AI 护城河为何仍然重要（以及它们如何变化）

The thing that is fundamentally different about this product cycle is that the software itself can actually do the work, and therefore the market opportunity for software today is no longer just IT spend. It’s largely labor. It’s not like all the jobs will go away. I actually think that’s not going to happen at all. There are a lot of things where, if I could hire somebody for $1 to do this task, I would 100% do that. I’ve never been able to hire somebody for $1. Now I can hire software for $1.

While it is important to understand model capabilities and what’s happening at the frontier, you still need to figure out how to apply that technology.

I think moats matter just as much as they did before. The one change is that, in this supply-and-demand equation, there’s conceptually more supply of software on the supply side because the barrier to creating this stuff has gone down dramatically.

I think AI is an incredible tool for differentiation. The idea that a voice agent can speak in 50 languages, fully compliantly, 24/7, is highly differentiated, certainly versus the humanness of that capability. In my opinion, the AI-ness of that capability is not a source of defensibility. It is just so consensus. Cloud was not consensus. Mobile was not consensus, and that’s why the incumbents kind of screwed up.

We’ve spent a lot of time talking about moats and how moats have evolved, and whether there are still even moats in this new era. So why don’t you reflect and share some of the conversations we’ve been having, or some of your perspectives on this broader moat question? Maybe, David, we’ll start with you.

Maybe just to jump right into it with a hot take: I think moats still matter.

Still matter.

Still matter. Exactly. And I think they’re largely the same. I often think about this in terms of differentiation and defensibility. I think AI is an incredible tool for differentiation, right? The idea that a voice agent can speak in 50 languages, fully compliantly, 24/7, is highly differentiated, certainly versus the humanness of that capability. But the AI-ness of that capability, in my opinion, is not a source of defensibility. It’s largely differentiation.

The defensibility of a software product resides, in my opinion, in owning the end workflow, the context in which it’s applied, becoming the system of record, having a network effect, and deeply embedding yourself within your customer. I think these were the heuristics that were always things we would look for when evaluating software companies.

I think the thing that is fundamentally different about this product cycle is that the software itself can actually do the work. Therefore, the market opportunity for software today is no longer just IT spend. It’s largely labor.

The challenge often has been that everybody can build something at small scale, and a lot of the—I wouldn’t call them network effects, but some of the defensibility moats—only become apparent at large scale.

Take an example from a long time ago, pre-AI era. If I’m building an anti-fraud company and I’ve seen lots of people, am I going to do a better job than a net-new anti-fraud company that’s seen a few people? The reason why this would be called a data network effect—although Martin and I did another podcast a long time ago debating whether or not data network effects are real—is that it’s almost like gravity.

Gravity actually—one atom exerts gravity on you, but you only really see it at very, very large scale, like the Earth. You notice the gravity of the Sun. You notice the gravity of Jupiter. You don’t notice it for that glass. It’s the same thing for a lot of these data network effects: at very, very small scale, when you have 20 companies all saying, “I’m going to stop fraud,” they’re all building the same things. They all have the same algorithms.

But when you’ve seen 4 billion people and you know these people are bad, now you can sell each incremental customer—each customer of your anti-fraud technology, to use this example—because you’ve seen more customers and you can get actually better results.

The challenge is that a lot of these moats only really become evident at mega scale. The same argument would apply: “I’ve seen 4 customers. David’s seen 3. I’ve seen 4; he’s seen 3. Pick my software.” But you’ve seen 4 customers. That means there are 8 billion customers you haven’t seen. There are 8 billion customers he hasn’t seen. What’s the difference?

Whereas at mega scale, it’s, “All right, I’ve seen 4 billion customers; he’s seen 1 billion customers.” It’s actually kind of easy to see that the results of my product will be better. But that’s at scale.

A lot of the question is: in the 0-to-1 phase, it’s hard to make the argument that I have better fraud underwriting. If it’s AI doing the work, I’ve done more phone calls to a particular type of customer and therefore I do a better job. It’s hard to make that argument at subscale.

This is often the challenge: it’s kind of self-evident that if you become the biggest company in the world, then you have a moat. But how do you get to the scale where you actually could show that? You can’t get to that scale if you have 9 million ankle-biters and you are yourself an ankle-biter just trying to get to scale, and nobody can because it’s so easy to actually produce software.

That’s the double-edged sword of AI: it’s very, very easy to produce software. Everybody can go do something that is a very obvious idea. Because it’s obvious, everybody’s going to go build it. But can you get to the type of scale where you actually could show a moat? That has gotten arguably harder because you have a larger count of potential competitors.

But if you get to mega scale, then you can show the moat. That’s kind of the 0-to-1 versus 1-to-N.

And maybe talk about what’s different about defensibility for even the bigger players today in the AI era than it was in, let’s say, the Web2 era. Are the companies today more defensible, less defensible, or how should we think about the strength?

I don’t know. I think the less defensible part—I mean, this is why a lot of enterprise software has gotten beaten up in the public markets—is kind of 2 things.

Number 1 is that, if you’re doing per-seat pricing, how do you come up with a pricing model that people feel is fair? A lot of it is just psychology, and for whatever reason, for the last 20 years, it’s been per seat per month with—you’ve heard my joke—the tall, grande, venti model of software charging. Somehow that felt fair. Whether that is fair or not, I don’t know, but people are like, “Oh, yeah, it’s $85 a seat per month.” “Yeah, okay, that sounds reasonable.” Whereas if you proposed that pricing 40 years ago, you would have been laughed out of town. So this just became the norm.

The reason why public software companies have been beaten up a little bit is, “Uh-oh, maybe you won’t sell as many seats.” Is Adobe going to sell as many seats if now you don’t have to hire as many graphic designers? Is Zendesk going to sell as many seats if the software just answers all the queries? The answer is no.

It doesn’t mean that the companies are toast. They might actually quadruple their revenue because now they charge per outcome as opposed to charging per seat. But that’s part 1.

Part 2 is, wait a minute, now everybody can vibe-code a Zendesk competitor. Maybe companies will just stop buying software. This one I don’t think we’ve seen at all. But I think there are these 2 risks.

To answer your question, does defensibility change? If you’re now able to code your own software, why am I paying? Your margin is my opportunity. Look at the margin of software companies. Salesforce has an 80% gross margin. They should have a 1% gross margin, or nobody should use Salesforce anymore. That would be the pro case for moats really starting to disintegrate.

But I don’t think we’ve seen that happen at all. It turns out that, on the one hand, 2 things are actually happening. One is that this is kind of Clay Christensen’s theory: the incumbents overshoot the market.

The amount of features in Salesforce, Zendesk, or NetSuite far exceeds the feature set that any individual customer needs because it’s meant to encompass all of these weird edge cases. You see this if you use Microsoft Word. When was the last time you wrote a book? When? Never, right? I haven’t written a book. It has all of these things. They probably have 50 software engineers to make it.

But if you do write a book, guess what? Microsoft Word has all these features just for book authors to make a table of contents or something. I don’t use that. They keep bundling more stuff in there, so they overshoot the market.

Theoretically, it’s going to make it easier for somebody. But going back to where I started with this topic, it turns out that this concept of “I’m just going to vibe-code Microsoft” has all these edge cases that you just don’t know about. Why don’t you grow your own food, weld your own aluminum, or build your own house?

It’s just kind of easier to use this concept of comparative advantage and say, “I’m going to buy something off the shelf.” So, anyway, I think moats matter just as much as they did before. The one change is that, in this supply-and-demand equation, there’s conceptually more supply of software on the supply side because the barrier to creating this stuff has gone down dramatically.

I think the flip side to that, too, is that while there will be more software—and, again, the marginal cost of producing software is declining asymptotically toward 0—the way these companies are getting more deeply entrenched within their customers has differed. Again, the software is doing the work and therefore, in many cases, it’s actually replacing labor. If you’ve transitioned a team out and that work has now become your software, you’re now much more dependent on that product to run your business.

Is it more difficult to replace that software with another piece of software or to rehire that team? I think it’s an open question, but again, the software is doing more of the work and therefore, I think, getting more deeply embedded within its customers.

Well, part of it is just the Goldilocks zone of pricing. I wrote an X thread about this a long time ago. I call it the janitorial services problem, because if I went to you—you’re the CEO of a giant company where you write your books in the future—so you have a 300,000-person company, and I said, “Erik, I can get your toilets 9% cleaner and save you 1% on your janitorial services spend,” not only do you not care, you don’t even care enough to exercise the mental energy to find the person in the company who does care, right? And that means that your janitorial services spend will never change.

The problem is, it’s hard to get in. The good news is, it’s hard to get out. Whereas if something is like 90% of my profits going to you—or if I’m now 90% of your profits as the CEO of GE, and they’re going to me—your number-one priority is getting the hell off of me and doing RFPs left and right.

Part of it is also just how relevant this is. There are some companies that operate in this Goldilocks zone of irrelevance, like these janitorial services, where even if you have 9 million competitors, they’re just not going to go anywhere. Which is why a lot of the strategy that we talk about internally is greenfield, right? Those companies are stuck for good.

Is there a high rate of new company creation that will not use the crappy old janitorial services company but will actually resonate? Your pitch of “I will get your toilets cleaner and I will charge you less money” really resonates, but that’s not going to resonate with the people that are using the old-fashioned stuff.

What are examples of companies or spaces in the Goldilocks zone, and what are examples of companies or spaces in the greenfield zone?

Well, payroll companies, right? ADP and Paychex—these are companies that are collectively worth hundreds of billions of dollars, very, very profitable. How does payroll work? You could do your own payroll. It’s kind of a good metaphor for software in general: Why is it that you have to pay me? You’re my employee; why can’t I just cut you a check? Well, because I have to withhold taxes. How much tax do I have to withhold? It depends, right?

There is this super-complicated lookup table: You live in this county, but you spend this many days in New York, and this, that, and the other thing. Oh, and you owe child support, and the IRS is garnishing your wages—all of these things that are very complicated. So it turns out it’s just cheaper to go to ADP, and ADP just charges you, I don’t know, $50 a month per person. It might be $100. It’s a paltry sum compared to the overall amount of payroll. So nobody really switches their payroll companies. That would be an example of one.

On the other side, I had a lot of companies coming out of 2022, when the market really went through a downturn, and they were like, “Wait a minute. I had 1,000 employees. I downsized to 200 employees. I had 1,000 licenses for Salesforce, right? What’s 1,000 times $100 a month times 12? That’s $1.2 million a year. Wow, that’s a lot of money because I only have 200 employees and I only have 6 months of cash. I have to save that.” And they didn’t do that for their payroll spend.

So you see it: A lot of companies do want to rationalize their overall software cost, especially for things where they recognize, in aggregate, that most people aren’t actually using the seats. I’d say Salesforce-type stuff. Some of the creative tools—Adobe is very expensive, and you might just do a wall-to-wall license saying, “Why not?” But then, if you’re asking, “How do I save $5 million?” and nobody’s using this, well, it’s $5 million.

Whereas for things where the delivery and the payment are inextricably linked, which is very, very different than a per-seat model—obviously, I’m not going to pay for payroll services unless you were employed here. We have 600 people who work at our firm; I think we have 600 licenses from Microsoft Office 365. We probably—I bet there are a lot of people here who have not opened Microsoft Excel in a year. So why are we paying for that? That would be the idea of rationalizing software spend.

It kind of depends, but I think per-seat pricing, where it’s just easier to pay for the entire thing wall-to-wall, your entire organization, is often the first to go versus things that are, again, inextricably linked to actual usage.

Yeah. So you mentioned earlier that we’ve seen—basically, you mentioned there was this concern that maybe, instead of Zendesk, companies will have a vibe-coded version of it, but we’ve seen none of that. So far, is your mental model that we’ll see it in examples where the cost is significantly high, or in which there are greenfield opportunities? What is your mental model for the types of software that will be replaced?

Yeah, I think the greenfield one is always true, but when you look at greenfield opportunities, you need 2 things to be true. You need the entrepreneur to be very, very patient and say, “I’m not going to try to sell to everybody.” If I’m starting a net-new payroll company, I’m not going to try to sell to GE because I recognize that they are hostages to ADP, and that’s never going to change.

One is that patience of the entrepreneur, and the other one is that you just need a high enough rate of new company creation to really make it work. Which is why, to pick on one space, electronic health records or electronic medical records: How many new hospital systems are created every day? It rounds to 0. So if I’m trying to build a new EHR system to compete with Epic or Cerner, I can do that. There are a lot of edge cases there, but I might have patience as an entrepreneur. Wait a minute: I need to sell $5 million deals to big hospital systems. Every single hospital on Earth is currently using an EHR system. It’s going to be really, really hard to make that work.

I think both of those need to be true: the right type of entrepreneur who’s willing to be patient, because it’s often a very lonely game. “I built this great product. Wait a minute. I don’t have any customers yet.” And you want to see high traction, because you’re seeing, in the rest of the market, some companies are just going like this and my company’s not. I’m in Silicon Valley, and I need to recruit the best people. They want to work at the company that has the graph like this. But this greenfield requires patience.

Yeah. So we’re talking about how moats still matter, and in many ways they look pretty similar. Let’s steelman the other side for a second. Why are we even having this conversation where some people say, “Hey, brand is shipping velocity,” or that this era is different? What’s the steelman of their argument?

Well, look, I think this market is noisier than ever, right? And so I think finding ways to stand out from the crowd probably matters more today than it has in the past, I would argue. I think the other thing is that the underlying technology is changing so quickly. As a founder, you want to be living on the frontier and understanding what model capabilities look like, because it can dramatically change the efficacy or the capability of your underlying product.

One of the things that’s changed, I think, that’s been really interesting in this current wave of especially vertical applications that we’ve seen is the type of founder. I think founders today are often younger and more technical than we’ve seen in prior generations. They’re less often native to the particular industry, but they’re fluent in the toolset, right? And I think that’s really important because, to the same point, you’ve got to stay on the frontier and understand what’s coming.

At the same time, I wrote this piece that I call “Context Is King.” While it is important to understand model capabilities and what’s happening on the frontier, you still need to figure out how to apply that technology. And so, while the founders themselves are maybe less native to the particular industry, they’re still hiring for context very early in a company’s life cycle.

A good example of this is a company called Eve, which I sit on the board of. The 2 founders of Eve were the earliest employees at Rubric, which is now a public infrastructure company. They built a legal AI company in the plaintiff law space. Neither of them had any particular background in employment law or personal injury, but they deeply understood how to apply document extraction technology, voice, and LLMs more broadly to this very particular workflow.

They’ve actually hired plaintiff attorneys on staff. Anytime a new model is released, they understand from people in the industry the impact that it’s having on drafting and their ability to reason through a case or a matter. Again, it’s this tension of building the brand, having momentum, understanding what’s happening on the frontier, and yet figuring out ways to apply that technology in the context of your specific customer, because I deeply believe that is where a lot of the sources of defensibility reside.

I’d love to find other examples of businesses where the technology reinforces their business model. It doesn’t compete with it. In lots of areas of legal, if you make your employee 50 times more efficient, you’re eroding your billable hour. In their business, they operate on a contingency basis, meaning they only get paid if they win. So there’s no limit to the amount of AI that they want to adopt.

If you can become 5x more efficient, you can take on 5x more clients. These are characteristics that I’d love to find more of, and hopefully that can be kind of a bad signal, too.

I think the other steelman is, if you believe that brand matters—which it almost tautologically does, because what do I buy? I buy the thing that I’ve heard of, right?—then there’s an advantage there. And if you believe that for a lot of companies and products, somehow having scale is effective, right? So, not a network effect, but a scale effect.

If I’m Honey Nut Cheerios and I know that people are going to buy lots of my Cheerios, I can build a big factory and not hand-crank out each Cheerio. I’m going to have these compounding advantages in terms of economies of scale. Amazon—is that really a network effect? No. It’s kind of nice that everything that I buy will show up the next day or in 2 days. How can they do that at low cost? Because so many people are buying things.

There are some things that have scale, and those things also benefit from brand. If you can move the fastest, right? If you can agglomerate capital and labor, so it’s like, I raise the most money—it’s a very generic idea, but somehow, like most other things on planet Earth, if it’s the biggest and really, really big, kind of gravitational scale, then it’s just going to work better. Can I get there the most quickly?

But there are 20 companies that are doing the exact same thing. At that point, I wouldn’t say that momentum is a moat per se, but momentum has the highest chance of getting you to gravitational scale, where you do have a moat. If you don’t do that, by contrast, you’re just going to get eaten alive because you can’t hand-crank out the Cheerios.

You have to get to the scale where you’re able to build a factory. With the biggest factory, you can crank out the most things at the lowest cost. So, what is the trajectory? What is the slope of you versus all of your competition? If you don’t have a good slope, you’re just not going to win that game.

Yeah. One of the questions for defensibility in Web2 companies was, “Would Google—would they someday build this?” Or Facebook, or name your incumbent. In the AI era, it’s, “Will OpenAI, or will some other major company?” How should companies think about that framework in the AI era?

I feel like 18 months ago, this GPT wrapper was on everybody’s lips, and I think it was largely used as a pejorative. To some degree, there are some spaces where the model capability and the application capability, if they’re very overlapping, I think you’re in a risky spot.

But the reality is that there are so many markets that were never particularly interesting to sell software into that are now radically interesting spaces to build companies in, again, in large part because the market is now labor, not just IT spend. Plaintiff law is an example. Alex and I have a company called Salient applying voice agents to auto loan servicing.

5 or 6 years ago, would we have backed a software company selling to nonbank auto lenders? Probably not. The company is doing incredibly well, again, in large part because the capability of being able to speak in 50 languages, fully compliantly, with customers in 50 states, working 24/7, is so differentiated versus the individual.

They’re finding that their ability to collect is meaningfully higher than that labor, so the cost-benefit trade-off is so dramatic. The company is getting a lot of revenue from customers who may not have had millions of dollars of IT budget historically and are now very willing to pay for a product like that, given the impact on the business.

And the way that we used to talk about this a long time ago is—and this almost had a pejorative slant to it—are you building a feature, a product, or a company? What’s the difference between the 3?

A feature is like there’s an existing product, and you tweak that product to make it marginally better. A product is not that. It’s some hopefully system of record or something that keeps track of something. A company is probably the most defensible of those 3, where you have a product and maybe you own a platform. The platforms tend to be the most valuable companies.

A feature is like, “I’ve built a Chrome plugin,” and that doesn’t mean you have a company. There were, by the way, a lot of Chrome plugins. Honey was a Chrome plugin that got bought for $4 billion. I wish I had done that right. That’s a good feature, but that was a feature.

A product would be like, “I built my own browser.” A company is like, “All right, well, my own browser company actually makes money.” You don’t actually have a company, even if you have 10 products, if you don’t have a sustainable path for that company to be around in 10 or 20 years.

Another way of thinking about what David just said is that now the features—the feature was the most pejorative and seemingly small of all of those 3, almost obviously. Some of the features can be incredibly profitable because it’s like, wait a minute, this feels like a feature because it could get added to Salesforce or could get added to one of these other things. But the amount of money that I can charge for my feature is orders of magnitude more.

It’s like, “Hey, I’m going to be the front-office receptionist for your orthodontic clinic.” That’s my job. That’s the feature, and it sits on top of whatever software you currently use. But the feature I can now charge $20,000 a year for because it is doing the job of labor.

But, uh-oh, will the existing product that my feature is riding on top of build those pieces of functionality, and/or will another company show up that just says, “Hey, we’re going to sell the greenfield with a new product that kind of has this feature set embedded”?

Feature, product, company—it still is out there. But I’ve just never seen a world where the features, if you will, can get to revenue scale as quickly. And, by the way, you kind of often have to start with the feature because a customer isn’t—think of it from the customer’s perspective, the customer being the business buyer of software—it’s like, “I know I want to be locked into a piece of software for 20 years. That’s what I’m looking for as a buyer.”

No. It’s like, “I have a problem to solve. My problem is I can’t hire a front-office receptionist for my orthodontic clinic, or I can’t call people in Mandarin or Cantonese to get them to repay their auto loans. What do I do?” Something shows up and offers that functionality. Boom, I’m a buyer.

That feature has to backfill product and backfill company as quickly as possible. So that’s still true today as it was 10, 20, or 30 years ago. The difference, again, is that the revenue for the feature is just so high, and the demand for it is so high, because in many cases you’re just responding to help-wanted ads effectively.

Yeah. And so I think the effect of that is that there’s been a Cambrian explosion of interesting markets to go after. I think it’s unrealistic to believe that OpenAI is going to go build the front-office assistant for the dental clinic as their core business. They’re not going to do that across every single market.

The other dynamic is that for many of these companies, part of the product value is actually orchestrating the work across lots of different model companies.

And so I think having one foundation model business going up the stack limits the actual impact of the application, potentially, as well.

Well, I think that if you think about this versus other platform companies, Facebook was the preeminent platform company of Web 2.0, from when they opened up Facebook Platform, which I think was 2007. People built their businesses on top of Facebook. Facebook would never do those particular things. They were like, “No, we’re going to have a platform that allows companies like Zynga to build these farming games.”

But what the platform normally does, if it doesn’t actually go compete with the underlying products, is say, “I’m going to tax it, but I’m going to tax it in ways that are kind of at my fancy. So this week it’s a 10% tax. That’s my promise. Oh, wait. I changed my mind. Now it’s going to be a 40% tax.” That’s why it’s always dangerous to build on somebody else’s platform.

So I think the 2 things to look at are, number 1, will the platform owner compete with what I’m doing? That’s also another Goldilocks zone question, right? I published this graph of VisiCalc versus Lotus 1-2-3 versus Excel. VisiCalc invented the spreadsheet in 1979 and had 100% of the market because it was the only player in town. Lotus built a better version of that. Lotus got to, I think, 70% market share by 1985, which was when Microsoft released Excel for the Mac.

And why is that? Because Microsoft owned Windows. The platform owner normally wins. But that’s because it was such a huge—why do I buy a computer in 1997? I want to use a spreadsheet. It was just so intrinsically linked. That was one of the main use cases for computers in business, right? Using spreadsheets. So that was a violation of the Goldilocks zone.

Whereas with other things, all you have to worry about from the platform owner is that they’re going to tax you, but they might tax you in very, very bizarre ways. Part of what David was saying, in terms of there being multiple model companies, is great. The problem with Windows was that it was 95% of the market. If 95% of your customers used Windows and you were going to build a competing spreadsheet, you were just toast because the platform owner was going to drown you.

Now there are 5 model companies, or more when you include all the Chinese models and open source, so I don’t have to worry about that. But I do have to worry about them saying, “Wow, this is so relevant.” Why is it that OpenAI got a public company CEO to quit her job and become the CEO of applications at OpenAI? Maybe because they have a huge application opportunity.

But this is the nice thing: a lot of these things are obscure, but they’re still big. I don’t think OpenAI is going to go do them because, if they’ve done that, I’d be short OpenAI. It’s like they’ve run out of good stuff to do.

That’s something they should do in 2029. And this is—I think I told you this story before—this changed my outlook on life. I pitched this guy Dan Rose at Facebook, who was running business development there. I said, “This is a huge opportunity. You should use us for payments. We’re going to do this. We can make so much money for Facebook.”

He was so patient and nice. I love this guy. I’m on a board with him to this day. He said, “Alex, that’s such a great idea.” I thought, “All right, I got the deal.” He said, “It’s a great idea, but we’re not going to do it because you’re pitching me a gold brick. We have gold bricks all around us.”

He was right. Facebook in 2010—how much has Facebook grown its revenue? They have more profit every quarter today than they had revenue per year in 2010. It’s just such an incredible company. He said, “You’re pitching me a gold brick that’s 100 feet away.” It’s real. I love that gold brick, but we have hundreds of gold bricks where I just have to stoop down at my feet and pick them up. So I’m just not going to do that one right there.

That’s how these big companies think. But the nice thing is that these gold bricks are bigger than they’ve ever been because you have software that can do the job of labor.

Yeah. Which, on that note, if you were running OpenAI and thinking about which gold bricks—or how you even think about what mental model to use for the things you should be doing first versus things that maybe you should let other people do—how would you be thinking about that question?

I think a lot of it is—well, it’s 2 things. Number 1 is: we want to be the backend for everybody, like the platform. Can we be the platform for pretty much everybody who’s building anything? We’re not going into these obscure spaces, like orthodontic care, at least not until 2045. Let’s make sure that every single developer is using us.

This is part of why Microsoft crushed Apple in the 1980s, because Apple made it really hard to develop software. What’s actually kind of interesting is that Microsoft started off as a compiler company. Their very first products were not Microsoft Office, and they were not DOS. They built a BASIC interpreter for the programming language BASIC, and they had a big business. Their biggest competitor was Borland, which only made compilers. The early rallying cry, if you talked to any early Microsoft employee, was “Beat Philippe.” Philippe Kahn was the CEO of Borland. Microsoft was focused on that and made a lot of money on it.

Apple was like, “We should make money on that too,” and they had a product called MPW, Macintosh Programmer’s Workshop. I remember using it in the 1980s. It was, I think, $2,000 in 1980s money to buy this IDE, or programming thing. How do you afford that? It was like, “We have to make money on that. Microsoft’s making money on this.”

Then, lo and behold, there were 10,000 times more DOS and Windows software products than Macintosh software products. Of course, Apple corrected that mistake when the iPhone came out. Xcode, which is the way that you build products for Mac, Macintosh, and iPhone iOS, is free. They corrected that mistake.

To answer your question, number 1 is: can we be the biggest consumer brand in the world? ChatGPT has 800 million weekly active users. Get that to 5 billion. Even if Gemini 3 came out today and it might be 5 times better, are the people using ChatGPT just as consumers going to switch? Maybe, but it’s unlikely, just because they make that their default.

Then be the backend for everybody who’s building anything. That way, all the gold bricks come to you. David Haber

I think the other thing that we should anticipate—we’re already beginning to see this from some of these big model companies—is what big horizontal applications they can likely sell to every large enterprise.

You saw today, with Google’s Antigravity launch, that the IDE is going to be one of those things. If there’s product-market fit for LLMs, coding is definitely one of the top categories. So thinking about the big horizontal applications in the enterprise is important.

I think there’s also, to some degree—and this will be earlier to play out—the Palantir opportunity. We’re still very early in the proliferation of this technology into large enterprises. At the same time, unlike prior product cycles, like the cloud, if I’m the CEO of a large public company and I’m asking myself, “Do I need to be in the cloud?” it was sort of an esoteric idea.

Today, I can plug a prompt into any one of these models and intuitively understand the impact it could have on my business: the efficiency gains in my customer support organization, in my engineering organization, and in all of my back-office functions. At the same time, many of them don’t know where to start.

So I think you will see a consultative, forward-deployed, Palantir-esque sale into very large enterprises from some of these big model companies. Again, I think we’re early in that, but you’ve heard inklings of this with Anthropic talking about wanting to build into financial services and other markets.

I agree: I think the biggest opportunities are the ones Alex is describing. But I think you will see these companies selectively try to build applications that cut across every one of those, and then they’ll probably choose a few lighthouse customers to build largely bespoke, custom integrations into these bigger enterprises. But where are the ACBs that really make sense?

In Web 2.0, there was a lot of winner-take-most. You were talking about one of the benefits in AI being that there are multiple winners.

To what extent is consolidation inevitable, or how do you think this plays out?

Well, I think if you have 20 companies that are all doing the same thing, what has historically happened is that it's a bad market if there are 20 companies doing it. Then, I don't know, the bottom 15 just go bankrupt. Maybe there's some consolidation where number 1 buys number 2, number 2 buys number 3, and, assuming that we have a functional FTC and whatnot, all of this gets approved. It's not like you're taking over an entire market; this is orthodontic clinic answering software or something.

Then what was a bad market becomes a good market. This kind of goes back to why momentum is important, because if you have 20 companies that are all at the exact same scale, then it's actually great for the customer: The prices go to zero, or they converge on the price of electricity. Whereas if you—and this is not saying you want to go build a monopoly in orthodontic answering software or something—but rather, you can charge more if you get to a certain scale, because whatever the quality of the product that you're delivering at the end of the day is just higher.

You have to get to the critical scale to get there, and sometimes you just need these markets to work themselves out. When I was running my company, TrialPay, we had, I don't know, 20 competitors, and it was tough because everybody would be pricing their product at a loss. This loss leader only works if you end up leading—you have to make money at the end—and nobody really had a plan for that, because the venture capital dollars were really subsidizing everything. That does not make a good market.

What does become a good market at the end? Sometimes this is what Vista, the private equity firm, would do: “We're going to buy 1 as our anchor, we're going to lowball, put the other 5 out of their misery, and now we end up with actually a pretty good product at the end, or a pretty good business at the end, a pretty good company at the end.” I think that will probably play out the same way here, because you just can't have a market where everybody is loss-leading and nobody's big enough to get any kind of scale effects. Is there going to be a world where the 19th player survives? Jack Welch would always say you have to be number 1 or number 2, and there's no value to being number 3 through 100. I don't think that's changed.

Right. Right.

Even in the model-provider example, I'm also curious if prices go down.

Yeah. I don't see how. There actually are—I mean, people know xAI, Anthropic, OpenAI, Gemini, or Qwen. They know the big ones, but there is actually a long tail of things that people haven't heard of, where they've raised lots of money. It's just not it; it works fine, but how can you survive? The model company is the most cutthroat, because unless you're state-of-the-art minus-minus-minus and you're trying to earn a living, it's just not going to work. So that game is super cutthroat.

I think the one area where that may have diverged, and Martin talks about this a lot, is that when markets are growing so quickly, you end up having specialization. In other modalities, in some of the creative tools, people have specialized to serve the upmarket. I'm producing movies; I want to create social-quality content. These are different markets that the models can specialize in. Time will tell how defensible those become over time.

Maybe that's the optimistic take: Early on, everything looks overlapping and competitive, but we're still so early that the market is growing, so everything can expand and people can specialize over time.

Earlier, when you were talking about the feature versus product, didn't Steve Jobs once tell Drew Houston that Dropbox was just a feature?

Yeah. I mean, that's why it's always been this pejorative thing, but that's the point that I was getting to: Nobody wants to be like, “Oh, I need this company.” No, it's like, “I need this feature.” Every now and then, you see a product that is not a feature because it's so far out of left field. Nobody was anticipating ChatGPT dominating their daily workflow in October 2022.

Once it came out, it was like, “Holy crap, this is incredible.” That's not a feature. You could argue it's a feature on top of your iPhone, but no, the iPhone is the delivery mechanism. That's a product, and they've obviously turned that into a company.

Whereas other things are like, why is there antivirus software? That almost doesn't make any sense. Shouldn't the operating system stop you from getting viruses? Why do you need a third-party tool for synchronization between devices? But it turns out the reason Dropbox has survived and thrived since Steve Jobs made that comment is that it's really hard to do well. Once you've built that feature, you can backfill with all sorts of other product, which is what Dropbox has done a pretty good job of.

But it is hard, because this is the danger of building on somebody else's platform: I'm going to build this thing that they should have had, right, if they had the foresight. If it doesn't operate in the Goldilocks zone, it's like, “Wow, this will triple Apple's profits.” Let's just say that Dropbox would have tripled Apple's profits. Would they have dropped everything and focused on building that versus the iPad or something—whatever Steve's last gizmo was? Sure.

But if it's in this Goldilocks zone of irrelevance, like janitorial services, it's like, yeah, they should do that. Platform owners get lazy. This is why half the things on my iPhone don't really work if they're built by Apple. Try Screen Time—any parent that's listening to this, if they've tried Screen Time, it's just an embarrassment upon humanity.

They don't have to go sell it, so they don't have to compete on feature. They compete on the fact that they're the platform. They don't even compete; they just roll it out, and it's going to be bad. That does create an opportunity for somebody to come up with the feature and actually outcompete the platform.

But you have to be careful, because obviously the platform owner is going to compete with you. That's why what I often find very compelling about entrepreneurs is when they know this. They've studied how, from every single platform shift—from AC versus DC current—there have always been these battles over who's going to be the underlying layer.

The best entrepreneurs have studied this and have a plan. Drew knew this. He was like, “I know that there's this stupid comment on Hacker News: ‘This is just like our sync with this, that, and the other thing.’” Of course Drew knows that, but he built this into a $10 billion company because he had a plan.

The best entrepreneurs often understand that it's not this naive: “Oh, I'm going to build this. There's no way that they're going to build it because they're too dumb and stupid.” It's like, no, they're not. These companies, if they get their act together, will marshal a lot of resources to go compete with you. It might take them 5 years, but they will 100% do it.

You have to backfill your feature with a product, and you have to have a moat for that product, as opposed to, “Oh, yeah, the big company will never figure this out.” It's like, that's not true.

I think what's also unique—I wrote this piece a while ago called “The Messy Inbox Problem”—is that it was sort of a wedge strategy that we've been observing across lots of different industries. It's just this idea that you hook into a bunch of your different unstructured data sources. It could be email, fax, or phone.

Tenor, as an example, has trained a model to extract all the relevant patient information from those data sources and plug it downstream into a system of record—in their case, an EHR, but this exists in a CRM, an ERP, or what have you. I think that wedge for that feature is interesting in large part because it lives up-funnel from software, right? You're replacing the human-level judgment of the individual.

Often, it's the secretary collecting the physical faxes and then plugging them into the EHR. Now a bunch of AI companies can wedge in and eat away at all the downstream workflows that might have been the point solutions of software companies. Tener is no longer just doing the messy inbox. They're now doing scheduling, prior authorization, and eligibility and benefits.

They've used that wedge to try to become the end-to-end platform. Eventually, maybe they become the system of record. But again, because you can replace human labor now with software, I think it's creating opportunities for these features to actually become products and, in their case, I think, whole companies.

Well, I think this is the thing that, in my mind, is very dramatically different than every other platform shift: The AI is just so consensus. Cloud was not consensus, mobile was not consensus, and that's why the incumbents kind of screwed up. Sometimes it was just completely—I'll use the Silicon Valley term—orthogonal to their business model, because it's like, I sell $5 million-a-year products, and wait a minute, I'm going to charge $100,000 a month.

That’s just hard. How do I pay my salespeople? How do I make my quarterly numbers? That’s why Workday beat PeopleSoft, or why Salesforce beat Siebel. All of these things played out, but behind it was this concept that the new thing—the iPhone—is stupid.

There’s no version of the famous Steve Ballmer clip of him saying, “Nobody’s going to buy an $800 phone with no keyboard.” There’s no version of that for AI. How do you find a big CEO, or even a small CEO, who’s like, “Nobody will use that tool that makes you 100 times more productive”? Of course not.

This is why it’s kind of a bonanza for most of the incumbents as well, because anybody who has a system of record will add a button or a feature, to use our parlance, that will make them more money. There are just gold bricks everywhere. The challenge, though, is that there isn’t this white space to occupy in the same way that there was for cloud, mobile, or a lot of the Web 2.0 things.

With cloud or mobile, the incumbents screwed up. They weren’t paying attention, and they scoffed at this new technology. Nobody’s scoffing at this new technology. Everybody’s just trying to embrace it. But the opportunity often exists in areas that seem too small and don’t have an incumbent at all. Those might actually turn out to be trillions of dollars of value.

That’s what makes it much more exciting than the last generation, where it was like, “I’m just going to copy everything that was on-premises and make it recurring-billing cloud,” and I’m going to do that at a time when the big guys say, “That’s stupid. I don’t get it.”

Some argue that mobile was ultimately sustaining, and that although there were net-new companies and use cases that were $100 billion, like Uber and Airbnb, some incumbents became trillion-dollar companies guided by mobile. When we look at the business impact of the AI era, what’s your mental model for thinking about the incumbent, startup, or net-new company in terms of value capture?

I think a lot of it is the same. Unless you really screw up the pricing model, or you’re all per-seat pricing, it’s very, very hard to get the market to adopt something that is violently different. You’re operating in the public eye, and your technology team is bad—there are a lot of things that need to happen. I have a hard time believing that incumbents will really suffer.

There probably are some exceptions. Take one example, and this kind of goes back to distribution versus technology: all of these business process outsourcing companies, these BPOs, are the largest employers on the planet. Tata, Wipro, and Infosys.

If I’m JPMorgan and I say, “I need a call center, and this call center needs to have access to customer records, it needs to be safe, everybody needs to be trained, and I need to have 100,000 people who can answer the phone,” you know who can do that for you? Infosys, right? Or Tata. Tata has already done the integration with JPMorgan. In this case, they might just add AI, and now they don’t need 100,000 people. They maintain that JPMorgan contract, and they operate in the Goldilocks zone where they’re going to make 100 times more money. That’s the bull case for Tata.

The bear case is that JPMorgan says, “Wait a minute. We should partner with the startup to do this, or we should do this ourselves.” Now Tata loses that relationship altogether. It could go either direction. A lot of these things are really up for grabs.

I think the default is that incumbents probably will do well. You can pick a lot of these cases. This is why you see the public markets kind of not know what to do. There is a case that is very, very bad for a lot of software companies. But there’s an alternative case, which is that if you operate in the right Goldilocks zone and you have the right momentum to actually build these things and embrace these new technologies, you’ll maintain all of your customer relationships, and you’re just going to have a more profitable business.

The most compelling thing I think about AI, which almost everybody gets wrong, is, “Oh, it’s going to destroy all the jobs.” Our beloved representative from Silicon Valley is trying to eliminate AI. That’s just so crazy—that our elected representative wants to turn us back to farmers of tangerines and whatnot in Silicon Valley.

It’s not like all the jobs will go away. I actually think that’s not going to happen at all. What’s going to happen is that there are a lot of things where, if I could hire somebody for $1 to do this task, I would 100% do that. I cannot hire somebody for $1. I’ve never been able to hire somebody for $1. Now I can hire software for $1.

A lot of these tasks are like taking taxis after Uber, right? Did you hear people say, “You probably took an Uber to get here today, right?” Would you have taken a taxi 20 years ago? No way, right? Where would you find the taxi? How would you arrange the taxi? It was just way too complicated.

Once you make it very, very abundant and less expensive, everybody’s going to use it. That’s what Ro Khanna and his ilk are missing. It’s not like I’m going to say, “I’m going to eliminate all the jobs.” Think of it in that JPMorgan example that I just mentioned. Wouldn’t it be cool if every single customer of JPMorgan Chase could have their own personal friend whom they could talk to every single day, who would help them with every single element of their financial life?

“I’m stuck downloading the app. I can’t figure out how to get it set up.” Talk to somebody in real time who will help you with that. Why don’t they do that? The cost is known. It’s high, and the value is probably low. As soon as you can bring the cost down to $0, now you’re going to start hiring AI in all of these different areas where you would never bother hiring a human, because you can’t train the human, you can’t find the human, and the human is too expensive.

I think it’s a good place to wrap, guys. Thanks for coming to the podcast. Most don’t matter.