AI Accelerators Enhance Computational Might Across Industries; Innovation and Open Standards Key for Adoption

Hello, everyone. Welcome to another episode of Experts Talk. It's a beautiful Thursday morning here. We've been going live Tuesday, Wednesday, and Thursday this week with experts discussions. And, boy, am I back to be or am I happy to be back, excuse me, in the hot seat here for some good discussion with some quality experts. Folks, I'm Daniel Litwin, the voice of b two b. Welcome to another episode of Experts Talk Market Scale's premier debate and discussion roundtable where we sit down with the top voices in your industry to talk shop on the technologies, trends, timely news, the big topics, the market movers that are defining the development of your industry. And, again, sitting down with the voices, the thought leaders, the experts, the professionals, the consultants, researchers, professors, you name it. The folks that are making it happen every single day, we curate the discussions, we bring them to the table, and we open up the floor to really get actionable with how these trends are defining the future of our industries and how we can get involved and make sense of them in real time. So thank you again for joining us live here this morning. Make sure you're heading to market scale dot com for previous episodes of the show as well as clips from previous episodes and, more info on upcoming broadcasts as well as we're going live every week with more Experts Talks. Alright. Let's get right into the meat of it. We've got four panelists today, which means we've got a lot of voices and a lot of ground to cover. So let's do a quick little intro here. We're gonna be obviously talking if you see below, we're gonna be talking AI accelerators. Right? So the AI accelerator industry, this need and this development of high performance parallel compute hardware for efficient processing of AI workloads, This whole ecosystem is currently experiencing significant growth, and this is being driven by advances in technologies and increasing demand across various sectors. AI accelerators are making AI powered compute at the edge a reality. Right? They're improving, for example, diagnostic tools and patient care. They're helping develop, advanced driver assistance systems. They are as most people, know them today where they're seeing significant growth, powering high, high compute workloads in data centers. And they're handling high frequency trading and risk management by analyzing vast quantities of data to make split second decisions. And those are just some of the use case examples. But, really, what we're seeing here is the adoption and the development of a foundational element of this larger AI revolution and AI compute revolution. So with AI accelerators acting as that crux of the AI revolution, naturally, it's a pretty competitive battlefield. So what we wanted to open up discussion on today is what is really separating the cream of the crop in this AI accelerator sector. Right? What does the development of future accelerators look like both in a competitive sense and also in a use case sense? Right? What is driving, future capabilities, for AI accelerators? Where is this sector most competitive? Where is there maybe room across the industry for a little collaboration, a little standardization, to support AI's use case development, obviously, supported by these accelerators? And what is really making these accelerators become the defining factor in AI's technology, but also market development. We're gonna get into all that here in just a second. First up, though, a little word from today's featured partner on Experts Talk. That would be Intel Network and Edge Solutions Group. Intel's advanced edge computing solutions are designed to propel your business into the future. Imagine real time data processing directly at the source of data collection from everywhere from manufacturing floors to bustling retail environments. But why stop at the edge? Intel's innovative network connectivity technologies also ensure that your data not only flows seamlessly, but does so at lightning speed with cutting edge advancements in five g, Wi Fi six, and beyond. So whether you're looking to enhance your manufacturing precision, you're securing your critical infrastructure, accelerating service delivery and telco. You name it, Intel is providing the architecture and the tech that forms the backbone of your enterprise's connectivity and computing. So don't just keep up with the industry, set the pace with Intel Network and Edge Solutions Group. Alright, team. Let's get to the discussion. I'm pleased to welcome our panelists here. Let's let the experts talk on experts talk. We've got a full suite of experts here today. Welcome. We've got some returning faces and some new ones, so I'm excited for this roundup. Let's go down the list. First up, we're joined by Mark Beque. He's a top AI market research analyst. Mark, welcome back to the show. How are you doing? I'm good. How are you? Doing very well. Thank you for joining us again. We're also joined by Grant Powell. He's founder at Curios. Grant, how are you today? Great. Thanks for having me. Absolutely, man. Thank you for joining us. We're also here with David Fellows. He's chief digital officer at Acuity Knowledge Partners. David, welcome to the show. Yeah. Thanks for having me on, and, good afternoon from London then. Absolutely. We've got an international presence here today. And we're also joined by and returning mister Joel Polanco, segment manager at Intel. Joel, great to have you on. How are you? I'm doing well, Daniel. Thank you for having me on, and looking forward to the talk today. Absolutely. Alright. Mark, Grant, David, and Joel, our panelists for today that are gonna help make sense of this evolving and highly competitive AI accelerator market. So let's just get right into it, folks. You know, I wanna start things off, a little elementary, but I want to just make sure that we prime our audience for understanding why we're even having this discussion in the first place. Everyone's talking about AI accelerators. Right? Not only as a hot market for investing in. There's always some stock games to play here, but it's also really the foundational element to AI's future development. More and more players in the space are realizing that, and they wanna get on that wagon. So let's get a primer for the audience here. If you could all give us your, analysis here and your take on why are AI accelerators such an important layer of the AI tech stack. Right? What what role will these highly capable AI accelerators play in growing and establishing AI use cases and AI compute as a standard? Give us that pulse check on why we are and really should be having this discussion today. I'll take a shot, Daniel, real quick Yeah. Since I'm the generalist. Let's do this. When I think there's some there's a little there's a lot of confusion about AI compute, but you can break it down into a couple of pieces that people need to understand. There's, there's AI there's training, which is a big load that typically you talk about parallel, compute, and that's a GPU, typically, a GPU, role. But there's AI inference, which is the everyday checking and, you know, working with a model, and that's a that doesn't discussion around NVIDIA and their GPUs, and they were limited. And if you kinda work backwards, the Gen AI threw all this into Flux. There was a reason for that. It was that, there wasn't really, an accelerator that was cut that was really purpose built for GenAI, the the workloads that we're seeing. So part of the background to this, in my opinion, is there was a scramble. You know, there's been a scramble going on about what to do. And I think it'll be interesting to hear what Joel and and David and, Grant have to say about this. But I would say what we've seen over the past eighteen months has been something I would kind of say is the parallel to the sixties space race and how well, some of our chip manufacturers have, answered the bell as far as getting, working on and putting into production, chips and accelerators that are going to address it. So I'll leave it there and let come back to run through that later. Yeah. So I I'd like to just add to that that you could kind of ask this question in subquestions because if you look at AI in real world viability and real world use, you have a few components. You've got the hardware. You've got the software, which is the intelligence, and you have the data, that's being pumped into this intelligence so that it knows more and more. And then, of course, you have the the applications of that AI. And so when you you ask the question of of accelerators and how important are they, I mean, fundamentally, because everything is going to either, you know, be an AI infrastructure business or is going to incorporate that AI infrastructure to change an industry. And that's just a little bit of how I would break it down. Yeah. Just to build on top of, on top of what what Grant said. Like, so, my company and I are in the game of building the applications that ultimately get put into people's hands. And, Mark's analogy about the space race actually is as good as any as I've heard, about the rate of change over the last, you know, year and a bit. Big thing for us is now, is economics, right? And compute costs money. And, you know, you've got one commercial model and there's a price for that. And now there's kind of a whole bunch of open source models and there's prices for those. So not only are we looking at the quality and output of the models now, but we're actually kind of looking at the cost of what it actually takes to actually output that output that quality from the models. Yeah. I would say, to David's point, it's looking at cost. Initially, cost probably was not, looked at as much because a lot of the purchases are being made by companies who, you know, really have very deep pockets. Right? And they're much more concerned, with with time to market, you know, the Microsofts and the Googles and and the Amazons, because they're in a race to get their services up and running. And, you know, to to Mark's point, things are evolving. This is still early days. This is much like the Internet in the late nineties. You know, there is a large infrastructure build out happening right now, and that is to get ahead of all the services and the applications that are gonna be delivered and all the amazing things that are gonna happen as a result of this. And, you know, I would I would say here, the at the core, these AI accelerators are like engines in a car. And today, you see several engines, at play. You have your internal combustion engines and you have your electric motors. And these these two types of engines, they operate better in different environments. And that's kind of what you see on the chip side. You have the GPUs, which arguably are your internal combustion engines. They are hungry for energy. They require a lot of power. And then you have your other AI accelerators, your neural processing units, your central processing units, and your FPGAs and ASICs, which don't require as much power, and operate like an electric motor, much better in other environments. And so we like to refer to this as heterogeneous computing. And that's your ability to run different, applications and workloads on different chips. And, you know, what we're trying to move towards is an open ecosystem where, you know, as you as a developer don't need to worry about what engine you're going to run on. You shouldn't have to worry about that. You're worried about, you know, what, application you're delivering and for your customer. You don't want to have to worry about this low level stuff, and and and that's that's the problems some of the problems that we're trying to solve. To follow-up on that, you just popped the question in my head here. What is the market education like today on AI accelerators? Like, when, the end users here, the b to b end users are looking at investing in AI accelerators, whether that's in their data center or that's in some kind of manufacturing operation or even to embed in a consumer device. Is the market education there to fully make sense of what you're investing in and the, impact that is going to have on whatever use case you're looking to deploy it for or not. Right? Where where are we in that timeline? Because as Mark said to start things off, there's still quite a bit of confusion in the market. So help us make sense of where where is that education timeline at today. I'll take a shot real quick. And you guys nod or or shake your head if you disagree. But, what we've got a little bit is I'm gonna focus back on what's called data center, you know, compute. So that's the really heavy load computes. What you have is cloud versus on prem. And, so you have a lot of enterprises that have their own data centers, and they they call that on prem. And they and so you ask where these things get bought. The those all of these sophisticated accelerators get bought in the data centers either on prem, so that's an enterprise buying it themselves, or by the cloud players being, you know, the Amazons, the, Azures, you know, Google, and those types of things. And what there's a debate, I guess. You know, you can say, it's not a real great time to start building your on prem if you wanna start from scratch. But, you know, compute is a a a narrow it's a limited commodity right now, I think, we're seeing in the marketplace. So a lot of enterprises are making decisions. If they don't have on prem, they're everybody's thinking about the cloud. Is this a good idea or not? You know, there are pros and cons to those things. The pros being you're up and running much quicker. There's some scale to that. The cons being, you know, how you handle your data. Security is always an issue for lots of companies, but that's where I would say is kinda one of the bigger challenges around what we're talking about. Yeah. And I'd I'd also like to add a a question about, you know, edge computing because we're talking about cloud versus on premises and, you know, incorporating, AI accelerators there. But at what point, you know, how far out in the future could the edge computing chips have enough processing capacity to not need AI cloud computing? And is that something that's taken into consideration, when educating people about either investing in AI accelerators or making a decision for their company and how they're gonna structure their their models and their requests and their hardware? For clarity, for me, I call that on device AI. So if you you're talking about more of a, personal chips and personal computers and, you know, mobile phones and edge devices. And to me, those are two different things, and I I'm not sure I would call those accelerators. I don't know what Joel and David think. Yeah. I I would, comment, like, you know, one one little known fact is actually Apple has had a NPU, a neural processing unit sitting in the iPhone unused, for several years now. And so it shows you the build ahead, the thought that's going into the future here with some of these companies. And, you know, now you see within the PC market large launches, for neural processing that is gonna go into our laptops and our desktops. And so, you know, I think with regards to edge computing or on device, it's really just getting started. And and it's always been something that's present where you have, central computing in the cloud versus your edge computing, and it's a constant push and pull. Right? And, on the edge side right now with AI, we're just like, it's really just getting started. And so you're gonna see more and more applications that land on these devices and start to pull some of that compute away from the the cloud. And, you know, our belief is that, there's gonna be value there for the user in terms of privacy and security for their data. There's also gonna be some value for, the the cost savings on the cloud side because we know that, you know, the energy costs are are going up dramatically. And so if you can run some of these workloads on your device, you're gonna save money in terms of energy and data transfer. Because if you don't need to send data back and forth, you know, you should you should try to avoid the cost of doing that. Right. And then, of course, you know, the other part of that conversation comes in is the model itself. So if you build an efficient model, it's gonna require less computing that could potentially run on a on a device's, you know, processing unit, which is gonna be smaller and less capable in capacity than, you know, parallel cloud servers. So the model and and how models are designed and engineered has to be part of the conversation as well. Yeah. I I agree with that. Just as a consumer, like, when we go to the supermarket, you look on the back of your, your your pork chops or your broccoli or whatever you've got, and there's a thing on the back of there that kind of gives you information about what it is that you consume in. And, the hyperscalers, GCP and Azure and AWS have started to do a good job of this. They've started to be able to tell people, you know, how much carbon are you actually using of these types of things. And to be honest, you know, my kids are super interested sort of like in the environment. So, you know, really, you know, the energy consumption really, really matters to them. So being able to sort of to see what is actually being consumed as you use some of these things, I think is really, really important. I've got a question for Joel, if Joel, if you don't mind. Do you think that, AI computing at the edge is going to be constrained by power consumption and, you know, what we can actually, you know, put in cars and how efficient we can get these chips? Or is there is there a lot of headroom still? Yeah. I think that, AI computing in general is gonna be consumed is gonna be, constrained by energy. So whether you're talking about the data center or on the edge. And what I mean there is is I you know, you you heard Mark Zuckerberg talk about, their build out of AI. And the reason that happened was because of competition with TikTok. Right? They they needed to spend a lot on infrastructure, in order to build out their ranking and recommendation, their bill ability to, generate videos, for their users. And he talks about, you know, where data centers today are know, hundred up to a hundred and fifty megawatts, potentially going to one gigawatt. And I had to go look this up. So, you know, I'm in I'm in Arizona. We have a nuclear, facility here in in in Phoenix, and it produces about twenty percent of our energy for the state. And it is a three gigawatt facility, right? And so now you're talking about building a one gigawatt data center, which is going to need its own nuclear power plant, which is just, you know, mind blowing, right, when you think about this. I think on the edge, you're gonna see this the same sort of challenges where, in mobile, we we had challenges trying to meet the power consumption requirements for these mobile devices to to have a long battery life. You know, if as you're running more and more AI, you can only assume that you're going to be consuming that battery. Right? You know, but to circle this back, I I would say I'm finding and Grant mentioned this, David mentioned this. You know, the market is the under the market understands that it's not sustainable to have these huge loads. They they they're just not it's not sustainable. It's not business sustainable. And what you've seen are multiple fronts. You know, it's across the board. The the the foundation models are getting skinnier. I call them you know, there's large language models, but there's really a huge slew of things where these are morphing into what I'd call small language models, which are more purpose specific. A large language model is really meant for one major certain kinds of things where they're not necessarily super purpose built. So we've seen this huge movement over a very short amount of time where these, in general, the the foundation models are getting more efficient at how they run. So that's that's one thing that's happening. You've got all sorts of hacks within the development community and how to work with these things and make them run more efficiently so that it's cheaper. And then you have all of our friends like at Intel. Part of the driver for them, if you look at their measurements, is how much faster can we run these things, how much more efficiency can we run out of the chips we're we're building, NPUs, CPUs, GPUs, whatever it is? And there's this is across the board. That's what I meant about a space race in that sense. It's all pushing forward. The the the market dynamics to me would not hold with the super idea that this would be too expensive. It doesn't make sense. Right? So we have to get better, and it looks like there's a lot of momentum in that. Yeah. I loved what you said about the models, and it's like having a a general knowledge AI versus, you know, an an expert. I actually I'm friends with the guys who started collectibles dot com. It's a great company. And they're they're using AI in every aspect of their business you can imagine, the ways in which they're using it. And, what they're looking at is building out individual models, that are that are, you know, miniature experts in certain verticals like coins or sports cards. And then, if you install their app on your device, depending on what you specify you're interested in, it could download that model and just, you know, use that that hyper specialized knowledge on your device versus it having to have all the knowledge of, you know, the world on it. And that becomes highly optimized, requires less computing, less data, less data transfer. So very much in line with what you just said, Mark. Yeah. Actually, I I really resonate I really resonate with that as well, Grant. Right? Like, twelve months ago now, again, just going back to mark's analogy of the space race right twelve months ago feels like years ago GPT-three came out. All right, and we're all running around here going we're gonna have to tune this thing. How do we tune it? And how much is it gonna cost? And we started to look at that and go this this is gonna get expensive real quick. And then we started well, okay, let's move on to a rag based infrastructure So we've done that, you know, and now you've got like a 7b Mistral model, Right? Like these things are getting smaller and smaller and smaller all the time. Yep. Yeah. And so then you're gonna be playing with the hardware. Sorry. Go ahead, Joel. Oh, I was gonna say, yeah, you bring up RAG, and, you know, that's really important concept here, which is just, retrieval augmented generation. And this is the idea that, you know, these large models you know, we talk about the Zuckerberg in in, you know, building out a one gigawatt, data center. Well, those are for these large very large models like LAMA three, which is just a seventy billion parameter, huge large language model. And a RAG model, you know, is is that small model where you can optimize based on your data. And so in the sense of, you know, conserving energy and differentiating your product, they go hand in hand because now you're building the small model. It's it's not just a general, you know, meta, or or OpenAI model, but rather it's a differentiated model based on your data, and it's tailored for your business. So you're not gonna be getting these, you know, really generic answers to questions that you need to answer for your business. And, to everyone's point here, things are moving so fast. I'm kind of coming back to Daniel, your original question, which is education's tough right now. Yeah. Things are moving so fast that you're constantly having to update your education as you go, literally every week almost, it feels like. Like. You know what's kinda cool about that though? And I'm interested, David Mike David's been touching on this a bit. But in my view, I've been watching what I call leaders, you know, that we have Intel's a leader in a sense. They're what I call an AI vendor, because they they compute and all those different things. But if you take people that have kind of done what David's role David's David and Grant's companies do, which is we're, you know, you're selling your things into the marketplace. You're not an AI vendor. You're selling your thing. Right? You're trying to figure out how to apply this. And I look at I'm gonna name four companies that I watch pretty closely that that have been invested in AI for a long time, and they're there's I think we're gonna look to those markets to say, did they do this in a cost efficient way? They're doing this, and it's making business sense. You know, those kinds of things. And you could we'll put Microsoft to size and an outlier because they're both a vendor and a seller of those things. But the ones that I'm gonna point to that I think we should watch are Adobe, Salesforce, Zoom, ServiceNow, big SaaS players. Right? And all of them have been invested for a while and I talked to them quite a bit. You know, they figured out the cost benefit analysis of doing this, and they're all leveraging generative AI. So in a sense, we're at a place where they are saying this makes economic sense to them, and they're able to use these things. They're not telling all their secrets, but it's interesting see that they are moving forward. Sometimes I wonder, specifically with with, Adobe because, you know, in my twenty years of doing digital, I've I've done a lot of graphic work as well. And, you know, how much of Adobe investing in AI is to keep up with competition like Canva, where they have AI that they use to, you know, delete a background from a photo. And Adobe saying, like, well, we have to do this to keep up versus it actually being a cost benefit to their company. It's kinda a bit little bit of a tangent, but it came to mind when you brought up Adobe. Well, when I talked to them, they've said that you think about their marketplace is about images. Right? So they're huge a huge part of their customers is how you manage creatives and do how creatives create content. And if you think about the image generation use case, they understood it. They were working on things in that sense. And so you probably a little bit of both. Right? But they clearly understood that that was something they could lean into if if if it worked out and it's it's working out very well for them. Yeah. The cost of integrating AI, the benefit of keeping their users are getting new ones because it's easier to generate images and render graphics. Yep. So since we're talking use cases, I wanna talk about the dynamic of how the use cases are also sort of impacting the development of AI accelerators. Right? So there are a lot of industries that are driving the development and the need for AI accelerators. We mentioned them, kind of already several times here on the broadcast, but data centers, for example, are now handling massive workloads. Those are clearly at the forefront of AI accelerator, adoption. We're also seeing them in consumer devices like smartphones where they function as co processors, in automotive or other sort of long lifespan applications. They're integrated within the system on chips, right, to provide a broader range of computational capabilities. But what I'm really curious about is how y'all are seeing these various industries actually sort of have that positive feedback loop, right, where they are defining the further development of AI accelerators because of the use cases that are defining those markets. Right? Kind of kind of removed from AI accelerators. It's just what are the use cases in automotive or, you know, network development, edge computing, even sort of smartphone capabilities, right, that are pushing the AI accelerator, industry to redefine their capabilities. And I'm I'm curious what y'all see as those use cases that are pushing the AI accelerator developers to, you know, further craft and define their AI accelerators and how involved are those downstream players in those conversations and in that development? Give us a pulse check on that. Are. Right? I think, verticals and applications are driving, infrastructure development and the accelerator development. My company, Kyrgios, I I I maybe mentioned, you know, is a blockchain company that's focused on what we call the creator economy. And so we work a lot with musicians and artists. And, a musician, if they have the ability to instead of hiring a bunch of musicians to, you know, compile an album to use generative AI to create beats or, you know, tracks, that that they can build the album with cheaper, then they can get to market with less of an investment, get their works out there, and, you know, try to get known as an artist. So I think, that's a specific example. And I think those kind of examples are a hundred percent driving the need for, the hardware and and and infrastructure development. So I think that, Mark, if if Joel said this, I I totally believe this. This is very nascent. I'm I'm picking on generative AI. So, AI has been around for a while, but it really what we're seeing now with, generative AI is maybe scale and possible scale in in making it cheaper to if you think about interacting with models, you don't have to have a data scientist necessarily anymore. That's how I would talk about generative AI and and, what foundation models bring to the table. But the use cases are nascent. I we don't have a killer one yet. Not that's not saying anything against Grant because, obviously, that's his business he's doing. When I'm looking at a larger market, I don't see it. There's there's a few that I think are are are kind of cool. Like, they there's a there's this goal of big data. Remember big data? Everybody wanted to, figure out how to mine all of their data and and do that. So, a lot of people are are thinking about what I'd call corporate search or company search or, you know, that kind of idea where you're able to realize the idea of big data, right, would be a really killer use case, I think. But we don't have a lot that's you know, because this is so new, there's no real proof points on, oh, that's that's gonna do it. That's gonna be everything. Right? The only other one I like that we really see some promise right away, was around, summarizations. They could use those for a lot of, collaborative, tools and things like that, but also, drug drug discovery. The the the pharmaceuticals are having a lot of luck with that. So just just for you a name. Yeah. So I'd like to throw in what I think is a killer use case would be, not pharmaceuticals, but medical. So whether it's surgery and I don't know if you would describe this as generative AI or not, but, I know that there are, you know, AI technologies already being used to, scan and then do a three d model of, like, an organ that's being operated on. And I see a phenomenal application for AI in general, again, whether it's specifically generative or or other AI, to improve medical practices. Yeah. And that's Radiology too. Yeah. And that's something that's meaningful, not just novel. Right? This is something that would change quality of life for humans. I'm gonna expand the, the concept of of an AI accelerator past just silicon, to talk about, you know, the software infrastructure that's been built around that as well. Like, you know, what the the guys over at NVIDIA have done with CUDA. They've, you know, built a whole software thing around that at Silicon. And there's there's other companies, you know, who've kind of built other platforms as well. I'm a twenty seven year veteran of building software solutions and trying to figure out how to optimize those software solutions. And Microsoft bought out, Copilot Studio, a little while ago, and and some people have started to pick it up. But, the the power of that thing, which is like a no code platform that can use things like GPT-three on the back end to deliver, use cases to market really, really quickly, is very, very powerful. And, you know, to be honest, I think that's got a good chance of disrupting a lot of, like, what we would call You're right. Traditional contemporary development. David, do you mean code code development assistance is what you're kinda getting at? Yep. Yeah. Copilot, you know, GitHub Copilot. The all all of these kind of suite of things. Yep. Yep. Yeah. So we're seeing sort of like, you know, a lot of that. I mean, those are I think those are killer applications. Do You know, one thing though that I would say, right? I speak to a lot of CTOs, you know, CTOs like us, right? And a lot of the, a lot of the, the, like the, the large language model vendors, right? You know, they, they talk about things like productivity and efficiency. And, you know, we're still struggling to kind of figure out what does that really mean, right? Like ai OpenAI put the whole Klarna use case all over, you know, the website. You know, that's really, really interesting. But, you know, we're in knowledge work, you know, so we're sort of like trying to figure out, you know, we take sort of like a model, we build something really quickly, we deploy it. And then how do we measure the, you know, the productivity and efficiency? It it's not easy to do, you know, in in in knowledge work. Well Yeah. Just gonna piggyback on that real quick. One of my pain points is that I've I've never met an AI phone prompt that actually helped me. So it's like, hey. Great. We've built these, AI phone assistance. They're supposedly making things better, but are they? And how do we benchmark that? Sorry. Go ahead, Joel. Oh, yeah. No. It's I was actually gonna touch on that same use case, which is, you know, presented, through the human AI pin and then the rabbit, device that came out earlier this year. Right? There's been other One of the things that really needs to happen on the back end is a connection to the APIs of the companies that have the workflows, ones that Mark was describing. Right? The Salesforces, the Slacks, the you know, all the tools and things that we're using and the ability for these, you know, AI assistants, to be able to communicate, interpret, understand, and then, send a command over. Right? So you wanna send an email to so and so. You can't necessarily do that today the way you want, or you can't schedule that meeting the way you want. We're but we're it feels like we're on the cusp. The question is, how how many years are we gonna be on the cusp for? Yeah. Because, you know, you know what I mean? I don't know if you guys saw this, and you talk about it yesterday. I think it was yesterday or the day before Amazon, introduced the agents that are now, GA on their on their platform, and it's very specific to what they this is kinda unique. It's what Joel said. It was like the idea. Right? We'll see if that I'm not saying they're gonna, you know, totally get this. But the idea is an AI agent kind of stitches together what you need done in code development. So it kinda understands pulling from different applications to do you're telling it what to do and it goes and does it in theory. That's in theory. But it's the first one I've seen that really talks that way, you know, kinda and that's interesting if you think about productivity and all those things. That might be really fun. Yeah. That is cool stuff. Yeah. Yeah. Good. Good. Sometimes I I I feel like the missing component is is like context or sentiment. So if I was to task, you know, one of my developers with, like, hey. How can we build this feature or this technology? And he or she would go research and pull together the different pieces. Hypothetically, an AI could do the same thing. But that, you know, CTO or developer of mine is gonna know the nuances of what I care about. Right? Like, well, I I won't work with company x. You know? And so if company x has a solution, pick somebody else. And, like, yes, technically, these two companies offer the same solution, but this one has better tutorials. So go with that one. And so there's this whole, like, context and sentiment, and and I I I feel like that's a big part of AI's viability. Don't don't get me started on that. And I will tell you, frankly, I come from natural language background, and GenAI does not solve the issue that is natural language understanding. Yep. I'll just say that Hundred percent. When we all these people are talking about what chat GPT can do and all these things. You're exactly right. I believe I agree with you a hundred percent, Grant. AI is never going to get the nuance, the sarcasm, all these things. I mean, how long have we been working with? Maybe not never, but certainly not now. We'll see. We'll see. I I saw an interesting, interview that was done. So Reid Hoffman, a, you know, a leading Silicon Valley investor, yeah, he interviewed himself, his AI on LinkedIn. Yeah. And if you haven't seen this, go check it out. It is it's awesome. He, and he actually, he he goes through these corner cases. Right? All the things that still need to be solved, and and it AI's ability, to not capture the emotion of the interview. Right? You know, he said some there's these these things you notice, but, you know, overall, just where things have gone in the last couple it's pretty pretty amazing. If you go check it out, I definitely recommend So Joe, it's like you said. I think think we're kind of, you know, on the cusp of things. You know, I keep on watching Sam Altman. Altman. Right? And the next time he says things like AGI has been achieved in the lab. Right? We're in a whole different world. Right? We're in a whole different ballgame to what we're talking about here. Right? That's right. It feels like we're on the cusp or something, but, like, how far away is the cusp? We've been away from that cusp for fifty years perpetually for the last fifty years. So now if if we talk about the market itself, and, I've I've just got a couple more questions for y'all before we wrap things up. Really great discussion so far. This is I'm gonna have to bring you the four of y'all back again here soon. Let's talk about, I guess, just, like, the the players in the space themselves. Right? We've been talking about the what's what of AI accelerators, but theme of the show is also who's who a little bit. Right? So, obviously, Intel and NVIDIA are top players in the space. But I'm curious what's really defining that cream of the crop in the AI accelerator sector today. Right? Like, what can today's top AI accelerators actually offer for high workload compute operations? How is that making them competitive? And how does that kind of define the ins, the use cases, the capabilities, the partnerships that are making new entrants or existing players continually competitive in this space? Yeah. I'll start real quick and just say I'd add to that list. Thankfully, there are other, chip manufacturers, particularly if you're gonna add AMD in there for, data center, accelerators. But also say there's some smaller players like, don't confuse it. It's called Groq, g r o q, not g r o k c k, not to be confused with Elon Musk Musk and all that. And then if you look at, to Grant's point, you know, about devices, Qualcomm's been very active in making, some on device AI stuff as long as as Intel has as well. So I'll just throw those out there as other players that are helping get us to the spot because we're gonna need everybody to get there. It's a lot of there's a lot there's too much compute and not enough making it. Yeah. I I I know we're we're talking about AI accelerators, like, you know, largely from a hardware perspective, but I I always feel like they have to play in the sand with the software, and the models. And, you know, I think an AI company an AI accelerator company that's gonna win is one who's not only looking, hey. How do we improve this hardware so that it can process the requests better? How can we change the type of requests so that the hardware can handle it better? And so they kinda go hand in hand. And I I believe the ones that'll be most innovative and successful embrace that work with the software developers as well. Yeah. I I would add, here is, we are still in the early innings of establishing open standards. You know? It it we're talking about the Internet in the late nineties, you know, before Linux took over. And, you know, in getting those open standards in place, you know, we believe that that will drive, a a raise all boats situation. Right? And if you are locked into a single closed ecosystem, you know, you you there's benefits, obviously, that's easy to use. You make the investment. Open ecosystem allows for the other companies to all participate. And so one of those ecosystems would be the UXL Foundation. And again, it it is the the mission there is to allow for these developers to not have to worry about what hardware you're going to, compile your code for. That really shouldn't be a worry for them. You should worry about delivering, you know, value to the end user, not not what chip am I gonna run on. Right? I mean, that, Joel's Joel's talking about, there was a lock there's kind of a lock in on CUDA, for NVIDIA. So it's like how they they did that not it was kind of by mistake. But, you know, what what, Intel and a few others a bunch of others have said is you shouldn't have you should have this ability to flow between any of that. It shouldn't be that, that that that software stack to work with the hardware should be open source. Thinking, you know, as a technology product developer, right, for twenty years, I always try to think about what does my end user want, not what do I want or what do my developers wanna build, what do they want. And so back to the question of how how could, an AI accelerator compete and and, you know, be successful in this space, they have to think about who their end user is and what their end user cares about, which is low cost, low energy consumption, ease of integration. Do those things, and you'll get the customers. I think there's another couple of things which people are having to start to think about. Well and, Dan, this is probably two shows in its own right. One is about sovereign nations and what they're doing. Right? You know, outside of what we're doing in the West, there's a whole bunch of other stuff going on over there. And then there's a whole bunch of things around regulation. Right? You know? That this thing is so powerful. And we have to have that conversation about how we're going to try to control this thing. And on top of what Grant said, you know, people are gonna need to sort of, like, think about those two things as we develop this technology further. You are seeding me for future conversations, so I appreciate it because you're so right. Those have already been swirling in the head. I actually had to delete a few questions from the rundown because I knew they would be twenty minute, you know, discussion points just in and of themselves. But but, yeah, last but not least, team, before we wrap things up, Joel brought up standardization. So let's talk just a smidge about AI accelerator design, I suppose, or just kind of how the development of AI accelerators is evolving in real time. How would y'all say the industry needs to balance this need for further innovation in a very nascent market still in the grand scheme of things? Balancing that with the push towards standardization in the AI accelerator space as we're just starting to dabble in these, established standards. Right? Do you see standardization in the near future stifling innovation in this space, or does it provide that necessary foundation for broadening technology adoption of AI accelerators and AI computing more generally? What do we think there as we wrap up the show here today? I'll just say quickly that I think standardization in a nascent industry is really hard. Right? Take a look at everything from laptop power cables to, like, HTTP encryption. Standardization didn't come until later when the people developing the technologies had a better understanding of how they'd be used. So it's a really tough topic. I go back to what Joel Joel hit the the he nailed it. The only piece of standardization to me, in my opinion, right now is that layer that he's talking about. So it when we're talking about accelerators, because they're gonna make what they make. You know, you can't that would mess up innovation. You you know, everybody's running as fast as they can with the smarts they have. The the layer where you need some standardization is how somebody uses that. Right? So the the layer he's talking about is the software stack to use hardware should be should it makes sense that that would move forward. It's gonna take a little time. I don't know what Joel thinks the time frame is on that. But to me, that's the only piece that needs to be. When we're talking about accelerators, what needs to be standardized? Yeah. How you data is gonna be regulations. Yeah. How you interface with the hardware, not the hardware Yeah. Exactly. It's it's, referred to as the hardware abstraction layer. So you're you're abstracting away the complexity. Right? And and that layer is, is, to get all these companies, you know, to come on, like Grant said, you know, to get them all coming on, moving in this direction, it takes time. And and developing open source communities take time. And so part of the education is, hey, this is happening. You know, they're the UXL Foundation, for instance, is, building this community. There's many companies participating, Intel, Arm, Google, Samsung, just to name a few. And as more momentum gathers, then your community builds, your standards begin to move forward, and then you have more choice available because that choice is needed for all these different use cases. You you don't want to put a, semi truck engine inside a Corvette. It just does not make sense. Right? And so you need to have this variety. You need to have this, availability of different options. And I think that's what customers are looking I I think I think that's right. I I would I would also offer that over the last eighteen months, the the amount of change and churn in our industry has just been unprecedented. And I include, you know, the data center shift and the shift of the hyperscalers. I think that if you're a business leader right now, you just have to be able to deal with the dynamics of the market. You know, what we're kind of talking about here, with the models and the AI infrastructure, it kind of rolls all the way through all the stacks. The models are changing all the time. The hardware is sort of, like, changing all the time. And I I think to be successful, you've got to get your head around. It's gonna be like that for a little while. Yeah. Alright, team. I think with those final thoughts, we'll go ahead and wrap up the show. This has been an incredibly insightful conversation. Per usual, our tech forward conversations are some of the most dynamic, so so I appreciate all four of you bringing your a game with your analysis here today. But more importantly, I feel like we laid the seeds, like David said, for another three, four, five conversations here each an hour in and of themselves. So we'll definitely be back. I'll be reaching back out to get this squad back together for further analysis as the AI accelerator market continues to develop, and we see new entrants in the space, new big news, and new use cases that are really defining how AI accelerators are going to, you know, be that defining crux, I suppose, right, of the larger AI compute revolution. But till then, we'll go and wrap things up. So thank you to the four of you again for your time and perspectives today. Folks, we've been hearing from Mark Beque, top AI market research analyst, Grant Powell, founder of Curios, David Fellows, chief day excuse me, chief digital officer at Acuity Knowledge Partners, and Joel Polanco, segment manager at Intel. Mark Grant, David and Joel, thank you again to the four of you for your time today. It's been such a pleasure. Thank you. Thank you very much. Thank you. And thank you everyone for joining us on another episode of Experts Talk. If If you like what you heard and saw today and you want previous episodes of the show, you know where to go. Market scale dot com. We have not only full episodes, but also snippets, little bite sized chunks of the hottest analysis from each of our shows available, again, on market scale dot com, as well as information on future broadcasts. And if you yourself consider yourself an expert and you wanna get in on the hot seat and join some of these roundtable conversations, hit me up on LinkedIn. Give me a ping on email, daniel dot litwin at market scale dot com. And who knows? You may just find yourself here in the hot seat with me for another Experts Talk. Alright, folks. Thanks so much for joining us, signing off for the week. We'll be back next week with more discussions. But till then, I'm Daniel Litwin, the voice of b two b, and we'll catch you on the next episode of Experts Talk.