The Cloud Native Stack Has Quietly Become the AI Stack
The cloud native stack took about 15 years to build. Now it is the foundation for AI. Peel back almost any AI platform, and Kubernetes sits underneath.
SUSE’s Cameron Seader and author John Willis examine that shift with Alan Shimel at KubeCon + CloudNativeCon in Salt Lake City.
Seader leads solution architecture for SUSE in North America. Willis co-authored “The DevOps Handbook” and wrote “Rebels of Reason” on the history of AI.
Why Kubernetes Won
Few people would have bet on Kubernetes in the early orchestrator wars. Docker Swarm and Rancher had strong followings. Willis, who sold a company to Docker, admits his horse lost.
He says Kubernetes won on scale. It grew out of Google’s Borg and Omega systems. That heritage now powers service meshes, gateways and AI infrastructure.
Willis once called Kubernetes over-engineered. Today, AI coding tools handle much of the configuration that made it so hard.
Seader points to developer simplicity. Tools such as Docker Desktop and Rancher Desktop let developers pull AI components straight from a registry.
What Comes Next for the Cloud Native Stack
Seader expects new schedulers, sovereignty features and stronger security. Kata Containers is getting fresh attention as teams look for ways to contain agents.
Willis credits the operator model. It lets teams inject policy, kill switches and escalation rules for agentic behavior. Neoclouds also run on Kubernetes, though Shimel warns that GPU scarcity is not a moat.
Data Centers, Edge and Governance
Audience questions turned to data center build-outs and possible overbuilding. The panel expects human creativity to win. Small language models, CPU-bound acceleration and edge computing all reduce the pressure.
Seader says managing thousands of clusters is the next challenge. SUSE is adding Liz, its AI assistant, to Rancher with human-in-the-loop validation.
Willis closes with advice on governance. Organizations should approach internal auditors about agents before auditors come asking. The cloud native stack can enforce those controls at scale.
Transcript
So our first panel is the cloud native stack is the AI stack, and I'm happy to be joined with me on this panel with Cameron Cedar of SUSE. Cameron? Is that right?
Yes. Sir. Tell people a little bit about yourself.
Thank you. I love that. Yes, I lead the solution architecture organization in North America for SUSE.
So I've been with SUSE for, what? Let's see, 21 and a half years now. So, a long time.
I've seen SUSE grow up in many different ways. So it's been very exciting. Yeah, I'm very happy to be here, and hopefully I can share some really great insights for you all.
Thank you. Thank you, Cameron. Yes, sir.
" John's written, I don't know, 7, 70, I forget how many books he- Yeah ... including "DevOps Handbook," "Rebels of Reason" about the history of AI. His newest book is, well, are we allowed to say yet?
Oh. His newest book that he's working on is on quantum computing, and John and I actually do a podcast on quantum computing every two weeks or so. So you want to catch that.
But John, welcome. Yeah. I just don't want to sound stupid.
Yeah, well, if I'm on the panel, you're not going to sound stupid. So there you go. I get beat out, that's all.
You don't have to outrun the bear, you just have to run faster than Allen. Okay. So, I wrote a special report on Cloud Native now about two months ago that the AI stack is the cloud native stack.
And it's a funny thing. We spent about 15 years building the cloud native stack, and I look out over this audience and I'm not saying anyone here is old, but most of you were around when cloud native, and Kubernetes, and Docker first started coming out, right? And if you remember, Docker was like a breath of fresh air for developers.
Like, oh my god, this is a game-changing kind of thing of how we're going to... Not just developers, in some ways ops too, right? But this is going to change the game for us.
And immediately, people started recognizing that, you know what, we're going to need something that kind of manages all these different containers, an orchestrator. And if I would've run Calci back then, there was no Calci back then. " I got to tell you that Kubernetes probably would've been in the bottom, in the bottom five or bottom- You really felt like Swarm was there?
Swarm was better. I think Swarm, we're not going to get into a fight about it, but a lot of people would've bet on Docker Swarm. A lot of people would've bet on what Rancher was doing.
Sal and Mike bet on Swarm. Yes. Well, John should mention, he sold the company to Docker, so he had a horse in that race.
But no- And the horse lost. Yeah. But SUSE, Rancher had a great- They did ...
alternative. Absolutely. Right?
And then, I remember coming back from DockerCon in, it was in Austin. And I came back, and I told the team, because we were looking to launch what today is Cloud Native now. " At the time I was trying to put Docker in the name, but they had all that locked up.
I said, "I don't know if we should pick on Docker. There's this thing called Kubernetes. Everybody's talking about it, and it's hard as s**t.
" But everyone- You know what I remember about DockerCon? What? Just sitting there in the audience and watching these executives show up on stage, and they're wearing the jackets, but they have T-shirts underneath.
That's the first time I saw that. " That was a cool look, though. I kind of went with that.
But anyway, look, here we are 15 years later. Kubernetes is dominant still. So I'm surprised that CNCF lets me come to KubeCon.
About, I don't know, about four years ago, I made this statement, not that everybody listens to me anyway, but I said that KubeCon is like 4, 5, 8,000 people running around with their hedge laws trying to figure out where to put semicolons and colons. And the truth of the matter is, back to the Swarm, not to get in an argument, but Kubernetes was completely over-engineered. I've been waiting for years for somebody to replace it.
I know this is crazy to talk about at KubeCon. But the interesting thing was, now we don't have to replace it. Because all those semicolons and colons, we don't have to worry about anymore, right?
We've got clear code. And it now is the fabric of AI, which is probably what gets you- And that's the important thing. Fast-forward, here comes AI, and we need a stack to build it on.
Why the cloud native stack, Cameron? Well, I mean, for one, if you look right and left, right, you're looking at from Kubernetes from the ground up, and even going back to OpenStack, God forbid. If you look from the ground up where it's instantiated from, we're talking cloud native, the infrastructure needs to be, all the tools and everything need to be instantiated from that infrastructure.
And if you even look at a developer laptop today, we're looking at Docker Desktop, you're looking at Rancher Desktop. Those tools are fundamental. They're spinning up anything that's container related.
You can spin up Kubernetes localized. You can even test all this stuff out localized. So why cloud native?
It's because it makes it easy. The developers go and grab their tools, their apps, everything needed for AI right from a registry, right? They can plug that directly into Docker Desktop, Rancher Desktop.
It makes it very, very easy. And we can go further and further down the road why that makes sense from a registry perspective and why containers, but it just makes it easy. Yeah, no, I remember when we launched Swarm and we basically, well, we bought Swarm, but at Docker, and then we combined the engine with Swarm, and we thought we were going to win the whole game, but at the end of the day, we lost because of scale.
Because as a historian, I like to write books about the history of technology, which is that it was always a scale problem or a scale opportunity. If you know where Kubernetes came from, it originally came for how Google was running internal containers, and so it was Borg and became Omega, and then it got released as an open source project called Kubernetes. So it had a heritage of massive scalable infrastructure around containers.
And so nobody else had a chance except the Rancher guys were pretty smart, right? Because they scaffolded around it. No, and I love those guys, but- They pivoted right into it.
Yeah. But it's always been a scale. And it's winning today.
Back to why is it winning AI? Well, the service mesh. The service mesh is now making a grand run at the gateway stuff.
Why? Because it's got the scale, it's got the time. All the stuff we did with containers and all that, so the egress, ingress, all the things we can do in the service mesh, Istio, Envoy, that stuff, right?
It may not be the best game, but it is kind of the only game if we want to scale this stuff. Yeah. Scale is definitely important.
If you think about from a scale perspective, even in high performance computing, where's that going now, right? You're adding on new schedulers that are allowing you to scale that even further for high-performance computing. Look at what NVIDIA's done with grabbing onto the Slinky toolkit, right?
And now you have scalability for high-performance computing with AI. So a bunch of things that are enabling that to- I mean, Dynamo is Kubernetes. It's the infrastructure.
It's every vendor. I do the tech field day, and every vendor, you sort of peel the onion, and everyone has basically got a Kubernetes fabric hidden somewhere in their structure. That's right.
So here we are today. Clearly we- AI is being built on top of this cloud-native stack. But things are happening so quickly, things are changing so quickly.
I'm not asking you to look three years out, because who the hell can look three years out today? But what do you see on the horizon? Does anything upset this apple cart, or is this locked and loaded at this point?
I think you're going to continue to see innovations, especially around the scheduler in Kubernetes. You're going to see some add-ons to that particular engine to help drive things faster, to help improve things, and tune things. And there's already tuning going on today in the Kubernetes stack and different add-ons that we're adding in there to make agentic AI better, and make that move faster and make it more portable.
There's a number of things there that are going to happen, I think. Even when you talk about sovereignty, which I think is a big thing when you talk Kubernetes. And as you look at sovereignty in the Kubernetes world and the security around all of that, you're going to see a lot of improvements there.
We're already seeing some of that happen right now, and especially with the containerization around Kata Containers. That's going to continually be a thing. It's going to improve.
That's been improving over the last decade, honestly. Kata Containers has been around for a while, and we're coming right back around to it. Why?
Because of security. Because of agents. Because of agents getting out.
Yeah. No, and I think anybody who answers a question that says they know what it's going to look like in three years is a liar, right? But no, I think Kata Containers a really good point, right?
Because right now, containment is the big issue with agents. And by the way, probably running a Docker container with agents is not a great idea right? There's a lot of escape routes in a Docker container.
Kata Containers, or even some of the OS primitives like Lighthouse and stuff like that, there's a nano project, the guy who wrote Six Store, Luke Hinds, he's got this project called Nono, right? And it's an implementation of a Linux. But Kata Containers, that's a great example because that's built into the fabric of a Kubernetes scheduler.
And one last thing as an historian is, where Kubernetes has gotten it right every time is, the operator model, right? Which became part of the-- That's what enables us to scale. Because now we can scale not only things like Kata Containers or new brands of containment, we can inject policy into the structure.
We can actually inject primitives for agentic behavior, like kill switches and sort of escalation policies into the- Right ... yeah. So yeah, again, I kind of wish somebody would replace Kubernetes, be honest with you.
But I don't think it's going to happen in my lifetime. Well, you bring up a good point in terms of policy and all that. There's this idea of creating blueprints, right?
Blueprints incorporate all that policy, the security, everything you need to actually run your AI stack on top of Kubernetes, all defined, right? That's powerful right there. That'll continue to be refined more and more.
No. So I almost feel obligated that, as I mentioned earlier, unfortunately, Nvidia couldn't be with us today. But the point I wanted them here to make was that, hey, they're building their stack on this cloud-native stack, right?
And they have made a big commitment to Linux Foundation, to CNCF, I think it's the AI Foundation as well, throwing in building their SPARC and everything on top of it. So again, as more and more companies buy into this is the AI stack, is the cloud-native stack, it becomes harder and harder for an outlier to come in here and kind of make a meaningful move on that territory, right? And so when you look at the leading AI companies and the model companies working on Kubernetes, you got to ask yourself, what would it take to displace them?
Yeah. I think there's no common sense to try to replace it, right? Even like, I know they're not doing great financially, but CoreWeave, to me, is this interesting Neo Cloud, right?
And they don't want to be called a Neo Cloud, right? That's fine with me. But it's Kubernetes under the covers, right?
Well, it's Dynamo, right? Yeah. And we recently saw a shout-out Techfield Day DDN, right?
By the way, if you peel the onion and ask some deep questions, it's Kubernetes, right? This makes sense. It's in the fabric of how we operate at scale because who cares if it's not scale, right?
Absolutely. Cameron? Yeah.
You brought up Neo Clouds. I've been looking for an AI expert to hire, actually, and so if you know any out there. Neo Clouds are really important in this space right now because they're all using Kubernetes.
They're all trying to get to that really high performance for AI. And so these Neo Clouds are popping up all over the place. Why?
Because can you even get a GPU out there? It's clear out to 2028 right now, I think. So it's a ways out before you can actually get your own hardware to run on premise.
And I think you got to be pretty special to even get some out there. So these Neo Clouds are very important in this Kubernetes space right now. I'm glad you brought that up.
So I wrote a book called "The Indispensability Trap," and I'll have some copies down on the floor tomorrow. I think Neo Clouds will be fine as long as GPUs are scarce. Sure.
But scarcity is not a moat to build a business around. No. And that's why I think to John's point about CoreWeave, you got to look at what they're doing beyond the- Yeah ...
scarcity of GPUs. Yeah. And again, from an investment standpoint, I don't want to go near them.
But one other point I want to make back to this fabric discussion is that, sometimes I talk about people talk about the miracle of a model. So the miracle of OpenAI or Anthropic, and whether you think they're miracles or not is another discussion, but the miracle isn't the model. It's the infrastructure that they've built to get that data ingress to the model.
Because guess what's at the core of it? Kubernetes. What's wrapped around that?
Kafka. Variants of Redis. They all have the same-- I've never worked there, but I know people who have, right?
So it is that fabric that is just, at this point, kind of irreplaceable. Yeah. Hey, I wanted to leave a couple of minutes for questions, though.
It's got a little technical. I like that. I like having it a little technical.
Anybody have any questions for John and Cameron? And I'll pass the mic over to you if you do. Anybody?
We got one over here. Yes. What's a Neo Cloud?
So Neo Cloud is a cloud provider that has essentially spun up. Just think of them as a third party, I would say, cloud provider that has spun up that is in the Kubernetes space that we're calling a Neo Cloud. Yeah, and I guess the only way I knew what they were is because a lot of them are providing accelerators.
So like the CoreWeave- Right ... and a couple other ones, they've gotten onto my radar because they have access to GPUs for large organizations that don't have capability. But in general, they're just smaller clouds.
But the thing that's making them very popular right now, again, is access to accelerators, right? Does that make sense? Something that's not a hyperscaler today or identified as the top- Not a hyperscaler.
So this may not be a fair question to ask, but since we were talking about scarcity not being a moat, and I'm curious with the elections and kind of pushback against data center build-out, the availability of land, power, water to go build these things out, do you see a slowdown, or do you see creativity taking place in a lot of the REITs that built out abandoned malls for rack space to build data centers, things like that? Are we going to get creative, or are we going to slow down? I'm going to take the easy route out, and it's the creative.
And I think there's some interesting discussions on CPU bound replacements for accelerator stuff. So, and I'm hearing those pop up periodically, well, certainly small language model stuff, and the barrier to entry training models has been obliterated. Like anybody...
If I can train a model, anybody can. And then I think there was some interesting discussions recently about some examples of nothing proprietary, but DoD is looking at CPU bound acceleration. So, that could be an interesting area for creativity.
The other question part of it, I have no clue. I certainly think there is a lot of creativity happening. Absolutely.
There's a lot of friction, in terms of data center buildouts. There's a lot of challenges with that. There's a lot of data centers being built right now.
And I know that there are several that have run out of money, and then other investors have picked those up. And like Google and some of these other bigger hyperscalers, have picked up those projects and continued on with those. But I would also say that look right and left to the chip manufacturers.
There's a lot going on there. They're really trying to figure out, and be creative themselves as well. Look at Marvell.
Look at what's going on with Qualcomm. You see the numbers dropping for their revenue, but they might be doing something behind the scenes there we don't know about. I would love to see them pick up some server level chips as well in the Arm space.
I would love to see that. But, I think they are going to get creative. Arm has been creative as well.
But Intel, some would say, oh, well, Intel's kind of been a little bit behind. But I don't think so. I think there's some sleeping giants here that I think might come out of the woodwork, and we might see some interesting things happen.
So... I just want to jump in on that. So again, this is right out of my book.
I don't think we ever needed $8 trillion in data centers. Which is what we were going to spend in the next four years. Let me come up because I'm not on video.
So you got $8 trillion in data centers running in the face of the creativity of humans. The creativity of humans wins every time. It was never going to be that we're going to have to run every single AI query back to some frontier model in some factory of 40,000 acres in some God-forsaken land somewhere.
Now, that being said, data centers aren't going away. I think, Robert, to your question, the real point of this election was not that data centers are bad, but that local people want to be involved in local decisions for local data centers that take place in their neighborhood. But you go to Northeast Louisiana where they build HyperOn, you got teachers there who got $50,000 bonuses from the data center.
You know what it is for someone in Northeast Louisiana to get a $50,000 bonus? That's a lot. That's like life-changing money for them.
Hey. So, I do think we're going to have data centers. We may not have the $8 trillion worth of factories.
We may run a lot of small language models locally in the closet, on your phone, on your laptop. Absolutely. And John, just I've got to get my quantum story in.
But there's an interesting story about creativity. In the quantum space, IBM did this interesting thing. They did a protein simulation of a 1,200 atom simulation, and the non-scientific Von Neumann architecture computation has probably 10,000 years, best case, right?
And even with a quantum computer that actually works and does the error correction and all that stuff, which could be five or 10 years away, it's still probably 50, 60 years to do this computation. But the creativity of this, and this is a signpost of how we humans get creative, to Alan's point, is they broke the computation up into three portions of math. They took the CPU bound math, the GPU bound math, and the QPO bound math, and they were able to solve the problem in eight hours or something like that.
And there's that interesting idea about humans can get very creative, because a binary look at this computation would've been, even with a fault-tolerant, capable qubit machine, it still would've been probably 40, 50 years. So... Absolutely.
I think another area that's ripe for innovation there is the edge. Yeah. Where you're kind of hearing about people talking about moving these smaller data centers into neighborhoods, and what that looks like and how that could be a possibility.
I live in the sticks. I live two hours north of Boise, Idaho, in a little town called Donnelly, Idaho, that has 600 people. And I have 5G fiber internet to my house out in a little town.
That's amazing. So I really feel like there are possibilities to bring edge technologies with Kubernetes, with a small infrastructure that can run some really powerful AI capabilities there at the edge. So I think we will see more of that happening.
Okay. I wanted to talk about overbuild, right? So we're talking about infrastructure.
Alan mentioned $1 trillion in spend. We're talking about data centers. Right now, people can't get enough, but I believe there's going to be tears before bedtime, and I think that we're going to find ourselves with huge CapEx exposure, which is then going to domino.
This isn't a unique view that I have, but if you have too much Kubernetes, no big deal, you don't have the big exposure. If you've got data centers as far as the eye can see with circular financing, that's kind of scary if things don't go exactly right. And I have more thoughts, but I'd like to hear what you all think.
I agree. We lived through that in fiber, too, right? Go ahead, guys.
You can answer. I certainly agree. I think it comes down to also how do you manage that infrastructure?
Yeah, there's lots of Kubernetes. How do you manage it all? SUSE certainly has some tools when it comes to Rancher.
That's what we've been all about all along is managing Kubernetes, multiple Kubernetes clusters. Thousands, hundreds of thousands of clusters. And that kind of scalability is something that we need to harness.
We need to harness not just the management of it, but the automation of it, utilizing AI and that tool set to actually manage all of that infrastructure. And that's what we're all about here at SUSE and figuring that out. We have tools today that are certainly positioned to doing that today.
We're going to continue enhancing that. Yeah, and I remember sitting in an analyst meeting, five, six years ago, maybe it was longer than that, where they talked about, somebody brought up how they were going to run Kubernetes clusters in each one of the Chick-fil-A outlets. And I thought, what is the person who creates the fries, is he going to take a call, reboot Kubernetes, and go back to the fries?
To my surprise, four or five years later, I ran into the guy who ran that project, and they actually do it. So... There you go.
Hey, last question, then we've got to go to the next panel. Thank you. Yeah, you just mentioned you guys are looking into AI automated, or AI management in Kubernetes.
How are you guys currently validating the actions and eventually tying that into scale and stuff? It's validation. So we have a tool in our Rancher stack right now.
It's called Liz. That's what we're calling it. It's our integrated AI platform inside of Rancher.
And so Liz is doing some validation based on common frameworks that we know and understand already. " for example. And we're still working on some of that.
We haven't gotten to the point where we're perfect at it yet. So we're still working on that validation. " It was before AI.
And we talked about how we create an audit, and we came up with a-- I won't bore you with the long version, but we came up with digitally signed attestations that were immutable. We used Sigstore. We did a whole bunch of cool stuff.
And where that stemmed from was, what are the auditors, what were the auditors asking at the time? The internal auditors, right? And we wanted to fix that.
Well, right now, the internal auditors, and maybe Ricky can validate this for me, is they don't know what to ask yet about agentics, but that freight train is coming down the pike, and once they figure out how to ask the questions about, how did you contain an agent? Did you have a kill switch? Did you have...
And I don't know what the answers are, but I'm trying to get that old group back together to talk about... And we didn't create a product. We created a book, we created a paper, and then we created a book, and that turned into some companies that have product.
But I think that's a legitimate conversation to have right now, and it starts your internal audit. Yes. And as I pontificate, one last thing.
The biggest mistake an organization can do is wait for the auditors to find out. You have to go to the auditors and tell them ahead of time because it's just not going to work the other way. We've learned that the hard way.
So what organizations should be doing is going to internal auditors and saying, "Hey, we're thinking around agentics. I know you don't know what it is, or we don't really understand it. These are the things I think are going to happen.
" And that's what it's all about. " Go check it out. Self-sovereignty.
Think about what that means in terms of AI. Self-sovereignty, you need encryption both ways, right? Your agents don't need to understand what the data is or what it looks like.
They don't. We need to figure that out. We need to figure that problem out.
There are use cases in healthcare. There are use cases in fintech where this is going to come right back around to being very, very important. I know you are, Chris.
And that's absolutely true, but the other problem is the whole rogue agent thing. In other words, what do you do when the inference is the task, which is an inference decision that is making these decisions, and that's going to be a hard problem to solve. And so we need, as an industry, to start having that honest conversation about how do you even think about a rogue agent in a bank?
Well, we saw this, DoD, what was it? Last week, right? The Pentagon, somebody literally ran code connected to an account that had clearance.
Oh, God. Yeah. No, trust me, this is happening in your organization right now.
Guys- Shadow AI. Yes. You know what?
We could just do one panel for the rest of the morning, but can't do that. We got other people here. Cameron, thank you so much.
You're welcome. Thank you for having us. John Willis, as always.
Hey, what a great kickoff. What a great kickoff today.