Free with a sign-in.
The best 16 min of 63
Anthropic at $2 Trillion? The AI Bubble, and Why the Best Startups Always Look Overvalued
Matthew Harris argues that venture power-law winners often look overpriced, while AI still leaves room for new services companies and undiscovered leaders.
How I Invest · David Weisburd
- Matthew Harris is a Partner at Bain Capital Ventures focused on financial services and technology.
The best minutes
- 1:58Valuation discipline can make venture firms miss power-law winnersPro pick 5 min
- 17:35AI labs may avoid competing with their biggest customers6 min
- 32:44Anthropic's growth could support its reported $2 trillion valuation4 min
Every moment plays free. PRO marks a write-up that's in Pro.
Also worth watching
- 11:53Bain's AI strategy centers on robotics and AI-native services6 min
- 8:51AI market leadership can shift despite consensus investing3 min
- 27:59OpenAI's open Codex strategy could win the application layer5 min
- 52:38Steady venture pacing helps limit cycle-driven risk5 min
- 39:48AI agents could expand the market beyond human users3 min
Skip the roughly five-minute parenting discussion, the sponsor reads and the personal AI-use examples.
Full analysis · 6 min read
Real Alpha · Nov 7
Matthew Harris says Anthropic's $2 trillion valuation is reasonable
- Matthew Harris said a reported $2 trillion Anthropic valuation looks reasonable if growth continues.
- But he flagged uncertainty about continued growth and seller supply when lockups expire.
- For founders and investors, separate operating growth from liquidity and test the assumptions behind scale.
Free with an account
Read the rest of the analysis
6 sections, the 5 other voices it tests Harris against, and what it means for you. Every episode we cut gets one.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.
Valuation discipline can screen out power-law winners
Harris's portable point is not that high prices are safe. Venture's payoff shape can make ordinary valuation discipline dangerous. Bain Capital Ventures, he said, carried some loss-avoidance habits from its parent into venture.
That is not a blanket case for ignoring price. BCV led Crusoe's pre-seed and seed rounds, and Harris said the unrealized company will surely return multiples of its fund. But that remains Harris's own confident forecast about an unrealized holding.
For example, Harris sketched a hypothetical AI application deal priced at what he described as roughly the average multiple, about 20 times revenue. Databricks, Stripe and Ramp, he said, were called overpriced at every round, yet became massively valuable.
The paradox is that investors may treat a high price as evidence of quality, then raise their ceiling again. If everyone believes winners must cost more, prices can rise without better selection.
“you will systematically miss the power law companies on average.”
Our read
Price sensitivity can be a selection error in power-law venture. But price insensitivity is justified only when evidence supports a likely winner.
AI leaders can lose consensus status
Harris sees capital clustering around one or two anointed companies per sector, but rejects the idea that discovery has ended.
He contrasts today's narrow winner-focused funding with 2021's broader rush into many companies. Consensus describes where money goes. It does not prove leadership will last.
For example, when BCV invested in its European legal software company, Harris said the consensus was that Harvey had won. People thought the upstart had landed a few European logos only because Harvey had not yet reached them.
That is not a call to search obscure cities or unusual founder archetypes. It is a reason to keep testing leaders after a category gets crowned.
“leadership is actually not handed down from on high and protected forever, but is actually constantly under siege.”
Our read
Consensus can make credible candidates easier to find. It cannot certify durable product advantage, so diligence should continue after funding concentrates.
Service businesses hedge against lab competition
Two themes explain 75% of BCV's investing over the past 18 months, Harris said. One is robotics and other physical-world AI. The other is AI-native services that perform industry work, rather than simply sell software.
For example, BCV incubated Reserve around insurance-loss processing, which Harris described as reading policies and synthesizing loss adjustments. He said it was on track for well over $100 million in revenue during its second full year.
One portfolio result cannot establish that this model repeats. Harris said BCV had made 20 more such investments, spanning law, earthmoving, home health and insurance. The breadth shows a strategy, not proof that every industry suits it.
Harris's rationale is that labs may hesitate to attack sectors where corporate customers generate revenue for them. His Google and Amazon comparisons make that conflict plausible, not certain. BCV also benefits if its service-company thesis works, so its investing experience is not an independent audit.
Ed Zitron said this on The Weekly Show with Jon Stewart in October 2026. He said 80% of OpenAI and Anthropic revenue came from 1% of customers, many loss-making startups. That does not show labs will enter those customers' industries.
“a massive TAM expander and allows you to have a more defensible business”
Our read
AI-native services hedge against labs building applications themselves, but they do not guarantee insulation. Customer relationships are part of the thesis and need evidence.
The $2 trillion case rests on continued growth
Harris called the reported $2 trillion valuation reasonable against Anthropic's reported revenue and growth. His condition was that growth continue, or that anything like its pace continue.
That is a direction-of-travel case, not a reproducible valuation. The exchange gives no revenue figure, growth rate, margin path or probability that momentum persists.
The host disclosed his Anthropic investment and said he had no insider information about the reported transaction. That context matters when weighing the question, though Harris's answer remains his own.
Liquidity is separate from operating growth. Harris did not know the cap table's seller composition and explicitly called his hypothetical 50% selling share speculation. Still, he explained how lockup expirations could bring meaningful supply when venture investors want distributions.
John Foley said Anthropic's growth rates are hard to extrapolate on Prof G Markets in November 2026. He also said the company's 2028 revenue target rests on uncertain assumptions. That does not prove Harris wrong; it shows how much his valuation depends on a path no one in this exchange quantified.
“I think the valuation based on the reported revenue numbers and revenue growth is actually reasonable.”
Our read
Reasonableness here means plausible under continued hypergrowth, not established fair value. Revenue momentum and shareholder liquidity require separate tests.
Agents widen demand while cheap models pressure pricing
Harris amplified the eight-billion-human frame by suggesting agents could mean 80 billion or 800 billion new entities. Those were possibilities he floated, not measured users or a demand forecast.
The same discussion contains a counterweight. Open-weight models and older models may make intelligence cheaper for routine tasks. Frontier models retain value only if they unlock new capabilities. But demand growth alone does not guarantee model pricing power.
On No Priors in October 2026, George Kurtz relayed a bank executive's view. Models three months behind were sufficient for most needs. That supports substitution risk for ordinary workloads, not every enterprise task.
Gavin Baker said on TBPN in October 2026 that securing a cluster a year ahead still required paying a lot. That supports a scarcity concern, but does not prove labs retain pricing power as supply expands.
So buyers should separate workloads that need leading capability from tasks where cheaper models suffice. Builders should track costs, capability and supply.
“this artificial cap of 8,000,000,000 as we think about market size is for the first time lifted because of agents”
Gavin Baker · TBPN · Oct 27Compute demand is keeping future GPU capacity expensive
George Kurtz · No Priors · Oct 22Enterprises may reserve frontier models for their hardest tasks
Our read
Agents could expand usage dramatically, but usage is not the same as durable revenue. Model capability and cheaper substitutes determine how much of that demand labs can capture.
Pacing manages risk, not the bubble
Harris would not assign odds to a broad AI bubble. He said a technology-wave analogy puts it in the eighth inning. A SaaS-like path, he said, could bring a correction within months or quarters. That timing was conditional, not a date call.
He said there was greater than a 10% chance of depressed venture returns in this vintage. But power-law outcomes make the overall result hard to predict.
BCV says it invests the same $500 million annually, a pacing rule intended to avoid chasing cycles. That can discipline deployment, but it cannot make an expensive entry price correct or protect every company from financing shocks.
Scott Galloway described Nvidia financing deals that help customers buy chips on Prof G Markets in November 2026. That raises another risk channel, though it does not establish a downturn. It broadens the stress test beyond venture entry prices.
“I do know from our perspective, we try to keep pacing even.”
Our read
A steady pace is a behavioral control, not a forecast. The investment bubble question remains distinct from whether AI adoption keeps delivering useful work.
What it means for you
- Stress-test the price. Write down the growth assumptions behind an AI-company valuation, then model a slower capability or adoption path. Separate evidence from the belief that high prices identify winners.
- Map provider conflicts. For each AI provider you depend on, ask which customers or industries could become competitors. Track customer concentration and whether those relationships could constrain product expansion.
- Split model workloads. Measure which tasks require frontier capabilities and which tolerate cheaper or older models. Compare costs and results by task rather than treating all AI usage alike.
- Keep pacing visible. Review investment or AI deployment commitments on a steady schedule. Record what evidence would justify changing the pace, rather than reacting only to market momentum.
What to watch
- Anthropic's growth against the assumptions supporting the reported $2 trillion valuation.
- The cap table's actual seller mix and the timing or structure of lockup expirations.
- Compute availability, frontier capability gains and the cost gap with open-weight models.
Send this to partners setting AI investment pace and founders presenting valuation cases.
Watch the minute where Harris assesses the reported $2 trillion valuation
Written by Real Alpha from the episode's transcript and the other moments it links, then checked against the transcript. Quotes are the speaker's exact words.
Minute by minute
A Pro write-up, open to you here. Try Pro free for 7 days
A Pro write-up. The minute by minute is in Real Alpha Pro; the episode, its moments and the transcript stay free.
Read it with Pro, 7 days free- 0:00AI could be late in its cycle
Harris says that if AI follows other technology waves, the market could be in the eighth inning. If AI is another SaaS-like chapter rather than a game changer, he says a downturn could be months or quarters away, though he does not claim to know the exact timing.
- 1:22Harris says BCV struggled to pay up
Harris says BCV has unrealized investments such as Crusoe that he expects to return multiples of its fund. He suggests risk-management habits from Bain Capital’s other strategies may have influenced venture decisions, and identifies historical valuation sensitivity as the specific weakness.
- 4:33Power-law winners can look overpriced
Harris says winners such as Databricks, Stripe and Ramp can look expensive at every round, even though their founders’ ambition and leadership were visible at the time. He argues that rejecting every deal deemed overpriced can systematically exclude power-law outcomes. He warns that treating a higher price as a signal of quality can push prices upward and inflate a bubble.
- 8:24AI leaders can still be challenged
Harris agrees that venture is in a consensus market, where investors crowd into one or two anointed leaders, but rejects the idea that there are no diamonds in the rough. He says leadership remains dynamic, citing Ligura’s gains against Harvey in parts of legal AI. He argues investors need an open mind about how AI companies can overtake incumbents.
- 11:42AI reshapes the investment lens
Harris says BCV evaluates investments on the premise that AGI is here or will be, and that abundant superintelligence could arrive in one to three years. He says physical-world businesses and AI-native services made up 75% of the firm’s investing over the prior 18 months, with robotics a longstanding area that includes BCV’s earlier Kiva Systems investment.
- 15:52Services firms can own the work
Harris says AI-native companies can expand a market and build defensibility by doing the work in an industry, rather than only selling it technology. He describes Reserve, an AI-native insurance claims business incubated with a design partner, and says it was on track to do well over 100,000,000 in revenue in its second full year. BCV also backed law firms, an earthmoving company, a home health agency and other insurance services businesses.
- 18:08Labs may avoid customers’ industries
Harris thinks foundation-model labs may hesitate to start law firms or insurance companies because major companies in those industries are customers for their models. He compares this with Google and Facebook’s reluctance to challenge financial-services advertisers, and Amazon’s restraint in banking because JPMorgan Chase is valuable to AWS. He says labs already have other areas to pursue, including intelligence, applications and consumer products.
- 25:54Compute could decide the AI race
Harris says differential access to compute may become the AI labs’ prime battlefield over the next five years, replacing talent and research leadership as the key edge. He says faster model training could compound an advantage. He also praises OpenAI’s move to let Codex users use open-weight models, describing it as an open-platform strategy.
- 32:44Harris says Anthropic’s $2 trillion price is reasonable
Harris says Anthropic’s reported revenue and growth make a $2 trillion valuation reasonable if that growth continues, while stressing the uncertainty of forecasting a hypergrowth company. He says the eventual supply of shares may depend on how many investors seek liquidity when lockups end. He notes that staged releases and price targets can help manage that supply.
- 38:54Agents could expand the user base
Harris agrees adoption may still be early and argues that agents could expand the market beyond 8,000,000,000 humans to 80 or 800,000,000,000 new entities. He says cheaper open-weight and older models can make intelligence more accessible, while demand and supply both grow. He says the value of frontier models will depend on how much their capabilities advance.
- 44:50New model capabilities shape demand
Harris says older and open-weight models may handle tasks AI can already do, but the next frontier models’ capabilities will determine what new work becomes possible. If models can reliably complete two-month reasoning tasks, he says they could change fields such as drug discovery and insurance underwriting. He argues established companies could use AI to launch businesses they previously considered out of reach, such as insurance.
- 52:38Steady pacing limits venture-cycle risk
Harris says nobody knows whether AI investment is in a bubble and sees a greater than 10% chance that venture returns will be depressed in this vintage. He says a power-law winner can still make a fund, and that BCV invests the same 500,000,000 a year. He argues steady pacing helps guard against market cycles and investor behavior.
- 57:30Agency matters as AI lowers barriers
Harris says agency—the ability to choose what to do and act—is a key skill as AI makes tools and knowledge widely accessible. Drawing on raising six children, he recommends Montessori preschool and answering children’s questions with questions of your own, which he says encourages them to find answers. He also advises avoiding privileges, such as private aviation, that could become a child’s baseline.
Published Oct 26, 2026.
Transcript · 46 min read
How I Invest, hosted by David Weisburd, with Matthew Harris.
From the transcript How I Invest publishes with the episode.
All it would take is one misstep to sort of blow the top off this bubble, like, you know, November 2021. Today, if I use analogy of a baseball game, what inning would you put us in terms of the bubble of AI? I'll answer one way, which is to say, if this is like other technology waves, we're in the eighth inning. If this is just like any... Like, the SaaS transformation, which I would say sort of started twenty twelve, twenty thirteen, and then ended in extraordinary fashion in November 2021, aided by COVID, aided by rates going down, all sorts of factors. If AI turns out to be like SaaS, which is to say, yet another chapter in the book of technology, but not a game changer per se, that probably won't be November 2026, but we are definitely, like, months and quarters away.
How would you describe that that so many of the winners are, quote, unquote, overvalued? They have extraordinary founders. They have extraordinary ambition. Those elements that led to them being these hyperscale winners were present at the time. Multiple thoughtful investors saw those characteristics, the extraordinary nature of the Colisey brother leadership or Eric at RAMP and said, you know what? I will definitely pay 50% more on a multiple basis to back these individuals because of the scope of their ambition and the quality of their leadership. But we also have this phenomenon where the winners are So, Matt, when I looked at your performance at Bain Capital, you guys are perennially top quartile, but you don't seem to have any power law outcomes.
How is that? I will say we we have unrealized companies now, Crusoe being one example where we led the pre seed and the seed that will surely return multiples of its fund. So plenty of examples where in the future, once they're liquid, we can talk about them. There's not so much what other returns, particularly in light of the fact that in recent years, have been fund returners just not yet liquid. But the why behind it. I imagine this is a pretty big issue. You guys must have discussed it internally.
Read the whole transcript
“you will systematically miss the power law companies on average.”
Early Bird, a firm firm in Europe, did did a study on this. They figured out that that this price persistence continues to persist. In other words, like, every round is overvalued versus the comps. Said another way is that the overvaluation is in relative to the competition set. But because in venture capital, the outcomes are so big and the markets are so big, even if another competitor could have gotten a 100 x on that return, but this competitor would get the 75 x if they win. It's smarter to go with the team that's likely to win than the team that would have theoretically given you the 100 x.
Very clearly. This is... The physics of power law mean that you should be extremely price insensitive if you have a greater degree of confidence that you're backing the power law winner. Now that's true, but obviously everyone can't be right about that. So if everybody's playing that same game, but most people are wrong, and everyone thinks they're backing a power law winner, then every price goes up by 50%. And then the next time you play that game, the starting point is 50% higher and everyone goes up 50% from there because they've convinced themselves they have the power law winner. That's what inflates a bubble.
When you look at what appears to you to be a bargain and your reaction instead of feeling greedy and excited is to feel suspicious and aversive, then you have a kind of a market where prices are going to trend up no matter what. I want to get back to this question on whether we're in a bubble, because I think it's an important question. But first, it seems undeniable that we are in a consensus market. Mark Andreessen says there are no more diamonds in the rough. There are only diamonds. Do you agree with that? I agree that we're in a consensus market. I disagree that there's no more diamonds in the rough.
How do you differentiate those things? I observe that we're in a consensus market, which is to say a company is anointed. This phenomena has a name. It is king made.
“And not for love or money would they invest in anyone who's not already anointed in the winner's circle.”
If you think about some new principles of investing that did not exist pre this AI market, what are some new principles that guide you as investor?
“And when we incubate companies, we don't take any common stock. We just put money in. We recruit a founder and then we give them all the common stock.”
“And our view, probabilistically, is that that will inhibit them in starting their own law firm.”
And we, having thought about it, resolved that they are unlikely to start law firms, for example. One thing I've learned from talking to hundreds of investors is that great investment firms aren't built on investment returns alone. The firms under are great at the things most people don't see, their operations, their relationship with LPs, and the quality of information they use to make decisions.
And here's what AI has changed. Every firm now has access to the very same models, so the intelligence isn't the edge anymore. The edge is what you could feed it. Firm with its fund operations and data in one connected record can actually put AI to work. A firm running on disconnected systems simply can't. That's why thousands of GPs run their funds on Juniper Square. Juniper Square puts your fund operations, data, and administration together in one connected record. That means less time managing disconnected systems and more time investing, working with LPs, and building your firm. This episode is brought to you by Juniper Square, the operations partner for private market GPs.
Learn more at junipersquare.com/howiinvest. That's junipersquare.com/howiinvest. Everyone I talked on the show is chasing the same thing, an edge, and more and more of the edge comes down to your information, not just having it, but being able to trust it when the stakes are highest. AI is doing more of that information gathering for you every day, and most tools are very good at sounding right. The summary reads clean, but can you trace it back to filing the transcript, the specific passage that drove the answer, or are you just trusting the confidence of the output? For investors, that's not a minor concern.
A miss filing, miss weighted source, context that got lost somewhere in the retrieval chain, those aren't edge cases, they're how decisions go wrong. AlphaSense is the AI market intelligence platform built specifically for this. They own the content over 500,000,000 curated documents from broker research and expert transcripts to filings and earnings calls, and they own the retrieval layer on top of that. So every answer links back to an exact verifiable source because the answer is only as good as what's underneath it. And with AlphaSense, you know exactly what that is. The edge goes to whoever can trust their information and prove it.
See it for yourself. Start your free trial at alpha-sense.com/howiinvest. That's alpha-sense.com/howiinvest. Many of the smartest people in the world actually think that industries are not the bottleneck for the large language models. It's actually gonna be compute. Even Elon has said this, which is why he's he's building the TerraFab and doing, data centers in space. Do you not think that compute itself is the bottleneck? They're separate questions. I think that you have a focus question within a lab like Anthropic about given a fixed level of compute, where are they gonna spend their time? It could be that the level of compute is too low for them to even entertain starting a law firm.
I don't think that will be the reason they don't start a law firm. But I do acknowledge that the differential access to compute is going to be perhaps the prime battlefield in the next five years. I would argue the prime battlefield in the last five was talent and leadership. That it was about having the best researchers, having the best research agenda and the best leadership that separated the wheat from the chaff. But I think it gets a little bit more mechanical here forward if you have the compute to train the most impressive model in the shortest period of time, I think that's going to create compounding advantage.
And so the chess move here about one of the three big ones getting ahead and then getting lapped by another. And by the way, here comes Meta and Google and their resources, I think, among the more fascinating things I've ever seen, particularly when you nuance it by all the external players, the private credit industry, infrastructure investors, insurance company balance sheet and what they're funding, the chip companies and how they're picking favorites or not picking favorites or buying companies to secretly compete. I mean, it it really... Wish I could buy the book that'll be written three years from now.
There's a lot of layers of complexity there, and I'm tempted to double click on a bunch of them. But let's just keep it to OpenAI in Anthropic today.
“I think it's genius, actually, and I think they're still going to sell a lot of OpenAI tokens into codex”
And I think I just as a student of both business history and military history, it's like, you know, catnip. I didn't wanna disclose, my investment in Anthropic because I wanted to hear hear
“If that growth rate can continue or anything like it can continue, they will grow into and through that valuation in a very reasonable period of time.”
I'm still shocked that the standard of no liquidity for VCs and then at six months, everybody gets unlocked.
That is the dumbest model I've ever heard that that has persisted. Couple companies recently had... Have changed that. Cerrebris, one of our other companies that it did this unlocking over over many tranches. Some tranches were as small as 6% Great. Of the holdings. SpaceX also did tranche. Do you think that this trend could continue, and do you see any of your portfolio companies speculating? For sure. Yeah. I think not only tranching, but price targets. We did an IPO in many IPOs, but one in particular in 2021, where we create a very thoughtful ladder of price targets and liquidity only comes with price targets and price targets that are hit with some duration and some level of volume.
And so you you can actually really mitigate the risk of drawdowns when you do that kind of careful titration of how stock gets fed into the market. And there's no doubt that feeding stock in sooner in a titrated, careful way is way better than not feeding it at all and letting it build up. So I think it's actually the underwriters can be a little hard to deal with on this, that the investment banks, you know, who run these IPOs actually, from a regulatory perspective, they are the underwriters. And so that that word has some meaning still, and they have to be very careful about this. They're often They're the counterparties to some of these transactions.
They're the ones who'll get sued if something novel gets tried and doesn't work. So that's actually I could see how the conversation went with Elon and his team and the banks. Well, this is what we have to do. Well, this is what you're gonna do. Okay. Yes. And if Dario wants to play But sometimes that's how standards are set. And then everyone's like, I want I want Elon's deal. And then the industry changes like magic. And I think that'll... I mean, it happened for every run of the mill issuer, but surely either Anthropic or OpenAI or both could could get a similar deal. Going back to Anthropic and Anthropic's evaluation, not just because I selfishly am focused on this.
Not because I I need to do some tax planning and portfolio planning, but also because I had... I previously had Lucas Swisher who runs growth for Co2, extremely smart investor. And one of the things that he said that I'm still processing, it blew my mind, is that despite these revenue numbers at these large labs, OpenAI, Anthropic, and others, he still believes we're extremely early. And the reason he believes that is because the majority of the revenue today is still coming from developers and engineers. And the penetration of those engineers is just so tiny, even in The US. So you're having a small part part of a small market, and that's leading the revenue today.
Now what happens when just all of engineers are unlocked and then all of other knowledge workers? I think there's only, like, 50,000,000 or so engineers on the planet out of 8,000,000,000. So you just have a small portion of them really doing a lot of token spend. What do you think about that? Could it could it really be that we're really that early in terms of token usage and AI adoption?
“And at a certain point, increasingly in our companies, there aren't even human users of AI. There are just agents doing stuff.”
It's interesting because we live in such a meme world. So many people, even with podcasts, they just consume this fifteen second clip, the the thirty second clip. So everyone tries to reduce these paradigms into these very simple binaries. But in reality, there's so many factors. Some are negative. Some are positive. Some play on each other. Some counteract each other if they start to compound to a certain level.
And then there's all these actors and all their game theory. And as one person makes a move, another person makes a move. Almost feel like you need an AI model to tell you what everybody's going to do and what all the AI models will do. It becomes complex quite quickly. Just today, was talking with Claude about data center CapEx. And the question I had was how much of the call it $5,000,000,000,000, which I think is sort of consensus CapEx spend next five years, how much of that, by the way, it could be 7. How much of that will be spent on expanding access? So more and more people using more and more A.
I. Versus expanding actual functionality. So training models so that A. I. Could do things that at this point simply can't do. At least Claude and it's funny in the caveat to the answer, says, you have to remember, I was made by Anthropic and that surely influences my answer. So that was a brilliant- It's now disclaimers. Yeah, disclaiming very specifically. I think in the IPO quiet period. That's exactly right. The chatbots in the in the quiet period. But it was it came down squarely in the on the functionality side of things. It said that, you know, it's a mix, but that the preponderance, the reason the number is 5,000,000,000,000 and not 2,000,000,000,000 or some other lower number is actually because of the intensity of training models to do unprecedented things, which was my intuition that this isn't just about the spread of A.
I. And the scope of A. I. It is actually about what A. I. Could do. And that's what's behind these CapEx budgets. And that's scary and just incredibly exciting. I want to bring up a contrarian thesis. There's now this meme in the market that AI has advanced so much that you could use old models to do virtually all of tasks. All the important... All the day to day tasks. A lot of people are are are saying this meme. You said it earlier that even the old anthropic models could do a lot of these tasks. I hear that. And I think also open source has a great role to play there. But at the same time, it conflicts with what you previously said and what is true about capitalism, which is this winner take all markets.
The number one player gets 70%. Number two player gets 20%. So are we overestimating these old models' capacity, do these tasks, or is now all these industries game theory and that if Hertz is using Anthropic 7.0, now Avis needs to use Anthropic 7.0 as well, and now the old models no... Are no longer have value. It really depends on this question we were just discussing, which is what can the new models do? So I don't think it's as simple as, you know, tit for tat with competitors having to keep up with the Joneses. I think it actually has to do with the functionality. So if you told me that Anthropics 2027 family of models could do two month long reasoning tasks reliably, then I don't think you have a choice for many of the things that are being attempted right now.
I don't think you have a choice but to use those models because they can accomplish things in drug discovery and in insurance underwriting and in all sorts of other historically human fields of endeavor that an open weight model will struggle to do in smaller chunks and will still be, you know, arming the human to do better versus actually replacing what has historically been exclusively human tasks. So if you told me that the leap forward in 2027 was massive in terms of functionality, then I would completely agree that for anything AI is doing today, a year from now, they'll be using cheaper, you know, open weight or old models to do.
But what will AI enable us to do tomorrow? And that tomorrow task will always be a frontier by definition, always be the frontier models. And so what's that mix of the value? Have we plateaued? Have we gotten the most out of it that we're going to get and it's all incremental from here? That's true. 2,000,000,000,000 is way too high. If the pace doesn't even have to steepen, if it just continues on its trajectory, the twenty five to twenty six trajectory, if that pace continues, then we're... Then absolutely both hurts. And even the car rental companies with their relatively modest reasoning tat.
I use those two companies because it's not the leading cutting edge companies, but they have a lot of data. They have insurance and all those factors that they are data intensive. Yeah. And I actually think the the wrong question is like, how can Hertz and Avis run their business better? The broader question we see and we see this on the private equity side of bank capital, and there's this beautiful interplay between venture and private equity now around it isn't merely how can you run a relatively old school company better. It's what can you now do? Like the idea that you could build a beautiful software product as Hertz or the idea that you could actually launch your own insurance subsidiary versus partnering with AXA, sort of laughable like that.
We don't we don't make software and we don't manufacture insurance products. We don't do any of those things. We don't know how to deal with the regulators and all the paper they require us to have. We don't have the actuaries in house to do that. We don't know how to price insurance. But then, you know, somebody like you would walk in that room and say like, well, don't you? Have there ever been accidents in Hertz's cars that you don't know about? Or do you have a comprehensive data set of every accident that's ever happened in a Hertz car? Oh yeah, we do. We actually know all that. And we know to the penny how much it cost to fix it.
We know who the driver was and how long the booking was and who that person worked for. We know all that. It's like, aren't you perfectly situated to launch your own insurance business? And just the fact that you have to file reports with an insurance commissioner in 50 states and another 30 countries, that's like trivially easy with AI. Also could be solved. Yeah. These things. It's like just take your blinders off and say, like, what are we unfairly positioned to do amazingly well? And relax the constraints that you've historically had about what is it feasible for us to do. And so I think this actually, particularly for kind of old school companies, is it's a really wonderful moment.
And it actually gives me confidence that when you pick Hertz and Avis, you did so. I think it's a it's a telling example. At first analysis, you'd say like, right now they're using AI to like figure out where the cars should be and manage their customer care costs. And their relatively few programmers are getting more efficient or they're pulling things from a BPO and whatever. So, they could do that with open weight models like probably today. But as the models get more performant and more capable, if they have talented management, they'll start quickly thinking, how do we use these models to do things that we never even contemplated?
Or asking their models what business should we be in? Correct. All of it. This is what people I mean, I remember going back quite a ways when I was learning how to use these agents and just trying to like by the quite a ways meaning like Q1 of this year. But I was like pushing myself in the deep end to like get good at this. And the guy I was working with would always... Every time I asked him a question, he's like, Why are you asking me the question? Ask Claude the question. Like, I am a crutch for you. Like, ask everything that you need to know is in the model. And anytime you pick your head up and ask me a question, it's just a mistake.
You don't know how to connect this to the box drive? Ask it how to do it. And it will tell you. And it was it's just an incredibly liberating thing to realize that you don't need a coach. You don't need advice. You don't need training. You actually have it all there. You watch, like, a 15 year old use it, and you can see. They just never pick their head up from it because that is everything they need. Going back to the 50,000,000 or so engineers in the world, up until 2026, coding was in code. Today, it's in every language. So in theory, up to 8,000,000,000 people can now code versus a year ago, 50,000,000 people could be code.
That's another factor that no one's really factored in, which is what happens when the number of engineers go up by a factor of a 100 or 500 x. We're gonna get a lot of useless software, but we're gonna get a lot of useful software too. I think it's amazing. I mean, this is why I like the dystopian stuff. I just don't have time for it. I I like nothing I can do about it. And I prefer to revel in the possibilities, because I think they're so abundant, frankly.
“I do know from our perspective, we try to keep pacing even. So we invest the same amount 500,000,000 a year.”
have, I think, six kids 18 or younger. True. So you have to deal with this this topic all the time. I'm having my first first kid on the way. What skill set survives in terms of...
In a post AI world for somebody that's growing up today? The one thing I see in my kids, because I have 18, 15, 13, nine, nine, seven. So I've got kids that are doing stuff. I certainly have kids who have this sort of ineffable quality that people in the valley refer to as agency. And just because it's commonly said doesn't mean it's not important. I find it to be actually the defining. Do you get credit for that or were they just born that way? That's more like their mom, I would say. Oh, in terms of our parenting. I think Montessori school is a good idea. So at a Montessori school, they show up and they can choose whatever they want to work with.
There's no program for the day. That's the characteristic of Montessori preschools. And I think that our kids all directionally then have this instinct upon a new situation of like, I get to choose. And you can see it's varying degrees. But I would say, you know, when you have more and more kids, every every belief you have gets shaken out of you over time because you're like, oh, this is true for one kid. Oh, it's not true for the second kid or whatever. And there's very few things that survive all six kids because sort of one by one, all of your priors get knocked out by the kid number four.
Turns out that rule isn't true, etcetera. And the one in terms of parenting? Correct. In terms of human nature in general, but certainly in terms of parenting. The two rules that actually survive. One is Montessori preschool is just good. Just do that. They're actually very, very consistent. Wherever you live, you can find one. Send your kids there. Two is don't answer their questions. Kids ask questions all the time. And as a good parent, you want to be like, oh, my my daughter wants to know why the sky is blue or whatever. It is not useful for you to answer their questions. They will never remember the answers to, you know, just say, why do you think what do you think?
Why do you think it's blue? And so our kids find it incredibly annoying, particularly as they get older. They just want to know, like, what subway stops to get off. But that I would just I just commend that to you as a new dad. And congratulations, by the way. Just you have no obligation to answer their questions and you're not doing them any favors. And that builds in them these processes to answer their own questions. Correct. And it's all about this agency. I mean, this is what the world demands right now is when faced with a blank screen. What are you going to do? And you have the ability of everything at your fingertips to do anything you want, anything you don't know how to do, you just ask it and it tells you.
So if you have agency, you start a business at age 12 or do whatever you want, write a college level thesis at age 12. But most people don't have that kind of agency. I don't hold myself out as a great parent. I certainly, though, have a pretty big sample size, and those are the only two things. You're an experienced parent, at least. Those are the only two principles that have survived the ringer of six kids. I'm sure your kids have grown up in quite a privileged background. Is there a way to account for this in terms of how you parent them? It used to be that choosing the venture capital asset class was a good way to keep them from being too privileged.
I mean, you know, I ran my own venture firm from, the late 90s through to 2011, 2012. And those pretty lean years. So I think at least the older kids understand what it's like to grow up in New York and feel like dad doesn't run a hedge fund or mom doesn't run a hedge fund. I think it's a little harder with venture capital having been a more productive. But at the same time, it's kind of an easy question. You remember how you were brought up. I remember a public school kid with two professor parents. And and I think if you think about it, you can remember what lessons you learned from whatever degree of denial and austerity versus whatever degree of privilege you were fortunate enough to have.
And it doesn't mean that you have to pretend you don't have means. It's the conclusion we've come to. It just means that you have to, you know, build a life that makes sense and could reasonably be available to them in the future, even as they stand on their own. And so that's set the standard too high. Yeah, you can't. In our view, you can't indulge in private aviation. That's like again, not that we would anyway likely, but that's surely something that once a kid experiences and feels is normal, that without it for the rest of their life, it's just going to be it's going to be weird. And there are a lot of things in that category around how you spend vacations and what kind of home or homes you have, etcetera.
So I actually think this is one where common sense is like pretty available and people just need to make, you know, good Well, on that note, this has been an absolute master class. Thanks so much for stopping by. That's great to be with you today. Congrats on getting to the end of my conversation with Matt. If you enjoy that conversation, you're gonna love the conversation right here.
