Inside the AI Economy: Competition, Compute, and the Next Phase of Enterprise Adoption

00:21, Chris Thonis

Hi everybody. Welcome back to William Blair Thinking Presents. Today is Wednesday July 1st, 2026. I'm joined by Arjun Bhatia, co-group head of technology at William Blair, along with tech analyst Jonathan Ho.

Arjun and Jonathan, thank you for joining us.

00:35, Arjun Bhatia

Thanks for having us.

00:36, Jonathan Ho

Thanks, Chris.

00:37, Chris T

So, today we're discussing your team’s Inside the AI Economy report which is a monthly look at key developments across the AI landscape. It includes everything from, you know, models, to infrastructure, to enterprise adoption, capital markets.

So, to start, I thought you could give us a high-level overview of what the report is, what mattered this month across AI, and then, why it's relevant right now.

01:01, Arjun B

Yeah, sounds great. Yeah.

I mean, you know, we have this report we put out on a monthly basis. It's just meant mostly to keep our readers, and investors, and clients up to date on what's happening in the world of AI. And, as you can imagine, you know, we're at this, sort of, rapid pace of change in the cycle. And every month there is so much news to keep on top of.

And this is something to just help organize it across all those key topics that you mentioned from infrastructure and build out, all the way to enterprise adoption and monetization at the back end, you know, which is still very early stages, but shaping up in enterprise right now.

This month, you know, I think there were a few things that really stood out. You know, first was on the model development front. We saw the key model providers, Anthropic and even some of the Chinese models, sort of, launch their latest versions, which, you know, had meaningful improvements over their prior versions.

So Anthropic, for example, finally launched Fable 5, which is, if you remember, last month there was a model Anthropic launch called Mythos, which they basically deemed too powerful to launch to the general public.

Now, Fable is a little, scaled down version of that, that they were hoping could be, sort of, deployed commercially. But, pretty much days after they launched it, the US government put an export control directive on it and basically restricted it for use. And today, as we're recording this, that restriction has been lifted.

So, there's a lot of, sort of, regulatory aspects of the model releases that are, you know, changing on a month-to-month basis. But elsewhere, you know, I think there were other, you know, important developments.

In China, as well, you know, they're, sort of, pushing to the frontier with competitive models. You know, MiniMax launched an open weight model that was competitive with Claude, OpenAI, and Gemini.

And we're seeing, you know, China really invest heavily behind model development themselves, which is raising, sort of, a competitive question for what it means for the US model providers and what it means for pricing at the model layer.

And, you know, even with US companies, we're starting to see them get a little bit more open to adopting Chinese models for inferencing workloads, which will be an interesting development and competitive angle for the space as a whole.

You know, I think second, you saw even Anthropic and OpenAI, sort of, ramp up their competitive intensity. There were reports that, you know, OpenAI is considering, sort of, cutting price on their models to win back some of the users that Anthropic had won from them and to become more competitive themselves. But, it does, again, bring up this question of, you know, is the model layer getting commoditized?

You know, I think it can still grow, but, you know, you might just see pricing pressure at the model layer, given how competitive it is between the frontier model providers in the US with China. You know, the Chinese models coming in at, you know, somewhere between a seventh to a tenth of the cost of the proprietary U.S. models. So, I think that's going to be an important, sort of, development to watch in terms of how pricing plays out.

And then, lastly, you know, we saw the build out of data centers and compute, sort of, continue to progress in the space to support not just the models, but also set up, sort of, interesting infrastructure that's, you know, going to be very compute intensive that uses these models for enterprise and consumer workloads, as well.

04:46, Chris T

You answered my first question which is going to be, you know, what changed this month? And, you know, what’s happening across everything?

But I have a second question regarding model competition, you know, economics regulation. So, at the model layer specifically, which you were just discussing, it feels like competition is accelerating, while economics and regulation are becoming bigger factors at the same time.

How should we think about that dynamic, both in terms of pricing pressure and the role governments are starting to play?

05:11, Arjun B

Yeah, pricing is going to be, sort of, evolving over time at the model layer. But I think one thing is going to be, sort of, clear is that we should see the price of tokens, or essentially the price of the models, come down over time.

And part of it is, right, like we're seeing so much infrastructure get built to support compute, and as more capacity comes online, you know, that should naturally have somewhat of a deflationary impact on pricing.

But the other part is, this is an intensely competitive market, right? It's a large market. It's going to be a huge TAM.

But competition is, sort of, limited between… in the US, at least, you know, there's only a few players that are competing at the model layer. And, if we exclude the Chinese open-source models, open-late models for a second, you know, it's Anthropic, it's OpenAI, it's Google.

And then you have Meta and SpaceX in there a little bit, but it's generally, you know, just a handful of players that are competing in the space. So, how are they going to compete against each other? And, you know, if you have one fall behind, what kind of levers do they have to win back share?

And that's what you're seeing a little bit right now is, you've had Anthropic, sort of, outgrowing OpenAI. And, as a competitive response, you know, it's not official yet, but the reports are that that OpenAI might cut price.

And so, again, that's kind of what I was talking about earlier, that you'll see these, you know, the price of these models start to come down over time.

I think, you know, that's great for the end consumer, and it's great for the applications that are built on these models because it should become, you know, less expensive to operate. So, you know, you should see maybe more businesses get created that are embedding AI into their product suite. It becomes easier for incumbents to embed AI into their offerings, and it becomes, you know, even more affordable for consumers who want to use these models more.

But for the models themselves, it means, you know, they have to grow through volume and not necessarily price. And, I think we're still so early in, sort of, AI deployments that that is very possible that, you know, volume growth should continue to pick up, while, you know, pricing maybe is going to come down at the model layer.

07:45, Jonathan H

On the government regulatory front, the last part of your question, I think, you know, what we found to be interesting is that, you know, the US government has stepped in when it comes to Mythos and Fable, with the initial launch.

And a lot of this is, you know, sort of, concern over, you know, whether these models can be jailbroken, you know, whether they can be used for unintended purposes. You know, it just highlights, you know, some of the risks around the newest generation of models. And, you know, basically the gain in function, the gain in capability that they have.

From our perspective, I mean, the danger is that, you know, the government, when they start regulating what these models can do, you know, it places the companies that have, you know, released these models at a competitive disadvantage. And the US government cannot, by itself, regulate all the models out there, particularly those that are not based in the US.

And so, you know, I think it's very challenging. It requires, you know, kind of, a different solution where, you know, basically, global governments have to get together to decide, you know, what's the right pathway going forward. This can't be, sort of, a unilateral, you know, government approach.

08:53, Chris T

You know, we're starting to see more transparency into the economics of these businesses and the potential for public market participation. How should investors think about the sustainability of the model and how the opportunity set may shift as more of these companies become investable?

09:08, Arjun B

I think the biggest thing is that AI is such a game changing technology, and it's going to have so many downstream impacts across different industries and on the economy, as a whole, that what you're really investing in is this large TAM that we are still very early in penetrating, even for the model companies themselves. Because, right now, you know, a lot of the use cases are consumer oriented, or, you know, concentrated, sort of, enterprise use cases like coding, for example.

But we haven't really seen these models get deployed, sort of, top down in a structured way in enterprise. And, you know, that's where you start seeing business models change. That's when you start seeing, you know, productivity increases that we're still in the very early stages of.

So, what it means when you're investing in the model companies themselves, it means, you know, there's a long runway for growth. And, as I mentioned earlier, it's, sort of, capital intensive to create a model or to become a frontier model company, which means, you know, naturally, your competitive set is going to be limited. There will be intense competition amongst those players that are in the space, but you're probably not going to see that many new entrants come in.

So, you have a large TAM. And these companies are, sort of, you know, infrastructure for…they’re platform companies, right? The model companies are platform companies. So, that means others can build on top of them. And that's how your use cases and your TAM keep advancing, is the beauty of a platform model. Your TAM is only limited by, sort of, what gets built on top of you.

So, you get a big TAM, you’re not going to have that many new entrants come in. And in the early stages, you, sort of, have to, which is where we are at now, you have to, sort of, weather this, you know, cash burn and maybe capital inefficiency that we're seeing from these companies.

But what you're, I think as an investor, what you have to bet on is that long term, you know, these companies will become more profitable. They will be of much larger scale, you know, which should eventually drive, you know, outsized cash flows to those that have invested in these model companies themselves.

And it will be interesting, I'll say, you know, last piece, it will be interesting to see how the model companies themselves advance, both up the stack and down the stack, right? Whether they move into, you know, apps, agentic use cases, or they move down the stack and vertically integrate some of the, you know, infrastructure that they rely on from external providers, whether it's chips or data centers themselves.

11:59, Chris T

You touched on this a little bit, but the scale of infrastructure build out right now is pretty unprecedented. You know, what's driving that level of investment? And where do you see the biggest constraints emerging? You know, whether that's compute, power, supply chain, whatever you think.

12:14, Arjun B

There’s a few things. So, mainly it is the model companies, right? And we started off this conversation talking about Mythos and Fable, and Jonathan touched on how the models are getting so powerful, but it is that the, you know, each successive, sort of, version of the models, they are improving or getting better. So, that means, you know, the model providers themselves have to, you know, have to have access to more compute capacity to train their latest models.

And then, the second piece of this is, once those models get deployed into the real world, that's when you have all these inferencing workloads kick in, which also require compute capacity. So that's what’s, you know, at the end of the day, that's what's driving the data center build out.

And that's why you're seeing all this CapEx investment from hyperscalers, who are selling capacity to, you know, Anthropic and OpenAI and others. I would expect that to continue. Like, the big question now and the big debate is ROI, all this spending and, you know, that's where we, sort of, have to, you know, we're in a little bit of a wait and see moment because we're in this build-out phase.

And the deployment phase, as we talked about earlier, is still, you know, it's still generally developing, meaning, you know, enterprises are still finding ways to do, they're still figuring out ways to do top-down AI deployments.

But that's where the ROI, ultimately, you know, has to come from. It's, you know, productivity or growth. And we see, you know, because you have the hyperscalers, there's, you know, again, relatively few compute providers, but they're all competing for the same scarce resources.

It's creating bottlenecks throughout this AI value chain, from chips, to memory, to, you know, the two that we hear most about is energy and, actually, skilled labor. So, think like plumbers and construction folks, right, to actually build the data centers themselves. So, you're seeing bottlenecks pop up, sort of, all over the data center supply chain.

14:07, Jonathan H

One thing I would point out is that, you know, when we look at, sort of, the size of models and how much they've changed, you know, the prior generation of models were trained on anywhere between 1 and 1.5 trillion parameters. And the expectation is that the Blackwell generation models are going to be over 10 trillion parameters. So, there's a significant step up.

It's not linear, but it's a significant step up in resources that are required from a memory standpoint to be able to service those models. I think the other thing that we've seen is that, you know, it's not necessarily easy to stand up, you know, these data centers.

Even if you intend to do it, you know, there's protests, there's, you know, definitely a lot of pushback in terms of just setting up the physical shell and the footprint of these data centers.

You know, I was reading an article where people were complaining about noise generated from data centers. Certainly, any type of energy cost increase, you know, where data centers have been set up, you know, is causing, you know, concern.

And so, I think the challenge is that you need a lot more horsepower to run these new models. But, at the same time, you know, the willingness to put, you know, the supporting infrastructure into place, you know, remains challenged.

15:18, Chris T

So, on the enterprise side, adoption is clearly moving forward but not evenly across industries, and we’ve seen some volatility in software markets alongside that. Where are we actually seeing real traction today, and what still needs to happen for broader enterprise adoption to take hold?

15:36, Arjun B

We're still very early in the enterprise adoption cycle. And that is part of, you know, what's driving the volatility in software markets is, you know, investors in the markets trying to figure out who are the AI winners, who are the app companies. You know, it's interesting. You can, kind of, listen to what, sort of, Jensen, CEO of Nvidia, says all the time, which is the ultimate economic value of AI is going to get created at the app layer, right? It's not going to be chips and energy. Those are all critical infrastructure, but, sort of, a means to an end, right? to ultimately build AI applications that can be deployed in the enterprise and that can, you know, drive productivity or drive growth.

But enterprises just move slowly, right? Enterprises are risk averse. They need to make sure they have, sort of, all their ducks in a row before they can adopt a new technology. They're not going to dive in headfirst, in most cases.

And, you know, I always try to remind folks, you know, we're in year four of this AI cycle. That's not that long, right? That's actually, you know, that's quite early on. You know, Cloud, when it started, it took ten years to become mainstream.

This should happen faster than Cloud, but still, you know, we're, you know, only four years into this cycle that's going to be many decades long, I think. But there are pockets, I will say, in enterprise where we do see more adoption than others. And coding is one, software development, coding is one where we've seen probably the most adoption.

And I think if you, you know, talk to pretty much any engineer, they have, at least, experimented if they're not, you know, in production with some sort of AI coding tool, whether that's Codex, or Claude Code, or Cursor, you know, one of the more popular tools, out there, so that's been one of the earliest use cases.

But then you see more experimentation in areas like customer service, you see sales. But, a lot of enterprises are still just trying to figure out, you know, where should we deploy AI first? What are the steps we need to take?  What's the change management that's required? Is our data in order?  Do we have the data pipelines built? Is our data cleansed? Because, you know, the output of AI is really only as good as, you know, the input, and the data, and the context that you give it.

So that's something that's mission critical for enterprises. And I don't think they want to, you know, deploy AI and have it, you know, there's a lot of downsides that can happen from, you know, pulling the trigger too early. And that's where the risk aversion comes in from enterprise.

18:29, Jonathan H

I would also point out that, you know, in addition to what Arjun is saying, the market itself is changing very, very quickly, as well. So, the notion of AI assistance, the ability to, you know, use true agentic AI to augment tasks, this is all changing, you know, on a month-by-month basis, in terms of enhancements and capabilities.

And, you know, the trust level wasn't previously there with prior generations of the models. As they've shown step function improvements, you know, I think that's shortened the time frame for adoption and caused people to really think about, you know, not just adopting at some time in the future, but, you know, the need or the greater impetus to do it within a shorter period of time.

19:12, Chris T

So, Jonathan, stay with me here on this question. One of the themes in the report is that AI is expanding the attack surface and reshaping how security is approached. At a high level, how is AI changing the cybersecurity landscape, and how should we think about the balance between risk and opportunity?

19:30, Jonathan H

Yeah, I think the notion of AI impacting the cyber environment, you know, comes from both the attacker perspective and the defender perspective.

So, first off, you know, the attackers are often times the ones that are ahead in terms of innovation. They're the ones that are figuring out how to utilize, you know, these gains in function, the improvements in the models, you know, for their purposes, which is oftentimes, you know, the ability to break into systems, the ability to send more effective, you know, ransomware attacks, the ability to deepfake individuals.

And so, what we’ve broadly seen, is that, you know, the adoption of the new models, you know, will likely, you know, create a situation where, you know, the attackers will have a temporary advantage, you know, where they can leverage these tools. The defenders then need to respond to that. And so, the asymmetric nature of cybersecurity is that the attackers, you know, we need to be successful in a small number of attacks.

And the defenders need to stop all of the attacks. And so, there's really no cost today for the attackers to launch their attacks. In the future, they'll be able to leverage AI to automate those attacks. So, we expect an increase in both frequency and sophistication. The defenders, you know, they can't afford to sit still. And so, they also will need to make investments in order to better defend networks and also to be able to respond in real time. So, the notion of AI versus AI, you know, from a cybersecurity perspective, is something that's very real.

I think it's also important to note, you know, sort of, the tension that a lot of the CIOs feel today. You know, they have pressure to deploy AI as quickly as possible internally. But users can, you know, inadvertently misuse AI and lead to the leakage of important data to frontier models or to the to the general public.

And so how do you balance, you know, the need to, you know, fully realize the productivity of AI with also the need to secure, you know, the use of AI within your organizations, to secure the AI models themselves. And so, I think that's a challenging situation where, you know, they're looking for answers, they're trying to, you know, understand, you know, how, you know, this is all going to unfold.

And, you know, the compounding effect is that with the Blackwell trained generations of models, it's not just Fable and Mythos, but any, you know, that have this, you know, these larger and larger data sets powering them. You know, they pose greater and greater threats over time. And so, you know, the perception is that that window is shortening.

22:14, Chris T

Monetization still feels like an open question. How do you expect pricing and business models in AI-driven security to evolve as adoption scales from here?

22:23, Jonathan H

Yeah, it's a great question. I think, you know, it is a moving target. One of the things that, you know, a lot of organizations are struggling with is, you know, how do you price, you know, traditional user or C based license models, user accounts, now that AI will gain access to them?

So, we’re entering this world where you will have agentic AI, in some cases it'll be 80 up to 200 different agentic AIs that work on your behalf. You know, is every separate agentic AI then a separate license? for a traditional business model or, you know, do you charge by asset, do you charge by outcome?

You know, we believe that, you know, the revenue models will change. But it's not yet, you know, determined exactly what, you know, the right model is for the right type of asset class and how much, you know, value you can extract, you know, as you add the value here.

Some of the challenges that, you know, some of the, you know, AI oriented solutions, you know, a lot of the value is in the large language model itself, and so, is the value captured by, you know, Claude or one of the frontier models? Or is it going to be captured by a platform player that has a large data set that is building the harness, providing the toolkits and toolchains that allow, you know, their AI models to be able to do what they need to do, to better secure these enterprise environments?

I think the final thing I'd point out is that there's additional cost that comes with running these models, in terms of tokens. And so, you know, providing, you know, “all you can use” models, is probably not sustainable.

Many organizations are going to have to charge additionally for additional, you know, token usage, particularly as these, you know, larger models become more and more expensive.

24:13, Chris T

All right. So, as we wrap up, if you look out over the next 12 months, what's the one area you'll be watching most closely as a signal for where this AI cycle goes next?

24:23, Jonathan H

Yeah. I think, you know, from my perspective, you know, when we look at, you know, the clear signals of where the AI cycle goes next, I think it starts with, you know, looking at, you know, kind of, the adoption of the frontier models, the capabilities that are driven there. And, you know, again, the hackers are going to be the first ones to take advantage of this.

And, you know, how those attacking tools are deployed will inform how the defenders, you know, then respond to it. I think everyone is running around, you know, scrambling, trying to find, you know, what the right solution set is going to be. And I believe it's going to be multiple solutions that are required within the AI world that really supports the view of platform vendors that have much broader, you know, perspectives with their data. They have contextual information. But it requires, in some cases, the repurposing of existing solutions, and in other cases, you know, brand new solutions, as well.

And so, I would look at, you know, the improvements that come from these models. You know, if they're realized, and we really do see that's step up in functionality, it's going to really cause a lot of challenges within the cybersecurity environment, particularly given how, you know, Mythos and some of these other models cannot just identify new vulnerabilities, but they can chain together previously benign vulnerabilities, you know, into something much more dangerous.

You know, it causes people to have to fundamentally rethink about the time they have to respond, as well as, you know, the complexity of what that response is going to look like.

25:54, Arjun B

Yeah, I think, maybe along the same lines, maybe there's two things that I think are super important here.

One is to watch the frontier model providers very closely. And I do think, sort of in line with what Jonathan was saying, that there's always going to be a decision point or maybe it's less implicit than that, or less explicit than that, but is it that the frontier model providers are going to say, hey, we'll be infrastructure and we want, you know, ten companies to build on top of us for X use case, or do we want to go and compete in this use case with our own product and our own agentic opportunity itself?

So, I think that's going to be one thing that's really important to watch, is how the model providers evolve their business models. You know, whether they want to stay infrastructure or move into the, essentially the application layer. And there's been, sort of, you know, there's been debate about where they go, and we've seen, you know, some of these companies shut down certain products. We've seen them start in other areas, as they look to move up the stack.

But I think that's, you know, these companies are still young. Just remember that they're not mature businesses by any, any means. So, they're still figuring out where they go, and how they best, sort of, fit into this ecosystem and maximize their economic value.

And then the second thing, I think, is, you know, obviously this is the topic of the day right now, but where are we in terms of data center capacity and compute capacity? And I think for that, you really want to watch, you know, what the hyperscalers are doing and what they're saying in terms of CapEx, because, you know, that has downstream implications through the whole data center supply chain, which is obviously a very important and maybe crowded trade right now, right? There's a lot of, sort of, dollars that are flowing into semis, and hardware, and anything that's required, to you know, expand compute capacity.

28:13, Chris T

That's all the time we have for today. Thank you very much for joining us. This has been great. The report is called Inside the AI Economy. For those interested in the full report, you can request a copy by visiting WilliamBlair.com/Contact-Us. Thank you for listening.

28:28, Jonathan H

Thanks, Chris.