# AI Token Pricing, Infrastructure Shifts, and Market Dynamics

**Podcast:** Another Podcast
**Published:** 2026-07-15

## Transcript

Hi, I'm Benedict Evans.
And I'm Tony Karen Brown.
I wanted you to say your name first today.
Yeah, switching it up.
Let's have another 20 minutes of pointless banter before we get to the point.
Absolutely not.
Let's do it.
What are we talking about today?
So yesterday, I probably something I've been chatting about to various people in various ways over the last couple of months, which is how we should think about token pricing.
And as hopefully everyone listening to this will understand, like in the last six months or so, the sudden product market fit of agentic coding means that demand for tokens has gone up by many orders of magnitude.
And the model labs have been sort of scrambling to adjust their pricing.
We have not been able to build capacity fast enough even before this.
And now we're definitely out of supply and demand is massively out of whack.
And so there are all sorts of involved conversations about token pricing and data center build out and GPUs and A6 and so on.
But the thing I kind of wanted to do was to step back from that and say, well, OK, where is all of this going to settle down?
Like, not what's going to happen in six months, but what is the long term steady state likely to look like?
And it struck me there's kind of two things you can say about where we are now.
One of them is like it's out of whack, it's out of equilibrium, but the other is it's transitory.
Because on the one hand, you've got...
half a billion, sorry, half a trillion, a trillion, $2 trillion of CapEx coming down the pipeline.
And we keep getting new models and the models are more efficient or less efficient.
And we, on the other side, we have all of this demand from really one use case that's got product market fit.
And what happens if we have more?
What happens?
What will the next use cases be?
What will the next models be?
Does Trump ban open source?
So all of the drivers of supply and demand are kind of in play.
And meanwhile, we sort of know that you have positive gross margins on inference alone of sort of 40, 50 percent.
But you've got the cost of building the next model and you don't know where the cost will move or the cost of the next model will move.
And you also don't really know what the ROI on agentic coding is yet.
And you don't know what the ROI on anything else would be.
Which is kind of a long way of saying, look, theoretically, token pricing should be a function of supply and demand between marginal cost and ROI.
But we don't know what any of those actually are or where any of those are going to be in five years time.
So like, who knows?
And I think it's kind of important to sort of step back and say, like, there is a stage in the S-curve in the development of a big new technology where...
Like at the early part of the curve, like it's suddenly very obvious that this is going to be huge and everyone's very excited, but you don't know how any of it is going to work.
And you have to know that you don't know how any of it is going to work.
And if you look at that and say, well, you're an analyst, it's your job to tell us like you were more like if you were in 1997 and saying, well, is it going to be portals or search or platforms or smartphones?
You didn't know how any of this is going to work.
Everything was wide open.
And so it struck me that.
What you kind of have to do instead is sort of step back and say, well, you know, clearly you can try and model GPU shipments in the next six to 12 months.
But you kind of have to step back and think, well, what are the kind of fundamental moving parts here?
And what are the questions for each of those moving parts so that you might at least understand what you're trying to work out, even if you don't know the answers?
So you're putting it apart and you're looking at the building blocks individually, essentially.
Well, what at least...
You don't know, maybe you don't know the answers, but at least let's try and work out what the questions are.
And so it kind of struck me like kind of running through it that the kind of the primary question is the shape of the cost performance curve, price performance curve, in that you've got absolutely the best latest frontier model.
And how many people are going to need that?
How many use cases will benefit from that?
Because clearly you can run translation on a phone.
and at one extreme.
And at the other extreme, if you want to do some big, complex, multidimensional cost company analysis, then the next model will give you better results, maybe a positive ROI.
And so there's like a curve there of which things will keep needing tomorrow's model and which things will be fine with the model from last year and what's sort of in between, depending on the ROI and the cost.
And presumably the older it is, the more commoditized it is.
And so then, well, how long is the frontier going to keep moving?
And that's like the primary science question where no one knows, but like it has kept moving so far.
And third, is the frontier going to stay competitive?
Because right now we've got Anthropics ahead at the moment.
We have OpenAI and Gemini 5% behind, like not far behind.
Grok just, SpaceX Grok Nazi model just came out with like a model that's in like the top.
at the top of some of the benchmarks on some scores.
Meta's new model is in striking distance too.
The Chinese models are kind of, depending on who you ask, anything from six to nine months, three to six to nine months behind.
So right now it's competitive.
Is it going to stay competitive?
Or maybe it'll stay competitive, but the models will diverge.
So there'll be one that's really good at code and one that's really good at analysis or something.
And so there's all sorts of kind of moving parts.
of are you going to get to, oh, yes, and kind of the final foundational point is presuming for the sake of argument you're doing this big, complex, multidimensional cross-company analysis thing and presume for the sake of argument that there's only two companies that can do that or three companies that can do that, does the model itself do that whole thing or is the model an API call and a set of tooling and infrastructure?
for some other much more sophisticated thing that's built by some other software company.
So even if your use case does benefit from the frontier model and the frontier models continue to be a thing and continue to be expensive, do they capture all the value or are they infrastructure for other people?
Because that's kind of what happened with cloud, for example.
There's only three hyperscalers, but and yet most of the applications are built by other people and most of the value is captured by people who run on AWS.
And kind of the thing here is like, you don't necessarily know the answer, but what this tells you is there's a wide range of possibilities.
There's kind of half a dozen, four variables I've suggested, all of which have wide ranges of outcomes.
And so this point, I kind of, it struck me, and this is a thing I've been talking about for a while, is you can kind of compare this with mobile.
The thing that people like to do is compare it with Fiber, because Fiber had this massive build out in the dot-com bubble, or massive overbuild.
The narrow problem with that, of course, is that that was built out ahead of demand, whereas this is being built out behind demand.
Although at the time, there were these kind of made up statistics about how much demand was growing that weren't true.
But in the end, the more interesting problem with the fiber comparison is that fiber is almost entirely fixed cost, that like 80% of the cost is digging holes in the ground.
So if your traffic doubles, your cost doesn't double.
You know, you swap out the DWM equipment, but you don't have to go and dig new holes.
Do you think with this, we're unsure still what exactly it is that we need to be done?
To grow this?
So, you know, in principle, most of the cost of the data centers is the chips.
Yeah.
And so if your demand doubles, well, you need twice as much data center, basically.
Maybe not literally twice as much data center, but you kind of need to spend the same money again.
Whereas with fiber, if your demand doubled, you did not need to spend the same money again because most of the money had gone into a hole in the ground and there was still plenty of space in the hole.
There was still plenty of space in the duct.
So the more interesting comparison or the more relevant comparison is mobile, because mobile does have marginal cost.
Like if your mobile traffic doubles, you kind of have to go out and build not exactly double, but kind of double the infrastructure.
And then there's a more relevant reason why it's interesting, which is that as smartphones took off, the networks collapsed because people bought iPhones and then the iPhones got 3G.
And then people started watching YouTube and everyone had flat rate plans.
And then in like 2010, 11, the telcos had to give up on the flat rate plans because it didn't work anymore, which is kind of exactly what happened in the last couple of months for Claude, that they had to give up on flat rate plans and move you to usage-based pricing of some kind.
Now, the really relevant point of the telco comparison, the mobile comparison, is that here we are 20 years after the iPhone, almost 20 years after the iPhone.
And mobile networks globally are about a trillion dollar industry.
They've changed the world.
Five billion people have a smartphone.
Maybe six billion people have a smartphone.
Like completely transformed the world.
Giant industry, trillion dollars of revenue, 200 billion dollars a year of CapEx.
Stocks have gone nowhere in 20 years and all the value is built by other people.
So everyone has an iPhone.
Everyone pays their.
In India, a dollar a month.
In America, a hundred dollars a month.
Everyone pays their money, but all the stuff you do with it is done by other people.
And all the actual money is made by Meta and Google and Amazon and Tinder and Uber and everybody else.
Well, not even that.
Uber is done by Uber.
It's not done by the mobile operator.
So all the value is captured by the people.
And of course, that's the other foundational question for LLMs is, will the model just kind of do the whole thing?
We're like, we'll chat GPT.
capture all the use cases or will that be hundreds of apps and will it be an API call inside lots and lots of other apps the other interesting comparison it struck me is semis because what happened with semiconductors is with each generation it got harder and more expensive so everyone knows Moore's law but there's also Rock's law and Rock's law says that the cost of a cutting-edge fabrication plant doubles every four years And so that meant that you went from having dozens of companies at the cutting edge to really just TSMC and then Samsung and Intel a bit behind.
And that's kind of the semi-model, create building models looks a little bit like that.
If it keeps scaling, then it will keep getting more expensive.
Although at the moment, it's not expensive enough to crowd people out of the market.
Anyone who's got a couple of billion dollars can make a foundation model, which is what DeepSeek showed us as well.
I mean, you can kind of keep proliferating examples.
I think one should kind of step back from this and say, like, this doesn't have predictive value.
Like, this is kind of the mistake the doomers make is they say, well, this is like nuclear weapons.
It's like, well, it's not nuclear weapons.
It's a model.
And when the iPhone was taking off, people said, well, the iPhone is closed and the Mac was closed and the Mac lost.
So this will lose.
And you can kind of be too deterministic about like drawing analogies.
The analogies don't have predictive value.
What they do tell you is there are lots of possible outcomes.
So you could also talk about cloud here.
Sam Altman talked about Windows, which is a really bad analogy because Windows had network effects.
He also talked about water companies and electricity companies, which are like completely the opposite.
They're regulated utilities with crappy margins, selling a commodity.
So completely different from Windows.
And the point of all of this is like, there's a lot of different ways this can work out.
And you can't just say you don't understand exponentials and talk about AGI a lot and show me the latest amazing thing it's done and say, well, therefore, the models will capture all the value.
You kind of have to like, you have to kind of look at this and say, well, there's a lot of possible ways this could end up, which is kind of my we don't know point.
You talk about in the piece, you talk about this desire of people.
needing or wanting to hunt for patterns to recognize.
Why do you think that?
Is it because that helps us understand?
Is it, do you think, because that helps us make sense of what's going on?
Do you think it just makes us feel a little bit better that we feel like we're in control when we're hunting for these patterns, even though there are maybe no patterns to hunt for?
It's interesting.
I mean, people always try and win arguments.
with a reference point.
I mean, this is Godwin's law.
There was this, you know, forever ago, like the 80s or 90s, there was a guy called Godwin who said that, like, this is probably on Usenet, that all arguments end with a comparison to Hiller.
So you're always trying to reduce the question.
The funny thing is, I then got into a discussion with him on Twitter and discovered that actually he is that guy who's impossible to talk to.
But the...
The point is you're always trying to bound it down, just take this unknown thing and connect it to a known thing, which is why do you say, well, it's like nuclear weapons.
And we know what we think about nuclear weapons.
Yes, but this isn't nuclear weapons.
It's something else.
You know, it's like Sacha Baron Cohen saying, well, talking about Facebook, saying, well, you wouldn't let somebody walk into your restaurant and say these things.
So why would you let people say those things on Facebook?
And the answer was because Facebook isn't a restaurant, dude.
I think the more relevant thing here is like you're at the beginning of this phase and you're trying to work out what's the market structure going to look like.
And you can wave your hands a lot or you can say there are 10 different variables and we don't know where they're going to end up, as kind of I've done.
Or you can say, well, this is what happened with mobile.
Or you can say, well, this is what happened with Windows.
None of those are particularly satisfactory.
All of which is, I hope, not a long way of sort of abdicating responsibility.
The thing that I keep coming back to here is if you go back through those variables.
So each of the variables, something would have to change for there to be market power.
Pick your term, market power, strategic leverage, value capture, dominance.
So does the frontier keep moving or does it slow down?
If the frontier slows down, then this stuff is definitely a commodity.
The frontier keeps moving.
But then look at the others.
So will the market become less competitive?
Right now, there's clearly a lot of competition.
Will the models be able to capture value further up the stack?
Right now, clearly they can't.
Will the models become differentiated?
Right now, clearly they aren't.
If you're actually going to propose that you're going to have...
Will there be network effects?
Will there be Winotakes All effects?
Right now, there aren't.
Will it be harder for a big tech company to jump in and build a model?
Well, we've just seen Elon Musk and Mark Zuckerberg jump in from zero in the last six months.
Each of them had a model in the mid 2025 that failed and then they built a whole new lab and built a whole new model.
Both of them did.
It was expensive.
It wasn't that expensive.
It wasn't trillions of dollars.
It was enough money.
They have enough money to do that.
The path to value capture to dominance is network effects would have to appear.
Barriers to entry would have to appear.
It would have to get much.
Everything would have to, something fundamental would have to change.
Because right now, there isn't a path to value capture because right now you've got a lot of people doing basically the same thing with basically the same technology and the same data and the same chips.
Do you think the network effect creates more competition or less competition?
Well, if there was a network, this is the thing, I wrote something in the spring about OpenAI.
I had this history professor at university who said that power isn't sophistication or complexity or doing clever things or being impressive.
Power is the ability to make people do something that they don't want to do.
And so in the days of the PC in the 1990s, if you were a developer and you wanted to make an app and get customers, you had to make it for Windows.
because that was where all the users were.
It didn't matter if you liked Windows or if you preferred the Amiga, if that was still around, or the Mac.
The users were all on PCs, so tough.
And the same as the user, it didn't matter if you liked the Mac.
Almost all the apps that you would need to use unless you're a graphic designer or consumer were on the PC, so that was what you had to buy.
Same thing for the mobile smartphone era.
It doesn't matter how much you liked Windows Phone in like 2010, maybe I forget when exactly was Windows Phone still around.
I bought a Windows Phone.
It was really cool.
There were no apps for it.
And as a developer, there were no users.
And so it didn't matter how much you liked it as a developer or as a user, this was not a viable product.
And so this is kind of my point is that you didn't have a choice.
You had to make for iOS and Android.
You had to make for Windows.
Today, like you would like people to, you know, you would like to use a different photo showing app.
Like, well, tough.
Everyone's on Instagram.
And so the point of this is, like, is there some mechanic that pulls ahead of just execution and just product quality and just preference and means that, no, that company wins?
Like, this is what happened for Google.
Search has network effects.
Google search has network effects.
So Microsoft can spend, doesn't matter how much money Microsoft spends, Google search is still better.
And so is there that mechanic here?
Like, right now we don't have that mechanic.
Clearly, would that mechanic emerge?
Are there other winner-takes-all effects that we don't know about yet?
I mean, it was not apparent in 1998 that search had network effects, for example.
I mean, it's obvious in hindsight.
I'm not sure that it was apparent at the time.
It was not apparent in 2008 or 2010 that there would be room for two smartphone ecosystems because in PCs there was only room for one.
It turned out that the market was so much bigger that there was room for two.
With rideshare or social, with social, people thought maybe one company will win in each country.
So like Facebook would win in America.
Maybe Bebo will win.
Was it Bebo?
Yes, maybe Bebo will win the UK.
Maybe the same thing with Y-chair.
Maybe it would be city by city.
So Uber would win this city and Lyft would win that city.
The point is maybe there will be other network effects that emerge.
Clearly what we would like to be able to do is...
is to be training the model continuously.
You have continuous learning, you have memory, you have all sorts of other things that are floating around that might happen.
But the point is, we don't have any of that now.
Something would have to change.
It may be that the model companies can build enough product further up the stack.
I was playing with a new chat GPT work product, which is clearly not that.
It's a complete disaster.
It's a chaotic, confusing mess.
But it does kind of point to the sense that people would like to try and build stuff on top of the raw model where you would have product execution on top of the raw model.
And I think you could go back to kind of Microsoft and Meta and say, well, they executed their way into a network effect.
There was a network effect, but they didn't have one.
Microsoft didn't win because of the network effect.
It stayed winning because of the network effect.
First, they had to out-execute everybody else.
They had to get everybody using Office first, and then everyone had to stay on Office.
Meta had to get everybody to switch from MySpace to Facebook.
Yeah, the network effect alone doesn't actually add the value that you need.
The network effect means that once you've won, you stay winning, but it doesn't get you.
After a certain point, it has a self-fulfilling.
effect.
But to begin with, like Meta had to get everybody to switch from MySpace to Facebook.
And Microsoft had to get everybody to switch from WordPerfect and Lotus 1-2-3 and all the other spreadsheets that were floating around to use Office and Excel, to use Word and Excel.
But you need to do that first.
You need to execute that first.
And, you know, Apple had to make the iPhone before the iPhone and make an app store and get people to do it before that had a network effect.
So there is that, you know, maybe we're at that ecosystem stage now where...
That execution stage now where they've got to build that.
But we don't know what that is yet.
There's a little bit of cargo culting in that what's it called?
OpenAI have done at least two app stores so far and possibly three.
I think we've all lost count.
And, you know, we'll do a super app.
We'll do this.
We'll do that.
But we are in this sort of fuzzy intermediate stage of it's very apparent that this is going to be a thing.
It's not clear what the levers are, what the winning model looks like.
It's not clear what the levers are, where the network effects are, or if there are network effects.
And it's also not clear, which is the point I keep coming back to, that the model itself is a product as opposed to the model itself being infrastructure that other people have to build product around.
And we were chatting about this yesterday and there's like another podcast in here, which is why is it that most people are not builders?
Most people do not make tools.
Most people don't think about how they could make their product better.
Most people are not innate.
That's a different skill.
And most people in most companies aren't in a position that they could do that anyway.
They don't want to and they don't need to.
They don't want to, they need to, but also they can't.
You know, they don't have the empowerment.
They don't have the right position in the company.
You know, if you are at Citigroup or JPMorgan Chase, that whole workflow that's touched by a thousand people might well get completely transformed with generative AI, but that's not going to happen because one of those thousand people build something in Claude.
Because it's a regulated product and there's a CIO and there's a process and it has to be done consistently.
And yeah, someone may come along and build a new thing that changes that workflow, but that isn't going to get done bottom up.
This is like the great fallacy of enterprise software that you can just sell to the users.
And it's like it only ever worked for Slack and one other company.
Maybe Notion, but even Notion has capped out.
So you've got this sense of like, is this infra?
Is it product?
Is it product with differentiation, with network effects, with winner-takes-all effects?
Maybe, but stuff would have to change for that to happen because we don't have that yet.
And without that, we have everyone building kind of the same thing with kind of the same technology and kind of the same chips and kind of the same data and kind of the same scientists.
So why would you have pricing power?
And, you know, I'm happy to be, you know, come back in five years.
Show me why I was wrong.
It's entirely possible that I'm wrong.
That's kind of my point.
We don't know.
But it will be wrong because something that isn't there today has happened or something that we're not paying attention to yet.
I do like all of this because even though there isn't an answer to all of these questions, it is a nice way of explaining why adoption hasn't been as far spread and widespread as we thought it might be.
I don't know.
There's something soothing about that.
Well, it is.
I mean.
I mean, that's sort of tangential to my point.
But yes, the LLM itself is not a great product for most people.
Usage is a mile wide and an inch deep.
And you have this kind of polarization between people where this really, really works and they really, really have product market fit, which is basically software developers and a certain quite narrow kind of knowledge worker.
And everyone else is using this a couple of times a week at most.
And many more people are using it a couple of times a month.
The analogy I always use here is, you know, imagine you're an accountant seeing the first software spreadsheets in the late 70s.
This is life changing.
Now imagine you're a lawyer seeing this.
Well, that's very clever and I should share my accountant, but that's not what I do every day.
And that's the challenge we're still in.
Is there something you build on top of that as a model company that gets you victory, that gets you value lock in?
Or is the answer, no, there will be 500 things that get built by 500 other companies on top of you that do that?
Some of this, I think, is a sort of an expansion and an iteration of what I wrote about OpenAI in the spring.
Like, what is the path for OpenAI to pull ahead?
But it's a point that generalizes here to all of these things.
What is the path for these things to become more than commodity infrastructure?
And the paradox is, like, right now they can name their price.
But that isn't...
where we're going to be in five years.
You can argue about how quickly the infrastructure gets built out and how fast the GPUs arrive, blah, blah, blah, blah, fine.
But that's a supply shortage.
That's not strategic leverage.
That's not a lock-in.
Feels like a good place to have.
Yeah, there we are.
That was a good monologue.
No, but it is.
I mean, look, nothing's going to change.
Our podcast episodes for the moment are all going to end with, there you go, there's another 15 questions for you to think about, which is good.
Speak to you next week.
Okay, speak to you next week.