NVIDIA GTC Washington, D.C. 2025 | 296 | Breaking Analysis | AI Factories - Data Centers of the Future

Clips
More from NVIDIA GTC Washington, D.C. 2025

Dave Vellante

Co-Founder & Co-CEO

SiliconANGLE Media, Inc.

David Floyer

Wikibon CTO & Co-founder

Wikibon

Paul Nashawaty

Practice Lead and Principal Analyst

theCUBE Research

Scott Hebner

Principal Analyst, AI

theCUBE Research

Bob Laliberte

Principal Analyst

theCUBE Research

play_circle_outline Transforming Data Centers into AI Factories: Embracing GPU-Centric Computing and Network Fabrics for Enhanced Data Processing

play_circle_outline Definition of an AI factory as a system for producing versatile AI outputs.

play_circle_outline Emphasis on the abstraction of complexities in AI through user-friendly APIs.

play_circle_outline Shift in data centers from cost centers to value-producing entities.

play_circle_outline The significance of scale-up, scale-out, and scale-across networking in AI.

play_circle_outline Unlocking AI Factories: The Crucial Role of API Adoption in Shaping Future Business Models and Economic Competition

play_circle_outline Adoption of open standards and protocols as critical for AI factory success.

play_circle_outline Developer implications and the need for modernized application development practices.

play_circle_outline Growth projections for accelerated computing and application modernization markets.

Info

296 | Breaking Analysis | AI Factories - Data Centers of the Future

Dave Vellante

Co-Founder & Co-CEO SiliconANGLE Media, Inc.

HOST

David Floyer

Wikibon CTO & Co-founder Wikibon

Paul Nashawaty

Practice Lead and Principal Analyst theCUBE Research

Scott Hebner

Principal Analyst, AI theCUBE Research

Bob Laliberte

Principal Analyst theCUBE Research

search

>> From theCUBE Studios in Palo Alto and Boston, bringing you data-driven insights from theCUBE and ETR. This is Breaking Analysis with Dave Vellante.

Dave Vellante

>> The data center, as we know it is being re-imagined as an AI factory. That means a power and data-optimized plant that turns energy and information into intelligence at massive industrial scale. Now, the tech stack is flipping from a general purpose CPU-centric systems focus to a GPU-centric accelerated compute, optimized for parallel operations and purpose built for AI. Network fabrics are very critical to this transition, and we're going to talk about that a lot today with other somewhat less important supporting elements, but still vital such as disaggregated storage, governed data planes, and a shift from app-centric operations to an agentic control plane that orchestrates models, tools and process workflows. Simply put, investments in general purpose computing are rapidly shifting toward extreme parallelism, what some call accelerated computing. And this is coincides with a networking shift towards scale up, scale out and scale across networking and automated governance that is built for AI throughout the stack. In this Breaking Analysis, David Floyer and I, along with Bob Laliberte, Scott Hebner and Paul Nashawaty delve into key aspects of AI factories. We're going to define what an AI factory is, we're going to talk about why now and the tectonic shift that's occurring. We'll put forth a reference architecture and take you inside the AI factory. And then we'll drill down into networking with Bob Laliberte doing a deep dive there. And then we'll look at market segmentation and updated forecast that David Floyer has done, and we'll importantly talk about the delta between CapEx spending and the revenue that is expected from things like token generation. We'll also briefly touch on the economics of AI factories and review Jensen's new law and then we'll explore the protocols and interoperability requirements along with the developer implications of the AI factory in this service as software world that we've been reporting on. And finally, we're going to close with a little view of some of the players that we see vying for dominance in this new era. First of all, David, welcome. Thanks for being here.

>> It's great. This is a fantastic subject. Very dear too my heart.

Dave Vellante

>> It sure is. All right, let's bring up the first slide and define what is an AI factory. I'm just going to read it for those people who are not watching along on the video and they're just listening. An AI factory is a system and that system is purpose built for AI production and that transforms... The purpose of this is to transform raw data into versatile AI outputs. Things like text, images, code, and other forms of intelligence including tokens. And it's doing so through automated processes that integrate data pipelines, model training, inference, deployment, ongoing monitoring and continuous improvement to produce intelligent at, as we said, massive scale. David, we're using this term of versatile AI outputs because it's more broad than tokens, so it's more than just tokens. The tokens are generally thought of as largely text-based, but images are typically pixel-based. And the AI factories, we want to be specific here that the workloads that they're supporting are AI-optimized, AI in particular. And so this definition goes beyond hardware, it addresses the end-to-end processes of creating intelligence, as we say at scale. And it's got the technical frameworks that we're going to talk about in some detail and maybe we'll debate that. But anything you would add to this definition?

>> I think is an extremely good definition. I would add video into that as a very important both source and output from these factories. And these factories are enormous and they have to do huge amounts of computing and they're very, very specialized in this process of creating intelligence.

Dave Vellante

>> All right, Alex, if you bring up the next slide, let's take a look at the reference architecture and some of the key points around that. As people often say energy goes in, we don't really actually show energy on this slide, but imagine power coming in, but it's really about ingesting data and then turning that data into intelligence. And so in order to do so... And again, we're talking about very specific AI workloads and infrastructure beneath it that we're going to talk about that's optimized for extreme parallel processing or accelerated computing. And you've got this process flow, today's data pipelines, you're doing training, you're building models, you're doing YOLO runs on those training models. Eventually you're doing inference. You're using specific tools today like TensorFlow and PyTorch and containers and container orchestration. David, I'm going to ask you about that. And you basically have these outputs and the ROI comes not only from training an LLM, but you might be building a computer vision system. You might improving fraud detection and lowering the time it takes to detect fraud. We're seeing a lot more AI recommendation engines. We're seeing use cases in healthcare, all kinds of different applications that are emerging in virtually every industry. David, I want to ask you in our prep call, you pointed to the second bullet here, this process flow, your premise is that these complexities are going to be abstracted away from the user that they are not going to have to worry about this, that software companies are going to worry about those primitives and hide that complexity from the user. Is that right? Is that your thinking here?

>> That's very much my thinking. These are specialized AI factories. They're not going to be things that individual corporations are going to put in unless they want to join in this providing these services. They're huge. And all of these things like TensorFlow and PyTorch will be under the hood as far as corporations are concerned. What they will see is APIs and software driven from those APIs and they'll be interfacing with that primarily. And that's the only way that in my view, that the complexity of this is going to be hidden and allow people quickly to get on with actually producing the things that matter, which are the real time outputs from this that will provide intelligence to medical diagnosis or real-time fraud or whatever the business problem or social problem that people are trying to solve.

Dave Vellante

>> Now, let's talk about the implications here on the stack. This isn't just a bunch of hype in our view. It's a plate tectonic shift. Alex, you bring up the collision slide between traditional data centers and AI factories. And really this is reshaping infrastructure. It's kind of an iPhone moment for data centers, so you can see that post late 2022 into 2023, the ChatGPT moment, the thinking around data centers started to shift toward these factories for intelligence. And that transition, it really isn't optional, it's really what's behind these exploding markets. GTC 2024 was in our view, one of the most important if not the most significant conference in the history of the computer industry. And you can see on this slide on the left-hand side you've got the traditional stack with x86 classic GPU centric compute. You've got traditional networks, you've got legacy storage and everything was built around the application. You would start with that application, you would build a snowflake for that application, build the infrastructure around that application, harden it, and then serve the business with it. And you can see at the bottom here, post 2023, and we'll quantify this, you really started to see gen AI breakthrough and all this CapEx pouring in and that's just been escalating. Everybody's talking about the power concerns and the limits of power. We have some thoughts on that. We think the world is going to solve that problem. And then this is the year of agentic hype. It's going to take some time to play out, but now you're seeing this AI factory boom, and all layers of the stack are changing. The compute you're talking about scale up and scale out and scale across networks. Bob Laliberte is going to talk about that today. Disaggregated storage and parallel file systems and object store becoming even more mainstream than it has been. And all of this is sort of governed and automated across the stack. And of course we talk about the agentic orchestration system quite a bit. And so gen AI is exploding. You're talking about 100X compute demand, you're talking about exponential growth. There's a hyperscale arms race going on, three, $400 billion being spent annually, and that's growing from hyperscalers and a complete revolution in hardware from the Frontier chips. The power demand inside of racks, 600 kilowatt plus racks, liquid cooling, everything's changing. But the economic shift is also we're shifting data centers as a cost center to actually a value producing entity. David, your thoughts?

>> Yes, I think this is the most amazing transformation. I think it's important to point out from a corporate point of view, the traditional data centers will still be there and they will not have AI factories themselves. The AI factories will be separate, they will be distributed, but they won't be things that almost all corporations will be running, so there's the factories producing the intelligence and then traditional data centers utilizing those looking for APIs into those API AI factories. That'll be the split between these two centers.

Dave Vellante

>> Let's take a pause here, take a beat and bring in Bob Laliberte to talk about the networking piece. It is one of the most, other than GPUs, important aspects of the stack. Take it away, Bob.

Bob Laliberte

>> Thanks Dave. It's really hard to talk about AI factories without talking about networking, especially when you look at the different dimensions as to how these AI factories are scaling. A lot of it, the talks so far has been about scale up, scale out, but now there's a third dimension that's scale across. Let's get started by talking about scale up, so this is traditionally what happens inside that single server or chassis. And it's all about that raw GP performance's an ultra low latency. You see things like Nvidia's NVLink, NVSwitch, maybe PCIe Gen 6 that are looking to connect the GPUs directly to each other and creating the high bandwidth fabric that keeps the data moving really fast for these large model training. Think of this as the intranode acceleration, really that foundation of every AI system. Next comes the scale out networking. And this is really about expanding beyond that single server chassis, interconnect dozens or even hundreds or now hundreds of thousands of GPU nodes within a data center. And this is where you start hearing about that discussion around ethernet and InfiniBand. Those fabrics are really dominant. Vendors like Cisco, Arista, HPE, Juniper, Nvidia, they've all built out their AI optimized fabrics using the Rocky V2, VXLAN EVPN, taking advantage of advanced telemetry. And this is the layer that delivers that east west bandwidth that keeps the AI clusters efficient and synchronized. And more recently there's been announcements from Broadcom around their Tomahawk 6 and they just did their Thor Ultra, which is an 800 gig network interface card that really enables organizations to get up to 100,000 XPU cluster environment. Things are moving fast, but you're seeing that technology keeping up, the network keeping up, getting into those 800 gig network bandwidth rates now. And then finally, the scale across networking is what's going to connect multiple AI data centers or in regions really into that one unified fabric, creating that fully distributed AI environment and regional AI factories. This is going to be super important as it helps organizations address those power availability requirements and really being able to reduce the burden on one specific power supply, so organizations can now start looking at distributing their AI factories across a regional area. And as I mentioned, really hot topic, everyone is bringing out solutions, Nvidia, Cisco, Arista all had announcements over the last month or two, and you're also starting to see startups pop up like LumaLens. And again, the whole idea is to enable secure, deterministic, low latency connection between these sites. And so it's really how organizations are going to deliver that federated learning, cross region inferencing and data sovereign AI. To sum it up, scale up builds performance, scale out delivers elasticity and scale across creates those regional solutions. And really together they're going to form the backbone of tomorrow's distributed AI infrastructure, connecting the compute data and agents across regions and potentially further, which of course will be determined by how much latency they can tolerate. Now, another key part that everyone likes to talk about in networking when it comes to these AI factories, is that challenge or that discussion around is it going to be ethernet or is it going to be InfiniBand? Well, I think what we've seen is that, and we've got research that you'll see here in the slide, that shows where enterprises are leaning towards and that's towards that ethernet solution. And if you look at the research, according to what we pulled last year, 59% of organizations prefer ethernet to support their AI workloads as compared to only 38 who wanted InfiniBand. Now, when you look at the reasons why one versus the other, you can see the top reason for choosing ethernet is pretty simple. It's already everywhere, so organizations have deployed ethernet, they're very familiar with it. Nearly half the respondents, 46%, said they use ethernet across the rest of the network, so it just makes sense that they'd want to use it there. 44% cited that they have high in-house skill sets that they want to be able to leverage. And then another 43% noted that the major cloud providers are relying on ethernet, and that's become virtually ubiquitous now. All of the cloud providers have focused on that ethernet fabric for their environments. If you want to flip the switch and look at InfiniBand and why organizations are doing that, the primary driver is a lot of organizations, almost 60% believe it delivers better performance. But again, ethernet's closing that gap, and obviously that's especially true for those deeply-coupled, closely-coupled GPU clusters. Others pointed to the fact that InfiniBand's packaged with many GPU systems, so if you're buying something from Nvidia, you're going to have InfiniBand likely, even though they have both, that's likely where you'll have InfiniBand packaged up as part of your AI factory. And of course some their partners do that as well, they'll package up that InfiniBand with it. And then lastly, the other place where InfiniBand, where it's traditionally used is in those high-performance compute environments. If you're an enterprise, you're already doing high-performance compute, you're likely to maybe say, "You know what? I'm going to use this as well because I'm familiar with it and I'm really comfortable with its performance." But at the end of the day, if I could give you my takeaway, now ethernet's definitely going to be the winner in the enterprise AI data center because it's familiar, it's scalable, it's cloud-aligned, and its performance continues to improve. And while InfiniBand will be the fabric of choice for the most performance-sensitive AI training clusters, end of the day I think we're going to see the ethernet winning out.

Dave Vellante

>> Thank you for that. Appreciate that insight and that data. I want to revisit David, something that you and I have talked about in previous Breaking Analysis and that's really this AI revolution and let's quantify what that looks like and revisit with this next slide, the data center build-out that's happening now. It's actually just remarkable. What this shows, this is our updated forecast that, David, you produced in the spring of this year, and it's all data center infrastructure. It's the core infrastructure, it's the power, it's the cooling, it's the compute, storage and networking, the core operating system and system software. But look at this, this is a business that is perpetually a $200 billion business through COVID and pretty much every year. And then all of a sudden in 2024 it exploded. It went from 222 billion to 350 billion. And you can see in the forecast here, the brown is the traditional general purpose computing and the blue is accelerated computing, what we call extreme parallel processing or AI extreme. And that blue just explodes and we're talking about nearly a trillion dollar market, David by 2030 with a mid-teens CAGR for the whole data center space. And then higher than that for the accelerated computing piece. And then I want you to comment, but in the context, David, of this next slide, which is worldwide revenue from AI factories. And this is of course in US dollars, but what I want you to... Well, first of all, I want to point out the delta between that nearly trillion dollar of CapEx end in 2030 and what you have here of about just under half a trillion in worldwide revenue, big delta there not a super surprise, but pretty big numbers and a pretty big gap, so the revenue doesn't match the spend. US companies, and I want you to comment on this, they look like they win despite China's formidable presence. But you have a theory and a premise on OpenAI that I want you to explore. And you've got an Anthropic on here, so there's clearly more room than just one of what are currently foundation model vendors, but they better find a business model beyond that, so we're calling them AI research labs. And you've got some concerns about the cloud. There's still big numbers here, but not nearly as big as some of the up and comers in particular OpenAI. Take it away, David, explain what your thinking is behind this, how you came up with this data and what it all means.

>> Yes. Well, this data is the revenue from that explosion, all of these AI factories. And the thinking here is that the factories themselves are going to put in that the APIs, the connections to it, which will allow revenue to be generated from it. And that revenue where we're going to talk about it in a minute and how to spit that out, but that revenue is going to not come from the models themselves, but come from software on top of those models. If we look OpenAI, they are in 2025, they have $6.6 billion, growing dramatically from 2024, and they are in the front seat of this usage by consumers, a lot of it at the moment, prosumers and obviously by corporations themselves, so that's a projection. And OpenAI has the lead at the moment and assuming that they don't screw up too much is likely they're going to continue to do that. And I have them at 150 billion in 2030. Anthropic have again got extremely good corporate interfaces. When we look at Google, they have tended to use AI inside their own corporation for doing things like adverts, as have Meta, so that's probably going to be harder for them to generate revenue from their models themselves. They're using it for their own good. There are software companies like Microsoft, which are obviously going to be fine, and Microsoft themselves are doing a lot of the actual running of these factories. And the other very, very strong belief that I have is that China will enter this market with a bang. They're very low at the moment, but they're going to enter it with a bang and they're going to be using their own hardware and software. They're going to create their own AI factories. They may not be quite as efficient, but they will definitely be producing a lot of revenue. And the revenue here is obviously worldwide revenue coming to them. Those are the main points I want to make from this slide, that the worldwide revenue is going to be growing, it's growing very, very fast and it's going to continue to grow.

Dave Vellante

>> Well, and basically you are putting forth the premise that OpenAI makes the transition to an enterprise provider via those APIs that you suggested, so that's going to be interesting to see because right now it doesn't appear that they have a viable business model. They're just out raising a lot of money and doing some great work. And one could say the same with Anthropic. I think they're still trying to determine what that business model looks like, but now it's because the market is so bubble-licious able to raise a lot of money. But the cloud guys are I think... Let's run through them. Microsoft, like you said, is going to be fine. They have a strong software business. Google's search is maybe under attack from the Perplexitys of the world and potentially OpenAI, but they've still got a big, big search business. And so that's interesting. AWS has a lot of, what is now general purpose computing running on its platform and it's transitioning to an AI world, so it's going to be interesting to see how Jassy and Garman navigate through that. But I want you to, Alex, if you could bring up the next slide. This is interesting to me, David, because you basically are saying that B2B is where the big revenue opportunity is. And of course we know public policy is very important here. You got to get the governments driving this, but it's important. This is revenue by sector, but it's just from AI factories that are US-based corporation owned, so it's US-based worldwide revenue. We don't have this ability at this point in time and all the long tail. For instance, China, I think if you go back, don't go back Alex, but then that previous side, remember there was an other line, other could be much, much larger than that. And so we'll revisit that over time. But it's interesting here, David, because B2B is where you see the revenue coming, even though consumer and prosumer is where it all gets started. That's the main spring of innovation. And again, this assumes OpenAI creates a viable business model, particularly for the enterprise, which is somewhat novel thinking. There are some people who believe that, but you are convinced that they are going to lead this charge, notwithstanding my prediction in early 2023 that they wouldn't maintain their first mover advantage. I debated John Furrier about that with Sarbjit Johal. You would say that John is going to be correct in that front.

>> I am not trying to pick sides there, but yes, if you go back for a second, if you remember on the first slide, the traditional data centers are declining and they will have to get all of the intelligence from the new AI factories, so unless people like ChatGPT and unless the others look hard at getting where the money is going to come from, which is corporations, they're going to fail. And I've got corporations software at the API, I think the APIs are going to be absolutely critical. And that's a very large amount of revenue that's going to come in from APIs. And then equally all the other services that corporations will directly connect to. Those are the areas where the growth is. Consumer and prosumer individuals improving will still be a very big market indeed. And my view of government is going to be very slow in the military, for example, so there's some interesting potential problems worldwide, but they have already been very slow in adopting any new function. For example, I think OpenAI ChatGPT was banned across the whole of the government for the last couple of years and so only just making its appearance there.

Dave Vellante

>> I think it's been in a lot of places, so people are concerned about putting proprietary data up into the... This is one of the... Your premise assumes that people are going to trust OpenAI, so we'll see. But let's take another beat here, David, let's take a pause and bring in Scott Hebner who's going to talk about some of the interoperability standards. Please take it away, Scott Hebner.

Scott Hebner

>> Hi everyone. I'd like to focus on what I consider to be a key success factor for AI factories and especially the generative AI layer that plays such an enormously important role in the whole concept of an AI factory. The key success factor is the adoption of open standards and protocols. About 25 years ago at the start of the internet era, if you remember back then we got caught up in what was known as the browser wars, Netscape, Internet Explorer, later Mozilla, so on and so forth. And when you look back at it, you know what happened, they standardized out. Today we're in the middle of the same kind of gen AI wars, and they're definitely dominating the discourse ChatGPT, Claude, Gemini, Grok, so on and so forth. But if you really think about it, they're both just gateways to something much bigger. The browsers were the gateway to the world of the internet and generative AI is the gateway into the world of AI. The real value doesn't lie in that gateway. It comes from what passes through it. And to pass through them and to be widely consumed and integrated into businesses, they need to be open. As history has shown, if you think about it, the browsers eventually succumb to open standards and protocols. And that's what made the internet ultimately accessible and scalable and brought on the whole transformation around e-business. Now, while generative AI is a gateway to AI, it's the AI factories that promise to power what flows through them. And that's the back end of the emerging intelligence economy, and that's what's really going to power this next phase of digital business transformation is these factories. Because most companies can't afford to do everything they need to do with AI. And there needs to be some standardized way to consume not only the AI capabilities, but also just as importantly the power that comes along with it. My view is, without open standards and protocols these factories risk becoming early mistakes of the internet era. You get fragmentation, you get lock-in, you get inefficiencies, you get untrusted outcomes. And our agentic AI futures index that we just released backs this up. If you take a look at the chart and bring that up, Alex, if you could, you can see a majority of enterprises are prioritizing open models, protocols and APIs, extensible adapters. They want to connect to things. They want interoperability among agents. They want to choose their LLMs, they want the ability to compose any agent into any workflow. It all comes around open standards. And this is clearly telling you that enterprises don't want to get locked in, they want choice, flexibility, and the ability to interconnect things. I think of it this way, just like our electricity grid, it's standardized circuits, breakers, voltage regulators, fuses, things of that nature that make it safe and easy to consume across a whole different ride of devices such as in AI world, agents or businesses, which could be the concept of a home. I think these open protocols play an absolutely critical role in allowing the intelligence pipelines to flow safely, reliably into the agents, the applications, and the enterprises that are so dependent on these things being successful. If you think about it, step back a little bit, open standards are what make the intelligence pipeline an actual pipeline, not a series of disconnected pipes that don't work together. What good is that going to be if you really think about it from a business perspective? I think this new agentic control plane that AI factories can actually deliver, where agents can communicate and coordinate and do things together and really bring to life this service as a software model for a next generation of applications and value, it's all dependent on that. And the good news I think, is that if you look at Nvidia and AMV and Intel and others, they're taking essential steps towards open standards. I think the bad news is there are partial measures. Most of today's AI factories remain semi-open at best. If you learned anything from the early e-business area, innovation compounds faster when you have leading vendors that go full stop in open standards. And I hope these companies are learning the lesson of history. To summarize all this, if you think about it, and as many learned back in the internet days, you can eat 95% health food, but if you eat 5% rat poison, you still die. You need to go full stop in open standards and without it we risk having these AI factories die a slow death over time. That's my two cents on this. I think the AI factories are a critical part of the long-term success of how businesses are going to consume and drive ROI from AI. And again, let's not forget the importance of openness. With that I'll turn it back to you guys.

Dave Vellante

>> Okay, thank you, Scott. Now, David, this is interesting because Scott, here he is, he surveyed a bunch of folks from traditional enterprises, and of course they're going to say these interoperability standards and open models and protocols and these matter, that they're critical because this is the lifeblood of how they get work done today. But your premise, again, going back to some of the earlier comments that we were discussing is that that all gets abstracted away. And the software companies like OpenAI and these leading AI model companies and potentially the cloud guys, they're not going to sit still here, have strapped that complexity away to compete. And we're limited on time here. I'm sorry to push you, but very quickly, what's the timeframe for that abstraction layer to actually emerge where OpenAI actually starts to show that kind of progress based off, I recall based on your earlier charts, we're talking about 27, 28, that really is in full swing when they become a 40, $50 billion company.

>> Well, we're seeing that already with the services that are provided to developers. There's a very rich set of services that have gone in already. And then we are seeing the beginning of those services with all the features included in, for example, OpenAI. And so they're putting in all of the building blocks already. Now, what's very important, I think very, very important indeed is that OpenAI put in the governance capabilities that corporations will need. But my simple point is this. This is not possible for individual corporations to do on their own. They have to have these huge data centers providing these services to them. And if an OpenAI fail, then it leaves the door wide open for somebody else to take over. I personally believe they won't fail. I think they've got really excellent leadership there, but they could go wrong. But the data has to get through them to the end users. And the only way it's going to be done is through simplicity and providing that wrap around it so the corporations can use it easily.

Dave Vellante

>> Excellent. Let's now take a look at the developer angle on AI factories and bring in Paul Nashawaty. Paul, please go.

Paul Nashawaty

>> Hey, thanks for having me on today. This is really exciting, and especially in the world of application development. When I think of AI factories in the data center of the future, today, we're really exploring one of the most transformative shifts in the technology landscape. The convergence of accelerating computing, application modernization, platform engineering, and DevSecOps is really a big factor, especially when I think of the CI/CD pipeline and the overall SDLC. But this isn't just another wave of technology changes. It's really strategically redefining and how software is being developed, secured, and delivered across every industry, whether it's compliance regulations or governance, it really takes a factor here. When we enter this new era of where performance, automation and intelligence are no longer optional, they really do become foundational. Let's take a look at, I have a slide here I would like to start on. And then in this slide I just really want to highlight that there's five segments here. And when we think about the first part, which we think about accelerated computing, this is leveraging specialized GPUs, TPUs, FPGAs, et cetera, that are unlocking computational capabilities that what really once thought is impossible on traditional architectures. And this is enabling breakthrough areas like real-time analytics, AI-driven decision making, and large-scale simulation. And according to my research from our AppDev summit that we did in September of this year, the research shows that cloud infrastructure, GPU-based acceleration market is growing nearly 47% in 2025, reaching $157 billion in value, driven by AI architectures and infrastructure investments that are growing on demand for high-performance computing. And so for developers this creates a really exciting challenge and it's really an opportunity. And when we look at it, we now have the ability to build smarter, more responsive applications that can demand new programming models, tools, and optimize for parallel architectures. In short, the accelerated computing part is really just speeding up things and it's really helping redefine what's possible next. The second part I want to talk about is application modernization and platform engineering. This is a big area that I think gets overlooked a lot. Modernization is a treadmill, it happens. It's a big market though. When we think about the service market here is valued about $19.8 billion in 2024, and it's expected to double to $39.6 billion by 2029. That's a big jump. And so this surge reflects how the enterprises are really re-architecting legacy systems. They're moving to cloud-native environments and embedding continuous delivery pipelines. But modernization is not enough alone. It really has to work with truly scaling innovation. And platform engineering really does become central, so platform teams are now tasked with building robust self-service environments that use infrastructure as code, container orchestration and advanced observability frameworks. This is really critical to the next generation of modernization. In 2025 the focus has shifted from simply providing infrastructure to delivering complete developer platforms that really enhance productivity and improve developer experience. Let's look at the numbers though that show this. In a research study that we did with Google Cloud shows that organization really has shown co-managed platforms give developers 47% of their production time to innovation and experimentation compared to 38% of companies managing platforms internally, so internal development portals and self-service capabilities are streamlining workflows, this is reducing friction and accelerating delivery. Platform engineering is no longer a support function, it really is a competitive advantage and a growth engine for a lot of organizations. And when we look at DevSecOps in the next part of this, what we really understand here is application scale and complexities, increased security also must have evolved around it, so that's where DevSecOps really does come in. And the global DevSecOps market is projected to hit $41 billion by 2030 and growing more than 30% annually, so there's a lot happening here. This is, not to reiterate a lot of things you're probably already hearing in the market, but the key trends that are shaping the evolution include that shifting left for security testing early in the CI/CD pipeline stops vulnerability before production, AI-driven automation. As we're moving into AI-driven automation, I'm still a firm believer of a human in the loop, but it's really meant to detect and respond to real-time threats. And zero-trust architecture is really continuously verifying every user and device across every environment, so DevSecOps is a big part of this. And it's not just a checkpoint, it's a continuous part of the SDLC, so that's really an important part. The other part I want to tie into is service as a software here, service as a software is that all the trends are converging into something bigger here. And we look at this new business model as services as a software. This model helps software become more a primary vehicle for delivering value experiences and services to customers. And when we look at that across these AI factories, enterprises are adopting this model by using agentic process-centric applications to accelerate innovation, optimize operations, as well as open entirely new revenue streams, so there's a lot of things happening here and those organizations that are delaying, they risk falling behind in the marketplace because defining by speed, intelligence and scale is really the new competitive advantage. In closing, what I would recommend here based on my research that we saw in the AppDev space, the analysis concludes that this converges accelerated compute, modernization, platform engineering and DevSecOps is fundamentally reshaping the software world. Where do you get started? Well, I'm thinking about this in the context of assess readiness by providing audit current... You have to look at it in audit current applications and infrastructures that are ready for modernization acceleration. This doesn't mean everything gets refactored. This means the things that need to be refactored are refactored. You want to invest in platform engineering. I can't emphasize that enough. It's a big part of the infrastructure. It's a big part of the internal build for developer portals and also that self-service model for environments to empower that innovation. You want to not slow down innovation here. And then of course, security is key. Embed DevSecOps that's integrated in the security automation throughout every stage of the development life cycle. Those companies that really embrace this shift will really lead the next generation of secure, scalable, and intelligent applications driving competitive advantage across these digital initiatives that are happening out there. With that, Dave, back to you. This is really exciting stuff. I think from the AppDev world, I couldn't be any more on point.

Dave Vellante

>> All right, thank you, Paul. Hey, let's wrap. Let's bring up the very last slide if you would, Alex. This is a slide from our friends at ETR. It basically shows net score or spending momentum on the vertical axis and overlap or penetration into a data set of almost 1800 respondents on the horizontal. And we just took company names, but this isn't any particular sector, it's all sectors. But some of the players that we see here, of course you've got Microsoft, you've got AWS, you've got your Google, OpenAI... I forgot to put Anthropic on this chart, apologies to those guys, I should have done that. You got Meta, you got the networking guys, Arista, Cisco, Juniper, and HPE of course the semiconductor guys with Intel and Broadcom, AMD, and of course Nvidia on the top of the heap here. And you got Uncle Sam and China, as David was pointing out before. You've also got some of the Neo cloud players, CoreWeave, Lambda, Nebius, other data center builders are going to be in there. And so energy companies, this ecosystem is massive. This is a multi, multi-trillion dollar market opportunity. Thank you Alex. And David, so we got to run, but I'll give you last word. Bring us home.

>> These are all the people that are contributing tremendously to this dramatic change. I think from a corporate point of view, they're going to be focusing on the people providing the services. They're going to enhance their own data centers with a high degree parallelism within the data centers, so people like Nvidia are going to be important, but the most important partners they're going to have are the people that provide them with the APIs and the software providers who are using those APIs. Those are the two sets of people in my view who are going to turn this fantastic operation of providing intelligence, turn it into a business profit.

Dave Vellante

>> Okay, we shall see. Interesting take, David. I really appreciate it. We got it wrapped there. Thanks, Alex Morrison and on production and manages our podcast, Ken Schiffman as well. Kristen Martin and Cheryl Knight, they help get the word out on social media and our newsletters. And Rob Hof is our EIC over at siliconangle.com. Thank you all. Remember all these episodes are available as podcasts, just search Breaking Analysis podcast, we publish each week on Thecubereserch.com and siliconangle.com. And you can email me at david.vellante@siliconangle.com. If you've got an idea, you've got a guest you want to bring on or email or DM me at dvellante, comment on our LinkedIn post, check out etr.ai for the best survey data in the enterprise tech business. This is Dave Vellante for theCube Research Insights. And David Floyer, Scott Hebner, Bob Laliberte, and Paul Nashawaty, powered by ETR. Thanks for watching and we'll see you next time on Breaking Analysis.

296 | Breaking Analysis | AI Factories - Data Centers of the Future

search

>> From theCUBE Studios in Palo Alto and Boston, bringing you data-driven insights from theCUBE and ETR. This is Breaking Analysis with Dave Vellante.

Dave Vellante

>> The data center, as we know it is being re-imagined as an AI factory. That means a power and data-optimized plant that turns energy and information into intelligence at massive industrial scale. Now, the tech stack is flipping from a general purpose CPU-centric systems focus to a GPU-centric accelerated compute, optimized for parallel operations and purpose built for AI. Network fabrics are very critical to this transition, and we're going to talk about that a lot today with other somewhat less important supporting elements, but still vital such as disaggregated storage, governed data planes, and a shift from app-centric operations to an agentic control plane that orchestrates models, tools and process workflows. Simply put, investments in general purpose computing are rapidly shifting toward extreme parallelism, what some call accelerated computing. And this is coincides with a networking shift towards scale up, scale out and scale across networking and automated governance that is built for AI throughout the stack. In this Breaking Analysis, David Floyer and I, along with Bob Laliberte, Scott Hebner and Paul Nashawaty delve into key aspects of AI factories. We're going to define what an AI factory is, we're going to talk about why now and the tectonic shift that's occurring. We'll put forth a reference architecture and take you inside the AI factory. And then we'll drill down into networking with Bob Laliberte doing a deep dive there. And then we'll look at market segmentation and updated forecast that David Floyer has done, and we'll importantly talk about the delta between CapEx spending and the revenue that is expected from things like token generation. We'll also briefly touch on the economics of AI factories and review Jensen's new law and then we'll explore the protocols and interoperability requirements along with the developer implications of the AI factory in this service as software world that we've been reporting on. And finally, we're going to close with a little view of some of the players that we see vying for dominance in this new era. First of all, David, welcome. Thanks for being here.

>> It's great. This is a fantastic subject. Very dear too my heart.

Dave Vellante

>> It sure is. All right, let's bring up the first slide and define what is an AI factory. I'm just going to read it for those people who are not watching along on the video and they're just listening. An AI factory is a system and that system is purpose built for AI production and that transforms... The purpose of this is to transform raw data into versatile AI outputs. Things like text, images, code, and other forms of intelligence including tokens. And it's doing so through automated processes that integrate data pipelines, model training, inference, deployment, ongoing monitoring and continuous improvement to produce intelligent at, as we said, massive scale. David, we're using this term of versatile AI outputs because it's more broad than tokens, so it's more than just tokens. The tokens are generally thought of as largely text-based, but images are typically pixel-based. And the AI factories, we want to be specific here that the workloads that they're supporting are AI-optimized, AI in particular. And so this definition goes beyond hardware, it addresses the end-to-end processes of creating intelligence, as we say at scale. And it's got the technical frameworks that we're going to talk about in some detail and maybe we'll debate that. But anything you would add to this definition?

>> I think is an extremely good definition. I would add video into that as a very important both source and output from these factories. And these factories are enormous and they have to do huge amounts of computing and they're very, very specialized in this process of creating intelligence.

Dave Vellante

>> All right, Alex, if you bring up the next slide, let's take a look at the reference architecture and some of the key points around that. As people often say energy goes in, we don't really actually show energy on this slide, but imagine power coming in, but it's really about ingesting data and then turning that data into intelligence. And so in order to do so... And again, we're talking about very specific AI workloads and infrastructure beneath it that we're going to talk about that's optimized for extreme parallel processing or accelerated computing. And you've got this process flow, today's data pipelines, you're doing training, you're building models, you're doing YOLO runs on those training models. Eventually you're doing inference. You're using specific tools today like TensorFlow and PyTorch and containers and container orchestration. David, I'm going to ask you about that. And you basically have these outputs and the ROI comes not only from training an LLM, but you might be building a computer vision system. You might improving fraud detection and lowering the time it takes to detect fraud. We're seeing a lot more AI recommendation engines. We're seeing use cases in healthcare, all kinds of different applications that are emerging in virtually every industry. David, I want to ask you in our prep call, you pointed to the second bullet here, this process flow, your premise is that these complexities are going to be abstracted away from the user that they are not going to have to worry about this, that software companies are going to worry about those primitives and hide that complexity from the user. Is that right? Is that your thinking here?

>> That's very much my thinking. These are specialized AI factories. They're not going to be things that individual corporations are going to put in unless they want to join in this providing these services. They're huge. And all of these things like TensorFlow and PyTorch will be under the hood as far as corporations are concerned. What they will see is APIs and software driven from those APIs and they'll be interfacing with that primarily. And that's the only way that in my view, that the complexity of this is going to be hidden and allow people quickly to get on with actually producing the things that matter, which are the real time outputs from this that will provide intelligence to medical diagnosis or real-time fraud or whatever the business problem or social problem that people are trying to solve.

Dave Vellante

>> Now, let's talk about the implications here on the stack. This isn't just a bunch of hype in our view. It's a plate tectonic shift. Alex, you bring up the collision slide between traditional data centers and AI factories. And really this is reshaping infrastructure. It's kind of an iPhone moment for data centers, so you can see that post late 2022 into 2023, the ChatGPT moment, the thinking around data centers started to shift toward these factories for intelligence. And that transition, it really isn't optional, it's really what's behind these exploding markets. GTC 2024 was in our view, one of the most important if not the most significant conference in the history of the computer industry. And you can see on this slide on the left-hand side you've got the traditional stack with x86 classic GPU centric compute. You've got traditional networks, you've got legacy storage and everything was built around the application. You would start with that application, you would build a snowflake for that application, build the infrastructure around that application, harden it, and then serve the business with it. And you can see at the bottom here, post 2023, and we'll quantify this, you really started to see gen AI breakthrough and all this CapEx pouring in and that's just been escalating. Everybody's talking about the power concerns and the limits of power. We have some thoughts on that. We think the world is going to solve that problem. And then this is the year of agentic hype. It's going to take some time to play out, but now you're seeing this AI factory boom, and all layers of the stack are changing. The compute you're talking about scale up and scale out and scale across networks. Bob Laliberte is going to talk about that today. Disaggregated storage and parallel file systems and object store becoming even more mainstream than it has been. And all of this is sort of governed and automated across the stack. And of course we talk about the agentic orchestration system quite a bit. And so gen AI is exploding. You're talking about 100X compute demand, you're talking about exponential growth. There's a hyperscale arms race going on, three, $400 billion being spent annually, and that's growing from hyperscalers and a complete revolution in hardware from the Frontier chips. The power demand inside of racks, 600 kilowatt plus racks, liquid cooling, everything's changing. But the economic shift is also we're shifting data centers as a cost center to actually a value producing entity. David, your thoughts?

>> Yes, I think this is the most amazing transformation. I think it's important to point out from a corporate point of view, the traditional data centers will still be there and they will not have AI factories themselves. The AI factories will be separate, they will be distributed, but they won't be things that almost all corporations will be running, so there's the factories producing the intelligence and then traditional data centers utilizing those looking for APIs into those API AI factories. That'll be the split between these two centers.

Dave Vellante

>> Let's take a pause here, take a beat and bring in Bob Laliberte to talk about the networking piece. It is one of the most, other than GPUs, important aspects of the stack. Take it away, Bob.

Bob Laliberte

>> Thanks Dave. It's really hard to talk about AI factories without talking about networking, especially when you look at the different dimensions as to how these AI factories are scaling. A lot of it, the talks so far has been about scale up, scale out, but now there's a third dimension that's scale across. Let's get started by talking about scale up, so this is traditionally what happens inside that single server or chassis. And it's all about that raw GP performance's an ultra low latency. You see things like Nvidia's NVLink, NVSwitch, maybe PCIe Gen 6 that are looking to connect the GPUs directly to each other and creating the high bandwidth fabric that keeps the data moving really fast for these large model training. Think of this as the intranode acceleration, really that foundation of every AI system. Next comes the scale out networking. And this is really about expanding beyond that single server chassis, interconnect dozens or even hundreds or now hundreds of thousands of GPU nodes within a data center. And this is where you start hearing about that discussion around ethernet and InfiniBand. Those fabrics are really dominant. Vendors like Cisco, Arista, HPE, Juniper, Nvidia, they've all built out their AI optimized fabrics using the Rocky V2, VXLAN EVPN, taking advantage of advanced telemetry. And this is the layer that delivers that east west bandwidth that keeps the AI clusters efficient and synchronized. And more recently there's been announcements from Broadcom around their Tomahawk 6 and they just did their Thor Ultra, which is an 800 gig network interface card that really enables organizations to get up to 100,000 XPU cluster environment. Things are moving fast, but you're seeing that technology keeping up, the network keeping up, getting into those 800 gig network bandwidth rates now. And then finally, the scale across networking is what's going to connect multiple AI data centers or in regions really into that one unified fabric, creating that fully distributed AI environment and regional AI factories. This is going to be super important as it helps organizations address those power availability requirements and really being able to reduce the burden on one specific power supply, so organizations can now start looking at distributing their AI factories across a regional area. And as I mentioned, really hot topic, everyone is bringing out solutions, Nvidia, Cisco, Arista all had announcements over the last month or two, and you're also starting to see startups pop up like LumaLens. And again, the whole idea is to enable secure, deterministic, low latency connection between these sites. And so it's really how organizations are going to deliver that federated learning, cross region inferencing and data sovereign AI. To sum it up, scale up builds performance, scale out delivers elasticity and scale across creates those regional solutions. And really together they're going to form the backbone of tomorrow's distributed AI infrastructure, connecting the compute data and agents across regions and potentially further, which of course will be determined by how much latency they can tolerate. Now, another key part that everyone likes to talk about in networking when it comes to these AI factories, is that challenge or that discussion around is it going to be ethernet or is it going to be InfiniBand? Well, I think what we've seen is that, and we've got research that you'll see here in the slide, that shows where enterprises are leaning towards and that's towards that ethernet solution. And if you look at the research, according to what we pulled last year, 59% of organizations prefer ethernet to support their AI workloads as compared to only 38 who wanted InfiniBand. Now, when you look at the reasons why one versus the other, you can see the top reason for choosing ethernet is pretty simple. It's already everywhere, so organizations have deployed ethernet, they're very familiar with it. Nearly half the respondents, 46%, said they use ethernet across the rest of the network, so it just makes sense that they'd want to use it there. 44% cited that they have high in-house skill sets that they want to be able to leverage. And then another 43% noted that the major cloud providers are relying on ethernet, and that's become virtually ubiquitous now. All of the cloud providers have focused on that ethernet fabric for their environments. If you want to flip the switch and look at InfiniBand and why organizations are doing that, the primary driver is a lot of organizations, almost 60% believe it delivers better performance. But again, ethernet's closing that gap, and obviously that's especially true for those deeply-coupled, closely-coupled GPU clusters. Others pointed to the fact that InfiniBand's packaged with many GPU systems, so if you're buying something from Nvidia, you're going to have InfiniBand likely, even though they have both, that's likely where you'll have InfiniBand packaged up as part of your AI factory. And of course some their partners do that as well, they'll package up that InfiniBand with it. And then lastly, the other place where InfiniBand, where it's traditionally used is in those high-performance compute environments. If you're an enterprise, you're already doing high-performance compute, you're likely to maybe say, "You know what? I'm going to use this as well because I'm familiar with it and I'm really comfortable with its performance." But at the end of the day, if I could give you my takeaway, now ethernet's definitely going to be the winner in the enterprise AI data center because it's familiar, it's scalable, it's cloud-aligned, and its performance continues to improve. And while InfiniBand will be the fabric of choice for the most performance-sensitive AI training clusters, end of the day I think we're going to see the ethernet winning out.

Dave Vellante

>> Thank you for that. Appreciate that insight and that data. I want to revisit David, something that you and I have talked about in previous Breaking Analysis and that's really this AI revolution and let's quantify what that looks like and revisit with this next slide, the data center build-out that's happening now. It's actually just remarkable. What this shows, this is our updated forecast that, David, you produced in the spring of this year, and it's all data center infrastructure. It's the core infrastructure, it's the power, it's the cooling, it's the compute, storage and networking, the core operating system and system software. But look at this, this is a business that is perpetually a $200 billion business through COVID and pretty much every year. And then all of a sudden in 2024 it exploded. It went from 222 billion to 350 billion. And you can see in the forecast here, the brown is the traditional general purpose computing and the blue is accelerated computing, what we call extreme parallel processing or AI extreme. And that blue just explodes and we're talking about nearly a trillion dollar market, David by 2030 with a mid-teens CAGR for the whole data center space. And then higher than that for the accelerated computing piece. And then I want you to comment, but in the context, David, of this next slide, which is worldwide revenue from AI factories. And this is of course in US dollars, but what I want you to... Well, first of all, I want to point out the delta between that nearly trillion dollar of CapEx end in 2030 and what you have here of about just under half a trillion in worldwide revenue, big delta there not a super surprise, but pretty big numbers and a pretty big gap, so the revenue doesn't match the spend. US companies, and I want you to comment on this, they look like they win despite China's formidable presence. But you have a theory and a premise on OpenAI that I want you to explore. And you've got an Anthropic on here, so there's clearly more room than just one of what are currently foundation model vendors, but they better find a business model beyond that, so we're calling them AI research labs. And you've got some concerns about the cloud. There's still big numbers here, but not nearly as big as some of the up and comers in particular OpenAI. Take it away, David, explain what your thinking is behind this, how you came up with this data and what it all means.

>> Yes. Well, this data is the revenue from that explosion, all of these AI factories. And the thinking here is that the factories themselves are going to put in that the APIs, the connections to it, which will allow revenue to be generated from it. And that revenue where we're going to talk about it in a minute and how to spit that out, but that revenue is going to not come from the models themselves, but come from software on top of those models. If we look OpenAI, they are in 2025, they have $6.6 billion, growing dramatically from 2024, and they are in the front seat of this usage by consumers, a lot of it at the moment, prosumers and obviously by corporations themselves, so that's a projection. And OpenAI has the lead at the moment and assuming that they don't screw up too much is likely they're going to continue to do that. And I have them at 150 billion in 2030. Anthropic have again got extremely good corporate interfaces. When we look at Google, they have tended to use AI inside their own corporation for doing things like adverts, as have Meta, so that's probably going to be harder for them to generate revenue from their models themselves. They're using it for their own good. There are software companies like Microsoft, which are obviously going to be fine, and Microsoft themselves are doing a lot of the actual running of these factories. And the other very, very strong belief that I have is that China will enter this market with a bang. They're very low at the moment, but they're going to enter it with a bang and they're going to be using their own hardware and software. They're going to create their own AI factories. They may not be quite as efficient, but they will definitely be producing a lot of revenue. And the revenue here is obviously worldwide revenue coming to them. Those are the main points I want to make from this slide, that the worldwide revenue is going to be growing, it's growing very, very fast and it's going to continue to grow.

Dave Vellante

>> Well, and basically you are putting forth the premise that OpenAI makes the transition to an enterprise provider via those APIs that you suggested, so that's going to be interesting to see because right now it doesn't appear that they have a viable business model. They're just out raising a lot of money and doing some great work. And one could say the same with Anthropic. I think they're still trying to determine what that business model looks like, but now it's because the market is so bubble-licious able to raise a lot of money. But the cloud guys are I think... Let's run through them. Microsoft, like you said, is going to be fine. They have a strong software business. Google's search is maybe under attack from the Perplexitys of the world and potentially OpenAI, but they've still got a big, big search business. And so that's interesting. AWS has a lot of, what is now general purpose computing running on its platform and it's transitioning to an AI world, so it's going to be interesting to see how Jassy and Garman navigate through that. But I want you to, Alex, if you could bring up the next slide. This is interesting to me, David, because you basically are saying that B2B is where the big revenue opportunity is. And of course we know public policy is very important here. You got to get the governments driving this, but it's important. This is revenue by sector, but it's just from AI factories that are US-based corporation owned, so it's US-based worldwide revenue. We don't have this ability at this point in time and all the long tail. For instance, China, I think if you go back, don't go back Alex, but then that previous side, remember there was an other line, other could be much, much larger than that. And so we'll revisit that over time. But it's interesting here, David, because B2B is where you see the revenue coming, even though consumer and prosumer is where it all gets started. That's the main spring of innovation. And again, this assumes OpenAI creates a viable business model, particularly for the enterprise, which is somewhat novel thinking. There are some people who believe that, but you are convinced that they are going to lead this charge, notwithstanding my prediction in early 2023 that they wouldn't maintain their first mover advantage. I debated John Furrier about that with Sarbjit Johal. You would say that John is going to be correct in that front.

>> I am not trying to pick sides there, but yes, if you go back for a second, if you remember on the first slide, the traditional data centers are declining and they will have to get all of the intelligence from the new AI factories, so unless people like ChatGPT and unless the others look hard at getting where the money is going to come from, which is corporations, they're going to fail. And I've got corporations software at the API, I think the APIs are going to be absolutely critical. And that's a very large amount of revenue that's going to come in from APIs. And then equally all the other services that corporations will directly connect to. Those are the areas where the growth is. Consumer and prosumer individuals improving will still be a very big market indeed. And my view of government is going to be very slow in the military, for example, so there's some interesting potential problems worldwide, but they have already been very slow in adopting any new function. For example, I think OpenAI ChatGPT was banned across the whole of the government for the last couple of years and so only just making its appearance there.

Dave Vellante

>> I think it's been in a lot of places, so people are concerned about putting proprietary data up into the... This is one of the... Your premise assumes that people are going to trust OpenAI, so we'll see. But let's take another beat here, David, let's take a pause and bring in Scott Hebner who's going to talk about some of the interoperability standards. Please take it away, Scott Hebner.

Scott Hebner

>> Hi everyone. I'd like to focus on what I consider to be a key success factor for AI factories and especially the generative AI layer that plays such an enormously important role in the whole concept of an AI factory. The key success factor is the adoption of open standards and protocols. About 25 years ago at the start of the internet era, if you remember back then we got caught up in what was known as the browser wars, Netscape, Internet Explorer, later Mozilla, so on and so forth. And when you look back at it, you know what happened, they standardized out. Today we're in the middle of the same kind of gen AI wars, and they're definitely dominating the discourse ChatGPT, Claude, Gemini, Grok, so on and so forth. But if you really think about it, they're both just gateways to something much bigger. The browsers were the gateway to the world of the internet and generative AI is the gateway into the world of AI. The real value doesn't lie in that gateway. It comes from what passes through it. And to pass through them and to be widely consumed and integrated into businesses, they need to be open. As history has shown, if you think about it, the browsers eventually succumb to open standards and protocols. And that's what made the internet ultimately accessible and scalable and brought on the whole transformation around e-business. Now, while generative AI is a gateway to AI, it's the AI factories that promise to power what flows through them. And that's the back end of the emerging intelligence economy, and that's what's really going to power this next phase of digital business transformation is these factories. Because most companies can't afford to do everything they need to do with AI. And there needs to be some standardized way to consume not only the AI capabilities, but also just as importantly the power that comes along with it. My view is, without open standards and protocols these factories risk becoming early mistakes of the internet era. You get fragmentation, you get lock-in, you get inefficiencies, you get untrusted outcomes. And our agentic AI futures index that we just released backs this up. If you take a look at the chart and bring that up, Alex, if you could, you can see a majority of enterprises are prioritizing open models, protocols and APIs, extensible adapters. They want to connect to things. They want interoperability among agents. They want to choose their LLMs, they want the ability to compose any agent into any workflow. It all comes around open standards. And this is clearly telling you that enterprises don't want to get locked in, they want choice, flexibility, and the ability to interconnect things. I think of it this way, just like our electricity grid, it's standardized circuits, breakers, voltage regulators, fuses, things of that nature that make it safe and easy to consume across a whole different ride of devices such as in AI world, agents or businesses, which could be the concept of a home. I think these open protocols play an absolutely critical role in allowing the intelligence pipelines to flow safely, reliably into the agents, the applications, and the enterprises that are so dependent on these things being successful. If you think about it, step back a little bit, open standards are what make the intelligence pipeline an actual pipeline, not a series of disconnected pipes that don't work together. What good is that going to be if you really think about it from a business perspective? I think this new agentic control plane that AI factories can actually deliver, where agents can communicate and coordinate and do things together and really bring to life this service as a software model for a next generation of applications and value, it's all dependent on that. And the good news I think, is that if you look at Nvidia and AMV and Intel and others, they're taking essential steps towards open standards. I think the bad news is there are partial measures. Most of today's AI factories remain semi-open at best. If you learned anything from the early e-business area, innovation compounds faster when you have leading vendors that go full stop in open standards. And I hope these companies are learning the lesson of history. To summarize all this, if you think about it, and as many learned back in the internet days, you can eat 95% health food, but if you eat 5% rat poison, you still die. You need to go full stop in open standards and without it we risk having these AI factories die a slow death over time. That's my two cents on this. I think the AI factories are a critical part of the long-term success of how businesses are going to consume and drive ROI from AI. And again, let's not forget the importance of openness. With that I'll turn it back to you guys.

Dave Vellante

>> Okay, thank you, Scott. Now, David, this is interesting because Scott, here he is, he surveyed a bunch of folks from traditional enterprises, and of course they're going to say these interoperability standards and open models and protocols and these matter, that they're critical because this is the lifeblood of how they get work done today. But your premise, again, going back to some of the earlier comments that we were discussing is that that all gets abstracted away. And the software companies like OpenAI and these leading AI model companies and potentially the cloud guys, they're not going to sit still here, have strapped that complexity away to compete. And we're limited on time here. I'm sorry to push you, but very quickly, what's the timeframe for that abstraction layer to actually emerge where OpenAI actually starts to show that kind of progress based off, I recall based on your earlier charts, we're talking about 27, 28, that really is in full swing when they become a 40, $50 billion company.

>> Well, we're seeing that already with the services that are provided to developers. There's a very rich set of services that have gone in already. And then we are seeing the beginning of those services with all the features included in, for example, OpenAI. And so they're putting in all of the building blocks already. Now, what's very important, I think very, very important indeed is that OpenAI put in the governance capabilities that corporations will need. But my simple point is this. This is not possible for individual corporations to do on their own. They have to have these huge data centers providing these services to them. And if an OpenAI fail, then it leaves the door wide open for somebody else to take over. I personally believe they won't fail. I think they've got really excellent leadership there, but they could go wrong. But the data has to get through them to the end users. And the only way it's going to be done is through simplicity and providing that wrap around it so the corporations can use it easily.

Dave Vellante

>> Excellent. Let's now take a look at the developer angle on AI factories and bring in Paul Nashawaty. Paul, please go.

Paul Nashawaty

>> Hey, thanks for having me on today. This is really exciting, and especially in the world of application development. When I think of AI factories in the data center of the future, today, we're really exploring one of the most transformative shifts in the technology landscape. The convergence of accelerating computing, application modernization, platform engineering, and DevSecOps is really a big factor, especially when I think of the CI/CD pipeline and the overall SDLC. But this isn't just another wave of technology changes. It's really strategically redefining and how software is being developed, secured, and delivered across every industry, whether it's compliance regulations or governance, it really takes a factor here. When we enter this new era of where performance, automation and intelligence are no longer optional, they really do become foundational. Let's take a look at, I have a slide here I would like to start on. And then in this slide I just really want to highlight that there's five segments here. And when we think about the first part, which we think about accelerated computing, this is leveraging specialized GPUs, TPUs, FPGAs, et cetera, that are unlocking computational capabilities that what really once thought is impossible on traditional architectures. And this is enabling breakthrough areas like real-time analytics, AI-driven decision making, and large-scale simulation. And according to my research from our AppDev summit that we did in September of this year, the research shows that cloud infrastructure, GPU-based acceleration market is growing nearly 47% in 2025, reaching $157 billion in value, driven by AI architectures and infrastructure investments that are growing on demand for high-performance computing. And so for developers this creates a really exciting challenge and it's really an opportunity. And when we look at it, we now have the ability to build smarter, more responsive applications that can demand new programming models, tools, and optimize for parallel architectures. In short, the accelerated computing part is really just speeding up things and it's really helping redefine what's possible next. The second part I want to talk about is application modernization and platform engineering. This is a big area that I think gets overlooked a lot. Modernization is a treadmill, it happens. It's a big market though. When we think about the service market here is valued about $19.8 billion in 2024, and it's expected to double to $39.6 billion by 2029. That's a big jump. And so this surge reflects how the enterprises are really re-architecting legacy systems. They're moving to cloud-native environments and embedding continuous delivery pipelines. But modernization is not enough alone. It really has to work with truly scaling innovation. And platform engineering really does become central, so platform teams are now tasked with building robust self-service environments that use infrastructure as code, container orchestration and advanced observability frameworks. This is really critical to the next generation of modernization. In 2025 the focus has shifted from simply providing infrastructure to delivering complete developer platforms that really enhance productivity and improve developer experience. Let's look at the numbers though that show this. In a research study that we did with Google Cloud shows that organization really has shown co-managed platforms give developers 47% of their production time to innovation and experimentation compared to 38% of companies managing platforms internally, so internal development portals and self-service capabilities are streamlining workflows, this is reducing friction and accelerating delivery. Platform engineering is no longer a support function, it really is a competitive advantage and a growth engine for a lot of organizations. And when we look at DevSecOps in the next part of this, what we really understand here is application scale and complexities, increased security also must have evolved around it, so that's where DevSecOps really does come in. And the global DevSecOps market is projected to hit $41 billion by 2030 and growing more than 30% annually, so there's a lot happening here. This is, not to reiterate a lot of things you're probably already hearing in the market, but the key trends that are shaping the evolution include that shifting left for security testing early in the CI/CD pipeline stops vulnerability before production, AI-driven automation. As we're moving into AI-driven automation, I'm still a firm believer of a human in the loop, but it's really meant to detect and respond to real-time threats. And zero-trust architecture is really continuously verifying every user and device across every environment, so DevSecOps is a big part of this. And it's not just a checkpoint, it's a continuous part of the SDLC, so that's really an important part. The other part I want to tie into is service as a software here, service as a software is that all the trends are converging into something bigger here. And we look at this new business model as services as a software. This model helps software become more a primary vehicle for delivering value experiences and services to customers. And when we look at that across these AI factories, enterprises are adopting this model by using agentic process-centric applications to accelerate innovation, optimize operations, as well as open entirely new revenue streams, so there's a lot of things happening here and those organizations that are delaying, they risk falling behind in the marketplace because defining by speed, intelligence and scale is really the new competitive advantage. In closing, what I would recommend here based on my research that we saw in the AppDev space, the analysis concludes that this converges accelerated compute, modernization, platform engineering and DevSecOps is fundamentally reshaping the software world. Where do you get started? Well, I'm thinking about this in the context of assess readiness by providing audit current... You have to look at it in audit current applications and infrastructures that are ready for modernization acceleration. This doesn't mean everything gets refactored. This means the things that need to be refactored are refactored. You want to invest in platform engineering. I can't emphasize that enough. It's a big part of the infrastructure. It's a big part of the internal build for developer portals and also that self-service model for environments to empower that innovation. You want to not slow down innovation here. And then of course, security is key. Embed DevSecOps that's integrated in the security automation throughout every stage of the development life cycle. Those companies that really embrace this shift will really lead the next generation of secure, scalable, and intelligent applications driving competitive advantage across these digital initiatives that are happening out there. With that, Dave, back to you. This is really exciting stuff. I think from the AppDev world, I couldn't be any more on point.

Dave Vellante

>> All right, thank you, Paul. Hey, let's wrap. Let's bring up the very last slide if you would, Alex. This is a slide from our friends at ETR. It basically shows net score or spending momentum on the vertical axis and overlap or penetration into a data set of almost 1800 respondents on the horizontal. And we just took company names, but this isn't any particular sector, it's all sectors. But some of the players that we see here, of course you've got Microsoft, you've got AWS, you've got your Google, OpenAI... I forgot to put Anthropic on this chart, apologies to those guys, I should have done that. You got Meta, you got the networking guys, Arista, Cisco, Juniper, and HPE of course the semiconductor guys with Intel and Broadcom, AMD, and of course Nvidia on the top of the heap here. And you got Uncle Sam and China, as David was pointing out before. You've also got some of the Neo cloud players, CoreWeave, Lambda, Nebius, other data center builders are going to be in there. And so energy companies, this ecosystem is massive. This is a multi, multi-trillion dollar market opportunity. Thank you Alex. And David, so we got to run, but I'll give you last word. Bring us home.

>> These are all the people that are contributing tremendously to this dramatic change. I think from a corporate point of view, they're going to be focusing on the people providing the services. They're going to enhance their own data centers with a high degree parallelism within the data centers, so people like Nvidia are going to be important, but the most important partners they're going to have are the people that provide them with the APIs and the software providers who are using those APIs. Those are the two sets of people in my view who are going to turn this fantastic operation of providing intelligence, turn it into a business profit.

Dave Vellante

>> Okay, we shall see. Interesting take, David. I really appreciate it. We got it wrapped there. Thanks, Alex Morrison and on production and manages our podcast, Ken Schiffman as well. Kristen Martin and Cheryl Knight, they help get the word out on social media and our newsletters. And Rob Hof is our EIC over at siliconangle.com. Thank you all. Remember all these episodes are available as podcasts, just search Breaking Analysis podcast, we publish each week on Thecubereserch.com and siliconangle.com. And you can email me at david.vellante@siliconangle.com. If you've got an idea, you've got a guest you want to bring on or email or DM me at dvellante, comment on our LinkedIn post, check out etr.ai for the best survey data in the enterprise tech business. This is Dave Vellante for theCube Research Insights. And David Floyer, Scott Hebner, Bob Laliberte, and Paul Nashawaty, powered by ETR. Thanks for watching and we'll see you next time on Breaking Analysis.