# AI Strategy: Missing Network Effects

**Podcast:** Another Podcast
**Published:** 2026-02-28

## Transcript

Hi, I'm Tony Cam Brown.
And I'm Benedict Evans.
You've finally written an essay that you've been thinking about for what, six months?
Yes, messing about for six months, exactly.
Yes.
Pressing publish today.
Yeah, pressing published.
Well, I've got to write the summary paragraph and then we'll press publish this afternoon.
Um having started six months ago saying, well, why are all these products identical?
What's the product strategy?
And then pushed forward and just about to publish something now about chat about OpenAI and ChatGPT and how it competes.
And it's worth just saying before we dive into the conversation, I think this has been the most fascinating thing about all of our discussions and with anything related to AI, is that it feels like we're going round and round in circles, but we're not.
It feels like it just the conclusion or having that aha moment of, oh, this is what I'm trying to articulate, just seems to be taking a little longer than we're probably used to, which I just find in of itself really fascinating.
Well, it's is the hard part is always working out the questions, not the answers.
Yeah.
And when something is puzzling and confusing and new, then the hard part is working out well, what exactly is it that you're trying to work out here?
Working that out is easy.
It's is working out what the question is.
It's hard.
Um and one of so as as we we were we we were chatting, I thought like a good place to go would be to think about network effects.
Yeah.
Which is, and the point here being that you know all consumer tech for since the 80s certainly has been based consumer computing has been based on network effects.
Yep.
That we had Windows and Intel, and then we had with the web, we had Google and then Facebook and Amazon, and then the things that Facebook bought.
And we had uh the iOS and Android, which again are based on network effects, their dominance is based on network effects, and Google Search is based on network effects.
And then we look at generative AI, and from ground zero, from from from day zero, you can't see network effects.
Yeah.
There are no network effects in building the models yet.
There may be in the future, but at the moment we don't know.
And so then you say, well, what is the basis for you there where one company could pull ahead of all of the others, or two or three companies could pull ahead of all of the others?
Where is there a winner takes all effect?
Where is there some mechanism whereby everyone has to use your stuff like it or not?
Everyone has to build for you, and also your product gets better than everybody else's, no matter how hard they try.
Yeah.
And that's what you see with Google is Google search is just better than everybody else's search because it has more market share, and that means the search gets better.
And everyone has to use iOS or Android, and every developer has to support them because those are the platforms.
And you know, every now and then someone applies another one and it can't break in, and Windows Phone fell away, and so on.
And Microsoft has spent God knows how many tens of billions of dollars on Bing and it can never catch up with Google.
Yeah.
And so that's been the kind of the base dynamic of every mature tech sort of sort of consumer tech we've had since the 80s.
And we don't see a way that you can have that in the model.
And so that raises a question.
Um, presuming it doesn't emerge, is we are likely to have scale effects to the extent, I mean, obviously, we already have scale effects in building the models.
It's billions of dollars, and the infrastructure is hundreds of billions of dollars.
Like the big four platform companies spent four hundred billion dollars on capace last year, and they've announced something in the region of six hundred and fifty billion dollars this year, maybe more.
We'll find out.
Yeah.
So there's a lot of the conversation we had last week of like that sheer amount of money and that scale is probably why we're continuously having these conversations as well, because the numbers are it is, yes.
So the scale there, but then there is a question of well, does that give you leverage further up?
After all, like TSMC isn't just you know, so so running back a sec.
So it seems quite likely that all of that will settle out at some point with some kind of oligopoly, where like the laws of gravity and the laws of financial gravity kick in, and well, this is the amount of money that the revenue that you can get, and this is what it costs, and this is therefore this is the number of companies that and this is how hard it is, and therefore this is the number of companies that can be supported in this space.
Yeah.
With this with this margin, and that's probably something between three and six, just picking a number.
Yep.
And they will reach some kind of price equilibrium, which is what's happened with the cloud, example, and some kind of margin equilibrium.
Fine.
Does that give you anything further up stack?
Because Intel had network effects, but didn't really have much control further up the stack.
TSMC has a monopoly, and people don't like TSMC apps.
Right, no developer in San Francisco had ever heard of TSMC before before the AI boom kicked in.
And so what generally happens is controlling the lower lows of the levels of the stack, the whole point of a stack is that it's abstracted.
Yeah.
Like Cisco didn't get much say in what the websites were, even though it might all have been running on Cisco readers.
And so, okay, great.
OpenAI manages to break in for the sake of argument, even though they don't have an existing business.
They get into the $100 billion, multi-hundred billion dollar CapEx Club.
Yep.
Five years' time, there's three, four foundation models.
They're each spending X hundred billion dollars a year on this stuff and managing to support that with revenue fine.
What's built on all of that stuff?
Are those companies effectively the new hyperscalers or indeed the existing hyperscalers selling this as commodity infrastructure at low margin?
Yeah.
At at marginal cost.
And then there's thousands of stuff that are built on top of it and run on top of it, the way you've got thousands of stuff built on top of and running on AWS or Azure or Google Cloud.
Or is it more like the thousands of iPhone apps and Apple's in control on the thousands of Android apps and Google's in control?
And they set the agenda and everyone has to use iOS.
And you look at the stuff that OpenAI have published and they say, well, yes, this is going to work like Windows.
We're going to be a platform.
And you think, well, that's a great thing to say, but it is a developer.
I had to build a Windows app because all my users are on DevWindows.
And as a user, I had to buy Windows because all the software was for Windows.
If I install Snap on my iPhone, I don't know.
In fact, I honestly can't remember which cloud it runs on.
Is it on Google or Amazon?
Maybe both.
Who cares?
As a user, you certainly don't have to know.
Yeah, yeah.
Is it as an enterprise?
When you buy enterprise software, it's running in the cloud, fine.
Who which one?
Like, why would you care?
Yeah, it's gotten to the point today where we don't care about that underlining.
Why would that mean?
Why would that mean anything to you?
That's not the abstraction layer you care about.
I mean, maybe you have like, you know, sovereignty issues or compliance issues or something, but otherwise, like that's not your problem.
Yeah.
Um so you're not Windows, you're AWS, except you're not AW AWS competing with Google Clouds that can't execute your can't execute enterprise.
You're competing with Google and Amazon and Microsoft, a whole bunch of people who want to build clouds.
Yeah.
Um and want to serve this stuff.
So it does sort of seem like in principle that yes, the you know, providing AI in the cloud will be the new hyperscaler thing, but that will reach a cost equilibrium, maybe high margin, maybe not.
But then all everything else will happen on top.
And so then that gets you, you know, this is gets you back to to looking at open AI or or indeed Anthropic and think, well, how what's your path to being Microsoft or Apple?
What's your path to being Google?
What's your path to being the actual platform?
Um, you know, the thing that's more than just you know a layer of layer of infrastructure.
How is it that you're going to comp out compete every entrepreneur in the tech industry trying to build cool new stuff on top of these models?
You don't have the existing feature set that Google and Meta and Microsoft and Apple and Amazon and everybody else have where they can add this and and Salesforce and you know Workday and you know, fake move and everyone else, some of the some of which will survive, some of which will be destroyed.
If you don't have surface area where you can make listed feature or where you can make it distribution.
You're not going to out compete why combinator of coloring it with cool new ideas.
So, what's your path to making yourself to breaking into the club of the big four fully integrated vertical stack companies?
What is it that you're going to be building?
And how is it you're going to invent that?
Well, it's interesting because you started talking about the fact that there is no network effect, but we've also started talking about vertical integration and the examples that you give, Apple being a perfect one, is that those overlap in practice.
And in the you know, in the case of Apple, it has both that vertical integration and that network effect.
Are you saying that if we don't have network effect?
Is the idea here that if we can't find network effect with the AI products today, that then the vertical integration and that vertical stack maybe becomes even more important for a company?
Well, I I was almost almost actually putting it the other way around, which is that the vertical integration doesn't seem to come with network effects.
Okay.
So you know, providing the full stack of tooling doesn't mean that anybody has to use the full stack of tools.
Okay, which wasn't the case previously.
Yeah.
Okay.
And so the challenge then is um you are out there competing, everybody's out there competing in this industry against each other.
Yeah.
Without sort of fundamental strategic advantage that nobody else has.
Yeah.
You don't have that growth and you don't have that adoption model because you don't have the network effects.
Well, it's not it's not so much it's not it's it's not so much it's not so much that.
It's that Apple is doing this thing that nobody else could do.
Um because they had this completely unique operating model.
Google was doing this thing with search that no one else could do because no one else had Google's once flywheel starts going, yeah, nobody else has the virtuous circle, the network effect that Google has.
So it doesn't matter how much money Microsoft spends on Bing and how many clever people they hire, Bing can never be as good as Google.
And so the the issue here is you know you can hire you have a whole bunch of really clever, really aggressive, really driven people, and you have to execute.
But then you're doing this thing, you've got this strategy, and you're following the strategy, and that is what's delivering fundamentally the defensibility and the sustainable competitive advantage of your company.
It's your strategy is not hire lots of clever people.
Yeah.
Yeah.
Because that's a you know, that's something you do in a commodity industry.
You know, your strategy is something else.
You know, Facebook strategy, you know, there's lots of ways you can talk about Facebook strategy, uh, uh a lot of which is about great execution, but you know, it's all there are other things going on there.
There are the network effects of social networks and the continuous willingness to disrupt yourself and try and jump on to the next network effect.
The same thing for Amazon.
Amazon famously published this flywheel diagram of you know, more volume, more customers, lower prices, better service, more selection, all of which are self-reinforcing.
Um the end of last year, OpenAI published a diagram of they said this is our flywheel, and the flywheel says more capex, more infrastructure, more revenue.
That's not a flywheel.
Yeah.
That is not a virtuous circle.
And it's interesting because in the same way.
Well, just if you can make go back compared to comparison with Amazon.
The Amazon flight diagram does not say more warehouses, therefore more inventory, therefore more revenue.
Yeah, yeah.
That's not the Amazon flywheel.
That's not a flywheel.
Yeah.
And all of this sort of gets you to, well, you don't have those lock-ins.
You it's not ordinary.
We don't know what the strategic, we don't know what the fundamental strategic differentiation and competitive advantages would be of running an LLM.
They might be other stuff you have, like you've got a search engine, you've got all these other user growth, but the model itself, it's not clear what you could do to pull ahead.
Because it used to be pretty clear.
I mean, especially in the example of social media of just that it's about users creating value for other users.
I don't see how but it should be for something like an AI product or an AI, you know, a Claude or a chat GPT, you would think that users creating value for other users is the this is the perfect scenario, but when that's not the same.
Well, at the moment that's not how the models work.
Yeah, that's it.
And there's a sort of theoretical question of, you know, with something like continuous learning or various other things, can you get to a point that more users would make the models better?
At the moment they don't.
At the moment, the companies all say we indeed we don't train, we don't use your data to make the models better.
Partly because the amount of user data involved is isn't enough to make the the relative to the the broader scale of the training data.
Um but whatever it is, this is the thing that I was puzzling over like last summer, is well, what's the strategy for why you would win, other than we're just going to be better?
Yeah.
Because historically that wasn't enough.
That wasn't, yeah, everyone's strategy is to be better, but you were doing something that other people couldn't do.
It's like saying, Well, what was Samsung's strategy to compete with Apple to be better?
Well, fine, but that's not a strategy.
You know, we will out execute is not a strategy.
Now, of course, you can take the counterpoint here and say, well, if you look at the reality of Microsoft or the reality of Meta, they would never thought they'd won.
They're always looking over their shoulders, they're for always paranoid.
And they're always jumping on to the next thing.
But they're always doing that from this position for within this strategic positioning that you know, this is what we've got, this is what we're going to do with it, these are the levers that we're going to pull that other people can't pull, and how we're going to do it.
And I don't know what those would be.
We don't see what those would be in in generative AI.
Now, at the point at the moment that you've got stuff that's built on top of it that's some other product, like if you could make some, you know, if SORA had worked, there is a strategy for how you could make something that nobody else can do.
But it's much less apparent what that would be in this field.
And the the other, and I was sort of looking at this and listening to um one of many podcasts with many people in in AI.
Like I sort of Robert Carey said turn every page, but he didn't have to deal with you know a million three hour podcasts to listen to or search through or read the read the transcripts off.
But there's a a comment from Fidji Simo, which reminded me of something that I heard last year, I heard both Kevin Wheel and Mike Krieger say, which is that basically you get your head of product at OpenAI or Antopic, but Mike Reger was both of those people have now left.
Um, your head of product, you get into the office in the morning, you open your email, and there's an email from the researcher that says, Hey, we've got this cool thing, what can you do with it?
What are you gonna use?
How are you gonna use it in chat?
And there's a quote that I put at the beginning of my essay from Fiji Simo saying, um, you know, Jacob and Mark set the research agenda, and then like six months later, researcher emails me and says, Hey, I've got this cool thing, what are you going to do with it in enterprise and consumer?
And I just think that's really, really interesting because what that's saying is the head of product isn't setting the product strategy and has no idea what the product is going to be in a month's time.
Yeah.
Because there's going to be stuff coming out of the research lab that they don't know about that the research lab doesn't know about, but they certainly don't know about, that's going to shape what it is that you're going to be building.
So you don't really know or control your roadmap.
Yeah.
Now I I've paired this quote from Fidji with um the classic quote from Steve Jobs from 1997 or so when he went back to Apple, and he said, You can't start with the technology and work to the user experience.
You've got to what start with the user experience and work back to the technology.
And what's happening with all of these labs is you don't know what the technology is going to be next month.
And the technology emerges and completely changes what's possible.
And so you're a strategy taker, not a strategy setter.
But it's an interesting one because if we also don't have the network effect, who's set so who's setting the agenda?
Well, the problem is that what's the problem is that's what's happened so far is that people leapfrog each other every couple of weeks or every month or two.
But because everyone is basically building the same stuff, um, nobody has anything unique in the models.
And of course, you know, as we we said before, if you're spending all day using this stuff every day, then you get very, you know, you have you'd think that's nonsense.
And yes, of course there's a huge difference between this and that.
Um and certainly if you're like if you're doing image generation or code or something, then there's differences.
But if you are, you know, the 95% of people who aren't paying for chat GPT, or the 80% of people who are using it only once a week, not every day, every couple of days, not every uh and only using it once a week or every couple of days and not using it every day, then you don't see those differences and they're all kind of the same.
And how do you do how do you differentiate, you know, it's the it's the the Netscape comparison again.
How do you differentiate it when the technology is the same and the product is basically the same?
Is it possible to make the product on top of the technology different?
Could you make your web browser different from someone else's?
And the answer was no.
Um so it came down to brand and distribution, which is why Microsoft was able to crowbar their way in.
I mean, I hate to say it this is having worked for Mark and Reason, but you know it Internet Explorer 3 was better than Netscape.
Netscape got really bloated um and slow.
Um but under fundamentally, bit the the reason that Microsoft was able to crowbar their way in was because the browser itself was just a browser, it was a commodity.
And so that becomes brand and marketing and distribution, unless there's some other thing that happens or some other layer of product on top of that.
And it does become interesting that then the market, the marketing, this is where the marketing teams are rubbing their hands going, and it's on to us now.
It's all down to us of how we market these products.
What I found absolutely fascinating is depending on who I talk with, what products seem to be top of mind.
And for people who are not in tech, I found it absolutely fascinating how anthropics clawed seems to be just top of mind in the way that they've built their whole branding and their marketing around this lifestyle product.
And it's I've been just fascinated of people that are not into tech that aren't using AI, who are just like, oh, I went to this Claude cafe, and I'm like, wait, what are you talking about?
This is this is the marketing, you know, experience that's been built by the marketing team of Anthropics Claude.
Um I just it's it's fascinating to see the different avenues that the marketing teams are going to take in terms of Well, Claude is interesting.
I was because Claude is an interesting case study here because I mean, as we were chatting about this earlier, that you know, if you look at the you know survey data on use it on usage, it's it's ChatGPT, and then at sort of two-thirds of that to a half of that is Gemini and Meta AI.
Even though Meta AI is a is a is a fiasco, Lama 4 was a fiasco, the new models haven't launched yet, but consumers don't know.
As far as the consumers are concerned, what's the difference?
They're all kind of the same.
And meta benefits from that sort of kind of network effect.
Well, it's also a distribution.
It's the distribution.
They'll have surface, they have they have surface area, they can use as distribution.
And then clawed is a rounding error, you know, it bounces between zero and one percent, even though the models are at the top of the benchmark charts because they don't have consumer awareness.
And the um the Super Bowl ads.
And the consumer, wait, and the consumer awareness that they do have is with people who seemingly aren't using this product on a day-to-day basis the way you've just explained it, which is fascinating.
There seems to be like a gap of just like there's these people who know about your product, but they're not the ones actually using it.
But you're I guess you're getting a good product awareness out of it.
Well, the ads thing, the ad thing is fascinating and and slightly hilarious because they they they come up and they say, Well, we're not going to do ads, ads are bad.
And they said, but but we might do them in the future.
Okay, so why have you just wasted my time?
Plus, um, of course, it's easy to say you're not going to do ads when you don't have a consumer audience, you know.
I mean, I think when they did that announcement, I posted a picture of um Don Draper saying Sterling Cooper will no longer take cigarette advertising after they've just been dumped by Lucky Strike.
Like, well, it's easy to like, no, I didn't dump you, you you didn't dump me, I dumped you.
Like, no, we're not, you know, we it's not that we can't get any advertising because we've got no users.
And it's there's something also interesting because I've all I uh weird tangentially little F1 parallel here.
I've always said the sponsorship in Formula One is a reflection of our time.
In the 70s and 80s, it was tobacco, then we had spirits, then it was oil giants, then we had crypto and tech, and now it's all AI.
And actually, now that I think of it, Claude um Antropic and Claude have become uh big sponsor for the Williams Formula One team, and it is just fascinating.
I'm very curious to see what they're hoping to get out of that advertising.
And is it just just like okay, is it just brand awareness?
Exactly.
Or wait, or and look, they've positioned themselves as the thinking partner, I think, the thinking partner of the Williams Formula One team.
So I think there's going to be a use case of how they embed and use AI and their product with the team.
But it was a fascinating one for me.
I was just like, oh, but again, very lifestyle, very of the moment versus.
Yes, it's a brand, it's a it's a brand of marketing conversation.
Yeah.
And I think, you know, this one should be clear here.
The technology, when the technology is, there's a sort of a slightly different point here, which is when the technology is basically undifferentiated, the product is undifferentiated, and you've got radically different market shares.
That tends to be an unstable situation.
And you know, even, you know, we know we talked been talking about network effects, even you know, the canonical example of the failed first leader of of MySpace.
At least MySpace had a network effect.
And it turned out that network effects in social are actually quite fragile, and you can have many of them.
So you can have Instagram and TikTok.
Um, but you know, here we don't have any of that.
And so you have these, you know, pick your, you know, Gemini and Chat GPT and Meta AI.
There's also Groc, which no one outside America was ever heard of, um, except as a place that you can get child porn.
Um, which apparently is free speech.
Sorry, we should make that clear.
That's free speech.
Oh, good God.
Yeah, I know.
Um, but you've got this diff very big difference between the technology being basically the same, the product being basically the same, the consumer awareness being radically different, the adoption being radically different for all that you know, open AI.
I'm sorry, Anthopic is now trying to do a do a combs push.
You know, you go and look at consumer awareness or even something simple like like Google Trends, and no one's heard of it outside of tech.
That may change.
It may not.
I mean, it's not clear whether Anthropic, I mean I'm slightly puzzled by Anthropic, because I'm not clear whether they actually do want mass consumer adoption or not.
I I really don't understand what what my Krieger has been doing there for the last year.
I tried to install Claude Cowork and it asked I I tried Claude Coworking and it asked me to install Git at the command line.
I'm like, yeah, you're not really looking for consumer adoption here, are you?
Whereas I think I use it in a very different perspective and I like the product and I like the UI and I'm just like this this works for me.
But I Yeah, yeah, the product's fine, but it's not but the point is the product is complete the product, you know, they have not attempted to get distribution or consumer awareness until now.
Now they went and spent a bunch of money on a on a um a Super Bowl ad.
So at least people in America have heard of it.
I guess my the question I had for you at the at the start when we were previously talking, and again, maybe it is the wrong question is so do we need to figure out network effects?
Us we not being us, but do these companies need to figure out network effect for us to need to or is there something else?
Well, some strategic some some form of strategic leverage wherein you're not purely dependent on turning up every single day and being better than everybody else because you can't plan for that would be good.
And I'm sure everyone would like that.
Yes.
And I'm sure that's on everyone's mind right now.
But at the moment we don't know what that would be, and this is sort of part was part of the kind of the th the thread of the piece that I'm that I've just will have just published when when people are listening to this, is if you're open your open AI, you've got, don't fundamentally have a technology lead.
You're first among SQLs at best.
You've got this giant user base that's not that has very shallow usage and engagement and isn't really locked, it isn't locked in and doesn't have a network effect.
It's very clear that there's a whole there's a whole bunch of other stuff is going to get invented around what all of this is and how it will, how all of it works.
And there's no particular reason why it would be you versus anybody else that invents that because you're competing with the whole tech industry to invent all of that stuff.
It's as though you're, you know, your Netscape in 1994.
Okay.
And then like you're trying to work that out without having the existing um feature um products that you can use as distribution and where you can add your stuff, it's features that Google and Meta and everybody else has.
Not that they're all doing a great job, but they're starting from a different place to you.
You're starting from a black completely blank, blank, blank, scrap, blank scrap, blank slate.
Um also you don't have the cash flows that they have.
Um, and so you're even though you were first and you've got all these users, you're still kind of back at the starting line, surrounded by five or ten other giant companies plus two or three thousand entrepreneurs who are trying to work out what happens on top of this infrastructure.
And meanwhile, as I said, you're a strategy taker, not a strategy giver, because in the end, like uh stuff comes from the lab, like you get an email tomorrow, you get an email next week, an email next month, maybe, maybe not.
I don't know.
Will something come?
Will something not?
Will it be a lot of things?
So building a bit of advantage is gonna look a lot different, potentially.
What is your strategy?
What is your plan for why your product will be better than everybody else?
Having better people, I mean, that's kind of a plan if you can't get a better one.
Yeah, I feel like that's an incredible place to end.
But having the how are you because what's your plan to have cleverer people than Anthropic and Google and Meta and Apple and Amazon?
I mean, you can sort of, I mean, we saw like half of Groc's founders have just left in the last month or two for a whole bunch of different reasons.
Um But what is how can you plan to and you know there's some of this is you can look at Google and say, well, Google clearly was a massive execution machine in the 2000s, and Facebook was a massive execution machine, and Apple was a massive execution machine, but that's not why everyone uses Apple devices.
It was not just we've got Tony Blevin and he's gonna do better procurement deals with Shenzhen, and it's not just you know Craig Federighi is going to be really great at having a regular cadence of shipping.
Yes, you execute really, really, really well.
But what are you executing?
You're executing something other people can't do, and there's no uh stuff other people can't do in this field yet.
And if there is, you're gonna have to work it out.
You don't have it now.
I like it.
It feels like good, it's that's become, I think, uh, the themes of our podcast.
It ends on a good good couple of questions to think about.
I mean, this is the thing we were talking about that um, you know, what do you mean when you say platform, or what do you mean when you say power ecosystem or leverage or lock in?
You know, do you have something that other people can't do that makes your product better and your company more successful?
And at the moment, I'm also not clear that anybody in this field really has that.
There we are.
Twenty eight minutes, not two and a half hours.
You see what happens when you don't have answers?
It's great.
You don't have to go round and round.
Um, I love it.
Fantastic.
This is very interesting.
Good to chat.
Good to speak to you soon.
Bye.