75B seed and early fund; Harry Stebbings estimated that seed fund at about $40M a year in fees. Jason asked how many $30M seed rounds such a fund can do with one-to-one reserves if it cannot count on recycling. He said the math makes sense even if the returns may be harder.
“Even if billion starts to sound small for seed funds, if you believe $30 million is a seed”
The leaked Anthropic draft S-1 contained no useful information apart from customer concentration, and that the upcoming Q3 revenue number matters far more. Listen
He called the 2025 figures ($4.5B revenue, about $8B compute expense) already priced in and old news. He said the one interesting fact was that two customers made up 25% of revenue, implying one customer spent about $0.5B on Anthropic. He said the Q3 revenue number is 90% of the data needed to decide on Anthropic's pricing.
“The only interesting factoid in that was that two customers did 25 % of the revenue, which means someone spent $0 .5 billion on Anthropic last year”
Jason Lemkin has a 'minor worry' that Anthropic's IPO will be heavily oversubscribed and then drift below the offer price as the negative disclosures come to dominate. Listen
He compared the risk to the Facebook IPO, where retail attention went to the negatives first. He pointed to the existential-risk factor, the large losses, data-center spending and credit concerns as items that could overwhelm the positives a month or two after listing, even with no real change in the business. He framed this as a hedged worry, not a firm call.
“I worry the IPO will be successful, they'll hit their number, whatever they ask for, it'll be more than over -subscribed than Oura. you know, 10X, but then a month or two in with no real change, we may see it drift below the IPO price”
Rory O'Driscoll sees a political and legal backlash forming, in which the labs' own risk disclosures are used against them in lawsuits. Listen
He cited a lawsuit in Florida seeking to stop OpenAI and Anthropic from releasing models. He argued that companies listing existential risks in legal documents to protect themselves invites the response 'you told me it wasn't safe, so stop', including from state attorneys general. He expects the AI IPOs to be a far more public, opinionated moment than a typical semiconductor listing.
“There's going to be a whole load of, you told me in a legal document that this thing was not safe. Therefore I'm entitled to believe you and if it's not safe you should stop.”
Jason Lemkin's test for which AI agent products win is whether people run them all day long, and he is unsure whether personal agents like Instinct or Muse have reached that yet. Listen
He says he runs coding agents 10 hours a day and points to lawyers using Harvey and Legora constantly as examples of all-day use. He called Instinct and Muse 'generation two' after OpenClaw, which few people could run safely. He said it is an open question, not a criticism, whether these become 'eight hours a day' apps, and that if they do, they win.
“Will we run Instinct Muse eight hours a day? If we do, I guarantee it wins, right?”
Legora and Harvey had light customer usage at $1M ARR and now have customers who run their whole working lives in the product, so early usage can understate a category. Listen
He said reference calls at $1M ARR would have shown limited use. Now customers say 'I run my whole life out of it', and the businesses are at hundreds of millions of ARR. He made the same point about coding agents and about early ChatGPT usage, and applied it to personal agents being early in their cycle.
“if you did the reference calls on their customers, when they were at a million of ARR, the And now you talk to them, and they're like, I run my whole life out of it.”
Jack Altman's framework for sizing a high-priced early-stage position is that the fund needs enough shots on goal (the lower constraint) while the check must still matter in the round (the upper bound). Listen
He said that however good you think your picking is, an early-stage fund needs enough investments to land something that really matters, which caps any single bet. Big rounds at high valuations, on the other hand, require more dollars to get meaningful ownership. He presented these as the boundaries Benchmark tries to stay within, and said that unlike sophisticated public investors, VCs do not spend as much time on sizing as on picking.
“the sizing is constrained by wanting the fund to have enough chances to get something great. And so that's the constraint. And then like the upper bound is, you know, these rounds are big, you need to matter in the context of the round.”
He would put 5% of his fund into a deal like Instinct at $10B, and that at his career stage a 30%-of-fund bet would alarm his LPs. Listen
He said he sees about 50x upside and downside protection of roughly 1x or more, because the company has raised only about $1.5B, has a strong team and is a likely acquisition target for a large provider. He noted that founders often miss that concentration tolerance depends on the GP's tenure and LP relationships.
“for me, at my stage of career, if I did a 30 % of the fund bet, my LPs would shit the bed.”
With valuations and outcomes both extreme, VCs are almost certainly investing either far too fast or far too slow. Listen
He described the current moment as having much more dispersion than other times: prices are extraordinarily high and so are traction and outcomes. Because both sides of the equation are so out of balance, he said the odds of getting the pace exactly right are zero.
“we're definitely not investing at the right speed. We are either investing way too fast or way too slow. But when both sides of the equation are this out of whack, the odds of having it right are zero.”
What worries him is not a high price, but three rounds in three weeks with no material change in the business between them. Listen
He said he assumes Instinct's cohort and depth of usage are exceptional, so its round does not worry him. He argued that dismissing '$10B for 14 people' repeats the mistakes of people who mocked the Instagram and WhatsApp prices. The red flag, in his view, is rapid repricing with no new data.
“What worries me is when you have three rounds in three weeks with no material movement in between and no data suggests there's been anything different.”
One panellist expects a wave of large acquisitions of world-model and robotics foundation-model startups over the next 6-12 months if markets hold, even though these companies lack a quick standalone path to revenue. Listen
The speaker said World Labs-type companies would not, on their own, follow the fast monetisation path Anthropic and OpenAI had through chat and then coding, which made them nervous about these deals. They now think the big foundation-model companies are probably in the market for a robotics foundation-model story and chipmakers want relevance against Nvidia. They noted there are still about 100 neolabs, so a company has to be in the top 10.
“I think there will be a bunch of these big -ass acquisitions over the next six, 12 months if the market continues to hold.”
Harry Stebbings cites a report counting 102 neolabs that have raised over $70B, and questions how many can realistically be acquired. Listen
His partner Paul wrote the report. Harry argued that only perhaps 10-12 acquisitions can absorb that many companies, though he expects the outcome to be okay overall.
“he wrote this report on 102 Neolabs, $70 billion plus raised. And my question is just like, just how many of them can get acquired”
About ten companies can now do $10B acquisitions and want to, which makes M&A a quicker, easier exit than an IPO. Listen
He said this is different from any time he has seen. M&A avoids long IPO cycles, and that has changed the calculus for infrastructure and lab-type companies. Given World Labs and other data points, he said he is hesitant to be too skeptical of neolab outcomes.
“There are like 10 companies that can do $10 billion acquisitions and want to. And that's just so different. And it's much easier than going public, and it's quicker”
His team says it cannot find relevant rounds under $100M, and that a top CIO questioned whether anyone can play in venture with less than a $1B fund. Listen
He noted World Labs' first round was about 65 (million). His team clarified they meant $100M round sizes, not $100M seed valuations. He also argued that talent costs are so high, because of talent's alternatives, that a $2-3M round no longer works.
“they we can't find anything under 100 million. I said, wow, seed prices are expensive. They said, no, no, 100 million round size.”
There are two capital worlds: neolabs that need hundreds of millions to start, and application companies built on those labs that ship on very little capital. Listen
He described app companies that raised $10M, shipped the product for $3M, still had money in the bank when customers took off, and then raised $50M because they could. He said the heavy upfront technology lift funded by the labs is what lets everyone else build on relatively little.
“we took ten million dollars, but we shipped the product for three million bucks. And then the customers took off and shit, we still got five million bucks in the bank, but we're going to raise 50 anyway, because we can.”
Entry capital now runs to a couple of hundred million dollars minimum for a neolab and about $500M for semiconductors. Listen
He agreed a $2-3M seed can still work in markets where a product can ship by building on others' infrastructure. He said a semiconductor company can need $100M before tape-out at a $2B valuation.
“it's a couple of hundred million minimum to enter the Neolab space and at five hundred million to enter the semiconductor space.”
Jack Altman probably believes the classic small-fund math (buy 10%, hope for a $1-2B outcome) still exists, but in slow-compounding companies outside the headlines. Listen
He described rounds like a $3M raise at $30M for companies that are in unfashionable markets, have an awkward shape, pivot, or get unexpected traction. His example was software that quietly compounds selling to police, fire departments or libraries through 2041. He hedged: 'I probably do.'
“do I think that there's a three at 30 round happening today, where in 2041, that company is just going to have quietly compounded in the market of police or fire departments or libraries? Like, yeah, I probably do.”
Rory O'Driscoll estimates today's venture opportunity set is about 80% fast-moving AI companies and 20% slow compounders. Listen
He said 2022 was a discontinuity that made what came before obsolete, so anything four years in is fast-action by definition. He wants to believe in slow compounders because his firm has made great bets on them. Harry Stebbings, by contrast, called slow compounders an anomaly that he cannot bet on.
“I do agree that the table at the moment is 80 % the fast action table and 20 % the slow action, which makes sense because in 2022 there was a discontinuity and everything before that became obsolete.”
A slow-compounder fund strategy suffers on opportunity cost, because its numbers look poor for years while LPs can back funds posting record returns. Listen
He said the slow model leaves a fund's numbers 'crap for quite a long time'. Meanwhile LPs can try to get into top firms such as Sequoia and Benchmark, which are posting never-before-seen numbers.
“if you want to go for that model, your numbers will be crap for quite a long time. And we always forget that we're in an opportunity cost game”
'quiet compounder' software companies are now unstable, because they must build far more software faster against many more funded competitors. Listen
He said he has pitched quiet compounding since 2012. Owner.com, where he and Jack Altman sit on the board, is past $100M in revenue but has a huge amount of software to build this year. Another of his $100M portfolio companies showed a competitor slide at its last board meeting where he had never heard of eight or nine of the companies, many backed by LP capital.
“They put up a competitor slide to last board meeting. I never heard of eight or nine of”
One panellist argues that a compounding business should come with low risk in exchange for lower growth, and that compounder-level growth with hypergrowth-level risk is a suboptimal game. Listen
The point was made in response to Jason Lemkin's instability argument. It frames the trade-off as compounding growth having to buy you low risk, which today's fast-changing tech environment no longer delivers.
“if you're making a compounding player, the quid pro quo should be low risk. And if the world is such that the tech environment is changing so much that you get the compounding, not the hyper growth, but you get the same level of risk that by definition is a suboptimal game.”
Current AI model costs are unsustainable for heavy users, which creates an opening for much cheaper, faster models. Listen
In a mock investment-committee pitch for an AI model startup seeking a $10B valuation a week after its seed (rendered as 'Jeff' in the transcript), he said it already has 17% of OpenRouter traffic and 20% through Vercel's router, at one-seventieth of the price and 100x faster. He said that even with newer model versions getting cheaper, spending at current costs 10-12 hours a day is unsustainable. He acknowledged it may not be the winner next year.
“These AI costs are unsustainable. It doesn't matter if Sonnet 5 .5 and the latest Opus is cheaper. It is unsustainable to spend these costs 10, 12 hours a day”
Rory O'Driscoll's rough math suggests a cheap inference-model startup could reach about $1B in revenue by capturing existing spend at a fraction of the cost. Listen
He estimated current AI spend at roughly $100B, of which 20% ($20B) is relevant to the startup. Compressing that 5:1 leaves about $4B of accessible revenue. Because developer adoption has been very fast, he said a $1B revenue line could be reached quickly by saving customers 80 cents on the dollar.
“Spend right now, today, is roughly $100 billion. And you do the analysis, and 20 % of that is relevant to Jeff. So that's $20 billion of accessible revenue. Say they compress it five to one.”
For the last 18 months the right venture bet was inference companies, which act as an index on everything outside the labs. Listen
He said there was a period when the right answer was simply to keep buying the labs at every round, and that more recently it was inference: Modal, Fireworks, Fal and Together have all worked. He quoted his partner Eric Vishria: 'it's all gonna work', meaning a crazy number of things are working, though not every company or sector.
“Inference has been a really great way to get an index bet on everything outside the labs.”
Many tasks are reaching 'intelligence saturation', where more model capability adds nothing, which pushes those workloads toward open-source and inference providers. Listen
His examples: a tax return is either filed correctly or not, and a hammer only needs to drive the nail. He expects more open-source usage and strong inference companies as a result. He added that the labs are structurally cost-advantaged in compute, users and ways to subsidise, so open source will be a big part of the market but will not dominate.
“Once you have filed it correctly, throwing more intelligence at that problem, doesn't do you any good.”
Open-weight models have peaked in market share and will decline, even if total inference volume keeps growing. Listen
He gave two reasons. Anthropic and OpenAI can price their non-frontier models as competitively as they like; he said a new Sonnet version was already about 20% cheaper. And enterprises fear lab training on their data but are even less willing to use Chinese-origin open models: at Dreamforce, no customer he spoke to was comfortable doing so. He expects C-level anxiety to grow with security concerns.
“I just got back from Dreamforce and I got to tell you, I know there are a No one wants to run on open source models there. At least Chinese China -based, nobody.”
Rory O'Driscoll suggests, without certainty, that the labs may be most willing to price-compete on commodity tasks when they have fewer frontier opportunities to spend compute on. Listen
He framed each lab's marginal decision as allocating the next chunk of GPU either to frontier work, such as a biology model, or to finite tasks like tax returns. He said he did not know the answer but called the trade-off very interesting.
“provided you have opportunity at the frontier, you won't waste time with the second string stuff.”
Jack Altman expects more US enterprises to post-train their own models on open weights and run them through inference companies. Listen
He said there is real anxiety about non-American models. Even so, he expects enterprises to build on open weights, use inference providers to make their own models, and then run those models themselves. He said he does not know how much impact this will have.
“I also think we are starting to see and will continue to see a lot of enterprises post train their own models. and draft off of open weights and use the inference, you know, companies to make their own models and then run them themselves.”
AI market share will largely come down to who owns compute, with compute share roughly tracking token share and revenue share. Listen
He estimated all inference clouds together at roughly a gigawatt, versus high single digits of gigawatts each for OpenAI and Anthropic. Efficiency and pricing differences matter at the margin, but in his view it comes down to who owns 'all of the computers firing all the time'. In the exchange that followed, he agreed compute share likely maps to revenue share within about 2x.
“on some level, this will also just come down to like all of the computers firing all the time and like who owns it”
AI economics forecasts multiply three numbers with large error bars, so the only actionable signal is how buyers actually allocate budget. Listen
He tried modelling falling token cost, rising token volume and token efficiency, and concluded honestly that he does not know the answer. He pointed to a customer spending about $0.5B on Anthropic as evidence that someone ran the numbers and is getting value.
“you're multiplying three numbers, each of which has an error bar that's pretty damn large. And if you have any intellectual honesty, at the end, you kind of go, I just don't really know.”
Rory O'Driscoll describes the public S-1 filing as the point of no return: before it you can do what you want, and after it pulling the IPO is the hardest move. Listen
He compared it to a bobsled run that, once started, has very few easy exits. He said he has been at IPOs that nearly pulled on the last day. He added that pulling is not fatal for a profitable consumer company like Oura, and would be harder for an enterprise company.
“the minute you unveil the S1, the minute it goes public, you're jumping in that bobsled and you're sliding to the bottom and there's very few easy way out.”
Jack Altman expects AI-native companies to start going public in 2027 if markets hold, and thinks they would trade well because public investors lack exposure. Listen
He said none has gone first because it is scary to list before the labs and before seeing how markets react. Many could go public tomorrow. For them it is a matter of choice and of whether management and boards are happy with the price.
“I think if the market holds on, I would expect that in 2027, there will be quite a few of these.”
Rory O'Driscoll tentatively argues that neobanks do best in markets with inefficient incumbent banks, which might explain Nubank's interest in Monzo. Listen
He framed it as more surprising that Nubank wants to buy than that Monzo wants to sell; Harry Stebbings added that Monzo is too small for US public markets and European public markets are poor. Rory contrasted Chime, trading at about $5-6B in the efficient US market, with Revolut thriving against old-school European banks, and Nubank's success in Brazil. He flagged that he was making the argument while expecting it might be wrong, and noted Monzo covers the UK, not Europe.
“the problem with the U .S. as a neobank market is we're just so damn efficient in terms of our banks that there's not a lot of fat profit to be taken”
One panellist argues that acquirers pay up to save time, which is how 'A-minus, B-plus' venture assets with a strong position get bought. Listen
On Nubank and Monzo, the speaker said the buyer is purchasing instant presence in the British market, not just revenue and customers. They said this dynamic is needed for venture to work, because companies that are not growing fast but hold a position can still find buyers.
“It's the only way people are going to buy the A minus B plus assets is to save time.”
Investing only your own capital, as NFX is moving to, frees a firm to make bets without explanation, such as $100K checks that are hard to defend out of a seed fund. Listen
He noted Homebrew made the same move and it worked out wonderfully. He said non-financial motivations, such as not caring about ownership targets, drive some of these choices. He also noted that LP capital means returns go to hospitals and endowments rather than to one principal, which some people care about.
“When you're not managing external capital, you get to do things without any explanation to anybody.”
If your goal is building a firm, you need outside LP capital even if you invest heavily yourself. Listen
Asked why Peter Thiel did not run Founders Fund purely on his own money, Jack said firm-building requires outside capital, salary to pay people and tension with stakeholders outside the firm. His suggested approach is to put in as much of your own money as you can and take as much LP capital as you need. Jason Lemkin added that junior people want careers, which also requires outside money.
“if your aspirations are firm building, then you would say, I'm going to put in as much of my own money as I can because I believe in the strategy, but as much LP capital as we need.”
He hasn't switched to investing only his own money because he wants to write $5M checks to be relevant, which is too much balance-sheet risk. Listen
He said Homebrew advised him early on to skip raising a fund and invest directly, and that advice still echoes with him. He gets little joy from $100K checks, and said that if he did, he probably would invest directly. Whatever the structure, he said, you have to be able to fund the check size you are best at writing.
“If I want to be able to write a five million dollar check to be relevant, that's too much for my balance sheet. It's too much risk.”
Rory O'Driscoll suggests founders often have voting control pre-IPO, but it typically expires when preferred stock converts to common at IPO, so control has to be re-papered. Listen
Explaining why Anthropic's founders are only now moving to lock in 50.1% voting control, he said voting rights in pre-IPO preferred structures typically expire on conversion. After that, control is based purely on ownership. He was speculating ('I'm willing to bet') rather than speaking from inside knowledge.
“it's when you convert everything to common stock that typically voting rights expire on the IPO. So what my guess is they had a pre IPO deal and now you got to recreate a post IPO deal because everyone's cap structure changes.”
Public-market investors will take a longer-term, more structural view of the AI labs than private investors do. Listen
He said private investors are 'hand wringing month to month' over the narrative around Anthropic and OpenAI. In his conversations with public-market investors, he found them thinking in quarters or even years and focused on market structure. He hopes this will make the discussion calmer.
“I actually think the public market investors will be a little bit more long term oriented than the private investors.”
Personal agents acting across the third-party internet are a paradigm shift big enough for independent players to win alongside the labs, as happened in coding. Listen
Benchmark (his partners Peter and Ev led the deal) invested in Instinct's $1B Series C at $10B despite competition from Meta's Muse. He compared the setup to Cursor and Cognition competing with the labs in coding. He said that when a category is that important, 'a lot of things can win', and that the labs will also have offerings.
“when something is that important, a lot of things can win. And so my view is basically something like there can be an amazing independent player like instinct.”
The agent form factor takes time away from chat usage, though that does not mean agent products cannibalise the labs as companies. Listen
Harry Stebbings said he had moved from ChatGPT to living in Instinct. Jack agreed the form factor draws on a fixed number of waking hours already spent online. He separated this from whether particular companies lose, since the labs may build agents themselves.
“the form factor cannibalizes. That doesn't mean that these products cannibalize the other companies, but it does mean that the form factor takes some amount of the space”
An economist told one panellist's LP meeting that AI will make venture returns more skewed and higher-variance, and many firms will fail. Listen
The speaker said they hosted the economist Tyler Cowen at an annual meeting. Asked about venture's future, Cowen answered that returns will be highly skewed, variance will rise with AI, and many firms will fail. The speaker also noted that VCs complaining in WhatsApp groups that the market is broken still wire $50M checks to neolabs.
“returns are going to be highly skewed, variance is going to go up with AI and many of you will fail.”
Price is sending a market-wide signal to pile into hot AI companies, and capital and talent will follow it. Listen
He invoked Hayek: money and price are signals, and right now they point everyone to the same place. He later applied the same logic to MongoDB's CEO leaving to run Meta's enterprise division.
“money is a signal, price is a signal, and price is sending a signal, everybody go right here.”
Jack Altman calls today an extreme talent war, and argues the unusually high movement of talent is probably good for society even when individual moves look strange. Listen
Discussing MongoDB's CEO leaving after less than nine months to run Meta's enterprise division (MongoDB stock fell 20% in a day), he said money matters but so does attention: the products in the headlines that people's families talk about. He said talent feels 'very unstuck' and is flowing to the most important opportunities, unlike past periods when great people stayed stuck.
“One of the benefits of it is that talent feels very unstuck right now.”
Momentum and a hot stock let a company make offers compelling enough to pull a sitting public-company CEO into a non-CEO role. Listen
He noted that MongoDB is a roughly $20B public company and that people who have been CEO rarely go back to not being one. He inferred that the Meta offer must have been extraordinarily compelling. His point was that whoever has momentum can take whatever talent it wants.
“When you have the momentum, you can make people offers that just allow you to take whatever talent you want.”