Cut through the AI hype: Get real infrastructure that delivers

The it kind of just sets up the rest of the discussion. Over the last few years, the language around our industry has changed really quickly. You know, uh workloads that once would have been easily described as HPC, uh you know, technical computing, modeling, or simulation are starting to get to get described as AI or digital twins, right? You know, that matters because language changes expectations. And uh if the category gets defined around the most visible AI use cases, then people start to assume uh the same infrastructure choices will work for everything else, too. So, starting with you, Addison, from your vantage point, what happened when the industry made that naming decision?

You know, how how did calling all of this AI change when enterprises and vendors expected uh it what they enterprise and vendors expected from the underlying infrastructure? >> I mean, I think it's a perfect setup, honestly, because this is exactly what we see going on in the industry right now. For [clears throat] context, I've been an industry analyst now for about a quarter of a century and 20 years of it the 20th year now at Intersect360 Research. And over that time, we always said that we were high-performance computing analysts, but that included things like what we started [clears throat] out calling ultra-scale internet and the scalable computing that would go into what are now known as hyper-scale companies.

Uh it included the whole big data revolution and it included uh machine learning and AI and cloud was going into that. But, they were all just different types of what we called high-performance or scalable computing and we could methodologically call them all different segments of the HPC market. Then hyper-scale grew to the point where we really had to take it out. Hyper-scale was dominating the conversation. It was more than the rest of high-performance computing. So, we said, "All right, the hyper-scale market is over here. Then the HPC market is over here." Then AI comes along, which really started in the hyper-scale market more than anywhere else.

Companies like Amazon and Microsoft and Google researching a lot of machine learning. Now, Meta is one of the biggest ones doing it. And because it was a high-performance computing type of workload, um we kind of put them back together again. But, AI kept growing and growing to where it sucks all the air out of the room. And the hyper-scale deployments of AI are bigger than anything else. And that's before we add in things like the AI-focused clouds or what the kids call neo-clouds these days. Uh uh the national sovereign AI types of data centers and a certain amount of enterprise AI that to your point could could be just about anything.

But, now because AI is driving all the budget decisions, I need to invest in AI, I need to get ahead. We can go visit the same high performance computing types of customers we've been viewing all along where here's my model of a jet engine and the streamlines going through it. And like here's a computational fluid dynamic simulation. You're doing HPC. Oh, this is my digital twin now, right? And I'm calling it AI. It's still really the same HPC application, but it's almost impossible to see what kind of computing I I can't relabel as AI somehow. The computer is like thinking for me now. So, it's all AI.

The budget is certainly AI. The boss wants me to do AI. But where that has an effect like you're talking about is now I would say I stopped saying I want a high performance cluster and now I say I want an AI system. But then the thinking is dominated by all of those consumer types of applications and it starts affecting the configuration because you're designing for one thing and maybe not thinking about the other thing. So, it's the can create a hardware to application mismatch in some of these cases. >> [clears throat] >> Absolutely. You know, that's that's something you and I have talked about that.

Basically all the points you hit we've been talking about this whole ISC conference. This will be a point to both of you if you'd like to discuss what the the other person said in their previous response. Please do, but take >> Yeah, Greg, what do you think? Do you run into this with customers at all? Welcome, Greg. >> So, it's very interesting. So, since I've been doing high performance computing, which is honestly started in the late 90s, like I I've always heard debates. What is high performance computing? And then you have on one side you've got the purists. You got the people who are like tightly coupled big iron systems and if it's not running at massive scale, it's not high performance computing.

And then you have the on the other side, you have well, you know, I'm I'm pegging my CPUs on my laptops or my workstations, and so I need all the high performance I can get. And then you have the mid-range computing, which is sitting in the middle, which is everything between your laptop and between the big iron. So you got all these different kind of ways of thinking about high performance computing. And since I've been in the industry just alone, like there's been multiple kind of I want to say false starts of taking high performance computing, however you're going to describe it, HPC into the enterprise.

And organizations and and companies that are like, "It's this year. We're now moving HPC into enterprise." And we've seen this like multiple times, multiple companies who have pivoted from HPC to now focusing on enterprise, and kind of landing a little bit flat. And I believe it's because it was it was a little bit premature. And I say that because I'm [laughter] going to say this tongue-in-cheek because AI didn't really exist as it does today then. We started thinking about data analytics, we started thinking about compute-driven analytics, and and then we started thinking about machine learning, and then AI is a is kind of a top-level you know, thing that we're we're we're thinking about in terms of a buzzword.

And then the LLMs came. And [laughter] it was like all of a sudden everything just just kind of pivoted. And so so when I think of what is HPC and and we think about nomenclature, um it's you know, I made a post on LinkedIn not too long ago where I said, "HPC equals equals AI." And I I and I just put it out there just for a debate topic. And there was there was some really interesting debates, but I think I said it wrong. I think I should have said it AI equals equals HPC. Because I think that when people think about uh uh HPC, there's a there's a lot of use cases out there.

And there there's a ton of stuff that we've been doing for decades that Addison, to your point, is now being called AI. And has it changed? Not really. It's just it's this new name being >> not yet. >> Yeah. Yeah. So, but but we have to be buzzword compliant. So, everything we were doing in HPC is kind of now being you know, put under the guise of of AI. And I've seen hardware vendors even that have clearly stated we we we don't do HPC anymore. We do just AI. But they're selling AI clusters to researchers doing modeling and simulation. >> Absolutely. >> Yeah, that's still HPC.

You're just calling it AI. So, I think the the the bigger question here is is is like what is the alignment between HPC and AI? They're not the same, but what where do those Venn diagrams overlap? How much do they overlap? And Addison, to quote you, and this was a while ago, you used to say, and I don't know if you still say this, but uh that AI is the killer HPC app. >> I I wrote an article about that uh about a year and a half ago now. HPC finally found its killer app. Can it survive? Right? And it was all about AI outgrowing its small-town HPC roots and moving on and like leaving HPC behind.

And that it's the AI market now and HPC has become a small segment really overall in what's the larger HPC market. and to your point, Greg, AI now is really doing what the promise of big data was 12 years ago, right? Where we were supposed to be doing analytics and have this huge revolution to use high performance technologies in order to get there. And certain analysts, not us, were saying, "Here's this tens of billions of dollars market of analytics workloads out there with this huge growth rate." We were studying it and we'd said, "Yeah, yeah, analytics is going to be an important workload, but as far as we could tell, people were using data they already had, storage they already had, systems they already had, people they already had, and there wasn't this big bump in spending.

There was a shift in software, but the software, number one, was in-house applications, shell scripts, algorithms that they were writing themselves, which didn't really move the needle much on spending. Number two was variations of of Hadoop and associated things like Cassandra that were mostly free. And number three was new modules of of enterprise software that they already subscribed to anyway. So, you get this new big data enabled SAP HANA, but it didn't raise your subscription price. So, the the total market value really didn't change very much. People were doing analytics, but it didn't move the market.

Now, because of AI coming in and the tremendous hype of what you can actually do different from productivity standpoint and how much it costs, we are really seeing new spending, a lot of which is tagged for AI, but HPC is along for the ride where they're part of the same budget, and there's net more spending out there. Some of it in the old HPC associated categories, maybe I rename the budget, but in separate pockets with uh hyperscale AI and neo clouds and sovereign AI. There's all kinds of AI out there. >> So, I'm really curious uh in terms of your research and and your background, um when you have companies that have never done high-performance computing before, uh it seems pretty clear that we have an when they they there's an AI team that is now formulating.

Um I >> Quick tangent. Um I've talked to many big companies now about their AI strategy, and most of them are like, "We're almost done with our AI strategy to create the strategy to create, you know, have a plan for AI." >> pre-call for that pre-school AI. >> It's a date for a date. Almost have their date for a date. Um but what I'm seeing is that there's there's >> AI promise bracelet. >> [laughter] >> They're spinning up like these new teams and the new new initiatives to focus on AI. But out of the companies that have historically done HPC, so automotive, aerospace, pharmaceuticals, etc., where are you seeing AI kind of hitting into those organizations?

Are they still spinning up from scratch kind of like an AI team, or are they spinning up the HPC team and enabling the HPC team to focus more on AI? >> Um you know, that honestly, Greg, is a mixed bag. It's I wish I could tell you, "Yeah, 100% of customers we've surveyed do it this way." And it and it's not. It's it's really just been catch as catch can out there. But what we can say is that the companies that are investing in AI, most of them, it turns out, also have modeling and simulation through R&D or engineering or some kind of scientific research where they're also doing things that we would have already categorized as HPC before that.

Not 100% but when we surveyed just any medium to large enterprise out there and this wasn't even an HPC oriented list that we own. We just like let's go externally and get any enterprise lists out there and we start surveying their IT departments and say here's definitions of machine learning and AI and those kinds of things. Do you spend any money on that? And 93% said yes, we spend money on that. But then we also said all right, then this is modeling and simulation and uh HPC types of applications. We define those. Do you spend any money on that? And it was like 82% said yes to that and in the Venn diagram, they're almost entirely overlapping that those HPC people are uh inside of AI.

So it's only the 10 11% of the market who are doing AI but don't also have some kind of modeling and simulation out there. Now that doesn't mean they spend the same amount. They might be spending way more on AI cuz that's the initiative but the HPC types of applications have to fit in there somewhere because when we do find both here's another thing is they're they're darn they're very probably sharing the same budget and often sharing the same infrastructure. Not 100%. I can point to some where it's not like that. Like oh yeah, the AI guys are all over there getting all the nice new toys and we HPC guys are stuck over here doing the old thing and they keep us separate.

It does happen. I can name companies where they're doing that. I would say it's more common uh if if the the most common thing I would say is they probably started with they spin up a team. Hey, let's go look into AI and how do we do that? And we start identifying here the easiest wins of things we can do on the analytics and enterprise side pretty easily. But if you try to bring it over into R&D, which is the top line promise of how AI is going to start benefiting our roots to market, our better products, our new discoveries, right? Then you better bring the HPC guys in because that's where the entire product workflow already is.

>> Yeah. Yeah. I I would validate that from what I'm seeing as well in in some of our customers and and early customer discussions. Uh it seems as though a lot of the organizations that are um driving is that are first to market, I guess that's the right way of saying it, first to market with AI, um building their AI infrastructure, seems to be the organizations that have a lot more uh HPC experience already. And sometimes, yeah, I am seeing overlap, but it isn't 100% at all. So it's interesting um the there's another perspective as well, which I think is worth noting, which is uh the nature of AI and the infrastructure that we use to drive it, the physical infrastructure, seems to have a huge amount of overlap with high performance computing, but the software stack historically has not.

>> Yes. >> [laughter] >> I I think you're really onto something. Kevin, Greg and I are just talking here. If you want us to talk about something, you're going to have to come in and frame us today. We're just going to talk for like 6 days and people will either enjoy that or not enjoy that. I don't know. I like talking to Greg. >> I I I'm not kidding. As you were talking, I was I was warming up the old vocal cords to jump >> All right. >> I I know you two can chat forever, but >> You go suck a lemon then tell me what it is we're supposed to talk about next.

>> You got to keep us keep us on track. >> Oh, yeah. [laughter] That's my job. All right. >> we doing so far? >> Something we've talked about before Addison, you and I. This conversation feels very familiar because we've been through a version of it before, you know. For years people asked why HPC wouldn't simply move to the cloud, right? And >> Looks nice. >> The way that question was usually framed made it sound like the answer should have been straightforward, you know, if the cloud was the future of enterprise computing then surely HPC would just follow, right? But what actually happened was much more instructive, you know, the technology matured.

The market expectation shifted and over time people had to reckon with the fact that some workloads impose constraints that just don't disappear because a category gets renamed or a deployment model becomes fashionable. So, starting with you Greg, you know, you were in the room during a lot of that cloud versus HPC debate. What feels most familiar to you between that era and the way that people are talking about AI right now? >> Well, there's there's a few different ways we can we can like go into this one. Um >> Sure, here's another three days of conversation. Let's go. >> So, uh I remember so I I I started off as a a bio chemist got super interested in Linux and open source and what do you do there?

I end up in HPC. Uh and so I did high performance computing uh systems work, programming, etc. for quite some time. And uh landed at the US Department of Energy. And I remember clearly when the management of US Department of Energy came back and said there's this new cloud thing where we may not have to buy all of our own resources and we can go leverage resources up in the cloud. And and and you know, let's go figure out what our strategy is around cloud. And so we started and of course being technologists, we started with the technology first, right? So we started looking at how would we theoretically go build or augment our HPC resources in the cloud?

And what are the the stumbling blocks that we're going to hit by doing that? And we spent actually quite a while kind of figuring out technical architectures for augmenting or moving our computing resources up into the cloud. And while we were doing this and we spent a decent amount of research and and time kind of focusing on this, somebody smarter than me came along and said, "Yeah, but it doesn't make financial sense." And then they ran some financial numbers and they're like, "It's 10x the cost." And don't quote me on the 10x. It's different for every organization, but for us it was massively like it would have cost much more money to run our our computing cost in the cloud.

And then the question is, "Well, where does it make sense? And and does it make sense?" And what almost every I don't want to say every, but I think the vast majority of HPC centers and HPC use cases found was that financially it makes more sense to buy your resources and run it on prem if with two F's, if and only if you are you are consistently leveraging that resource. Meaning you have constantly utilizing that resource. If you're bursty, sure, it makes financial sense to run it somewhere else. But if you are consistently hitting that resource and doing a lot of computational work on that, it makes more sense to run it on prem and to own it yourself.

And and so that's what most HPC organizations found. So most HPC was run on prem historically. And so uh if we now take a fast-forward jump and we look at where is AI mostly getting used today? It's very similar nature of work. Workload, the application and the software stack is different as we alluded to before. But most of it for most organizations is actually being done in the cloud. They either they're using uh uh partner providers like Open AI or Anthropic or or Gemini or or whatever. Uh and they're just leveraging an API or a an interface to that cloud resource. Or they're actually ins- instantiating their own AI infrastructure up in the cloud.

Um and as a result of that, we're starting to see uh uh an interesting dichotomy of infrastructure where traditional HPC has been done predominantly on prem and AI has been done predominantly now in the cloud. And in organizations that I've spoken with and I'm really curious uh if if you both have seen this as well is that organizations are now starting to recognize kind of what we already learned in high-performance computing about infrastructure costs. And it does make more sense to actually run this on prem and or in in your own resources in some way, shape, or form. Um and then even to start leveraging some of your own models to do this, whether those are public models, open weight models, uh or or commercial models that you can purchase and then run on prem.

Um but that's the that's the the the interesting perspective that I've seen kind of between cloud and on prem. Now of course there's other and Kevin, I think you were kind of leading into this. There's other areas as well, other other metaphors, but that's immediately where my brain went. Was kind of just the dichotomy of infrastructure and are we starting to see a convergence back towards on premise? >> Absolutely. Absolutely. And you know, Addison, passing this to you, you know, from the analyst side, there was also a period where clients wanted the story to move faster than the infrastructure reality allowed. You know, what did that decade look like from your perspective and when did the market really accept that the workload requirements hadn't changed just because we named it something else?

>> Well, I mean, this is relevant what Greg was talking about before with the applications and that's what really matters here, right? We brought a technology to the table, whether it was AI now or cloud then and said, here's this great new paradigm. Let's go take advantage of it. And at an industry level with AI, if there's a danger right now, I think it's that we're in danger of repeating the mistake of just assuming all workloads are the same and that doesn't matter. Right? And that's certainly where we were with cloud. Cloud was going to be this great enterprise tool that was going to benefit everybody because cloud was magic and uh and you know, you only pay for what you need and therefore it automatically is cheaper.

And uh and it also automatically balances supply and demand. So, every cycle gets used and you can always get a cycle when you want one. Some of that stuff really wasn't true, right? There's an economics behind cloud infrastructure. Um but you know, cloud became all the rage in like the late aughts, around 2010. We're supposed to be moving everything to cloud and then we spent 10 years answering the question of what's wrong with HPC and why isn't it moving there. Economics, which Greg just mentioned, was a huge part of it because if I could fill I mean this is classic rent versus buy, right? If if I could fill the resource and then it was cheaper to buy it than it was to rent it.

And the first time we did a really deep economic study of that, we found that the break-even utilization was about a third was about 35%. If on average if you can get over a 35% utilization hurdle, it became cheaper to buy than to rent. It's it's higher than that now. It's more like 50% now, but still if you're going to use more than half your cycles economically, it tends to be better to uh to buy than to rent. And the other part of it was that HPC is just hard and requires a lot of special resources that the cloud in the early teens wasn't ready for yet.

I mean it wasn't zero. We were measuring HPC in the cloud as early as 2009. It started heating up around 2013-2014 kicked into a higher growth rate, and that was because some of the software that would be common in high-performance computing, the licensing of that went from being really cloud punitive to being more cloud adaptable, right? pay some huge penalty to move my license to utility computing. So it started growing. And and that was through the middle of the decade. It really kicked into high gear in 2020, not because cloud was so amazing that everyone suddenly woke up to it. Did anything else happen in 2020 that affected the market, right?

And what happened with COVID was it changed the market and this is the biggest thing that affected the market that we did not predict as analysts, right? We didn't see it coming and there it was. We did measure the effects of it. And it had people pushing projects out. It had people pulling projects in. Do I want to buy something? I don't know. Do I have budget? Do I not have budget? And one thing cloud is really good at is flexibility. So it was worth paying a premium even to get flexibility. But then by the time we came back economically out of the COVID era, a new wave was taking over and that was AI.

And AI was a new capability that I was looking for that I might not have in house already. And again, I'm willing to pay a premium to do some testing of that out there and and let's go. Now here we are today and AI is amazing and everyone's waking up to it. Why don't we do everything in AI? Well, again, the high performance computing stuff, a lot of that is kind of a special use case. And in this case, there's an additional irony because that's what we're being sold on.

We're sold on the idea of, you know what's great is AI is going to predict the weather and climate change and cure diseases and find new energy and uh you know, we'll be able to make drugs and all these kinds of great things that are the types of applications that we're used to associating with science and engineering and modeling and simulation and HPC. But then you say, "Great, go do it." And most of what you get is what Greg was talking about with chatbots and bio coding and well, most of it is really consumer oriented things. It's you know, making funny pictures and memes and augmented search and Instagram filters, pornography, video games.

And if if you're freaked out by that, don't be. It's like the internet, right? The I mean, does the internet help science? Uh-huh. Is it mostly what you use the internet for? Not really, right? Mostly for the the internet is consumer entertainment, and AI is going to be like that. So, we start treating all workloads like they're the same. Look at all of AI can do anything. So, sure, it'll do all these things, but it does the easy stuff first, the relatively Not that, you know, AI is easy. But, the more straightforward consumer applications, the horizontal things first, and even enterprise AI, it does the horizontal things first.

Help me write an email. Help me write a business plan. Vibe coding, chatbots, the things that any organization might do. But, help me optimize an airplane wing, or um change the design of this plastic bottle so that it uses less material and is also more shatter-resistant if you drop it. That's an HPC kind of application, and it's not as easy to just say, "Hey, you know, whatever your favorite AI agent is, I'm not going to name-check any of them, go redesign this plastic bottle, go redesign this airplane wing." It has to do all the HPC stuff also as part of that, and we're just not there yet.

Does it have that promise? Yes. Is it there now? No. The organizations already doing AI. They've got these applications, but if you think, "Well, I'm just going to do this over here," that's the mistake. It doesn't work, not yet. >> Absolutely. >> If if if I can respond real quick to that, there there's I think a lot of people are thinking about it very much as Addison described, which is it it it's AI is going to replace a lot of these these HPC like applications. And and who knows what the future is going to bring. Maybe it will. Um but what I see is the instant wins today is a hybrid environment where the AI knows how to run these simulations, knows how to manage these simulations, and then evaluate the results, and then iterate.

And then it can do that constantly. And then you end up with that iteration that's constantly occurring. And then if you can have a human in the loop or human monitoring this, you can end up in a really great spot. Without the human monitoring it, you you end up in a really potentially bad spot. It's it's >> Absolutely bang on right. We're going to get this AI-augmented HPC loop where I can do it all better, but not without the HPC. It's AI for science, not AI or science. Right? You can't just do away with the scientists. You you need the science guiding it. So you're you're absolutely right, Greg.

>> I'm going to have to jump in again. I'm I'm sorry, Greg. We got to keep >> Go for it. Go for it. >> [laughter] >> We'll never We'll never finish this if we don't jump in. But uh what what I was going >> We have We have to We have a finish on this? Like we have to stop? >> getting started. >> [laughter] >> Yes, this is an 8-hour webinar. Strap in, everybody. No, but uh So so far we've been talking about fit, right? You know, whether the infrastructure matches the workloads uh that we're putting on it. But But something that uh we've been You and I and again Addison have been discussing, something that I've been hearing all week at ISC is sovereignty, right?

It raises a completely different issue of control, all right? Uh as AI gets closer to corporations, proprietary data, internal decision-making, the question is no longer just what works technically, how well the technology is working, uh but it's about what organizations are comfortable putting on infrastructure that they flat out don't control. And, you know, what has to stay closer to home. So, Gregg, I I cut you off. I'll give you the first response here. Uh when you look at organizations leaning more heavily on external AI infrastructure, where does that exposure become more most dangerous in practice? >> I I know you called on Gregg, but I'm going to give a little bit of context on this to Gregg's answer because I think I'm hearing sovereignty used two different ways.

We hear a lot about national sovereignty stuff, which is we got to protect country's assets and make sure that we're not cross-boundary, whatever. But, I think the bigger discussion is this idea of enterprise or data sovereignty, where every organization wants to think of protecting itself and shepherding my own intellectual property and core competencies into the future, which is in some ways a little scarier cuz it's more personal. So, which ones do you see? I That's what I wanted to kind of dig in on. >> I focus much more on the latter, uh in terms of how I'm thinking about it, and in terms of how I'm seeing organizations kind of falling into uh the mix.

With that said, recent events has kind of made the former um kind of a a a a a major issue, a major concern. >> know what you're talking about, Gregg. >> [laughter] >> So, we're we're we're seeing we're seeing Okay, so this may turn into an 8-hour Joe Rogan, hold our beer. Um >> [laughter] >> we may we may end up with Cheers. We may we may end up with um uh the first actually driving the latter. Uh so, uh and I think that's what we're we've seen quite a bit. >> sovereignty drives the the data or enterprise sovereignty. >> Because I I'm I'm not going to discuss my my opinions on this.

Maybe over over those beers that we were talking about in in person. Um, but yeah, the fact that now we are being limited in terms of what models uh, we can have access to and what types of models that providers are able to um, provide to their user base uh, is actually changing the dynamic considerably. Uh, I think I think it is now a fairly well agreed upon fact. AI is big. AI is going to change how most of us work and operate. Uh, I I don't think that's debated anymore. Uh, I actually routinely say that I think AI uh, is as big of an innovation as the internet.

And since the internet, I don't think there's been anything as big as as what we are currently talking about as AI. Um, now, if we're going to have access to the latest and greatest and best models, if we're going to have access to do um, to leverage this technology and this resource to its fullest, I think it's now starting to become clear that you have to control it yourself. You have to own it. And you have to be able to um, control your intelligence. It has nothing of course to do with our uh, the name of our company, Control IQ. Um, but you need to control and and own your AI infrastructure.

And again, this is something that HPC centers have have known for quite some time. Uh, there's large numbers of organizations out there, everything from media and entertainment even, going through pharmaceuticals, going through aerospace, that still as of today won't run any of their proprietary models Uh, and research in in the clouds. They have to run it on prem because of data and controls. And we're starting to see that exact same scenario playing out with AI right now. So organizations that have very very rigorous controls on their data, they have to manage compliance, they have to manage who has access to what it is that they're working on, trade secrets, et cetera.

They are not necessarily going to provide that data to a prompt. And I'm going to level set here real quick. Uh for an AI for an LLM to to give you a good response based on some question you have, you have to be able to provide it relevant data to do that. >> give it data and context. >> Data and say again? >> And context. >> And context, yes. And so I'll use a coding example cuz it's easy to to to see this. When when you're running cloud code, code X, pi, whatever you're running. Uh for it to be able to provide you a patch to some software or some additional code into your software or fixes, you have to give it enough context for it to understand what it's working with.

So you are taking a snippet or groups of snippets of your code and uploading it through the prompt through the API so the LLM now can go process it and give you an informed intelligent response coming back. Well, you've just leaked a little bit of your data. If you do that enough, you're going to continue to leak enough data to where somebody else now has your trade secret. Somebody else now has access to pretty much everything that you are working on. And guess what? AIs are pretty good at reassembling content from context, from bits and pieces of context. So, it's just a matter of time before the AI providers are able to to do this.

Now, I'm not saying that this is a bad thing. I'm not trying to state that there's any malicious intent. I don't believe that there is. But, if you have to manage controls, if you have to manage compliance as an organization, this gets to be a very difficult thing to manage. And the way that in HPC that we've done this is well, you run it on prem. You control everything that you're doing, where your data is going, what has access to your data, and what do you do with the results. And in AI as of today, very few people are doing this, and most people are still kind of going for the easy win, which is let's go leverage a provider's AI interface.

Uh and and that's great for easy wins. But, what happens when you start to scale to production for your organization? Can you continue to do that? And that's not even including the cost of tokens uh by doing it that way. So, just the the when I think of sovereignty, uh I think of owning and controlling your own infrastructure in this case. >> I I think you're you're absolutely right, and it is an issue. I mean, just like we talked about security concerns with respect to cloud or even the internet. People were afraid of, you know, why am I going to use my credit card for something on the internet?

I would talk to someone over lunch, and they would say that. Meanwhile, the waiter's been gone with their credit card for 5 minutes waiting for him to come back, but [laughter] I can't possibly use it on a website. Uh ultimately, we start building trust in these things. I think we'll get there on trust, but the cost uh remains. Is it best to move all these bits around all the time, or do you want to do it on premises? How much of these resources you're going to use? And you were talking about the small language models or or being able to build my own AI models internally.

Most of the organizations we talk to would have a preference for moving that on premises if they can, but there's also now a supply-side push to, you know, the the big companies kind of want to guide me into the cloud and and have the data over there. If everything's going into the cloud, I better worry about how much data sovereignty I have there. And I I do think there are uh enterprises, practitioners who are looking for who are the vendors who can help me manage these things in-house without having to worry about who has my data, how much is that going to cost, do I have a budget for it?

We we've seen this story in cloud plenty where I enable my people to use cloud, and then I'm shocked, shocked at what my cloud bill is this month. AI is now repeating this. Like, okay, let let's go use the AI, and then wait, I paid how much last month? No, no, wind it back. It's like I left my kid alone with an unlocked cell phone, and I got $800 of in-app purchases on on some game, right? And you know, that absolutely happens, and uh and you know, how can I control it? How do I control my intellectual property? How do I control my spending? How do I shepherd this into the future?

And how do I leverage AI across all these applications? So, I think it is the same rhythm that we've seen with HPC vis-a-vis big data, vis-a-vis cloud. In a lot of these patterns of this is the application I really want to accelerate, but I also want to do it thoughtfully with a with a view out to the future, not just what's best for the next 15 minutes. >> I agree Addison and uh you know, we're going to have to keep this moving along. Again, you know, it's an 8-hour discussion, but uh something I want to talk about here is the orchestration gap, you know, uh Uh a lot of enterprise AI doesn't stall because the model is bad or the use case is unimportant.

It's it uh stalls because the life cycle is fragmented. Right, you know. Training sits in one place, inference in another and uh the scheduling just doesn't line up and the operational burden ends up being heavier than teams ever expected as as kind of what you were just talking about there. So, you know, at that point the bottleneck is no longer enthusiasm, it's coordination. Addison, when you look at the organizations moving fastest, what did they understand earlier than everybody else about infrastructure abstraction? >> Oh, I want Greg to answer that first. >> Oh, yeah. All right. All right. So, So, people use the tools generally and I'm I'm sorry.

I'm going to approach it from a very pragmatic perspective in terms of, you know, focusing on operations and production infrastructure, uh which is where I spent the majority of my career. Um uh people use the tools that they're familiar with. Um this is one of the reasons why I think cloud is being used so much for AI. This is what people are familiar with. This is one of the reasons why uh we've been using the same HPC architecture and model for the last 30-plus years at this point, which is an incredibly advantageous and impactful architecture. I'm not dissing the architecture in the slightest, but it hasn't changed in 30-plus years.

Uh while the rest of the ecosystem has moved miles, uh the the core fundamental architecture of what we what we do HPC with has not has not changed. Uh so, we're we're looking at two predominant use cases when we think of AI. Uh we're thinking of AI orchestration, let me be more pedantic. We're thinking of um uh types of workloads that are more HPC-like, like like training. Right? That is a job. A job is something that has a defined starting point and it ends when you finish doing whatever it is you're in your algorithm to do. So, that is a job. So, it has a defined starting point and an end point that you want it to be as close to that starting point as absolutely possible.

Right? So, you can fit more in. So, you want more efficiency and to drive that to be as fast as possible and then for that job to end. Uh and then you stack all of these workloads together, so you have many jobs that can potentially run, some in parallel, some uh sequentially. And and that's the idea of orchestration of of an HPC type resource. And that's very similar to what we do in in training. And this is one of the reasons why in in AI training most people are using a traditional Beowulf architecture to do this. Then you're on the other side of the house, uh AI also you take your trained models and now you go and inference against them.

And the inferencing is a completely different type of of job. Uh it does has no defined end point. It is a service that is very computationally focused. And as a result, the orchestration framework that you need for that is sufficiently different than the orchestration framework you need for the training side. Now, the infrastructure, the hardware infrastructure that you need for both is actually incredibly similar, but this goes to the point that Addison and I brought up earlier, which is the software stack is different. And so, on the training side, you have a traditional Beowulf. And on the inferencing side, well, enterprises are now trying to solve how do we do inference inferencing at scale on their infrastructure?

And they're in many cases, just like HPC has been doing the same thing in HPC for so long, we're seeing now the same sort of thing occurring on the enterprise side. Let's use the tools that we know and the tools that we are familiar with and the tools that we are tooled to to go and roll out and manage. Uh so, Kubernetes, which is a fantastic microservice platform, is now being used for something that is arguably not microservices. Can it be used for it? Absolutely. And people are doing it as of today. But, is it the best architecture? And that's the question that uh I I've been asking a lot of other technologists right now.

And understanding, well, what does it do great and what are the pain points that are still yet to be solved? And And to be clear, there's pain points on both sides of this that still have yet to be solved. On the HPC side, the architecture is now held us back uh considerably in terms of new and more modern workloads. On the AI side, on the inferencing side, we're now having to kind of rethink how Kubernetes is operating with a lot of these different types of services. And And to be blunt, these are completely now separate systems because the software stack is different, not because the infrastructure's different.

So, the hardware infrastructure being different. So, software stack above that has to be different in order to handle both of these two different types of workloads. There's still a massive gap that needs to be solved in my mind. Most organizations are not going to go build up two separate disjoint systems to manage these sorts of of resources. And if you look at what uh what the nature what what the both of these workloads have in common. I'm going to call them performance-intensive computing. They are They're both of these are performance-intensive. Uh whether you're running jobs or you're running services that are doing inferencing.

And then the question comes down to is, well, if we're solving the same nature of the problem, then do we really need another Do we really need separate orchestration fabrics, or can we start building a infrastructure that pulls everything together, makes it easier, and makes it more robust? >> Well, I think the way I think about that, Greg, I think you gave a really solid technical explanation of it. From an analyst standpoint, the way I think about it is just through workloads and how people think about orchestration.

One of my favorite questions I've gotten from any audience the last couple years is actually a group of students, and one student asked, "What's the best way to ensure that their enterprise AI uh program is successful in its adoption and rollout?" And I said, "Well, the number one way to ensure its success is to start by knowing what it was you were trying to do to begin with." Right? As opposed to just AI for the sake of AI, like it's a parlor trick. Look, I did an AI, and isn't aren't I successful, right? Is that successful or not? I don't know. What did you want to do?

Right? It's hard to know whether you're successful if you don't know what it was you wanted to do. So, if what you want is to have a customer service chatbot, and you have a goal associated with that, I want to reduce the price of customer service by this much, then you can measure that success. Where it gets complicated is if you start saying things on the R&D side. I want a faster time to market, or reduce defects, or a new product in some way, uh better discovery of something. And those can also be reasonable metrics, but they're a different workload. And here's where the orchestration comes in and becomes important because I think at cross enterprise computing, most of the time when I've heard the term orchestration, it's tied to things like virtualization.

And virtualization most of what we would do with virtualization software is I'm making my infrastructure more efficient by taking one resource, one processor, one server and breaking it up so it looks like 10 servers or 10 processors, right? So I can fit more things on there. High performance computing, the heritage is kind of the opposite. I have a big problem and what I want to do is take 10 processors or 10 servers and make them look like one, right? So I'm parallelizing. And it's in a way maybe the converse or the opposite of our virtualization problem. And now where this brings us full circle around to why the HPC applications are different cuz you brought up Kubernetes and containers, which was kind of the virtualization topic on steroids, right?

We're going to hype it up. Now not only are we virtualizing it, but we're going to put every job into this container that then fits on the virtualized thing so it always runs the same, which is a great idea except when the container has to be bigger than the resource in question. So we got different container discussions for high performance workloads going back to Docker, right? And when Docker was the technology of the year and and a lot of enterprise went with Kubernetes, but the high performance applications needed special kinds of containers. So what is it you're trying to do with the AI? If it's strictly on the cost management analytics, that's one solution or one orchestration uh path.

But, as soon as you want to include anything on the top line, the discovery, the R&D, you need to have a different concept of what the orchestration looks like because you need different containers, you need different deployment, you need different federation in order to get these things to actually work because it's not just I'm going to run this on a segment of a process of a GPU somewhere. No, right? I it needs to scale. There's this notion of parallelism. And if I don't solve the parallelism, it doesn't work no matter how much AI I throw at it. Uh you haven't solved the orchestration problem. So, it's not even that you have to solve the infrastructure first.

You have to know first what you're trying to do. And then if what I'm trying to do normally takes a high-performance cluster, I shouldn't expect that because I use AI, it now fits into something much smaller. AI isn't magic that way. It doesn't make my job take less resource. So, you need a partner who understands how to do a scalable application and federate it in a way that will incorporate the AI, get you the benefits, but not run into it an infrastructure wall. That's the way we see it in terms of what's been successful. >> Great point. Great point. Uh once again, we kind of turned that singular question into an hour-long talk, but keeping us on on schedule here.

By the way, I'm going to ask a couple of questions here and then we'll toss it over to Q&A. So, if anybody's got any questions, please feel free to throw it in the comments and we'll we'll get to that in a second. But, you know, we spent most of this time talking about where organizations get stuck, you know, the the the naming shift, the repeated infrastructure pattern, sovereignty, orchestration. Uh so, at this point, we need to shift from diagnosis to execution, from problem to solution, right? If the real barrier has historically been the complexity of making all of this operational, then the useful question is, what changes when the orchestration layer is actually built to handle those hard parts?

Greg, this is for you. Given everything we've talked about here, performance-sensitive workflows, sovereignty, orchestration, all that, where does fuzzball change the equation for organizations trying to move from AI pilots to something they can really operate? >> Well, [snorts] that's embarrassing. I don't know what just happened to my camera. >> [laughter] >> Oh, no. >> Um, I will >> you fine, so go ahead and talk, and then we'll solve the camera later. >> It it should be a pretty quick >> our attendees know what you look like. We should have just done a freeze frame. I'll draw something on the screen that looks like your face while you talk.

>> [laughter] >> I can fix it real quick, Addison, if you want to take a first stab, and but I'm also happy to just run with this. >> I think you should run with it. He asked you the question. >> All right. All right. So, uh, first off, sorry for everybody who's who's doesn't get to see my my smiling face over this. Um, but it's probably for the best. >> [laughter] >> So, in terms of fuzzball, uh, this is going to be the a little bit of a shameless plug for what it is that we are what it is we've been doing. I'm going to tell a little bit of a story as part of this.

So, coming out of high-performance computing, >> minutes in, and Greg's going to tell a story. Here we go. >> Oh, I I thought we had eight hours. I'm sorry. >> [laughter] >> So, uh, coming from high performance computing, Addison already talked about kind of Docker and containers and what was happening through that whole through I I want to call it hype, but it was much more than hype. It was a it was a absolute paradigm shift in terms of how do people think about distributing, moving, and making their workflows more portable. And we saw an opportunity because we started doing kind of more modernization of high performance computing on top of Kubernetes.

And this is where we started. And we recognized that Kubernetes was just not designed for this. And very similar to how Docker was not designed for high performance computing and thus Singularity came to be, which I've now moved to the Linux Foundation and is now Apptainer. But very similarly, the the ecosystem that people were trying to move HPC into, which is Kubernetes at the at the very beginning, turned out not to be very efficient. And that's because it's not designed for those sorts of workloads. And as a result of that, we started thinking about, well, how do we make a more modern computing architecture that can both on the HPC side modernize how we are thinking about doing the compute, lowering the barrier to compute, and then making it more reproducible, easier for organizations to leverage, more scalable, etc.

But on the enterprise side, how do we take all of the lessons learned out of HPC and HPC infrastructure? How do we take that and bring that into an enterprise organization in a format that that is very similar to how they are already building and maintaining their infrastructure. And Uh, that architecture, through that shift, Fuzzball was created. And Fuzzball gives us the ability to kind of reunite those two halves of orchestration that I was talking about. On on one side, you have the training and you have the jobs. On the other side, you have the inferencing and you have the services. And leveraging the same physical resources, how do we pull all that together in such a way that gives us absolute capability to to run on a single infrastructure and lower the barrier.

So, whether you're just starting off with AI and you need some easy wins or you are ready to run this at scale for your entire organization and you need an infrastructure that can do this. Fuzzball was specifically designed for these purposes. And gives you the ability to go from something as small as a NVIDIA DGX that if my camera was actually working right now, I'd be pointing right behind me as I have a DGX sitting there on my desk on myself. Taking a single computer and turning that into an AI powerhouse which actually we just demoed at ISC, the International Supercomputing Conference. And it was actually funny on a LinkedIn post, there's a picture of a DGX and on the back in the background you can see the monitor that's that it's running and you can see that it's running an AI inferencing process, a service.

It's running Jupiter Notebook and it's running our studio all at the same time through Fuzzball. And I want to reiterate, these are push button clicks in Fuzzball because of template-ized workloads. So, we already have templates to do all this. So, easy wins with the DGX Spark uh or uh any small form factor uh GPU-focused computer going all the way up to scaling up to to giant GPU systems, rack scale, and um data center scale systems. And fuzzball can take you from one, take your workflow that worked on your development system, and scale it all the way up to thousands upon thousands of GPUs, and then run that in production.

And again, balance training along with uh the inferencing services. Uh fuzzball was designed specifically to solve this this need and to make this easier uh for for organizations to to leverage uh running their models locally and sovereignly. Is that a word? Sovereignly? I think it is. >> Maybe. It is now. You just did. We decided we could coin terms on this call, and I'll I'll just say words long enough, Greg, for you to reset your camera and show off that you have a DGX Spark. You know, way to go. But, I'm not going to talk about fuzzball as much. That's your job, other than I agree with the trends that you just highlighted.

And I I think whether or not you need something like fuzzball comes back around to that premise of why is it you want to do AI? What do you want to do, right? And we've heard discussions about Oh, cool. You can point to your DGX now. I see it there. It's just over your ear. There it is. Um you know, we've heard all these discussions about how GDP is going to supercharge the economy, and we're all going to be so much more productive. That's the argument that we're going to drive more top-line productivity somehow out of these things, which is different from saving costs. Saving costs doesn't deliver all that.

And even then, there's a question of how much money will you spend in order to save a million dollars, right? And can you repeat the trick and spend it again and save another million? Not really. So, that that's a limiter in AI. The boom in AI is if you start really pushing your business forward in ways other than than cost efficiency. So, what promise of AI do you believe in? Are you going to do the AI for science, for engineering, for R&D, and the HPC associated workloads? If so, then you need these orchestration tools that are going to help bring that to market. You need someone someone with experience in managing HPC workloads, federating HPC workloads, containerizing HPC workloads, right?

That's a critical software piece that can't be solved just by saying, "Hey, AI, go do it." Right? That doesn't exist not now. Will it exist in 20 years? I guess. I mean, 20 years is a long way out, but is that the time frame that you're on for this AI deployment? If you're talking about what's going to make this positive ROI for your organization over the next measurable time frame where you're going to go to your board and say, "Yeah, Bob, we did this, Bob, and here's the payoff, Bob." Right? That's the kind of conversation you have to have. Or if you don't care about that and you just want to be in chatbot world, that's different.

Maybe you don't I mean, it helps anyway to be able to federate and scale. That's it's never bad to have those tools. But to me, what really falls down is if you say all workloads are the same and it doesn't matter. And then I'm going to try to do the R&D stuff in an environment where that's not true. I That's my thesis. Workloads matter. We have proven ourselves over and over for decades that it matters what it is you're trying to do. The workload, the application matter, and you should have a partner who understands that. >> I I I agree, Addison. And now that we've uh we're seeing your lovely smiling face, Greg, I uh I just want to ask you another fuzzball question, you know, uh Uh Greg, uh pick a real pattern you see all the time.

Enterprise that's already standardized on Slurm or PBS or, you know, plus some Kubernetes for AI pilots, right? What does their day-to-day look like before fuzzball, and uh what does it look like after they move those same workloads onto fuzzball? >> So, that's a great question. So, um as I kind of alluded to before, a lot of organizations are still like I like I I've been I've been like totally deep into AI at this point for about 2 years. And uh CIQ as a company has moved very much and very aggressively towards AI. I don't think there's a person at the company who is not using AI on a daily basis as of right now.

So, we have fully embraced AI. There's a lot of organizations who not only have not embraced it yet, but not even sure where to start. They're still trying to figure out what their strategy is to develop the strategy. And that's where So, so, there there are easy wins that everybody can get today going and leveraging Anthropic or OpenAI or Gemini or whatever your AI provider of choice is. You can get easy wins from that. 100% and many portions of your organizations may be able to continue leveraging that for the long term. There's going to be sections and areas of your organizations and data that you're not going to be able to do that with and you still are going to want some easy wins.

This is where Fuzzball really helps. So, you can again, super easy win, go grab a DGX Spark or whatever your mini your mini uh GPU system is going to be uh or or or a very well outfitted workstation or server. And and you can you can install Fuzzball into it in uh one or two commands and you can start running your AI models at the click of a button. So, what is the what is the delta look like? Well, you're going from not knowing where to start with AI to now actually having a pilot. And the coolest thing about Fuzzball is you can take that pilot and you can now go and scale it up.

You can go run it on you can buy more systems and you can go run it on larger infrastructure. You can go run it on clusters. You can actually create hybrid environments and scale up to whether you're running in clouds or you're running on prem uh or multi-prem and you can run this where you can find the GPUs you need wherever they are with Fuzzball and and and start running your workloads against this. And because Fuzzball started off being a uh a computer a very general purpose computing platform, it is not an AI specific platform. It is not an HPC specific platform.

It is a general performance intensive computing platform that you can create workflows that are very highly defined for exactly the applications uh to quote Addison on this, the workload matters, and you can encapsulate your workload into a fuzzball workflow that you can then submit it to any fuzzball system, wherever it exists, and have the same reproducible outcome. And that's what fuzzball brings to the table is it is a computing platform built from the ground up in order to make your lives easier and to to scale and drive security and compliance. So, that's that's that's the the big pitch for fuzzball if I were to put it into a into a short nutshell.

Um and and we're seeing early adopters of of fuzzball on the HPC side and on the AI side. And what you can do with it, again being a computing platform, you can pretty much do anything you want with it. And it is best at doing things that are performance critical. >> An exciting piece of tech. Thank you so much for your description there, Greg. That's all the questions I had, but you know, I'd I'd like to open this up a little bit to Q&A. Again, anybody listening, please put any of your questions into the comments here. I see a question from Richard. I think that goes back to the beginning of our conversation.

Is it safe to say AI isn't always HPC but AI at scale should be considered HPC? >> I mean, that's something I would like to be pedantic about. I'm an industry analyst. I am a professional pedant. I I get to like make methodologies and definitions and define things. And there's what the pedantic answer is. I would like to say, "Yeah, the AI at scale technically is an HPC workload, and we're going to count it all in the HPC envelope." But then I'm also the boss of a really small company and I don't get to control the zeitgeist of how the world thinks of the of of AI.

And the fact of the matter is people are going to look at powerful computing and point at it and say AI regardless of what it's actually doing. This is AI, that's AI, it's all going to be AI and for me to sit on the sideline and say, "Actually, that's HPC." It's not going to do anyone any good. I have to get on board and say that's the AI market and and now underneath that I can define this is the segment that we call HPC, but you know, to say AI at scale, that's all HPC. It's just not going to fly. Politicians aren't going to think that way.

>> So, you know, it's interesting. I've seen some fairly significant um knee-jerk reactions to this. And uh I think a lot of it is because the HPC side of the house sees AI as a workload for an HPC-like problem. The AI side of the house says, "This is something brand new and it's never been done before and we're This is not HPC." So, it depends on which side you're looking at and who you want to offend most. Uh so, if from my perspective, and I kind of generally call it now just performance-intensive. This is a performance-intensive or performance-critical workload. Uh anything performance-intensive, well, kind of fits into the high-performance computing.

That's the whole point of high-performance computing. Uh so, I would absolutely agree with Addison. This is an HPC problem. Whether somebody would call it that now becomes more of religion and and what they're comfortable with. Uh but I think the point that nobody can argue and nobody would debate is we have decades of experience in high performance computing of how to parallelize and scale applications uh as Addison was saying. Instead of like on a virtual machine uh in a virtual infrastructure, you have a virtual machine to to uh oversubscribe your hardware. So, you can have 10 virtual machines all running on one machine all thinking that they're the only ones there.

High performance computing and in AI, we do the opposite. And we have a whole bunch of computers that are pretending to be one, that are all pretending to work on the same workload, the same job. And uh that is a high performance computing workload. Now, are people going to call it that? Probably not. Uh but it is what it is. Uh we are talking Yeah, go ahead, Addison. >> Yeah, this answer comes with a dad joke, right? You know, where does a mansplainer get his water? Well, actually. >> [laughter] >> You know, nomenclature's actually a very funny thing to talk about as well. So, um uh I've talked about this before, but I'll I'll I'll summarize it a little bit.

Uh so, I'm also one of the co-founders uh of CentOS. And a lot of people know of CentOS, little Linux distribution that kind of ran the world at some point. >> Heard of it. >> And um uh but the name, CentOS, even as I say it, is CentOS. Uh I was running the project that gave way to CentOS. It was called Chaos Linux. And when we first created CentOS, we first called it Chaos EL. And that name did not land well for a number of reasons and people wanted its [clears throat] own name and there was a person in in the community who proposed the name CentOS to me.

And I said actually I really love the name. I said but I don't like it when we say it sent OS. Because it it yes it is an OS but it kind of emphasizes or separates out the sent part of it. And sent is commonly referred to as money. And even though it's a small amount of money, it's this is always designed to be a free operating system. So it never sat well with me. So I said yes we can call it that but let's always make sure it's called CentOS not sent OS. Well, we all know how that one landed and I'm still on my hill alone calling it CentOS.

Well maybe not alone but it's not as big as the hill of everybody else calling it CentOS. And for a little while there when I was young and stupid now I'm just for the record old and stupid. Um when I was young and stupid I used to try to die on that hill on the nomenclature. I've given up at this point. Call it whatever you want. I'm cool. But if you go back to you know titles versus roles title doesn't matter so much the role does. The role is performance intensive computing. It's high performance computing. That's what we're doing with it. In terms of name whatever anyone wants to call it I'm cool with and there goes my camera again.

>> Now [laughter] we must be done. Your camera is signing off. >> Yeah, you know we're about an hour and 15 minutes in. That's a a good signal I think. Uh may maybe uh >> Good signal it says no signal right there. >> [laughter] >> A signal that says no signal. Maybe just some closing remarks from you too on this one. I'm going to start with Addison. >> [laughter] >> I I've made my closing remark. My closing remark is workloads matter. AI is great. Workloads matter. Yes, do AI. The workloads matter. Know what it is you're doing the AI for. And then figure out what solution fits that.

Not all workloads are the same. That's my closing statement. >> Greg, you're muted. >> muted >> Yeah. >> Did I just did I just now? >> [laughter] >> Um From from from my perspective is and okay, my camera is just not happy now. So, that's just that is so embarrassing. I've never once had a problem with this camera for like the last 6 months that I've been using it and now all of a sudden, yeah, it can't it can't manage itself. So, >> Greg's camera doesn't work and Addison's telling dad jokes. I think this webinar is going off the rails. Do we have any viewers left?

What are we doing? >> [laughter] >> Um from my perspective is um AI is is a is a new way of thinking about applications or it's a new type of application running on performance intensive workflows in infrastructure, excuse me. My camera has me all flustered now. Um So, leverage what we've already learned in high performance computing. And let's not reinvent the ground the the the wheel from the ground up. Let's leverage what we already know. Let's move forward in a way that's going to bring easy wins, scalable wins, and sovereign wins. And the last point I'll make is on local models. We have seen a tremendous amount of innovation and capability coming out of the local models or local open open weights, however you want to describe them.

We've seen a tremendous amount of capability enhancement coming out of that sector of the ecosystem to the point where you know, the larger GLM models are now being compared to Opus. It's just a matter of time before the open models will be able to do pretty much everything that an organization is going to need to do around AI. And we built Fuzzball to help you roll that out whether it's on a small scale or it's on a giant company-wide organizational scale. That's that's the direction of where AI is going and I think the writing is pretty clear on the wall at this point. And we're here not to do another shameless plug, but CIQ is here to help.

>> I'm here to talk to you about it, Greg, anytime you want. I can talk about that topic for another hour, but we should save that for the next webinar, my friend. >> That would be deal. And I'll get my camera working by then. >> [laughter] >> Yeah, so I think we may want to sign off before we lose any more cameras, audio. >> [laughter] >> Addison, Greg, thank you both for taking the time and thanks to everyone who joined us here for this webinar both here physically at ISC 26 in Hamburg, but also online. On behalf of Intersect 360 Research and CIQ, you know, we appreciate you spending this this little more than an hour with us closing in on an hour and a half.

We'll follow up with resources from today's discussion and you know, we just look forward to seeing what you build as you move from pilots to production. Thank you. >> Thanks, everybody. Take care.