20VC: Is Anthropic's Coding Business Worth $2 Trillion? | Should American Enterprises Work With Open-Source Chinese Models? | Why 80–90% of Neo-Labs Die in the Next 18 Months? with Eno Reyes, Co-Founder @ Factory

I see a world where the smartest model is actually the cheapest. People are thinking about outcomes in AI, and they're looking at 20, 30, 50, and they're saying, “That's ludicrous. That's crazy.” That is underestimating by an order of magnitude how massive a transformation this is going to be.

The TAM of frontier models is frankly overweighted right now. $8 billion and $10 billion is the new $1 billion. Two of the largest companies that provide models today have explicitly said, “We're going to go after every single one of these industries and businesses that we provide intelligence for.”

I think it could be 80 to 90% of neo labs die in the next 18 months. Calling open-source models Chinese models is a psyop by the frontier labs to basically trick people into thinking that they're scary and otherize them. In 3 years, 99% of workflows are going to be done on open models.

I am fed up with speaking to visionary, insightful leaders who aren't actually building today in the trenches at the cutting edge of infrastructure and AI. Today, we have CTO and co-founder of Factory, Eno Reyes. He is one of the most articulate and insightful thinkers about the value stack of AI that I've interviewed. Factory is one of the leading companies specializing in autonomous software development.

I co-invested in their round with Sequoia, and Eno is incredible in the show today. There are going to be a lot of notes taken in this discussion.

Eno, dude, it is so good to have you in the studio. I obviously had Matan in the studio, and I was chatting to Keith Rabois over the weekend about you, so thank you so much for joining me, dude.

No, thank you for having me. I'm super pumped to be here.

Now, I hate background stories. I'm sure you listen to podcasts. It's like, “How did you get into this?” and you're just like, “I'm bored of this.” But downstairs I asked you how you got into technology and fell in love with it, and your story was heart-wrenching and compelling. So I have to ask it. How did you first fall in love with computers, do you think?

Yeah. I think the beginning was my parents. Both my parents went to art school. They were always in love with technology and how that intersected with creativity and media, and I think that's where it started, especially with my dad.

He was born in San Francisco in the late ’60s. When he was 6 years old, he actually got hit by a bus. That changed the trajectory of his life in a pretty crazy way. He wasn't playing sports, and he wasn't able to do as much of that traditional 1960s kid stuff. Instead, he went to technology and computers, like the early days when you barely had screens and were instead tinkering.

He spent most of his life embracing technology as a way to extend his own reach beyond what I'd argue his physical body could do.

It's so interesting how life delivers you a hand, so to speak, and how consequential that hand is to who you are today.

100%.

It's very hard to transition from a father being hit by a bus to margins.

Right.

Only a venture capitalist could make that transition so swiftly.

Well, some margins in AI might make you feel like you've been hit by a bus.

Yes, absolutely. We chatted before, and you said the cheapest model isn't necessarily the cheapest system.

Mm-hmm.

I read this when I was doing the work over the weekend, and I was like, “Huh. Can we just unpack that? The cheapest model isn't necessarily the cheapest system.” What does that mean?

1. Outcome Based AI Pricing

Yeah. It really comes down to this idea that when you're thinking about price, you should not be thinking about the inputs to the price; you should be thinking about the outputs. So I think about the price of the outcome.

Let's take software as an example. How much a code review costs is far more interesting than how much the tokens inside of that code review cost. If you take a very sophisticated model and it's able to do that code review immediately without making any mistakes, getting the right outcome right away, searching the right phrases, and you use 1,000 tokens or whatever, it will be significantly cheaper than using a cheap model that spends time running and uses 50 million tokens.

Ultimately, that price difference for the full outcome makes the higher-quality model cheaper. That's not how all tasks go, but for many of the most demanding tasks, I see a world where the smartest model is actually the cheapest.

I totally hear you there, and it kind of goes against the “a token is a token” theory. But will we then have millions of specialized models, with every company having specialized models operating on its own data? Because, to your point, it'll be able to work much more efficiently.

2. Specialized Models Democratize Intelligence

Yeah. I think there's a real world where the speciation of models increases very rapidly. This is the world where Fireworks and the people who help make models possible, I think, win, because the alternative is that you only have a very few specialized providers that have models.

In our view, there is probably going to be a difference between the commodity task executors. This is just your everything model, and we think that will be dominated by open models.

Then you have businesses that will say, “We do a lot of commodity tasks, but there are a couple of very high-volume specialized tasks that only we do.” For those, your commodity model won't be good enough, and your frontier model will be too expensive. So they'll want something in between, where they can take a commodity model and make it good enough via post-training.

Ultimately, they'll run that, and they probably will be the only consumer of it. They won't even give it to the rest of the world; they'll just keep it entirely internal. I think that will lead to a lot of models. Not millions, but it'll definitely be quite a lot.

When you look at the post-training required, and when you look at the implementation required, I look at that and think that company structures and teams today are simply not equipped to do that. How will we solve for that? Is this just moving into an incredibly services- and implementation-heavy world where we have these insane AI teams coming into every company? How do we solve that?

Yeah. I think this is one of those things where, right now, the recipe for post-training and building models lives primarily in the heads of a specialized group of people. But that's also how software development was 20 years ago.

So in my mind, the same tools that are currently democratizing access to software are actually the tools that will be used to democratize access to intelligence in general. I see a world where you're a company, an enterprise, and you say today, “Well, we don't have the knowledge or the skill set to build our own specialized models.” Well, very shortly—and already, to a certain extent—you can open up a platform, go to its webpage, click a couple of buttons, describe the task you care about, point it toward those workflows that happen in your business today, and out comes a model. That model will be really good.

I think that right now there are a couple of companies that claim that recursive self-improvement and model training will be only their domain, and I think in reality, many businesses will have access to that technology via software services that other companies sell.

You said focusing on the outcomes and not the inputs. That is a great idea when it's a very verifiable output.

Mm-hmm.

Code review.

Yep.

When there is ambiguity to something, which could be a marketing conclusion—did it come through X channel or Y channel? My girlfriend's a lawyer, and different legal notes are ambiguous. Some people like it one way, some people like it another way. How do you think about the importance of verifiability in determining outcome quality?

3. Verification Defines AI Success

Verifiability is ultimately the single most important property of success with current AI systems, and I think that the way that we'll progressively address this is by building new ways to verify the work that we do. I'll be really concrete about that. What a bunch of the eval creators and model trainers have done in some of these domains, like healthcare and legal, where you don't really have a concrete set of verification strategies, is take experts, bring them in, and have them create effectively their own new forms of verification. Maybe they show 2 examples side by side and say, “Which one, based on your judgment, is better?”

Being able to then take intelligent models that have a lot of the grounded reasoning and knowledge that these domains have, combine them, and say, “Now you, as the model, go and build this similar form of verification”—I actually would argue the frontier right now of AI is AI systems that can build verification where there is none. Thus, they can progress into tasks that today humans consider to be too difficult for AI to resolve.

AI systems that progress into verification. What does that actually mean?

To make it as concrete as possible, imagine that you walk into a room at a law firm and they say to you, “Hey, you're going to start basically judging how to determine whether or not these new hires are good.” What does a very novice manager do? They go by their gut. They're looking, and if they are actually intuitive, they'll go by their gut and make good calls. They'll say, “That new grad is going to be big at this firm. I like them. I'm going to continue to promote them.”

But when you have a giant firm, or you systemize, you realize that that approach to management is very rare. In reality, you need to go to the firm and start writing stuff down: “Here are the things that we like about great new hires. Here are the things we don't like.” Any company that starts to learn what managing at scale looks like has to write down the things that it cares about and build a framework or a system to analyze the job to be done.

I think that great AI systems are basically relearning this management strategy of saying, “You can go by your gut.” Honestly, a great AI system can be right very often without this sort of structure or framework. But in reality, what you'll need to do is write down, “This is what good looks like, this is what bad looks like, and here's how we judge.”

What's interesting is that, yes, this makes the system better at determining what good looks like, but it also changes the incentives. When you write down, “This is what good looks like,” people read that and then start to act more like what good looks like. You have to be careful, because if you say, “Good looks like A, B, and C,” you're going to get a lot of A, B, and C. But if you built the wrong incentives, then that system will end up following that pattern regardless of whether it's actually good or not.

That is a brilliant statement: “Show me the incentive, and I'll show you the outcome.”

Mm-hmm.

It's the hardest thing about actually running venture firms as a business, because if I set you the goal of 3 deals per year, you'll give me 3 deals per year.

Yep.

I don't want 3 deals per year. I want a great deal. I want a factory. I don't care if it's 3 or 6 or 1.

100%.

I think I made a big mistake because I'm in Mercor. I think we did it at, like, $2 billion or $3 billion. My memory should be better, but I'm older than you. I didn't do the latest round at whatever, $20 billion, because I thought, “How much bigger can it be?”

Maybe $100 billion, but that's a 5X. It's just not that exciting. A 5X, and that's with no dilution. I'm now thinking that I'm completely fucking wrong and that there is a pathway to $200 billion or $300 billion in the data requirements that will be needed. How do you think about what I just said?

4. AI Outcomes Create Massive Value

No, I think that's totally true. People are thinking about outcomes in AI, and they're looking at $20 billion, $30 billion, $50 billion, and they're saying, “That's ludicrous. That's crazy.” That is underestimating by an order of magnitude how massive a transformation this is going to be.

However, I think that what people underestimate is that the types of businesses that are going to become massive do not look like businesses 20, 30, or 40 years ago, where they had a technology moat or some sort of key capability that no one else could replicate. Instead, it's basically collections of people that understand what the future looks like a little bit more clearly and more clear-eyed than other people.

These data companies, like you mentioned Mercor, yes, they sell data, but every person at that company understands how AI is going to look much more clearly than the average human, and that makes them worth significantly more than even what investors will say.

Okay. Going back to what we said about lots of specialized models and companies working on their own data, which is obviously proprietary, I'm confused. How does this not reduce the TAM for frontier models?

I think it might. I think that the TAM of frontier models is frankly overweighted right now. The world basically assumes that there are going to be 1 to 3 companies that have total domination over the intelligence era.

I think that is a silly proposition, because generally people don't like that sort of strong dominance by a couple of small companies or large companies. But the real question is, how do you defend your margins if you're a model lab when there are so many options? I think that shrinking margin profile is going to change the expected value of these businesses.

Baked into $2 trillion, $3 trillion, and $4 trillion valuations is an assumption that you can basically 2X the price of those tokens and people will buy them.

Is the margin profile looking shit? What I mean by that is Anthropic just produced their first quarter, I believe, of profitability. Margins are seeming to be ripping, and they're throwing off cash now. Is that not going counter to what you said?

5. Applications Capture Better Margins

No. I think also that one of the biggest drivers of that is actually the applications on top of the models. One of the things that I think is quite clear to all of these model businesses is that the model itself may be a fairly rough trade-off. Rather, the margin profile of the models is definitely worse than the applications.

In my mind, if you are a model provider, you're basically looking at 2 options. A, you want to dominate the platform era, in which case you want to be 1 of N companies that get really good at selling inference. Or B, you just want to move up to become an application-layer company that has really good models.

Anthropic seems to be following the application path, while OpenAI seems to be dipping its toes in both, but the platform commitment from them seems much stronger.

Which strategy do you think is right if you were to bet on 1?

I think that the application layer is going to be a much harder battle, because being model-locked is actually a huge disadvantage if you're trying to sell outcomes.

Why is that?

It's bad incentive alignment. If you are a model-locked provider, then you are inherently selling those tokens in order to make sure that your business gets the margin it needs. If you are going to a company and saying, “We can give you the best outcome,” you have to do that with only your models. So they can really give you their best model.

Meanwhile, someone who's not model-locked can give you the best model, right? That difference between their best model and the best model can be massive in the pricing. We see this right now in real time in coding. Anthropic can basically only deliver its model's outcomes.

And the best model is also highly subjective, depending on the consumer.

Oh, yeah.

It's totally different based on the task, the profile, and the risk-taking that people want to have. There are tons of different options.

Can you help me understand? Again, I'm very thick, but I don't like cynical questions. I like to be optimistic. I think it's fantastic we're seeing Anthropic potentially go out at $2 trillion, but you're essentially placing a $2 trillion price on Claude Code.

Mm-hmm.

Which is not that difficult to switch off of.

Yeah.

What am I missing? What should I know that I'm not getting? How should I think about that?

I think that is fundamentally the risk for an investor: you're making a $2 trillion bet on one of the most competitive application markets in one of the most finicky segments of the market, which is dev tools.

I do think that part of what needs to happen in order to make companies like Anthropic and OpenAI realize their value is that they either have to pursue regulatory capture—which they are—or figure out a way to build applications and outcomes that match the true Pareto frontier of cost and quality. I think that means opening up to more models.

It's at odds with the 2 strategies. You either capture it and keep the model, or open up to everybody. This is a very hard decision, and one that I think you can start to see OpenAI actually grappling with as they've let more models into their harness. They're not making it official, but they're clearly supporting an open model ecosystem in a more direct way.

Do you think Dario's marketing message has been mistaken?

I think that the marketing of AI in general was probably one of the worst marketing jobs done by contemporary capitalists. It basically did the opposite of what you want: scare every single person, tell them it's very unreliable, and basically threaten their well-being and livelihood with the technology while you roll it out at scale.

I think that the challenge is that the things Dario brings up are not only well-intentioned, but there are very real threats from unregulated and dangerous AI. I think there's a way, though, to communicate about this without maybe embellishing the economic ends.

The carrot-and-stick here can just be: one, the technology will be incredibly transformative; two, humans will have a huge role in that transformation, and you will have a huge role in that transformation; and three, if we don't do this well, then, like all other technologies, there are going to be risks. But the moment you start talking about the singularity and AGI and create this godlike mythology out of AI, you're going to scare a lot of people.

We're going to be the last remaining private company. Say, "Well, you flew here on United." I don't know how that's going to work.

Yeah, there's going to be a lot of change if people think that there's going to be 1 company. I think Sam Altman just did an interview where—and kudos to him, because it's hard to go back and say, "I was wrong"—he basically says, "I totally underestimated the momentum of the economy, the momentum of existing businesses, and so I predicted this future that actually has not come true."

I think that's a great reckoning, where you go and say, "I didn't think that this was going to happen, and it's clearly not going that way, so I'm revising my prediction."

What did you not think was going to happen where you had to revise your prediction?

I love that question. I think the biggest thing that surprised me, where I've had to go back and seriously correct my priors, is that building a business is much more reactive than planning.

Almost every one of our best decisions as a company has been in reaction to some information and a split-second decision, rather than a master plan that we forecasted 6 months ahead. Knowing that, you start to look at the rest of the world, hear from other business leaders how that's also how they make decisions, and realize that the world is just this constantly reactive feedback loop where people are talking to each other and no one actually has an answer to what the future is going to look like.

You are basically defining it in real time in group chats and in the actions that you take as a business. Learning that in real time has made me, first, question the existing world and structure. Basically, nothing's guaranteed and anything could change.

Second, it's quite empowering. It makes you realize you're basically a couple of decisions away from even greater outcomes and an even bigger business than you had prior.

I totally get that anything could change. One thing that I hope changes is margins and margin profiles. Help me understand. When we look across the wave of incredible businesses in AI, the margin profiles are still lower than they were previously.

Yeah.

They're at 30 to 35%, say, as a barometer, compared to 70 to 80% with SaaS.

Right.

Is that a momentary period in time where we're in a build-out phase and they expand over time, or is that just a net-new model where the revenues are going to be much larger?

Not all businesses in AI have bad margin profiles. I know we've got good margins.

Okay, great.

That actually comes in a way that I think is also customer-aligned, in that we are so focused on thinking about these outcomes themselves as the thing to be valued. We've gone away from a lot of common paths that we see other AI companies taking.

We don't subsidize consumers in the dev tool space. That hurts us in a lot of ways. We don't have the mindshare from self-service users.

So you have no PLG?

I would separate self-service from PLG. We have a lot of focus on PLG within companies that we've deployed to, to allow the adoption to increase. What we don't have is a consumer-facing plan that is, I would say, super-rational for a current consumer unless you're optimizing for quality.

In a land-grab environment like today, should you?

I think it's a really good question. In my mind, the biggest reason for this is that there are 2 players with effectively infinite money who are trying to flood the market, and their intent is: if we flood the market, we keep you.

But remember what we talked about just a second ago: you don't keep them after you pull back the subsidies, as we've learned. I think the consumer will continue to follow the most cost-effective solution.

What does that look like in 1, 2, or 3 years? I think it's open models. The most cost-effective solution for a model is going to be the cheap one that you can run locally on your computer. We want to make sure that we are the product that best fits that type of experience.

That's why we're so optimized for local and open models today for on-prem, but in the future it's for all consumers. We won't have to subsidize as much as we'll have to create an amazing experience for self-service.

I think of this as a long game where eventually self-service will come to us, but not because we subsidize, because we have the best product on the market.

If you're advising me as an investor on how to think about margin and how that plays into my decision to invest or not in a business, what would you say?

For us, this is a part of our strategy. It doesn't mean, though, that it's the only way to win. I do think that there are probably going to be businesses where they're able to lock you in because of a workflow or a system of record that they produce, and the margins—or the subsidies temporarily reducing margin—can be a route toward gaining a customer base.

This is a classic strategy, right? This isn't even new to AI. What I would say, though, is that if there's no path to increasing the margin profile, that is very risky.

A lot of investments are being made in businesses where the promise is simply that they will raise prices, but you won't see a commensurate increase in the value of the platform. It is so competitive in AI right now. If you are not also raising the outcomes and the value that you get out of the product while you raise that price, people will churn and move to another thing.

I do think it's quite tricky, and the margin profile actually matters a lot, but it's not the be-all and end-all.

Will you have that churn in enterprise sales? You work with some of the biggest companies in the world. I'm sure you sign year-long minimums.

Yep, yep.

You have pretty sticky client bases there, no?

I think so. People also see these year- or multi-year partnerships as just that: a partnership. Part of what makes it interesting to build right now is that a lot of what you're selling is not only the technology, but your knowledge about how to best use that technology.

I would carefully differentiate that from consulting or professional services. You don't actually have to go in and do all of the implementation. I honestly think if you have a product that requires 100 FDEs to get it deployed, you just have a bad product.

But instead, I think that if you have the advice and the knowledge of the direction that you think the world should go in, you're selling that with the product. People are willing to go and buy that. And I think that if they see it from you today, they sort of know that in a year you'll also still have that same knowledge and forward-thinkingness.

Now, obviously, lots of people can give that, but if you combine that with a product that then acts a little bit more as a platform or a system rather than a tool that people use, you can also get stickier by just being something that you build on top of.

We spoke about the different frontier model providers essentially having this really challenging dynamic of being locked into their own models—

Yep.

—when serving, say, Claude Code or any application that they choose to serve. One then thinks that the value becomes in the routing of models: the tasks to the model, what's optimized for each use case, cost, latency, function, whatever it is. And OpenRouter gets bought for $8 billion. All the value's in the routing. Great. And then everyone is doing routing. Ramp has a routing provider. One of my companies, Merged.dev, has one. We're in another startup, Requesty. And I'm like, "Well, the routing's completely commoditized."

Right.

Help me understand: What world do we live in?

6. Routing Is Not The Moat

Yeah. Well, I think that routing is a really interesting technology in that I don't think the technology itself is necessarily that differentiated. And so if somebody comes up and says, "Look, Stripe bought OpenRouter for $8 billion because of the technology," then I'd say either, A, if they have insider knowledge, that was a bad decision, or, B, if they're just assigning that to it, I think they're missing what I read when I read the letter to shareholders, which was that they see this as a bet on where the direction of capital allocation is going.

Think about it like this: What are tokens other than intelligence? And how do you get tokens? Well, you pay for them. How does the infrastructure layer get it? Energy. It's literally translating energy into intelligence, and you're just trading dollars along the way. All of this is basically converging toward one thing. Whether you call it allocating energy, allocating intelligence, or allocating money, businesses need to allocate whatever this is in order to grow and expand how they operate and grow and expand their bottom line.

If you're Stripe, you already control the flow of 1 of the 3. With OpenRouter, rather than the routing technology, you actually just gain the information of where these models are going, right? You start to understand: What are people doing with intelligence? How are they allocating it? What models are they using? So now you start to control the second of these 3 things. Maybe they'll make a play into energy infrastructure or data centers at some point, but just ownership over those 2 is a massive bet on how to think about where companies allocate resources.

And so for them, I think this is very reasonable. But what's interesting is you wouldn't pay $8 billion for the same company that had no users with better technology. That difference, I think, is really important, and it gets toward that broader idea that the technology's just no longer the moat.

Do you think it was a good buy?

At $8 billion, it would have to be really foundational to the team that becomes whatever this next bet on allocating capital is for Stripe—or, rather, helping the businesses that Stripe has as customers allocate capital. I would say that if they think that data gives them insight into how to run Stripe better as well, that could also potentially make it worth it. $8 billion is quite steep, though, so stranger things have happened. It feels like $8 billion and $10 billion are the new $1 billion.

I'm intrigued. You obviously have a routing product within Factory.

Mm-hmm.

What do you see? What insight do you get from that that maybe the world doesn't see?

Well, one of the most interesting things—and I've been talking about model routing for these products that you've mentioned—is gateway routing. That's sort of how it's referred to because, ultimately, the routing effectively happens outside of where the task is being completed.

A lot of companies will do this. They'll look at Ramp, Stripe, and OpenRouter, and they'll put in a model gateway. That gateway is just how all of the different tools and products of the company route to LLMs. What's interesting is that we've seen you can definitely get some nice cost savings doing this, like 10% or 20% from these types of products.

But you really need something fundamentally different when you have agentic workflows because, to actually take the most advantage out of models, you need to do something that's a little different from just routing. You need your agent or your system to dynamically understand the task it's working on and understand how to allocate intelligence in a much more stateful way.

And when I say stateful, all I mean is that you need to know not only what's going on today, but also what just happened and what's going to happen in the future. That can't happen outside of where the task is being completed. You have to be in there in the task.

Is that related to the importance of context-window expansion that everyone talks about?

I'd say that, in a way, taking the context window and expanding it was solved not at the model layer, or outside at the endpoint, but instead inside of the agent with something called compaction. It's exactly similar. People really want the problems to be solved somewhere else, like in the model or in the gateway, but more and more, we see it's the harness that solves these problems.

The thing that's underestimated about why the harness continues to solve this is that the harness is effectively the new application. It's just where all the logic happens. It's where the state is maintained. It's the easiest and best place to do work with AI.

And so, as that starts to accumulate more advantage or technology benefits, I think we're starting to see people ask questions like, "Should I be building my own harness? Should I get a really great harness? What is a harness?" That's something that I think at Factory we're trying to spend as much time as possible educating people about: What does well in the harness versus what can be done outside, and how to think about building versus buying a harness.

The context-window expansion is 1, and then continuous learning and the rise of the first truly continuous-learning model. Does a continuous-learning model help or hurt Factory's business?

Well, it's interesting because there are sort of 2 directions for this. I would say that there was an idea of continuous learning from over the last couple of years that said you would have a literal LLM-like model where all of the learning happens internal to this closed-loop system. That technology has not been developed. It doesn't exist. There are attempts and early looks at it, but in general, anybody who wanted to try and hold all of the learnings behind an API would be able to potentially accumulate an advantage that would make it harder for others to use that model in their product because, ultimately, they would be accumulating all of the learning.

But in reality, what has happened is quite the opposite. Basically, model providers have even acknowledged that all of the continual learning happens at the harness layer. That learning is basically something that businesses are going to find very critical to own, that they are the sovereign of. And I think that that is ultimately the question of the next 5 years of AI: Who is the sovereign of your intelligence? Is it you, or is it some other company?

Sorry, can you unpack that for me? What does that actually mean? Who owns the data outcomes that are generated from the tasks that you do?

7. Businesses Must Own Intelligence

Basically, who owns the learnings and the workflow that successfully achieved the outcomes for your business? A great example of this would be if you were a law firm and all you did was outsource every single one of your cases to some other company—in fact, maybe it was 10 different companies. Then, 5 years later, those other companies can just turn around and screw you over because they know exactly how to do your entire business.

What's interesting is pretty much every company in the world is thinking to themselves right now, "What's going to happen in a couple of years if I outsource all of my intelligence to someone else?" And what's interesting is that at least 2 of the largest companies that provide models today have explicitly said, "We are going to go after every single one of these industries and businesses that we provide intelligence for."

Do you see with large enterprises—the biggest companies in the world—are they scared of OpenAI and Anthropic coming after their business?

Not all of them necessarily think they're going to come after the business, but the largest companies in the world are very wary of the model labs coming in and promising them intelligence and sort of luring them into a trap.

That's something that I know many businesses are starting to become aware of. Palantir has been quite loud about this idea of owning your intelligence. Microsoft as well. Satya wrote a great piece about this. I think that all of that opining is very spot-on.

At the end of the day, if it's not your intelligence, then there is a real risk that either, A, they come after your business, or, B, if they disagree with what your business is doing, they have a little bit more leverage and control than I think the typical business owner would like.

And when we talk about sovereign intelligence and owning that outcome, is that why on-premise is so important?

I think that's a huge part of it. To most businesses, on-premise isn't even about the technology. It's just about the idea that, if I need to, I can take control and ownership over every dimension of this software, and we get to stay in control.

That element is interesting because, for us, we offer an on-premise offering. It's, in fact, one of our most popular offerings, Factory Private. Even though we offer this, a lot of the businesses that we talk to actually go with our SaaS model because they have the peace of mind of knowing that, if they need to switch, not only do we have it available, but they understand exactly how it would work.

I think that part of this story is just being able to share with people that we are incentive-aligned. If you need this, we have it, and you won't lose anything. On-premise and owning your intelligence are very similar stories.

How much does it help you or hurt you that Cursor was bought by SpaceX? It gives them some amazing scale benefits in terms of access to compute, but it does make them model-biased.

Yeah. That outcome for the folks at Cursor is obviously amazing.

Sure.

Needless to say. I think where it helps us is that it's going to be a very hard story to become model-independent—or rather, stay model-independent—when you're attached to a model lab. They're going to want to push Grok. The products are going to become increasingly oriented around Grok, and that, I think, will become a challenge.

There are also, to be honest, trust and enterprise-related concerns that they're going to have to deal with under their new brand. Ultimately, the team there is obviously incredibly competent, so I don't discount them as a player in this market. However, I do think that most enterprises are going to have a second look at the idea of ceding their software development life cycle to a provider that is, one, likely to be model-locked and, two, has an existing history or pattern of maybe struggling to operate in these larger and more secure environments.

Can I ask you, when we look at the cadence of model development today, it's just so fast? I actually use arena.ai as a discovery mechanism for new models.

Right.

I'm suddenly using these weird models that I've never heard of and would never have used before—

Right.

—and I'm loving the output. My question to you is, will we see the cadence of model creation sustain in the way that we are today? In other words, will the rate of new models keep coming, or is this a momentary period at the start of a new cycle?

8. Model Creation Keeps Accelerating

It will likely sustain for quite a long time. This actually gets to another interesting property of model routers. A lot of people treat model routers as effectively an information or news stream about which model is next. It's kind of a free advertisement every single time a model drops. You know, now Stripe can tweet, “New model on our router,” and you see Stripe's name with this news cycle.

I think it's very common for people to use new model drops and all this news as a way of keeping up with AI in general. New models will likely continue to drop as, one, it gets easier. It's just going to become fundamentally easy to build models. And two, with this idea of sovereign intelligence, just like humans, we're going to have models with tons of different opinions and tons of different perspectives.

That, I think, is going to play nicely into how people operate in today's world. A lot of the time, you don't go with a business because it's purely the best-performing. You go because you like the person who started it, or you want to buy from someone you saw on the news or on TV and agreed with. Models are going to be like that as well. They'll emit opinions and have takes that are different from the ones that are most popular, and people will gravitate toward those.

I see this getting much faster and even broader before it shrinks.

A lot of guests on the show before have made bold statements that 70% or 80% of the Neo labs we have today will die in a given time period—3 to 5 years, whatever you want to choose. Do you think that's true? And how would you advise me and other investors on the AI labs that will thrive versus die in this next wave?

I think it's plausible that it's even more. I think it could be 80 to 90% of Neo labs die in the next 18 months. “Die” is going to be a funny word to use because, for a lot of them, it'll probably mean incredible outcomes. I don't know if it's necessarily doom and gloom as much as these businesses may not make sense as independent businesses.

A lot of what I think matters for an AI lab is that you should ask questions like: one, is this business attached to a durable workflow? Two, is that workflow going to change if new frontier models get better? And three, if this workflow were to be introduced to a new business, would that new business figure out something even better?

Basically, is it durable to an entirely new way of thinking or a new way of working? Legal is a great example where I think, one, new models won't necessarily get better without access to the data; two, it's obviously a very proprietary workflow; and three, we're still going to have a legal system in 5, 10, or 20 years. So probably all the Neo labs focused on legal are going to have great outcomes.

By contrast, I would argue that there are some places, like a lot of knowledge work related to intermediate tasks—people operating in Excel and Jira—that just aren't going to be differentiated. The workflows are very common. I think we may not use a lot of tools like that in 5 to 10 years, so general computer use and all this other stuff may not be as valuable as an independent business.

The thing that strikes me is just the misalignment in capability progression. When you look at coding and customer service—

Yep.

—amazing, undeniable. Legal is good, but not at the same level as coding and customer service. And then other things, honestly, like marketing copy and visuals, are so far from being there. If I wanted to use AI to clip this show, it misses both of our faces because it goes through the middle.

Yep.

It has no understanding of how to align clips between an audio edit and a video edit. It's so far off. Will we see a real, multiyear time lag between different sectoral capabilities progressing in the same way that coding has?

Definitely. The biggest reason why, for example, clipping a podcast is still such a hard problem for models is likely because the handful of businesses that deal with media haven't devoted 100% of their time to taking the knowledge that lives inside people's heads and bringing it into AI.

The moment we start to see businesses capitalize on that delta, I think the progression will happen extremely quickly. It's purely a matter of time before most workflows become something where a business capitalizes on that first bit, which is a workflow that currently has proprietary data or proprietary knowledge.

You don't normally think of clipping a podcast as proprietary, but I think, for the most part, it's a real skill. A lot of people couldn't even describe how they know when to make the right clip, and that intuition—writing it down—is hard.

I totally think so. Knowing the hook, if you started 10 seconds earlier and the hook was 10 seconds in, your chance of virality goes down significantly. The skill of knowing what is a hit kind of goes to taste.

100%.

We always hear about the Chinese open-source ecosystem and the questions around security and everything in between. Do you think those are justified, or do you think we should leverage it and thank them for their capabilities?

9. Open Models Challenge Frontier Labs

Calling open-source models Chinese models is a psyop by the frontier labs to trick people into thinking that they're scary and to otherize them. In reality, in my mind, the open models that come from a bunch of different places are not any different from an existing frontier model from a lab. They just happen to have been created by people a couple thousand miles away.

There are some very real challenges with taking in models from really any business. What's interesting is that they're actually the same challenges as with models from OpenAI and Anthropic. You should ask questions about all of your models: one, what is potentially being censored by the creators of these models? Two, are these models going to be able to solve the problems that I care about? And three, if this model goes away in 6 months or 12 months, will I be able to switch to something else?

If the answer to all these things is yes, yes, yes, then I do think that there will be concerns about using those models. For me, the Chinese models specifically have demonstrated no examples where they have some sort of security risk or backdoor compared to American models. Instead, they are just biased in a way that American models are biased toward their own creators and preferences. These are things you have to be aware of, but they typically don't change the day-to-day.

So you don't think that American companies should be concerned about using open-source Chinese models?

As of today, the current Chinese frontier models should be analyzed by American companies for the tasks that they care about.

If you’re, for example, working on national security in the United States, you definitely should not be using Chinese models. But I do think that we have to be clear-eyed and say that for a code review, it’s very likely that a Chinese model and an American model will give you the same result.

I’ll give you a great example of this. If you are writing a 10-K that expresses your business’s current state and some of the examples in preparation for sharing financials with investors, let’s say that part of your strategy is about introducing recursive self-improvement to models, and you care about AI and your business is leaning heavily into it. If you use a model from this provider, it will block you. The answer is Anthropic.

And so if you are writing a memo about American defense or preparation, I highly recommend against using a Chinese model. But all of these are basically contextual, based on the preferences of the creator of the model. It’s just like any other technology. You have to be aware of who created it, and you have to be careful, because if the person who created it doesn’t want you doing the things that you’re going to do with that model, it is going to be harder. That’s something that I think people need to be aware of for all models, though.

Guillermo from Vercel tweeted last night, or yesterday, about the weighting or usage of open models significantly increasing at a much faster rate than tokens used on closed or frontier models. What percentage of workflows will be completed with open models in 3 years’ time?

In 3 years, 99% of workflows are going to be done on open models. But 1% of those tasks is probably going to be 30% or 40% of the economic value of the future of intelligence.

Wow, but you still think that 60% will flow to open models?

I think almost all usage of models in 3 years is going to be primarily open, but that difference between frontier and open is going to become actually larger than it is today. I think that this is actually pretty aligned with how—maybe that percentage is overweighted—but you’ve listened to how Sam and Dario talk. The use cases that they’re talking about their frontier models being used for are incredibly niche.

They’re talking about bio research at the frontier. They’re talking about super-advanced LLM and AI development. They’re talking about security and defense. These are use cases that really only fit a very specific profile of effectively the frontier of science and technology.

This is where I think labs like OpenAI and Anthropic can actually be incredibly differentiated, because they already have the muscle to work on those very frontier problems. But if you go into any business in the Global 2000 today and ask any random person, “What are you doing today?” it’s not something that needs the true frontier of intelligence 99% of the time. And so cost will dominate.

OpenAI and Anthropic could release an open model that they then run inference on, and I think that that could actually be a great business for them.

Given the commoditization of the model layer, like we’ve spoken about, people seemingly chastise Microsoft. Given that commoditization, do you think Microsoft has actually played a great hand in having a little bit of a bet through OpenAI, but not being tied down with an extensive model investment layer?

10. Microsoft Wins Model Independence

I think Microsoft might be one of the best-positioned hyperscalers, honestly, with respect to AI, because of this independence. Right now, Satya has played a masterful game of getting huge upside from the OpenAI investment and work. Kevin Scott as well, who I know was sourcing a lot of that deal. That team found a lot of the potential of what AI was going to be, but they currently realize that one provider is simply not sufficient to cover what intelligence is needed in the enterprise.

And so now they’ve moved towards more concretely expressing, one, that they support all AI developers, and Azure should be a place for inference on Anthropic, OpenAI, and, most importantly, open models. But two, even if you do want that frontier intelligence, you can come here.

I think that that duality, that model independence, is going to be massively valuable for a business that wants to accelerate work for all other businesses. I think it’s a very clever positioning, because they captured the upside with a bet, and now they’re capitalizing on the market as a whole.

Is Zuck wrong, then, to be putting as much money as he is into Spark and building out that program?

I think he’s right for humanity in that we need more open models like that, especially American-made open models. I think that a lot of people, even with respect to what I said earlier on Chinese models, are going to be biased. And so open models are great because it just increases adoption in America and abroad.

But two, I think that as a business, they’re going to have to build on top of that and take what they did with Meta, the monstrous consumer business that it is. They need to power all of their operations with these new models, and the investment will be well worth it.

So would you buy Meta or Microsoft today if you could only buy one?

If I could only buy one, Microsoft for sure.

Wow.

The biggest thing that Microsoft has going for it is that infrastructure. They own so many of these data centers. They’re spending so much on build-out. No matter what model runs on top of that, Microsoft is going to win.

Do you worry about the debt cycle? What I mean by that is, there is so much cash being put out into the data center build-out, and we’ve just never seen levels of debt like this before. You’re seeing that in bond pricing for Meta and other companies.

Do you worry about that sustainability, or do you just think, “Fuck it, we’re still so early”?

I think that it actually should start to concern investors. If you do not have a huge amount of free cash flow, then taking on massive amounts of debt is bad. It’s dangerous.

And so for the Microsofts and the Googles of the world, they have access to businesses that are durable, existing, and have huge barriers to entry. As a result, those businesses may take a hit from a future collapse in value, or even if it’s not a collapse, just a minor hit in the projected future cash flow from AI. Those businesses are still going to be around.

I think it’s a risk. But to be completely honest, if you’re OpenAI or Anthropic, the hundreds of billions in free cash flow that you need in order to pay back the debt that you’re taking on to accommodate these data center build-outs and get the next big training run—it’s totally existential for them.

And so they need to become the single greatest free-cash-flowing businesses in the history of technology in order for them to just live. It’s a—

It’s a high bar.

It’s a pretty massive bar, yeah.

Do you think they should be moving into the chip layer as they are? I mean, we’ve got the wonderfully named Jalapeño, and then we have Anthropic now reportedly working on its own chips. Do you think that is the right move?

I think so. I think verticalization is clearly the strongest way to free yourself from taking on a massive amount of debt and burden. In the future, if they become multitrillion-dollar companies, they will simply have to enter this market and own more of that infrastructure layer. So it seems quite obvious that they want to play there.

I think the biggest question is how this changes their relationship in the medium term with their current vendors, because ultimately their explicit goal is to replace their dependency on them. I think that’s going to be an interesting challenge to navigate.

I think it’s a little bit like competition in VC, which is, no one really has any loyalty anymore.

Yeah. No.

I’m so sorry to say that—

No, it’s true.

—but you know what I mean?

It’s true.

Yeah. And Jensen’s like, “I’m working on Nemotron, I’m buying Poolside.” Sam knows that fully.

It’s coopetition with everyone.

Yeah. Listen, this is our job. We have to survive.

Yep.

Do you worry about where we are in the market today? I have older, wiser friends being like, “Harry, this is peak froth.” And then I’m also like, peak froth, but also Cursor just sold for $60 billion after 4 years. That is cash that’s coming back to hospitals and foundations. That ain’t froth or IRR; that’s cash back.

Yeah. No, I think that what I am less concerned about is a 2008-style financial crisis or massive bubble or asset crash. I think that seems disconnected from the true reality of where this technology is and is going.

We’ve already started to see outcomes in science and in serious, real human-prosperity-style changes. It’s very early, but the technology pretty clearly, to almost all experts in the fields of science and technology, is on a trajectory towards accelerating human prosperity. That is hard to deny from an outcome perspective.

Now, what’s more interesting is which businesses are going to be the backbones of that transformation? We are probably in the Yahoo era, where we don’t actually have, or at least widely recognize, the Googles of the world—or the thing that comes after.

And so today, when I look at OpenAI and Anthropic, there are, I think, more analogies to Netscape and these technology companies that were first. But ultimately, it’s hard to navigate being first.

I think Steve Jobs always had a really great strategy at Apple—not being first, but being the best at almost everything they did. That’s something that we also like a lot.

You know, being first or best, but we heavily lean toward being best.

I’m continuously changing my mind on outcome sizes.

Mm-hmm.

We’re also an investor in a lot of the builders of the world. Suddenly, it’s a $13.5 billion business at $600–700 million of ARR. Fuck, dude, the trajectory of company growth is just unparalleled. Do you think investors need to change their mindset on outcome expectations and company growth expectations?

Yeah. I think that it is hard because there’s a current space where people are experimenting and buying a lot of technology in a fairly speculative way. And so it’s hard to index on AI for every other possible industry.

But you start to look at some of the other industries that are being influenced by this wave, even CPG, right? It feels like every other day you hear about a massive brand that got bought for billions of dollars that started 2 or 3 years ago. So I think that maybe it is just true that the world gets faster and grows bigger and better than ever before, and we’re starting to see the early days.

I think that a lot of people say this is what the singularity will feel like. Things will just move faster, things will grow bigger, there will be more, and we’ll start to normalize it and build models and say, “Oh yeah, that’s just the way it is today.” But I think it might just be us expanding as an economy.

Before we move to internals, which I do want to touch on, because you’ve got really interesting takes on hiring, we saw Airtable go for $2.5 billion, give or take. Listen, it’s a fantastic outcome, incredible, but it’s just a reduction from the $11 billion price before. Will we see a generation of SaaS companies sell and exit before we see this wave of cannibalization that could occur?

Yeah, I mean, I think that certainly we will. I heard this from a fellow founder who told me this, and I totally resonate. Basically, contemporary SaaS businesses are more like movie studios now, where you have to hit a blockbuster, and you have to keep hitting blockbusters in order to keep the attention of the world.

And if you are Airtable, you made that one movie that was a hit, and people loved it, so no doubt it’s a great business and a great outcome. But if you rest on that, then yeah, Bending Spoons will come and eat you. And I think that the new world is much more about continuously getting bigger and growing larger.

I won’t be surprised to see a huge wave of M&A of these businesses because they’re still good businesses fundamentally, or you can make them good businesses. They just aren’t going to be Stripe or these massive things that capture fundamental pieces of the economy.

I think a very good parallel is also gaming companies, where you have a banger of a game and a hardcore user base that will sustain for 5–7 years, and that life cycle is great, but you need another big, big hit.

100%.

I totally get you there. Listen, your hiring is candidly different. When we were chatting before, you said, “We expect 100% of our future hires to come through acquiring companies and bringing their founders and teams into Factory.” I read this and I was honestly like, “Wow, that’s hard.” Because often founders make bad employees. I would suck as an employee. How do you determine whether a team will thrive internally within Factory, or whether that’s an opinionated, pretty arrogant, egotistical founder who’s not a team player?

I think that’s a great question. To me, the most interesting thing about this is that the profile of an organization has changed so rapidly that acquiring an organization is no longer what it was 10 years ago.

What I mean by that is, if you are someone who just created an open-source project, then you are quitting your job and spending 5–8 months just building that one thing, and that shows so much more conviction than you can track in an interview. And so it’s much easier, actually, to spot these talented people, who are sometimes companies of one, who are ready to execute.

They want to be a part of a mission, or they’re already operating toward a mission, and basically what you’re offering them is more resources to do that. And so I think that a lot of what we’re looking for when we try to bring a team in is: Are you mission-aligned? Are you someone who’s going to operate independently and be able to take on a huge amount of responsibility?

What’s interesting is that all of these people are also aware of this lack of technology moat, and so they’re pretty willing and ready to either integrate everything they did in a couple of days or scrap what they’ve been working on in order to build something even bigger. And so for us, this profile is just such a match made in heaven for a team that already operates with a ton of former founders and a ton of people who were previously working at startups.

Can I be a dick?

Please.

“Mission-aligned.” Of course, no one wants someone who’s not mission-aligned.

Right, right.

And then also willing to take on responsibility. It’s not groundbreaking. What—do you know what I mean?

No, I totally get what you’re saying. In my mind, mission-aligned for us means that you’re literally working on the exact problem that we’re working on and doing it very well.

I think that when people come in and say, “I loved that we were working on payments,” in a way, payments is just like AI for software development. You’re kind of like, “All right, I’m sure that you absolutely could work on this,” and in fact, maybe we’ll actually hire that person, right? So I’m not suggesting that if you worked on payments, you can’t join Factory.

But the people that we’re looking at are already deep in the weeds of building harnesses for software development, where they’ve already built something that tens of thousands of people are using on a daily basis, and maybe they even say, “That’s better than the stuff we’re getting from Factory.”

Or they’ve thought so deeply about the problem of outcomes in AI and measuring that, and they’ve already sat down with business leaders to say, “I want to solve how to translate these AI inputs into AI outcomes.” And so mission-aligned to me isn’t a property that you can suggest or say. It’s actually extremely evident in the work of the founder.

That has made it very easy to stress-test whether you’re going to do well at a company, because you’re basically doing the same thing, except backed by us.

“Mission-aligned” goes against what Chamath has got a lot of heat over the weekend for saying.

Yep.

Which was that Silicon Valley has become too money-centric, and that’s now a problem. Do you think he’s right?

That is something that it sounds like Chamath would say because, frankly, that’s how he’s made his money. I think about what we’re doing and how we got started. When we created the concept of the software factory, something that he loves to use that term for as well, we said to ourselves, “We’re looking at this very hard, fuzzy, ambiguous problem. We’re eating so much glass. We’re going up to people saying, ‘This technology is going to exist. Here’s how it works.’”

And people would say, “Leave. Go away.” It has nothing to do with the pursuit of an outcome, because at that point, you’re literally trying to be right about a technology. And so it’s very producty, very engineering-heavy, and contrary to what the world is saying, or at least people are saying: maybe at some point in the future, they’re going to dismiss you.

In my mind, that sort of mindset—“I’m trying to build a thing. I see the way the future might look. I’m going to get super in the weeds. I’m going to have people tell me I’m wrong every single day for 3 years straight until they eventually agree with me”—I think that’s actually felt in San Francisco.

Everywhere you go in Silicon Valley, there are tons of people who just care about technology and about doing something that might change the world. And I think that what you get around that, though, are people who do want to profit on top of that, and their only way of participating, because their background might not be in building, is to try to build financial instruments around it.

That’s actually a healthy and important part of the ecosystem, but it’s not the only thing that happens in San Francisco.

I actually think it’s the paradox of what Chamath says, which is that I think the influx of money has led to 1% realizing, “I’ve got plenty of money. Whatever I do, I can always go back to a big company, to a great company, and get paid a lot.”

100%.

“So I’m going to choose to work on something that’s really interesting.” Do you see what I mean?

Yeah, and there’s just no shortage of people who are purely trying to realize a dream in San Francisco. It is amazing to see, and I think that it’s actually quite harmful to that culture that people really want to create a narrative that it’s purely financially seeking.

Because in today’s world, the words that you say and how you portray something become part of the story that the intelligent systems we’re building ingest. They’re world-model LLMs and the tools that we’re going to use to do work on a daily basis for the next decade. That technology is built on the stories that we tell.

So I’m always trying to share a little bit more of the optimistic side of how I perceive the world to be, because I think that actually helps make that world occur with a higher probability.

We’ve talked about team additions.

Mm-hmm.

Cognition places a lot of emphasis on the chess champion and the math prodigy. Do you think we are overweighting the importance of traditional certification, or do you think that is the right thing to focus on in a more verifiable, engineering-heavy hiring process?

Yeah, I think that it is conventional hiring wisdom to look at pedigree and achievements and say that's the right way to pick people who are going to be smart. I think, though, that in many ways the least agentic path that you could take is to only try to hit the goals that other people set in front of you. That tends to look like you go to the right school, do the right competitions, and follow the rules well enough that you then get recognized for how well you follow rules or operate within the system.

There are plenty of smart people who are going to do that because that's also how you almost guarantee a great outcome for your life. To be clear, you can still find many smart people who follow that path. But in our mind, the most important trait to hire for is how capable you are of operating outside the bounds of what today the system calls the rules. That is something that is very hard to measure for.

And so ultimately, I think you need to look outside of that traditional pedigree and start to look for people whom that system might have overlooked. I know that I'm a big proponent of this. I went to an Ivy League school, and I learned firsthand that that is barely a signal for competence. There are plenty of idiots who went to Ivy League schools, and I think the clearest signal for me that someone's done something great is that they have built something that they care about and that they want to tell the world about.

I think that you can see that all over, and that's why I say it's not just companies that we think will make up the people we, quote-unquote, “acquire”; it's also 1-person shops that just built something in their spare time that demonstrates they are going to go outside the boundaries of what traditional systems would reward.

How do you think about placing a value on those 1-person and small teams? We're seeing Poolside being bought and a lot of the employees moving over to NVIDIA at a rumored price of $12 billion. How the fuck do we put a price tag on heads?

No, it's a great question. This is ultimately probably one of the biggest challenges of capitalism in general: this desire to place a value on humans and talent, which is challenging. Some of the best outcomes in history have effectively come out of 1 person making a gut check or the right call.

People say this about Jeff Bezos. Is he worth $100 billion, $200 billion, or is it the company that built it? In my mind, there is something magical that happens in the connections between people. It's not that any 1 node is worth $100 million, but when you put all of these nodes together, the graph they make can be worth tens of billions of dollars.

I think that's actually one of the interesting parts about a talent strategy: you have to constantly be thinking about the graph you're building. Those nodes can come in, and there are traits and properties that you want them to have, but the best people are people who make that graph much stronger than it was before.

That, I think, can easily be worth $10 billion, $20 billion, $30 billion or more, especially to companies that are operating in a model where their talent was built pre-AI. Any business whose graph was decided before this insanely game-changing technology was created has to update its graph very rapidly. Bringing people in can be the difference between a $2 trillion company being a $4 trillion company. Almost anything's worth it.

Penultimate one before we do a quick fire. What do you see other founders make in terms of mistakes on hiring that makes you go, “Oh, no. Sarah or Simon, I wish you hadn't done that”?

I think the biggest thing is performative work culture: founders who look for people who say, “I'm going to grind 24/7, I'm going to be unstoppable, and I'm going to work myself to the bone.” I think a lot of founders see that, and they catch a signal that this person is going to be really productive. But what we have found is that this sort of performative work culture, this 9-9-6 attitude, is almost always correlated with making up for some other detractor or trait that basically means this person might not be a great hire.

I think that actually applies at the company level as well. All the companies that, for the most part, say, “We work people to the bone Saturdays and Sundays, and have to have people in all the time”—at different points in your company's life, you will work on weekends. You will work 15 hours in a row. That's going to happen. I think trying to make that your culture points out that your business doesn't make a ton of sense without it.

I get in a lot of trouble in the UK and in Europe for being the 9-9-6 guy.

Yeah.

I think people take 9-9-6 from me too literally. I definitely do not mean 9:00 a.m. to 9:00 p.m., 6 days a week. I mean a culture where, if I ping you on Sunday morning saying, “Hey, a big client has a problem. We need to jump on a call with the buyer there,” you jump on Sunday morning. It probably won't happen, but it's not, “I'm sorry, it's my weekend, and I will resume on Monday.”

Yep.

I am constantly talking to people on weekends, and we're doing stuff that indicates that the business operates outside of Monday through Friday.

Sure.

I think, though, that there's clearly a difference, and that's why I say “performative work culture.” Any time you create an incentive to show people that you're working rather than to actually do work, you're incentivizing the wrong thing.

Really focusing the business on outcomes is important. It's interesting that you just mentioned jumping on a call with a client in order to achieve something. It's pretty easy to see that you're not talking about the act itself; you're talking about the outcome you want to achieve.

Yeah.

That difference is pretty massive when you're hiring. I see a lot of people think that they're making the right call by only selecting for people who have that trait, and I think you miss out on a lot of great talent who know that that's bullshit.

I think the other thing is senior engineering talent, especially those with families, are instantly put off by the performative, often young hustle culture. I've learned that the leverage you get, especially when it comes to infrastructure engineering or architectural engineering, is very real.

Right now, being able to point a set of agents in the right direction and knowing from the beginning which direction to go in is worth not only more because it gets the job done faster, but it now translates into real dollars. If you spend 100 times more tokens trying to get an outcome because you just don't know as much, it doesn't matter that you worked harder. It basically means that you missed the ball the first time.

I had Brandon on the show from Macaw, and he said that they spend more on tokens than they do on engineering headcount. My dear friend Jason Lampkin from Sasta, who we do a weekly show with Rory, said, “We'll give $100,000 of tokens to our best engineers.”

Yep.

Where do you sit on that today, and how do you think that changes?

We actually don't even think about allocating tokens or credits toward people like that. I think that's a very weird way to think about it, and I think it gets at the fact that people are ultimately looking at inputs.

What we think about is how many tokens, or how much spend, we allocate toward projects and outcomes. I'll give you a great example. There's an evaluation that we've been hill-climbing against in order to see if we can build a system that beats it. It's incredibly difficult. It's called Program Bench.

One of the things that we've done is effectively allocated almost 7 figures of credits in 1 day on this benchmark. The work was currently being done by 1 person, so I guess you could say that we allocated 7 figures of credits to that person. But in reality, what we're trying to do is see whether our research pans out on this project. Of course, we're willing to spend that much in order to see if that outcome comes true.

Similarly, for a lot of our engineers, what we do is scope out these projects. Once you know the scale and scope of the project, all you have to do is share the bid for what you think it's going to cost, and then we send it off.

We have products called Agent Effectiveness that let you allocate and look at how many credits were spent on a given project in order to measure whether you're achieving the outcomes you'd like in your business, given the spend you're putting toward those areas.

I think that in this new world, the relationship of 1-to-1 mapping between agents and humans—or saying that an agent has a name—is a weird way to think about it. Really, you have an agent system, and you allocate capital toward projects.

I see that number for some businesses approaching 8 and 9 figures easily.

Final one before we do a quick fire.

Yeah.

What really good idea did you say no to that was very hard to say no to?

I think self-service is by far the hardest thing for us to continuously say no to. It is actually quite painful as a builder of products. I want more people to use our product, and there are certain types of adjustments that we could make. I think many of them would unfortunately come at odds with either making our business more successful or improving the experience for the enterprise.

That’s something that I don’t see as being permanent. I do think we’re going to hit a certain scale where we’re allowed to pursue many things at once, but it constantly nags at me that I can’t just hit a button and then have 10 million people using the product. When we look in the market at comparable solutions that have millions of users, basically the biggest difference is economics, and that’s it. We know from a product perspective that we have a lot there; we just have to make the hard decision not to subsidize.

If you own the outcomes of those consumers and you get that data back, could you not make an argument that the improvements that data would provide to the core product would outweigh the cost to serve those free self-service users?

That’s actually how we use self-service today. We have a product that, obviously, you can download off the internet and try. To be clear, we have a pretty steady stream of tens of thousands of people who use the product every day from that segment.

Those self-service users are giving us feedback and helping shape more of the experience for the individual user. At this point, I think a lot of that can be achieved with fewer than 250,000 people. Once you start to hit critical mass, you get that feedback loop and you have everything you need. You don’t need 5 or 10 million people to get those bug fixes in.

However, there is something really special about seeing the community build media, content, and storytelling around your product. That’s a part of the experience that we have to work really hard to create a comparable version of.

Yeah, a grassroots community—

Yes.

—that kind of grassroots brand is harder if you don’t have that identity.

Yeah, 100%.

We’re going to do a quick-fire round.

Let’s do it.

In the UK, we have a game, and I’m probably going to get in trouble for this, called “Shag, Marry, Kill.”

Right.

Brutal, but you get the theory, which is buy for the short term, buy for the long term, and sell hard.

Yep.

Meta, Microsoft, and NVIDIA.

Ooh, this is a fun one. I think I would have to say marry Microsoft, shag NVIDIA, and kill Meta, and I’ll tell you why.

Microsoft, to me, represents a software company that has become an everything company that really does sit in the lifeblood of almost every Fortune 500 company. There’s not a single business that doesn’t have something from Microsoft, and I think that means that company is going to be here for a very long time. I say this also as a former Microsoft employee for a year and a half.

NVIDIA is the current kingmaker of technology. They get to decide who is currently even sitting at the table, and I have no doubt that’s going to continue to grow massively. As for how durable that is, I think it’s just a matter of whether they accumulate power at the right rate. If they do it at the current rate, then at some point people might ask, “Are we going to let one company control the whole supply chain?” But today, I think the answer is probably, “What other choice do we have?” So they’re going to keep growing.

And Meta, I say this with love for the company, but I think that ultimately, from a technology perspective, they’re often quite accurate. I think Zuck’s push into VR was technologically correct. That felt like the right move. I think the push now into open models is technologically correct.

They have one cash cow, which is their ads business, and I think that means they’re going to have to work really hard to figure out if there’s anything other than that that can sustain the business. That makes it the weakest of the 3.

Will NVIDIA be a $10 trillion business in 3 years?

In 3 years?

Yeah.

I think there’s a real, serious chance that if we let SpaceX be worth $2 or $3 trillion, then NVIDIA probably is worth $10 trillion. So I think the answer is likely yes.

It depends on the buoyancy of multiples.

Yeah.

You said something about the entrenchment of Microsoft in businesses. Would you be a buy or a sell on Salesforce, given that?

Oh, I’m a buy on Salesforce.

Really?

I believe that the businesses that are going to be most durable are the ones that have a workflow and a system of record they’ve defined that everyone agrees is consensus. You look at the Salesforce of the world, and today, at least, you look at Atlassian.

I think the biggest thing is that when people say, “I hate that software,” and everyone buys it, that’s probably a pretty good business. They’re not buying the software; they’re buying what’s underneath it. To me, that’s actually much more durable than the technology itself.

I’m an investor in Linear. Linear are absolutely crushing, though.

Yeah.

Atlassian are crushing, and so are Linear. It just goes to show, again, that the market size is so much bigger than anyone comprehends. I think we think too much in zero-sum terms: I take from you, so you lose.

Yep.

But we’re both just crushing, actually.

I think that’s really true. One thing that’s important is that Linear and Atlassian are selling the same thing: the same workflow, agile, and a system of record that represents agile.

I will say, though, that one thing that’s pretty clear to me is that the way we build software is fundamentally changing. I think agile might be one of the things that gets hit by this new way of developing. That makes me think there’s a huge opening for what the next system of record and the next workflow look like.

I do think it’s going to have to be much more radical, and it’ll be very challenging for both Linear and Atlassian to transform into that new way of building.

Codex, Cursor, Cognition, Claw Code. Rank them 1 through 4 in terms of the threat level you feel from them.

Threat level is interesting. I think the current ranking would be—actually, I have to qualify this by saying I don’t feel an impending threat from the rise of these players. They’re all going in a different direction from what we’re building, and that’s really important. I’ll touch on that in a second.

In terms of relevance to the conversations when I’m talking to enterprise buyers, number 1 is Claude Code. It’s brought up in every single conversation, and we always have to share with people why we see ourselves as largely complementary to the Anthropic platform and suite.

The second is Codex, which has increasingly been referenced in deals and conversations where people are saying, “Look, we were on Claude Code, and now we’re switching to Codex.” That’s actually the greatest news for us because it shows how unsticky this is and it gives uncertainty. However, they’re coming up much more frequently now, and I think it’s because their work platform is better than Anthropic’s. It’s not Codex for coding, but rather Codex for Work that’s coming up more frequently, which is fascinating.

Then I would say Cognition, because they’re basically the only other model-independent vendor in the enterprise. One of the consistent feedback points we hear from them is that they’re building this cloud offering. It’s very futuristic, based on the idea of imitating a software engineer as a human. I think that makes people ask, “Is that different from or the same as your strategy?”

Then Cursor is present in a lot of these businesses. I don’t think anyone really perceives Cursor to be their primary enterprise software development strategy so much as an IDE, which is still a great business because they’re still going to get a lot of usage. That’s how I see them.

This is mostly informed by the idea that almost all 4 of these businesses are coming to enterprises and saying, “We’re going to build you an eventually human-level AI replacement for labor, and then we’re going to do an Indiana Jones swap and replace the people in your business.”

I think that’s so different from what we’re saying to them. You’re not going to replace humans with AI so much as you’re going to build a new system for developing software. Humans are going to build that new system alongside AI, and that new system is going to look very unfamiliar.

It’s not so much a 1-to-1 labor mapping as it is an entirely new development methodology, and that’s something they’re really only hearing from us right now.

And I think it sort of contextualizes the many tools in the space compared to Factory.

Will Chumath be successful with his... I can't remember, the 1809 or, but whatever it is?

Yeah. I think my answer would be that it depends on how real the software is. I haven't seen any examples of it working in an enterprise environment. If they're very focused on building software that works and delivers outcomes, I believe that he has just as good a chance as anyone and is very well connected. But ultimately, I do think that part of this is about building with the enterprise. So I think the biggest question is: Can they get enterprise traction in the markets that matter with people who take them seriously as a full-time software development opportunity?

Single biggest piece of advice on selling to large enterprises in today's world?

I think the biggest thing that I've learned about selling to enterprises is to stop treating it like persuasion, where you're trying to convince them that you're right, and instead treat it as a discovery opportunity to learn about what's currently the biggest problem they care about.

That approach difference is unique to new markets. So if we're selling something where the market's established, it's finite, zero-sum, and everyone knows it's a commodity, like databases where there are a million options, I do think persuasion is the strategy. You're trying to convince them that, all else being equal, they should buy from their friends.

I think that in this market, it's much more about trying to understand just how big of an opportunity it is and learning with the customer. People actually put a huge amount of value, especially in software, on people whom they perceive to be trying to problem-solve with them. If you're on the same team and you're both trying to problem-solve, then you're going to land on a real problem that that business has not yet solved, and that means almost 10 times out of 10 you're going to bring value if you can figure out a solution to their problem.

So I think that's something that people underrate. You're not trying to trick them or persuade them. You're just trying to help solve a problem for them.

Totally get you. Age-old enterprise sales doesn't change that much.

No, definitely not.

Final one. I like to ask the question: What seems ludicrous or strange today that will be incredibly commonplace in 5 years' time? I can give examples of finding your husband or wife online. Bizarre.

Yeah, that's a great one.

Putting your credit card details into your phone. Of course I'm not doing that. That's so dangerous. What today do we think is crazy that will just be obvious in 5 years' time?

I think the biggest thing that we're going to be surprised by is the fact that we let a sort of priestly class of maybe 2 million people decide the fate of all software for all of humanity. In 3 to 5 years, it'll be unthinkable that you couldn't just generate the thing that solved your problem with software on the fly, in the moment, for nearly any problem you have in front of you that can be solved by information manipulation.

Today, we see a little bit of that with Lovable and Bolts and these tools that let you build personal applications. But I think this will look quite interesting when you think about what problems not just an individual consumer has, but really what society at large has.

You'll be on vacation in Belize, and the boat operator that gets you from point A to point B will have software and an interface fully customized to them that looks better than your HR or IT software back at home. I think this total disbursement and distribution of amazing software to the entire world is going to make everything just feel way more futuristic, and that's going to happen very, very quickly, on the order of 3 to 5 years from now.

I love doing what I do because I genuinely get to pursue my own curiosity in a very natural way. So thank you so much for entertaining my curiosity, and you've been an amazing guest.

Thanks for having me. This was a fantastic conversation.