Industrial AI at the Edge: How Red Hat & Edgescale AI Are Making It Work

Speaker A: Foreign.

Speaker B: From Boston, Massachusetts. The ARC Digital Transformation Viewpoints podcast is the only podcast dedicated to all things related to digital transformation in energy, industrial and critical infrastructure applications. The podcast is the creation of the ARC Advisory Group Digital Transformation Practice. ARC advises leading companies on technology trends and market dynamics that affect their business business. To engage further, please like and share our podcasts or reach out directly on Twitter @arcadvisory or please go to the website at www.arcweb.com.

Speaker C: hello and welcome to another episode of the ARC Advisory Group Digital Transformation Podcast. I'm Colin Masson, Director of Research for Industrial AI. If you follow my research. You know, we talk a lot about the growing intelligence divide. And according to our brand new industrial AI pacesetters report for 2026, only the top 13% of industrial organizations are truly leading the pack today. And I think we were generous at making the cut at 13%. Uh, these uh, paces have uh, stopped treating AI as a siloed IT experiment and have weaponized it as a core OPER strategy. Meanwhile, the mainstream majority and the laggards remain stuck in what we call pilot purgatory. And our research data tells us exactly why this is happening. Our, um, just slightly earlier Q4 2025 industrial AI robotics and energy survey staggering 63% of respondents stated that decoupling data from monolithic software app applications is critically important to achieving their AI objectives. Yet integrating the OT data with modern AI frameworks remains a massive complex hurdle. Today we're going to talk about the heavy lifting required to cross that divide. I'm joined by two guests who are actively solving the OT integration nightmare that keeps so many AI projects grounded. Welcome to Brian Mengwasser, CEO of Edge Scale AI, and Cole Wangsnes from Red Hat. Gentlemen, welcome to the show. Brian, perhaps you can tell us a bit about Edge Scale AI first. And I said you're the CEO, but maybe you fulfill multiple roles, so why don't you uh, introduce yourself first?

Speaker D: Absolutely. Thanks Colin. And uh, thanks for the invitation. It's a pleasure to. Pleasure to be here on the show with you. As you mentioned, I'm the um, CEO, co founder as well of the company, which uh, means, uh, whatever needs to be done to uh, help overcome this gap that we see holding back the industry. We're a physical AI company which has, uh, set ourselves a mission, uh, to bring AI to operators everywhere in industry so that we can benefit from the massive productivity gains that we're seeing in AI in practical, uh, real world environments.

Speaker C: Great. And then over to you, Cole.

Speaker A: Yeah, thanks for having Me, Colin. So I assume a lot of the listeners here are going to know of Red Hat, but really where I sit and where my team sits within Red Hat is our edge group. So it really breaks down into three P's is what I like to call it. The first being like, how do we take the enterprise IT products and have deployable patterns, repeatable solutions that make it easy for people to consume at the edge? Partners like Brian that make it easier for those then customers to consume the bits and pieces maybe they don't understand. And the last is product enhancements. Right. The core product we're going to need to do different things at the edge. How do I get the engineering teams to do what I need to do?

Speaker C: Okay, great. Well, with that introduction, let's jump in. I think one of the characteristics that we see from the pacesetters in industrial AI is very much that they are deploying to the edge. And in fact we see a, uh, significant swing back to the costs of moving everything to the cloud and, you know, is greater than expected, shall we say. Right. And so, uh, a lot of the leaders are actively working to deploy back down to the edge. Some of the workloads they did move, or if they haven't moved them yet, they're kind of reappraising that. So one of the areas where we saw significant, um, evolution is really with kind of local execution or process optimizers as we call them in our industrial AI models taxonomy. We've kind of looked at what are all the M patterns we're seeing and one of them is process optimizers, which is a boring term. So we've just renamed it autonomous, uh, execution agents. Right. To be in line, I think with where the AI modernization efforts are going now. The software exists. So I was at a vendor last week, right, where we were talking about deep learning, process control that very much needs to run on the edge. And I think that's the real nightmare that people are having. The software exists, but how do you deploy it to the factory floor and how do you do that at scale? Now Brian, I know last time we talked you shared a story with me about an EV battery plant, if I remember correctly, a top gap weld inspection, uh, use case as well. Walk our listeners through why in those cases, a seemingly simple AI vision task, that computer vision is not new. It's a highly trusted technology and it's one of the things we are seeing actually successfully deployed at scale because it is trusted, highly repeatable. So why did it take 18 months to deploy such a. Well Known technology. Right. And the value is there.

Speaker D: Yeah, great, great question, Colin. Firstly, I would, I'd like to echo the commentary at the beginning about the edge is where the data is, right? It's where the actions are being taken. So even if we could sort of like suck all of the meaningful operational insight out of the edge and put it somewhere else, then you have to come back and you know, that's called a hairpin and telecommunication is not a good idea. So we're really focused on how do we make it practical and easy and fast to deploy systems that help us take intelligent actions where we need to take them. And so in this particular example, you know, as you mentioned, we've got enormous software capabilities. Those are only um, let's ah, say growing in sophistication and ease now that we have AI that can help us write and manage software. So it's really not a software problem, uh, in our view it's a deployment problem. Uh, so for this particular customer, you know, they had the ability to uh, collect images of their welds and in a relatively straightforward fashion train a machine learning model to detect whether a weld was good or bad. The problem came in, in the complexity of the environment, to get the model there and to run it repeatedly and reliably. Reliably. The complexity of the stack is only growing when, when you say something like, well, we need to, we need to compute, we need to run it on the network, we want to do it securely. So how do we do things like key management and hardening of the platform? Uh, then there's a GPU which has drivers and other kinds of things that need to be put in place there. And so this 18 month period was just series of like trial and errors. There's nothing really magic there, but there's so many pieces that need to come together well that it takes a long time. And obviously this is a operational environment running 24 7. So it's not like we've got, you know, three weeks of downtime to kind of like work on this. It's all like while we're running the business trying to like add some stuff, you know, bits and pieces here and there. So sometimes I call it a, um, or some people sometimes call it a jigsaw problem, like bringing it together. I like to think of it more like a Tetris problem because, you know, you, you have to get all the pieces working there and then you have to keep doing it like, like anything else that, you know the AI itself is an operational system and so you have to stay on top of it 24 7, just like the rest of your business.

Speaker C: Okay, Carl, bring Red Hat into this. We know hyperscaler clouds are incredible for training these models, but why is it so difficult to push these cloud native AI workloads? So to Brian's point, there are ways to train locally. Right? But right now I would say that obviously the hyperscalers want to ingest all that data and take it to the cloud and, and then you have to work out how do you push the trained models back down to the edge, because that's where they need to. The inference needs to happen. As kind of Brian was laying out, what's Red Hat's role in all of this?

Speaker A: Yeah, no, that's a great question. So I think, kind of alluding to what Brian said, it starts from this difference of what we're talking about when we talk about it and what we talk about when we talk about ot. I think really, if you want to simplify it, obviously there's a number of differences. Obviously it focuses on infrastructure, data management, security, ot primarily thinking about how do I actually keep making this product, keeping uptime of a system. You know, historically disconnected processes mean there's different behaviors for buying stuff. I think that breaks down to kind of what Brian was saying is like, when I talk to IT buyers, they want to buy a product. When I talk to OT buyers, they want to buy a solution, right? They don't want to think about. They're like, I don't know what I want from a key management platform and I don't know how to talk to it about that. I want the vendor, in this case Red Hat and Edge scale, to have an opinion on all these. And I think historically part of this, as we see the emergence of more physical AI between. The disconnect between physical and cloud is a bit of this idea of dysfunctional development processes. So a team, but I see a lot of CTOs doing is they're going to spin up a digital transformation team, sitting that somewhere in it with some level of connection to ot. But the reality is, and I sold AI hardware going back many years now, you know, in previous roles was this idea of, okay, well, we're going to build everything centrally and then we've really not thought about optimizing or what this actually looks like at the edge. And what ends up happening is you try to force that centralized cloud infrastructure to a plant. What you end up seeing is not only issues with like latency data security, all those, you know, typical things you might see when you try to cram a lot of stuff at edge, where you need maybe a mix between cloud and edge. But the reality is like sometimes it's just a matter of space. I've been to steel plants where the idea of the IT or the data center room, whatever you want to call it, the local zone, there was like a uh, in window AC cooler and they had a big like industrial shop fan on the floor, pointed at the racks. And that was the secure server room that had like some guy in it, had a key. So it's almost like there's a more of an organizational or institutional disconnect where I think we've seen a lot of progress in the last couple years. But the reality is you kind of have to do both. You have to satisfy ot, you have to build with the idea of uptime and that, you know, we're not going to spend a million dollars on bandwidth, you know, sending this up to a, uh, to the cloud when it comes to a lot of the data operationally, um, we're not going to rip out the process control systems, but we need to adapt some of the cloud native IT practices, but in such a way. And that goes kind of back to during my intro, it goes back in a way that we need to tailor them for these experiences. People who are used to saying if I need to stamp sheet metal, I go buy a stamp sheet metal machine. I don't go buy all the parts and build them myself. So I think it's adopting this kind of ethos. When you think about how do I make my IT stakeholders, how do I make my ot and then these digital transformation people in the middle, how do I make them all happy in a synergistic fashion?

Speaker C: Yeah, look, I think we have made progress in IT ot, if not convergence, collaboration. Right. I think we've seen some modernization of at least the idea. Some of the ideas that IT have uh, been using for some time, like containerization, uh, have now become kind of standard approaches, practices that are used both on the cloud and the edge. So I think that there's a, as you were talking through it, and I don't disagree with anything you said Kael, but I do think that's part of Red Hat's um, proposition here. Right. Is that you can take a lot of that cloud native advancements and uh, some of the things that are accepted as ah, new best practices in the OT domain and you can enable them to be taken to the edge. Right. Am I oversimplifying it?

Speaker A: No, I think you're totally right. Right. It goes Back to saying, let's help people understand where there may be, you know, there's an understanding of it, but there's not always a, there's, there's this skill gap, um, in training. So I think it really, you know, the industry is so diverse. Like I still help people with stuff around digital transformation about Industry 4.0, you know, that's, that's going to keep doing, they're going to keep, we're going to keep having to deal with brownfields that have, you know, physical ch boxes on paper. That's going to be a thing for a while. I think for those organizations that are ready to adopt something new, it's just about saying here's a tool. And maybe it kind of sounds a little bit scary, but let's make it a bit easier for you to use.

Speaker D: Yeah. And I would add on the AI front, the pace is accelerating at an unbelievable rate. We used to think about software, like modern cloud based software doing regular releases. You have like a major release every quarter and then we got a, you know, DevOps and Dora metrics, you know, people kind of competing like how many times per day can I release new software. But we're looking at AI now, you know, with entirely new models and the like capabilities and the behavior of AI changing like every time you use it. And so now it's like you cannot keep up anymore until you find a way to harness that and in a way that fits into your operations all. Uh.

Speaker C: Right. I think this is a good segue to Brian. What Edge Scale has created something called the cube. Right. You've described it to me at least as looking like a cloud on the inside but an appliance on the outside. What exactly is the cube?

Speaker D: Yeah, uh, I love that kind of the visual.

Speaker A: Right.

Speaker D: It's like we're talking about operational environments. So it needs to work, work. Right. So we're putting something there. It is cubicle in shape, but fundamentally it's about an appliance you can plug in. We reduce all of the complexity to a cable and IT and it works. But in order to keep up with the pace, uh, of change in the enormous capabilities of modern software and AI, we brought all of the best things that we could find from the cloud environment and put it on the inside. So not something that an operator ever needs to think about. Right. It's just you're getting the best of what's coming, you know, from, from AI and software that you can possibly get and it works and we stand behind that. Uh, but from an IT perspective now we have something that we can, we can work with, right? We can get like vulnerability scanning out of this thing and plug it into our scene. We can do you know, some of the things from an IT perspective that they really like to do. So we think of this as a, as a physical AI appliance, something that, like I said, we've shrink wrapped what we needed to uh, into something that's easy to deploy. I think you know, at the outset you talked about the real challenge here being like, okay, we can put something there, but how do we integrate it all together and do that in a way that um, you know, doesn't actually add more things to keep track of. And this is I think where we're really novel and stand out compared to anything we've seen before. That when we bring all those pieces together that means that we've got AI that can be at the shop floor level. And something that AI is really good at is integrating to other systems. So we use AI ourselves as part of our solution to do integration, to write potentially hundreds of data pipelines and manage the data and organize that and so forth in such a way that it can serve a use case which oftentimes is AI related, doesn't have to be, but we're essentially using AI to be able to use AI. The trajectory we need to be on. Honestly when we look at the volume of work that needs to be done and how we've been able to compress something like 12 to 18 month timeline to something that can be online in weeks, it's because we are harnessing for the benefit of our partners and customers, AI to uh, address that problem.

Speaker C: Yeah, and Brian, it's been ah, interesting last few weeks for me because I've been talking to quite a few, the full cross section, you know, from startups to the uh, big enterprise software and industrial automation companies and this idea of that we might need industrial grade data scientists who understand all the complexities of data science and also understand manufacturing I think was rife maybe a year ago and now it's like, nope, don't need that. The uh, data ops and to some extent the AIOPS is being taken care of by AI M and so it's becoming simpler and simpler to do that and we're seeing lots of companies now start to talk about um, agents that are doing a lot of the integration uh, steps. So I fully can envision how you're uh, you're doing this but you're also taking away a lot of the complexity of, you don't have to understand whether you need GPUs or NPUs or the memory or the networking and all of that you're taking care of with the Cube, right?

Speaker D: Yeah, you do need all those things, but you don't need to custom design it and custom figure it out.

Speaker A: Right.

Speaker D: It needs to run in a reliable fashion to, uh, echo on that. That is the mission of the company, to empower people who understand their business and their process. Right. Nobody understands that better. AI definitely doesn't understand it better than the people who are doing it. Um, the difference here, which I think is the most exciting thing that's happening in AI is, is that everyone is a quote unquote software engineer now. Because if you can talk to AI about what you're trying to accomplish, AI can do those things. And then we extended it all the way down to the physical stack that needs to be there.

Speaker C: And to a large extent I'm starting to reassure lots of people that domain expertise is more important than ever. Right?

Speaker D: Yes.

Speaker C: Understanding the outcome and what you're trying to achieve. Now, Cole, the engine inside the Cube is Red Hat. Uh, walk us through why an appliance like the Cube needs a platform like OpenShift and Red Hat Enterprise Linux running under the hood to be viable for an enterprise IT department to accept it.

Speaker A: Yeah, absolutely. You know, I almost find it funny with the amount of conversations I have, um, about the topic of how are we going to do this at scale? Like that's the number one question I ask people, um, in, you know, for my entire career is how do we manage all these devices, how do we manage all these applications? And the amount of times I get a blank stare from someone who should know better, um, is somewhat concerning really. It's all about so many people try to do a POC and say we're going to stitch together these services. And then they realize, well, when I have 600 of these, this really isn't going to work. It's all about that ability to do enterprise fleet management or management of both devices and applications inside the Cube. It's running Enterprise Linux or RHEL and OpenShift, you know, providing that common baseline of enterprise grade software to be able to give the IT teams the same tools that they're used to in a cloud environment, um, whether it's public cloud or private cloud, allowing them to do declarative automation via GitOps, handling model drift, handling application failover, everything that you'd expect of, uh, applications running in the cloud, giving you that consistency when it comes to the edge, and of course giving you security and governance, but also making sure OT has that same level of security that you're not mucking with their process control systems and exposing vulnerable networks to the wider network. So it's all about making it as kind of Brian was alluding to, making it drop in ready, making that ROI quick and making that ongoing maintenance a breeze.

Speaker C: Okay, let's maybe switch gears a little bit and talk about, because we've been kind of building up to it, let's talk about the outcomes, the business value, because at the end of the day, as I think you've both said, um, our uh, listeners care about yield downtime. Brian, once the Red Hat and edge scale infrastructure is in place with the cube, what does this enable? I think you mentioned digital twins in a, uh, previous conversation with me.

Speaker D: Mhm.

Speaker C: Production schedules. How do you get a plant manager trust an agent enough to let it take closed loop action?

Speaker D: I've got a couple of examples for you that happened this week. Um, we're starting to deploy this in a lot of places. We're starting to get more of these case studies and I think that's part of the story here, that it's doable. Now we're getting the results. So I want to share with you a couple. One of them is in the direction of Digital Twin, but I love to you know, not, not be too sci fi about it.

Speaker C: Right.

Speaker D: Like this isn't like some magical like three dimensional universe of your facility. We actually think about it more like a, like a digital lens. It's a, it's a unique view that an operator or a line coach or a supervisor wants to have about what's happening in their own facility. We installed our, our appliance. We've uh, actually replicated it to multiple facilities. Now we do, as we mentioned before, all of the, the data integration work. So within a couple of weeks we had all of the production data of the facility running on the cube and then we, we serve it to this, to this um, digital lens application. And uh, feedback that we got this week was um, while using this for the first time, it was the fastest shift they've ever run at the facility. The operators said, well, I'm not just like running from, from job to job or station to station. You know, I'm able to see like what's happening at all of the stations and so I can go to the one that really matters for me to jump in on or take some action on. It's uh, because we're using AI, we're able to kind of adapt. So it's Obviously not the same view for an operator versus a line coach. And it didn't require that much more work for us. You know, it's kind of like a tailored lens for what they need, but we're adding more layers to it all the time. But initially it's just about understanding what's happening at different stations and cycle time and so forth, and adding now layers, like what's happening with each of the pieces of equipment, sort of like statistical process control at each location. And um, I think what we're seeing is that because we've set the foundations and that's what we do as a physical AI infrastructure company, we solve the infrastructure problem, the data integration problem, and now it's kind of like Sky's the limit on what operators want to see. We can deliver that and iterate on that at cloud speed in an operational environment. And another one, because it's, it relates to Colin the kind of scenario that we talked about earlier, um, about computer vision, because I do think there are plenty of use cases out there like that and plenty of know how that that should be harnessed. Uh, we heard from another customer this week that they actually developed the model themselves. So it was an operator who understood the process extremely well. You know, she's been a mechanical engineer in this environment, so nobody knows that, uh, process better than, better than her. Created a model using Pytorch and showed that it was effective. And now we're talking about using our stack to deploy that model. And the uh, estimates are unbelievable when you look at the more accurate and real time quality assessment, something that used to take hours and hours with a dyno, now could be done in real time and then make a change to the way that the process is operating that results in less scrap. That business outcome here is tens of millions of dollars of reduction of scrap. And it's all back to this idea of like real time intelligence. We know, uh, we know the process, but we can't always see into the equipment and into the materials.

Speaker A: Right.

Speaker D: That's why we have quality and other testing frameworks. But if we stitch those things together and we can go from like know how to solution to practical implementation and bring that cycle time down to uh, weeks and days instead of 18 months or never, we're really having an impact. Cole, to your point about scale, this is what really sort of keeps motivating us in an increasing fashion, that we've got a scalable way to do this now. And so we're pacing ourselves, how quickly can we do it and roll it Up.

Speaker C: Yeah, it's not difficult to gain that um, support and trust. If you can move from uh, 18 month project down to weeks and you can show the ROI to the people that actually operate the plant in that kind of time frame, then I mean you can build that trust really quickly. Right. And I think that's the definition of crossing the uh, digital or of course now the intelligence divide. You're moving operators from kind of being in the loop, maybe getting some uh, advice from co pilots to on the loop, uh, where you're actually orchestrating physical intelligence rather than just reacting to dashboards. Right. So I think it's a fundamental shift that you're uh, describing here, Brian and

Speaker D: Col, I would add Colin if I can to that just that this is about um, like harnessing the knowledge and delivering that knowledge into having operational impact. Right. It's about like not having to adjust the knob or select the right recipe or something like that. You know, at a low level it's more about how do I improve that. And that's what we see as, as really exciting because it's um, it actually makes you, you know, our customers and, and operators and their, and their IP more valuable. Like we were talking about the domain knowledge, that's the real thing that we need and then it becomes directly linked to the operational outcome.

Speaker C: Uh, look, I think this is probably a good, we could probably talk for another hour or two but I think in the interest of time, let's stop there. But I need a closing thought, a takeaway from each of you on if you had one thing to say to people that are maybe listening to this and saying, why should I talk to Edge, scale and Red Hat about the cube? Frame that out for me. I don't know who wants to go first, but I'll give you both an opportunity to kind of close with. Ah, some meaty thought for those listening to the podcast.

Speaker A: I think what Brian said there in the last bit really sums up what this is about. Right. We spent 20 years digitizing systems making analog processes capable of being read via a uh, computer, via laptop, via an hmi. But the reality is these so called intelligent systems still rely on people to make the decisions. Can we democratize these systems? Can we make them more intelligent? It's just another evolution. The challenge is a lot of people get trapped in this game of we need to build everything in house and we need to do this and that and there's all these different teams, let's just do it and do it simple. And that's the idea here, is you drop a cube in your facility, Brian and his team help you set everything up, and you're off to the races getting value immediately instead of spending, as we said at the beginning, 18 months.

Speaker D: Yeah, I would add I've been a buyer of systems as well and I've uh, directed internal teams on things. And I can tell you from, from that perspective, what, what really deprioritizes something, even if it seems valuable, is the uncertainty. We don't actually know is it going to take 12 months or 18 months or 24 months or we, you know, because there's so many things that we, that we need to figure out before we can actually get to the value. We're still kind of like stuck on the, on the plumbing and we're not really sure like how we can get there. We're just trying to flip that entirely around, you know, Cole, to your point, and say, let's solve the plumbing problem for you. Things that are not, that have to be there, but you, you know, you don't really want to have to have to deal with them directly so that there's something that's actionable instead of uncertain fast to try and then to iterate and get to something that's um, of unique value for your business. And that's I think, where we're going to have success and be able to keep pace with, with the way that AI is evolving.

Speaker C: Great. As a final thought on my side, uh, I kind of want to look back at our Q. AH4 2025 industrial AI, robotics and energy survey where we found that um, almost 50% of executives in the energy and chemical sectors, and we did look at um, discrete verticals as well. But this one, I can remember off the top of my head that 50% of executives in the process industries, if you like, identified AI as the single biggest enabler of autonomous acceleration. But then they go on to cite the cost and complexity of OT data integration as their primary challenge and risk. So I think you've both presented um, a pretty compelling argument about how you can help them address that to our listeners. If you're struggling with that exact integration nightmare and feeling stuck in pilot purgatory, uh, what you've heard today is the, is an architectural blueprint to break out of it. It requires bringing enterprise grade software to the physical edge and doing it in a way that respects the physics and safety of the factory floor. Brian Cole, thank you both for joining me today and sharing how Edge, Scale, AI and Red Hat, uh, are tackling this together.

Speaker D: Likewise. Thanks very much, Cole. Thanks very much, Colin. Appreciate it.

Speaker A: Yes, thank you, Colin.

Speaker C: Great. For those looking to dive deeper into the frameworks we discussed today, including our industrial AI models taxonomy, our pay setter report and the survey, I've been sharing a lot of nuggets from Head over to, uh, arcweb.com or follow me on LinkedIn. Thanks for tuning in and we'll see you on the next episode.