The Next AI Moat—Network Effects Beyond the Model

Julia Nimchinski:

And… Now we want to welcome our special guest. Welcome back to the show, James Currier, General Partner at NFX. James has been building and investing in startups for over 30 years, and at NFX, he spent the time studying the patterns behind pattern-breaking companies, from network effects to technology windows, and to what he calls the one metric that ultimately matters, speed. Welcome back, James. Super excited to have you here. How have you been?

James Currier:

Well, thanks for having me. I’m here in the NFX office of Palo Alto.

Julia Nimchinski:

Amazing. Well, I have a million questions, as always. And I’d like to start with your essay, Your Life on Network Effects. Specifically, one quote, and the quote is, the world seems chaotic, but it’s not. Underlying all this apparent complexity is some wonderfully simple math. So AI seems to be creating a lot of chaos, a lot of new kind of chaos. And just seems that sovereign AI harnesses or open models, they kind of try to structure it all. What does the network effect math tell us?

James Currier:

Huh. Well, the, the chaos that we’re in is actually something I think we should all be… Deliciously eating up, because for every 14 years, we seem to have a big technology revolution. And after the first 5 or 6 years of that, it gets pretty boring. We’re all, you know, most of the low-hanging fruit’s been taken, and then we’re just repeating the same patterns. And up until 2022, we were in that situation for about 10 years. And it was getting pretty boring, and so the chaos we see now is just opportunity. It should be seen as a wonderful, time to live through, and not as an enemy.

I think the chaos can only be good for all the people on this call. And for people who are flexible and intelligent. So, that’s the first thing, is not to fight it, but to roll with it. Not to see it as an enemy, but the chaos is actually really positive, and we’re very lucky to have it. Number two, the, The network effects will tell you that each of the nodes is going to end up doing what is mathematically in its own interest. And so, about 3 years ago, we published, maybe 3 and a half years ago, we published an article called the 5-Layer Tech Stack of AI.

And… We just said, here’s what all the nodes are gonna do, here’s what the layers are gonna be, and that’s… exactly what’s happened, because it’s pretty predictable, honestly. And what we are calling a harness these days is really just that second layer of the stack. We called it the OS layer or the API layer. Now we’ve… you know, for every new thing, we have to create a new word, like cloud. And now we’ve invented this word, Harness, to describe that layer, which sort of is the connector between the workflows and the models. But it was obvious that there was going to need to be that layer, even back then.

And so, it is wonderfully simple, once you understand what everybody’s mathematical needs are, and where they’re sitting, and where their position is, and where they have to kind of have to react. So, it’s not that hard to predict how people should behave and what’s next.

Julia Nimchinski:

I’d like to touch on your concept of technology windows, and about 2 years ago, you mentioned that we’re, I think, in Phase 4 of this technology window, meaning the competition. Where are we now?

James Currier:

I think we’re… we’re in between 4 and 5. We are… so if you look… if you go to NFX, and you look at… or if you just go to Google and type in technology windows, you’ll get an article, and there’s a diagram showing a curve that comes up and then flat lines at about 10%. And basically, it points out that there are six phases to every technology window. That’s automobiles, and that’s cell phones, and that’s cable, and that’s… Browser and everything. Mobile. And we’re gonna have the same thing with AI. It’s… it’s been true for all of human history.

And the, you know, the first, the first part is just when the technology are just hobbyists. And then the second is when there’s a success. And then it’s a third when everyone starts to notice, hey, there was a success over there. And they’re not hobbyists, but they’re looking for money and status. And so suddenly everyone gets interested. So then there’s the most intelligent people in Phase 3 are saying, wait, wait a minute, what’s going on over there? Someone just had a success, maybe I want some of that. And then in the fourth phase, we get to the fact that everyone sort of… it now becomes legible, and everyone floods in.

And that’s the competitive phase, and that’s where we were two years ago, everybody was flooding in, all the VCs moved, all their investing to AI, all the founders got off of crypto and onto AI, and, you know, everyone moved to AI everything. And that was 2 years ago. That has continued to be the case. There’s really no other game in town. And now we’re just getting to the point where we’re starting to see some of the incumbents. So the incumbents emerge. Anthropic, remember 2 years ago, was nowhere. Now they’re emerging as sort of even the leader.

But this is the phase where the incumbents start to, reveal themselves. OpenAI, Google, Microsoft, Anthropic, blah blah blah, for this next new phase. And then, we’re all going to be creating sub-segment, you know, sort of sub-window companies within those incumbents. And then… and then those incumbents will start to shut things off. But I don’t think we’re there yet. I don’t think we’ve even gotten to 5 yet. I still think that… XAI could end up being one of the incumbents in this area. I still think that Meta has a role to play. There’s many companies who could still become some of the big incumbents.

There might be a Harness company. I think Open Router’s getting bought, but that’s what’s happening. OpenRouter could have maybe become a harness, sort of incumbent, and control all the models and control the workflows, but they didn’t. They decided to sell, maybe. I don’t know if that transaction’s taken place, but… There was an opportunity for them to be one of the big incumbents, but Maybe the management wasn’t good enough, maybe their vision wasn’t strong enough, maybe their investors were fighting, who knows? These are all very random conditions that create the outcome in the end, but that’s sort of where we are between Phase 4 and 5 in the technology window.

Julia Nimchinski:

Then this framework, you have a concept of sub-windows, as you mentioned, and sub-sub-windows. Do you see sovereign AI and partnerships?

James Currier:

Say that again?

Julia Nimchinski:

within the concept of, you know, technology windows, you… you’ve mentioned you have a sub-window, or sometimes some sub-window. Do you see harnesses and sovereign AI being the sub-windows rising now, or how do you see it?

James Currier:

Yes, that’s exactly right. So there will be sort of a battle among companies and this word, harness, to say, we are a harness company. What does a harness company mean? You know, what are the features, what are the elements of a harness company? So again, what happens is, we get a technology. We then put words around it. And then people put their companies and their technologies around the words.

Julia Nimchinski:

Nope.

James Currier:

in order to win markets, and there’s this dialectic, there’s this conversation between the technologies and the words we use to describe things. I’ll give you an example, which was, back in the early part of the internet. You know, we got this browser, and then, people are like, well, when you turn on the browser, where do you start? And everyone’s like, oh, my start page. And they kept using the word start page, but then one of the journalists used the word portal. And as soon as they used the word portal, then everyone dumped in, what is a portal?

And then they started adding all these features, and then they said, well, these 3, these 4, these 5 are now all portals. And the value of those companies exploded. Same thing happened when we had, sort of software until, you know, software for getting revenue. And then someone called it a Customer Relationship Management, a CRM. And then they came up with CRM Magazine. And suddenly, the value of all the companies that were doing that Quadrupled, because now people had a handle, a language handle on it. So, in terms of the harness, yes, this is a sub-window of a piece of language that we invent, and then we build out what the features of a harness company is, or a harness function is.

And then we’ll see if the models buy the harnesses, or whatever, to see how long that sub-sector stays, or which of them make, you know, make money, which of them are successful, because they got in early enough to this sub-window. And then sold, early enough. And then the people who come late, there’s not really much juice left, and they really struggle.

Julia Nimchinski:

There are companies now, only names, but, they’re essentially commercializing this ability of a Harness to collaborate around code, and, you know, bringing enterprises in. Do you see that as a separate kind, new kind of network effect?

James Currier:

Unfortunately, I don’t see it as a new one, right? I mean, if you go to NFX.com, or you just type in a, you know, network effects map, you’ll get the 16 network effects we’ve identified. That’s a lot of network effects that have different mechanics. This one, you know, is, you know, Very much a direct network effect. Combined with a data network effect. And the data network effect, if you read the article, if you type in data network effects to Google, you’ll get the article, that we’ve written about data network effects and how they work. They really work best in real time.

And one of the things that, you know, we’ve been, you know, this new word, harness. Is, really on the backs of… A word we were using a year and a half ago, or a phrase we were using a year and a half ago, which was, Instead of a database of record, it was going to be a database of action. Right? That was a word people were trying out a year and a half ago. It doesn’t seem to have stuck. But Harness maybe will stick. And people can agree on the value of a harness or whatever.

It’s a more elegant, phrase. But, you know, the idea there was that we were going to have all of the agents and all the people coming to a place where the real-time data was existing, so that they could make up-to-date decisions, and better decisions because it was real-time. And so having a… a system of action. I’m sorry, it wasn’t a database of action, it was a system of action. Harness is really a system of action, as far as I can tell, right now. I mean, it’ll evolve into something different, but that’s sort of where we’re going.

And so, no, it’s not a really new network effect, it’s a direct network effect, and it’s a data network effect, those two combined.

Julia Nimchinski:

Speaking of data, you mentioned many times that data, the founders especially, would tend to kind of over, you know, estimate the value of data. And more data doesn’t essentially mean better data and a network effect. and it tends to asymptote quickly, so I’m just curious, with that concept of systems of action and, you know, the It seemed like, after the trillion dollar graph opportunity, that article, just… I… everyone started to try just to, capture data, enterprise data, and just many companies created around that concept. Do you see that, as a fundamental… as a fundamental change now?

James Currier:

No, I don’t. I continue… I continue to see that, you know, every… every model, has been able to get access to all the data they needed. They could either synthesize it, or they could steal it, or they could buy it, or… they could cobble it together from various sources, and that everybody is moving toward… well, I mean, why do we think we have, you know, 6 models that are all almost identical, and they leapfrog each other every 2 weeks? It’s because there’s no data advantage. And this proves out again and again. And so, my point that I’ve always been making for, I don’t know, for the last decade, is that the only data advantage is in real-time data.

Right? It’s perishable knowledge. And if you have valuable perishable knowledge, like a stock price or traffic for Waze, then you can have the opportunity to have a pretty robust data network effect. But in most cases. maybe 99% of the cases, it’s just not the… it’s just… it’s not true. The data network effect isn’t strong. You can find other data network effects, or you can find embedding effects to give you the defensibility you’re looking for, but the defensibility isn’t coming from the data. And I would say that 90% of entrepreneurs think data’s the way I’m gonna have their… and it’s wrong.

It’s, like, literally 1%. Everyone’s like, oh, I’m in the 1%. Bullshit. you’re not in the 1%, you’re just deluding yourself. Now, if you delude yourself into giving you the energy to do your work, and you’re deluding the VCs because they don’t understand it, and they give you the money, then maybe it doesn’t matter. Maybe we should… you should all just go continue to pretend that the data network affects the real thing, but it’s… it’s not.

Julia Nimchinski:

And speaking about embedding, just curious yourself, with the rise of sovereign AI as a concept. And obviously, we have the classic example of Oracle being the strongest network effect. Do you see Sovereign AI and Palantir changing, that embedding value or future of this network event, per se?

James Currier:

Yeah, so Oracle is not an example of a network effect. They are the best example of the embedding defensibility, where they embed into your business, and you don’t care if another business uses Oracle or not. Right? What it means is that for me to rip out Oracle from my company would just… it’s gonna be giving me so much brain damage that I’m gonna retire before I do that. I’m just gonna kick the can down the road and keep paying Oracle their additional 20% every year. That’s why it’s such a good business, is the embedding, but it has nothing to do with network effects. the sovereign AI movement, I think, is more of an embedding effect than it is a network effect.

It’s… it’s, you know, hey, we’re gonna give you your own box in your own headquarters, where you can run your own models. You can split up, you know, if you have some 10% or 20% of your tokens that you need to send to Anthropic or OpenAI, that’s fine, but they’re not going to get the full picture and steal all of your data, because you’re going to do 80% of it, sort of, on your own servers. I think that’s an embedding effect. Who keeps the hardware up? Who monitors that hardware for you? Who, you know, what service allows you to then have a better workflow software on top of that?

I think those are interesting businesses, but they’re not network effect businesses, they’re embedding businesses. And that’s a fine defensibility. We coach our founders that there’s four, and if you go to NFX.com and just type in NFX defensibility, you’ll get the articles. They explain the four real defensibilities in the digital age. The best ones are… are embedding and network effects, because those are available to startups, and those are still available in the age of AI.

Julia Nimchinski:

James, for your concept, we see… oh, we see 3.0, You mentioned that AI will automate sourcing, decision support, portfolio support, analysis. I’m just curious, how do you build that structure internally? Are you just even concerned about outsourcing your alpha to, you know, the Frontier Labs? Is it a combination of open models, closed models, or do you prefer one or the other?

James Currier:

It’s interesting, I think that the alpha in venture capital has been diminishing very rapidly for the last decade already. And I think we need to remember that in the venture capital community, only 25% of Even 13 years ago, only 25% of VCs beat the S&P 500. So the VC asset class is a small asset class, about $200 billion a year, $250 billion a year, that’s the size of one big hedge fund. And it is an increasingly efficient asset class, so that only 25% are beating the S&P 500 even 12 years ago. I bet it’s… and it takes us so long to measure a venture, because it’s sort of a 12 to 14 year horizon.

I bet the percentage of VCs who will beat the S&P 500 today is Probably closer to 10%, maybe 8%. So, the alpha’s going away anyway, just through competition, before AI ever arrived. Alright. On top of that, we now have AI, which is going to further democratize access to the deals. Right? So, right now, if you lived in Palo Alto and you hung out at the Stanford campus, you could meet Sergei and Larry and give them their first 25 million bucks, and own 20% of the company, and… You know, make a lot of money. Now, the people at Stanford know that they can talk to 35,000 different VCs.

You have… they’ll raise $25 million, and they’ll… you only buy, you know, 2.5%, you know, of the company. And so your returns just systematically drop because of that. And I think AI is going to make it easier for more and more funds to get access to those people at Stanford. And so I think that, yes, AI is furthering the efficiency of this small asset class we call venture capital, and will further diminish the alpha for everybody. And it’s, eventually we will have a Bloomberg of private companies, meaning there will just be hundreds of data inputs.

And every venture firm will spend $100,000 a year to get access to the data feed, and as a promising… you know, there’s only about 80,000 promising founders in the U.S. I mean, everyone wants to think it’s more, but it’s really only 40,000 and maybe 80,000 who know how things work, who are aggressive, who have the connections, who are in the right cities, da-da-da-da, all the things that lead toward the big outcomes. There’s only about 80,000 people that need to be tracked, and there’s already about 6 or 8 of us venture firms who are tracking all of them on a daily basis, and in the future, it’ll be… you know, 300 firms will be tracking them on a daily basis.

They will make a change, and they will immediately get 68 emails, 68 phone calls, whatever, saying, let me give you money. So, I think the AI is furthering the diminishment of alpha in this small sector, and it’s going to continue.

Julia Nimchinski:

It predicted that LLMs will asymptote quickly. I’m curious, where do you see this value moving next in terms of EC money? Is it, the harness layer, or, you know, compute layer, or where do you see it?

James Currier:

Look, I think the LLMs are asymptoting. I think the fact that we’ve got open source models from China that are, you know, 3 months behind at this point is showing you that. It’s also the case that I can use these open source models to do stuff That only the… you know, most of my… most of the VC work that we do and we use models, we can do with the open source models now. Right? We just… we’re just paying electricity at this point. So… we are asymptoting, we’re going to be able to do… now, of course, we have all these frontiers about what can be done in physics, or in chemistry, in biology, and we haven’t got… and that’s the next frontier.

That’s, you know, Anthropic announced that they’re Last week that they’re hiring someone to start doing M&A for data and for other companies in the bio area. So, we are going to see, sort of, scientific edge stuff that normally humans in America have been doing. Humans in America and China and Japan have been doing. for the last, you know, 30 years, we’re gonna start seeing the LLMs do that, so that’s a frontier. Clearly, the data center build-out is continuing, and a lot of venture capital money’s going in there. Clearly, new types of chips are going to be used and needed, to lower electricity costs and to speed up certain types of calculations.

And and just, I think we’re still very early in the, what do I even use AI for? companies are trying to figure it out, consumers are trying to figure it out, like, I know that my wife, at every dinner party, she’s like, so what are you using AI for? Like, it’s still a big question on everyone’s mind. So I think that there’s tons of frontiers still, I still think we’re in the, there’s gonna be a lot of sub-windows, that are gonna be created underneath the overall AI window, and I think it’s still the frothy, wonderful next 3-4 years for investing around all of these things that are around it.

I think the LLMs are just the beginning of it. Just like… Just like hardware was the beginning of computers in the 70s and early 80s. Just the beginning.

Julia Nimchinski:

James, in terms of the debate between open models, closed models, and hence the open systems, closed systems, do you see fundamentally different network effects forming in each of those?

James Currier:

Yeah, I mean, look, I think the only… Yes, definitely. You know, if you’re OpenAI or XAI, you do want to build some sort of a platform, like a marketplace, like Salesforce did with Force. Or Apple did. Remember, Apple was worth $40 billion market cap until they launched iOS and got their app store. Alright, so that becomes their network effect, is all these apps going through there. And then they have the embedding because they’ve embedded into your life. Right? Because this is the thing you carry around. You’re not carrying around both an Android phone and an Apple phone.

So they did some embedding, and then they added their… their marketplace. Force added their marketplace. Their market cap was at, like, $17 billion. They added Force. It jumped to, like, $100 billion over the next 2 years. So, if you’re a… if you’re a big enterprise, like OpenAI, or… or XAI, or Anthropic, you want to build in that type of a network effect, a marketplace network effect. It’s not clear how you do that. But they’re trying to figure it out, and they should be trying to figure it out. If you are… a harness company, you’re gonna have different types of, sort of, real-time action, network effects that, can kick in.

And, each of these each of these companies is basically going to need to play with both embedding and with network effects, and they’re going to take flavors, and so that’s why we drew the network effects market map that you can get on NFX.com. It’s like, these are basically colors you can paint with. Right? And there’s… there’s actually, I have a really interesting article on NFX called Reinforcement. And the reinforcement article takes you through… and then there’s the other one on Facebook and Uber, showing you how you add different defensibilities into your business, and you paint with these different colors, essentially, to make a defensible business.

I mean, Uber is not a defensible business. The network effect they have in the marketplace is very weak. For instance, if you… if I got 400 cars in San Francisco. they’re going to arrive almost as fast as any Uber or Lyft at this point. They are vulnerable, but they have made their business defensible in other ways, and so they’ve painted with different colors. And so, yeah, I think everyone’s going to choose which colors they paint with, and there’s going to be lots of different beautiful paintings that founders can make.

Julia Nimchinski:

bringing this back to B2B, and, you know, if it’s a subwindow, technology window, You mentioned there are currently something like 1.8 million SaaS salespeople in B2B. And it’s obviously they all have relationships, and they are, in a sense, embedded, virtually into those relationships, and obviously assist in for being network effects for all of these incumbents. For AI-native disruptors, do you see FDE as a role and concept being a wedge?

James Currier:

Yeah, so the FDE thing… It’s interesting. Yes, of course it’s a wedge, but only if your ACVs are 10 million. Per customer. It’s just too expensive. To have this level of mind that you have to pay so much, because they have so many options. To embed them in your customer, unless you’re charging huge ACV. So, you know, obviously Palantir is charging those types of ACVs and more. And so, the FDE approach has really dovetailed well with their ACV and their product. Product, ACV, FDE, all works. But you can’t, as most companies, just pick up an FDE and make it work for you.

It’s… it’s… It’s, very rare that a B2B software company, even an AI SaaS software company, has a high enough ACV to support it.

Julia Nimchinski:

You’ve also written that, a new technology window can just reopen a certain window that got closed. With just industrial windows in the company that obviously getting a lot of attention on Adams. What is your take? Is it… is it reopening, the industrial window, and what is your take generally in building a company in stealth for something like 8 years, I think?

James Currier:

Oh, you’re talking about Travis’s thing? Look, I… Okay, this is two separate issues. One is industrial tech and what’s going on with that. I think there is a new opportunity for industrial tech, mostly because the U.S. government is willing to give out such huge loans for people who are doing industrial-type things. But the returns on the… on that money is going to be far less than it was for software, just fundamentally. Same thing is true, I think, for defense and for bio. It’s like, the returns characteristic were very special for software between 94 and 2014.

You know, and today as well, if you’re doing an OpenAI type of a thing. I think someone, however, did a calculation about somebody who invested in OpenAI very early, and they’re still only making 100 times their money, right? It’s, It. These industries that we’re going after have much worse returns characteristic than software did between that beautiful window between 94 and 2014. And so I think that’s the first thing to say about industrial tech and these types, you know, these types of things. But I think, in general, the story that they’re spinning about cloud kitchens. And saying, I’ve been in stealth for 8 years.

I think that’s just a fun… they’re just having fun. They’re… they’re seeing what sort of outrageous thing they can say, and see if it sticks. At this point, they’re just having fun. No, they weren’t in stealth for 8 years. It’s just… it’s just they’re sitting around, and they say, oh, wouldn’t it be amazing if this is what the story was? Yeah, let’s try it. he… he started building a cloud kitchen, he thought that would work, it didn’t. He tried something else, it didn’t, he tried this, it didn’t work. And then in the end, he says, well, here’s the story.

You know, I was actually thinking about building AI, but in… in his own… in his own interview, he says, I wasn’t really thinking about it this way. So he admits it. you know, he’s… he’s a very smart guy, and he’s also wealthy enough that he doesn’t need to be completely inauthentic, but I think it’s more of a fun game they’re playing by spinning that story, that they were 8 years in stealth and an industrial tech thing.

Julia Nimchinski:

James, I’d like to start to touch a little bit on the topic of militarization somewhat, though. It seems to be reopening as well. Can you comment on that, the reasons, and how do you see the impact of that window towards all the other windows?

James Currier:

Yeah, I think that, back in 2000… I don’t know what it was, 14 or 13, Elon Musk got the U.S. government to change… Their procurement rules. So that the government would be more capable of allocating Capital and contracts to startups. That aren’t one of the five big… Sort of defense contractors. You know, which had started out as 60 defense contractors, then consolidated over 80 years into 5. And the procurement system had metastasized around these five. And Elon Musk came in and broke that, just by explaining to people, look, we aren’t NASA, but you should still be giving us some capital.

And some contracts, because it would be good for America, and they believed him, and he turned out to be absolutely freaking right. And on the basis of that success. The system has opened up to startups, and so that is what has changed the technology window and reopened it for startups. Plain and simple. Andrew jumped in. Andrew, because of… because of, you know, Founders Fund’s intimate relationship with, with Elon, they knew that this had now changed, and they saw the window was opening for defense contractors directly, you know, weapons builders. And so Anderil came from that, and that was, that was, you know, again, prescient and, based on… on… on prior successes.

You know, and so, yes, the defense window is open, And, I would say, what we’re gonna see is just… I mean, more innovation in… weapon systems than we’ve seen between, let’s say, 1950 and 2020, because we’re going to apply AI, and we’re going to apply drone technology, and we’re… There’s been a bunch of changes in, systems controls and software and AI that have enabled a whole new type of warfare, so everything’s gonna need to get redone. So there’s a… there’s a redo. That is happening, so there will be a lot of opportunity for startups I think.

In the next… 15 years. It’ll be… it’ll be a big window, and… And, hopefully we’ll see a world that has fewer wars. And that, you know, the Ukraine war and this Iran war, Will be the exception, not the rule.

Julia Nimchinski:

Fingers crossed. Thank you so much, such a pleasure. Yesterday, you announced a teaser to your new masterclass, Speed. Where should our community go? And when is it coming up?

James Currier:

Yeah, if you go to NFX.com, there’s… I think we’re the second most popular VC website after Andreessen’s because of our content, so it just turns out that we like making content, and we’re good at it, and… And hopefully you guys will like it, and we’re coming out with our second masterclass. The first one was, of course, on network effects, which is why we’re called NFX, Network Effects, but the second one is called Speed, because we think it’s the number one indicator of success, and so we’ve got a whole… a whole video series on that, and that’s gonna come out in the next couple weeks.

And, and all the other content is there, as usual. So, feel free to hit NFX.com and sign up for the newsletter, and I think we have 200,000 people on our newsletter. And we clean it out, so we’ve, over time, we’ve had maybe 300,000 people sign up, but really 200,000 active people, so it’s pretty popular, so enjoy.

Julia Nimchinski:

Thank you so much.

James Currier:

Thanks, Julia.