The Rise of Neoclouds: Powering the Next Phase of AI

00:21, Chris T

Hi, everybody, and welcome back to William Blair Thinking Presents. Today is Thursday, October 15, 2026.

And we are taking a closer look at one of the fastest-growing areas of AI infrastructure: the rise of the neocloud. I'm joined by Jason Ader, William Blair's co-group head of technology, media, and communications. Jason and his team recently published an extensive report titled “Rise of the Neoclouds: From GPU Specialists to Long-Term Hyperscaler Challengers,” which takes an in-depth look at specialized AI cloud providers and whether companies that emerged in response to GPU and compute shortages can develop into durable competitors to the traditional hyperscalers.

01:02, Chris T

William Blair estimates that the global neocloud market could grow from roughly $25 billion in 2026 to about $250 billion by 2031 in its base case. So today, we'll discuss what a neocloud actually is, what's driving the market's growth, how the economics work, and the opportunities and risks investors should be watching. Jason, thanks for joining us.

01:26, Jason A

Absolutely. Good to be here.

01:28, Chris T

So let's start with the basics. Neocloud is still a relatively new term. What exactly is a neocloud, and why have these companies emerged so quickly?

01:36, Jason A

Yeah, I would think about it like AWS, which is the largest hyperscaler in the world. They have a lot of infrastructure in data centers around the world.

They serve a variety of customers, large and small. The hyperscalers can offer a broad range of services and technologies to support pretty much every application, workload, and customer under the sun. When you think about neoclouds, they were really designed to serve AI workloads. So AI applications, whether training models or running models, which is called inference.

So they are really specialized around AI workloads, in particular large scale AI workloads, not necessarily large scale, but certainly the bigger neoclouds, like a CoreWeave or a would be designing their infrastructure around serving the needs of very large scale workloads where you have very dense clusters of GPUs, you know, made by Nvidia or others that are optimized for those workloads, for in particular for training and inference use cases.

They really came onto the scene a few years ago. Interesting, a lot of them were Bitcoin miners initially, where they had already experience with large GPU clusters, which is how Bitcoin is mined, and they kind of pivoted towards AI as they saw, you know, the market size is, you know, is massive. And just some of the some of the capabilities that they already had.

You know, just you know, there was more revenue opportunity, more growth opportunity in the AI space. So they initially stepped into what you would call like a capacity gap because the hyperscalers themselves, which would be the natural destinations for a lot of these AI workloads, they just didn't have physically the capacity, the data center capacity and the GPUs to serve the demand to meet the demand.

So these neoclouds stepped into that vacuum and have offered, you know, really fast deployment times and enabled, you know, some of the largest, you know, model companies the Anthropics, the OpenAIs, the Googles, the Microsoft's, the Metas to leverage this infrastructure that they've built. Ironically, you know, hyperscalers are big competitors here, but they're also big customers of the neoclouds.

Think about the hyperscalers have a whole, you know, chunk of business from some set of customers, and they just can't meet the demand themselves. So they basically offload it to the cloud. So they call that hyperscaler offload. And we can talk more about that. But that's definitely kind of a kind of a weird thing here where the hyperscalers are major competitors but also major customers of neoclouds.

04:29, Chris T

The growth you're projecting is significant. What gets the market from roughly 25 billion today, as I mentioned, around 250 billion by 2031. And then what has to go right for that opportunity to materialize?

04:44, Jason A

Yeah, I mean, like, it's not just conceptual. Like if you look at the combined backlog for, for the top six neoclouds, you know, it's already like $250 billion. So it's it's huge. So you know there is a you know there is some like visibility there, you know around this opportunity continuing to grow. And I think it is tied obviously to very, very strong demand for those AI use cases.

Training I think is a smaller part of the mix today than it was a couple of years ago when that was like really most of the capacity was being dedicated to training models, but still an important element of the customer set that are that are training models. And again, Anthropic and OpenAI are two of the largest customers here for the neoclouds.

But then you've got inference, which is running the models and running the applications. Right. That's where inference kicks in. And inference has just exploded over the last 12 months or so. You know, you've got, you know, the models that are doing more reasoning now. You know, you probably use a ChatGPT or a Claude. You can see that it can browse the web now.

I mean, it's just it’s consuming a ton of, of tokens because it's thinking, so the models have become much more attuned to, to thinking. And therefore that drives a lot of, of inference so that the, the models just consume a lot of resources. And that's good for infrastructure providers like like CoreWeave.

And then then you move from reasoning into like workloads where you have these agents that are not just answering questions but actually executing tasks for you. So like Meta, Muse is the latest kind of personal agent that you probably heard about. And then the enterprise AI, we're just in the very, very beginning of of the adoption curve for enterprise AI.

You have, you know, Microsoft with its Copilot products. But certainly Claude is Anthropic. Claude has had a huge success in the last couple of years and starting to penetrate enterprise, whether that's for, you know, coding or other use cases. So I think we're still in the early, early, early innings of enterprise AI. And when you put all that together and, you know, I would add like government as well, the government's got huge application here for AI, you know, not just military but all branches of government.

And that's not just the US government, but think about like governments around the world and the need for, you know, sovereign data centers in some of these countries where, you know, you don't want to have it outside your borders, there's just a huge amount of data center capacity that's going to be needed to fulfill all this demand. And that's how we get to the, you know, the sort of base case of of $250 billion. But I think it could be, you know, quite a bit higher than that, depending on, you know, how pricing plays out and power availability. Power right now is really one of the biggest constraints on sort of, you know, governing the pace at which, you know, the revenues will grow.

07:52, Chris T

So you've touched on this already. We just talked about it a little bit, but let's dig in a little bit more to, you know, AI moving from primarily training models towards much more inference and agent activity. How does that change the infrastructure requirements and the opportunity for these companies?

08:08, Jason A

Yeah, training is primarily going to be leveraging like GPUs. So what the neoclouds are building is not just sort of the GPU clusters, but they're also building now a bunch of higher layer services and software. And I think they all, again, want to be more like a hyperscaler that has a kind of a full stack and a full menu of services for customers.

So training is somewhat limited in terms of like the part, you know, the kind of the, the, the infrastructure resources that it that it's going to consume. When you move into inference, first of all, it becomes much more recurring because you're not doing like a model training run, which is like, you know, eight months or something, and then, you know, you're waiting another six months and then you're going to do another training run.

But inference is just going to be ongoing. So there is a recurring element to it. And as those applications and models, you know, move into production, and I think it creates a lot more kind of predictability, you know, for the for the infrastructure, you know, for, for the infrastructure provider, the neocloud or the hyperscaler or whoever. And then what I talked about earlier with sort of agents and reasoning models and thinking, you know, these models just are going to be consuming a lot more overall, like a lot more resources than, than you would have just with training.

And there's a need for other kind of software elements, you know, you think more storage, more memory, more CPUs, and, you know, more like orchestration software. So there's a there's an infrastructure opportunity that really extends up the stack. And those become much more important, you know, revenue drivers for the neocloud beyond just like, you know, what is what is the GPU price per hour, you know, which is kind of what training was, was, you know, was was kind of limited to.

10:06, Chris T

Another question I have gets to the heart of of the investment debate, which is, you know, if if many of these companies ultimately have access to similar Nvidia hardware, where does the real competitive differentiation come from?

10:20, Jason A

Yeah, I think it's easy to kind of to think about it as, okay, everybody's got everybody can buy Nvidia GPUs. I think that's fair. I think standing up, you know, thousands of GPU clusters, you know, in a cluster and connecting different clusters. I mean, that's where it gets really complex.

And then optimizing, you know, how the GPUs operate, where you're getting better efficiency out of the actual, you know, GPU hardware.

So there's some proprietary software that the best companies in the space have, have built to really I would say optimize, you know, that that GPU layer, as I talked about earlier, then it becomes, okay, where do you go from there? How do you move up the stack? How do you how do you add more software and services? I think that's going to be a key element of differentiation.

And then you start getting into, okay, putting those things together, you know, which customers can I sign? And the more you know, large customers, important customers that you sign, that obviously creates a flywheel effect where customers see, you know, where the most demanding customers in the world are going. Right. And that that's a that's a great marketing tool.

And then just from a like a physical deployment standpoint, how fast can you secure power? How how fast can you, you know, deploy a rack of like the latest Nvidia chips? How do you can you test that? Can you make sure it's reliable and you know it's not going to fail? How do you finance all this stuff?

You know, you know, in some ways these neoclouds are like as much financing companies as they are technology companies because is so important. They’re such capital intensive businesses. You need to be able to secure the capital at a, at a reasonable, at a reasonable cost. They're really good at that. Those these companies are really good at that.

There's some question marks and we can talk about some of the risks. But but definitely like interest rates are rising right now and inflation is rising. And so there are some, you know, some flies in the ointment a little bit that, that, that, you know, the markets are kind of grappling with. And this week in particular is a great timing for us to be doing this podcast because the neoclouds got whacked yesterday.

I mean, the two biggest neoclouds, CoreWeave, were both down 15% yesterday. So the bond markets are definitely spooked because of cost of capital. That's a I mean that's an issue. And you know the the companies are going to need access to capital. So so how are they going to get it. So yeah. But you know like we've seen a little bit of a reversal today in, in the in the neocloud.

So it does seem like, you know calmer heads or, or cooler heads are prevailing at least for the time being. But it's going to be a very volatile space because it is so capital intensive.

13:04, Chris T

You talked about this fascinating relationship developing, you know, with the traditional hyperscalers, because they are simultaneously customers and potential competitors. How should investors think about that dynamic?

13:16, Jason A

Yeah, I think the fear is that, you know, the hyperscalers are kind of giving sloppy seconds, you know, to the neoclouds. And at some point they'll have enough capacity themselves at the hyperscalers. They'll just kind of pull it back, you know, pull the customer back and sort of insource, you know, right now they're outsourcing and they can insource, you know, once, once they're like five year agreements with the neoclouds expire, right.

You know, I think that's a risk and certainly something that that I get asked about. But if you just look at the the demand side and you look at like the physical power gap, you know, like the ability to have enough power to serve the token demand that's going to be out there and like 3 to 5 years. You know, my view is that it's probably pretty low probability that the hyperscalers are going to have enough capacity to serve all of the needs of their customers. So, you know, I think it's really good probability that we'll continue to see the hyperscalers and the neoclouds kind of partner the way they've been doing. You know, the flip side is if the hyperscalers, you know, indicate a year before the contract expires, I like hey, guys, we don't need you anymore.

At least I think it'll give the neoclouds enough time to go out and find other customers to fulfill the demand. So. So I think it is a somewhat of an unknown right now. And I think that, you know, that that's part of the, you know, the risk factors that we talk about right now. It's definitely a marriage of convenience.

I mean, both companies kind of need, you know, both sides need each other because there's just such massive demand out there.

14:56, Chris T

That brings us to the economics. You devote a lot of attention to utilization in the report. Why is utilization such an important number for investors to understand?

15:07, Jason A

Yeah. Well, it just has to do with GPU depreciation cycles, which, you know, the neoclouds usually some somewhere between like 5 to 6 years for the, depreciation cycles there. And you know, as you get later on in the age of the GPUs, you know, can you still utilize them, you know, or do they become kind of worthless. Right? Then it becomes a problem for the business models.

One of the things what we've seen is these companies, these neoclouds, have done a really good job of redeploying older GPUs. So, for instance, you may have had a model trained on the latest and greatest GPU, H 100, you know, three years ago. And you would appreciate that on a, you know, on a, on a like a, say, a six year cycle or a six year curve, you know, after a year or two training the models, you know, can you redeploy them for other use cases. And I think what they're showing is not only are they doing a good job of redeploying those GPUs, but actually the prices are going up in some cases on the older GPUs because there's just so little supply out there.

I think about it as like fleet utilization or fleet management. It's like a fleet utilization challenge. And the companies are all I think, showing, you know, really strong success in, in utilizing the older GPUs.

And obviously the newer GPUs are going to be used for the like frontier models and the, the really the high performance use cases where you need the, you know, the best, the best chips. But I think, you know, utilization on the GPUs is very important. Then you think about utilization of the data centers themselves, right. And that speaks to how think about it like a hotel.

Like what's your vacancy rate? Right? And so obviously you want to have very, very low vacancy rate. So you want to have very, very high utilization of your infrastructure to maximize your revenue and your margins.

17:03, Chris T

So we spent a lot of time on the size of the opportunity. But you know, these are also highly capital intensive businesses. What are the biggest risks investors need to keep in mind?

17:13, Jason A

Yeah, there's definitely a bunch. I would say the demand side is probably the single biggest risk that if demand slows down for AI, then you start to get potential overbuilt. Like basically you have too much supply of data centers and GPUs.

17:28, Chris T

Is that a big worry right now, would you say?

Right now? Would you.

17:29, Jason A

No. Because the demand is the demand signals are incredibly strong. So I don't think that's a big concern like today. You know, maybe in a few years it'll be a concern. Certainly the supply is, like I said, somewhat governed by like power and data center, you know, approvals. You know, there's a lot of community opposition in data centers right now in various states.

So I probably worry a little less about that. And that's tied to like pricing and things like that. Right. If you're if you have an overbuilt situation, right, pricing is going to plummet. Right. And that's not good.

 I think the the access to capital and cost of capital is a major risk, as we're seeing, you know, with the debt markets this week kind of going, you know, going haywire, I think deployment timelines is a major risk just from like in terms of the capacity. Right. And so that's like the flip side of supply being constrained is that, you know, these companies have put out certain forecasts. Are they going to be able to, you know, execute and deploy the capacity and the timelines that they told investors that they can and some of that may be out of their control. So I worry about that.

Technical obsolescence. We talked about the, you know, the GPU depreciation cycles. Like right now things are all looking good, but that might change. Hyperscalers, you know, like that. That's definitely a concern. Some of the biggest neoclouds have good good or high customer concentration with the hyperscalers. So I guess customer concentration would be the other with other risk. And they're all, you know, they're they're all working hard I think to try to diversify.

And one interesting stat is that CoreWeave has talked about 10% of their backlog is actually financial services customers. So I think we are starting to see the diffusion that people have talked about with AI. You know, beyond just the model companies, beyond just the hyperscalers, you know, even beyond just like the AI natives, like the, you know, the startup, you know, AI startup ecosystem, it's also starting to diffuse into the enterprise, financial services, healthcare, manufacturing, obviously, like defense is a big vertical for AI.

So so it feels pretty good, I would say in terms of like the opportunity to diversify and you got to just, you know, strike the right balance because these large customers are also like awesome customers to have.

19:56, Chris T

As we wrap up, if investors remember just one thing from this report and also where this market is heading, what should it be?

20:04, Jason A

Yeah, I would say it definitely feels a little like the 1849 Gold rush and not everybody is going to survive. So I you know, we've talked about it in the report, but I do feel like there will be a handful of winners at scale. And I think there is a significant opportunity for them to to turn into, to evolve into long term challengers to the hyperscaler oligopoly, which is, you know, a massive market opportunity.

And, you know, these are going to be, you know, very, very large companies, if that if that happens. But there's also going to be a lot of carnage. I think there's a lot of like companies that are, you know, just have like shaky balance sheets that don't have, you know, the customer, the customer base that, that, you know, is going to that they're going to be kind of durable businesses.

So I think there's going to be a lot of consolidation over time in the neocloud market. And, you know, we're trying to focus on the names that we think could be the consolidated and and the companies that come out the other end. But, you know, the demand is there. And that's really ultimately why, you know, I'm positive on the space overall and that these guys are experts in a part of the market where, you know, the demand is exploding.

21:23, Chris T

Well, that's all the time we have for today. Jason, appreciate your time and insights. For those interested in learning more, Jason's report is titled “Rise of the Neocloud: From GPU Specialists to Long Term Hyperscaler Challengers.” Thanks for listening to William Blair Thinking Presents and we'll see you next time.

21:38, Jason A

Thanks, everyone.