theCUBE + NYSE Wired: AI Factories - Data Centers of the Future | Alexis Crowell, Axelera AI

More from theCUBE + NYSE Wired: AI Factories - Data Centers of the Future

Alexis Crowell

CMO & President of the Americas

Axelera AI

Info

Alexis Crowell, Axelera AI

Alexis Crowell

CMO & President of the Americas Axelera AI

search

(INTRO)

John Furrier

>> Welcome back to the Cube, live here in Palo Alto for the third annual Cube AI Summit for the leaders of infrastructure build out. Of course we have our party tonight, 180 people reception, where they'll be sharing their stories about what's going on, sharing, comparing notes, sharing their stories, but it's the biggest build out we've seen in the industry ever, and you're starting to see it happen, enabling all kinds of use cases. We have a returning CUBE alumni, Alexis Crowell, CMO and President of the Americas Alex Zahra. Welcome back to the CUBE. Thank you, thank you so much for having me. Thanks for coming back on. We had you guys on in Mobile World Congress, which was smashing because everyone was talking about Edge, which fits in nicely with what you're doing. and GTC, obviously that's the mega event. And then two weeks ago at the RACE Summit, you guys were on again as robotics and agents kick in, which highlights a lot what you're doing. It does. Give us an update on the latest and greatest.

Alexis Crowell

>> Happy to. So as you know, right, we started intentionally at the edge so that we built really efficient performant compute. So we've got customers that are putting us in drones to go search and find people and help prevent against wildfires. So there's some really interesting, fun stuff happening there. But the product line scales into servers and it scales into on -prem enterprise multi -user LLMs that are really solving a lot of what people are concerned about, which is what happens when I throw all of this compute into the cloud and then I lose a little bit of control or I don't know what my token costs are going to be. So we're helping bring some of that back on -prem with the server solutions that really really maximize infrastructure capacity without having to completely rebuild your data center.

John Furrier

>> Talk about that token cost equation, how that relates to enterprise, because this is the biggest conversation happening, because compute and GPUs and the role of pre -filled decode, disaggregated infrastructure, disaggregated inference is coming up too, so starting to see smaller footprints, different use cases, what's the impact to the enterprise?

Alexis Crowell

>> Well, it really comes down to what are you trying to solve for? My favorite example is if you're a legal aid and you're trying to help build a contract, you don't necessarily need to have a language model that had French poetry trained as well, right? Like that didn't need to be part of the source material. So you can use a much smaller model and get really accurate, really important productivity out of that scenario. And that's where solutions like Accelera really play in because we can do all of that at the edge without having to leverage big frontier models. Those have their place. I'm not trying to knock them, they're amazing. I use them. But they're not necessary all the time. You can use an open source model and really solve for what you need.

John Furrier

>> Talk about where you guys have the momentum right now because as the market develops, we see the AI factors in the service for sure. Rack scale systems, those are the monster racks. but they're starting to see different levels of clusters, of systems, the edge is out there. We're seeing, you mentioned drones. Yes. Sovereignty is about control. Yeah. Enterprises love control and they have their deterministic workloads and the agents, but talk about where you guys are winning your deals, where momentum is.

Alexis Crowell

>> Absolutely, so there's a few different places. When we were founded, Vision was the entirety of the market, so we are still really winning in Vision for sure, but now it's not just Vision. It's vision, search, and summarization. So companies that are trying to understand what's happening in their retail chains, right? We have a huge convenience store in South America that's deploying us to be able to not only monitor inventory, but then to real -time update and understand what's happening and then use reasoning to actually go action on that. And you can do that both out at the edge within the store with a single chip, or you can do that back in a centralized location on a few chips across the entirety of your stores.

John Furrier

>> So people want to bring their LLMs to their own LLMs for the company. They also use the general purpose, large scale. So you've got general intelligence, that's the internet, open AI, cloud, that's great. But then you start to see the specialty AI. How do you guys talk about those two forces? They're not necessarily mutually exclusive in some ways.

Alexis Crowell

>> No, they're not. We very much see them as complementary. It's very similar to tools. If you're going to go build a house, you need a hammer and a screwdriver and a saw and all these things. right? Each one of them has their purpose. Where we find ourselves is not so much just a hammer in that, you know, you can hit anything with it, but it's more of, as you know what your use case is, you're going to want hardware that is really purpose -built and optimized around that. Now, the good news is, is we're completely programmable. So we can, if it's inference and it's matrix math, I can run it. But more importantly, it's highly, highly optimized for power, meaning energy costs, and performance from the edge all the way through to where it allows customers to say, okay, maybe I don't need that much performance today, but it gives me some headroom to move into what's coming next.

John Furrier

>> You know, the inference conversation has shifted a lot. I want to get your thoughts. You had mentioned earlier at the top of the interview about cloud and relationship with the drones. Pulling stuff from the cloud to the device is something that's happening. But AI inference has many different use cases. Where do you guys see that, and how do companies that have a lot of cloud pull that down and bring that capability, that intelligence to the device?

Alexis Crowell

>> You know, there's a number of different ways, right? There's the, I have a local model that's only trained on my data, mine is a Lexus, like it's got my bank accounts, my things. I don't necessarily want one specific company to have access to all of that in one place. You're starting to see more of an enterprise grade solution actually end up in homes for those types of things, right? On the enterprise side, you see companies companies that have trained entirely on their data. In our case, as an example, we built a tool called Wingman. And Wingman is an entire LLM. It uses the backbones of some of the major models, but it's built on all of our documentation, all of our GitHub repo, everything to make programming really, really simple. And now in minutes, you can get something up and running. So the reason I say all of this is it really depends on what problem you're trying to solve, right? And that's always the first question is, what problem are we trying to solve? and then let's go figure out what tech stack makes the most sense.

John Furrier

>> You guys have the AI accelerator, you've got the systems and the software. That's right. It's interesting, I was having another conversation with another company in a different area and they make silicon, but they said, no, no, we don't want to be known as the silicon player, but they also have cybersecurity in silicon. Yeah. I don't want to be known as a cybersecurity company. He's actually a controller, the software is driving it. You guys have a very similar pattern. How do you tell that story? Because you don't want to get pigeonholed into one bucket. Right. Because you're doing all, but the software tooling is critical for developers because they have to do the work. Exactly. I want to get my stuff in there, get intelligence injected in. Talk about that dynamic, because you have a lot of piece parts, but they work together with the software. Why is that important?

Alexis Crowell

>> They're designed together very intentionally. I mean, we have as many folks working on the software side of our business as we do on the hardware side. And it's not that I'm trying to monetize the software. I'm trying to, to your point, make sure that developers have a frictionless experience. variants, right? Because if I go in and I try and disrupt what a developer is doing, that doesn't work. You and I have been in this industry a really long time. That doesn't work, right? So our goal and the reason we co -design our software is so that as somebody goes and buys one of our cards or one of the systems, they can just start, right? It's all PyTorch. It's all Onyx -based. It's all the same stuff they're doing today. It's just landing on a much more performant per watt, performant per dollar solution as a result.

John Furrier

>> One of the things you guys, I'm impressed by, and you guys talked about it in some of the interviews, is the ease of use. The ability to stand up something so compelling, so fast. With the coding assistant agents out there. So all this is coming to the table. So if I'm a developer, explain to me, I already kind of know the answer, but explain to me if I didn't know, what's in it for me? Because right now I'm coding maybe on a box that's Nvidia, and then okay, I write software, but I want to move it into say a robotic system. Okay, does the right chip work there? So now I got to port that code, cause I'm developing here. This has become kind of, I won't say, it's kind of a friction point. For sure. How do you guys address that? Cause the developers just want to, I don't mind coding on a super device or the box itself, but I got to get it into smaller form factors.

Alexis Crowell

>> Yeah, the good news is, is if you've coded it on any platform, you are likely already using PyTorch, right? From an inference perspective, that is by far the dominant platform. That is what we support. We support Onyx as well, but we support the major solutions. And then our tool chain allows that port really fast, right? You literally have a chatbot that is hyper -optimized on our system and our documentation to where you can have 20 % to 30 % better accuracy than just running Cloud Code alone, right? Cloud Code is amazing, but it's not hyper -tuned on what we do. So we built a tool to make that port really easy and not make it hard for developers.

John Furrier

>> And they get what they want.

Alexis Crowell

>> They get what they want. They can test it. We've got cloud deployment so they can test it. Try and make it as easy as possible.

John Furrier

>> Okay, so a lot's changed since we last talked when you were on. I'm expecting it to be just as crazy speed -wise. What's your focus now as you go to market? The market's hungry for solutions. It is. What's your strategy? What's the plan?

Alexis Crowell

>> Well, the good news is our strategy hasn't changed. So that is fabulous. We already have products in the embedded space, the edge space, into the server space, and we have deployments with customers with that. But our focus, candidly, right now is just making sure we continue to deploy those and support the customers and feed the demand as best as we can. As you know, we announced Europa last year. It's on track. It's coming to market this year. We've got good news coming soon. So all of that. We're executing. We are. We're absolutely executing.

John Furrier

>> What's next? What can we expect without giving away the farm? Yeah. Sharing too much information. But you're a private company. You're not public yet. Yet.

Alexis Crowell

>> Yet. That's right. Yet. Exactly. So we have always had the strategy and the vision from before we were funded to go edge all the way into cloud. Right. And we we've announced that we are part of a program called DARE, which is the digital autonomy for RISC -V in Europe. We are part of that in so much as it's helping us architect what a cloud solution would be like based on the hyper efficient edge architecture that we started with. Right now, obviously there are constraints, you can relax, you can get more power, et cetera, in the cloud, but it's that same architecture that will flow through, and that's what we're working on.

John Furrier

>> All right, well, put a plug in for what's going on in the Americas. Yeah. You can see I'm also in charge of the Americas. What's new, what are you looking to do? Hire, the goals, people watching, customers.

Alexis Crowell

>> We're constantly looking for really good talent. If there's folks that are interested, please call me. We've got a couple of really big shows coming up where you'll see even more of what we're doing, both at AI Infra Summit in September. Please come join us there. Yeah, exactly, Santa Clara Convention Center, and then Super Compute in November. We will be there in full force.

John Furrier

>> Yeah, we'll be there as well. Alexis, thanks for coming on. Appreciate you coming on.

Alexis Crowell

>> Yeah, thanks for having me.

John Furrier

>> We'll see you at Super Compute, and maybe OCP, maybe in Santa Clara.

>> Absolutely, that sounds great.

John Furrier

>> Thanks for coming on for today.

>> Thank you.

John Furrier

>> I'm sure for a third year in a row, doing the AI leaders coming together here for the AI Summit here, the CUBE and the NYSEY, of course the Pool Party. Be right back with more after this short break. Thank you.

Alexis Crowell, Axelera AI

search

(INTRO)

John Furrier

>> Welcome back to the Cube, live here in Palo Alto for the third annual Cube AI Summit for the leaders of infrastructure build out. Of course we have our party tonight, 180 people reception, where they'll be sharing their stories about what's going on, sharing, comparing notes, sharing their stories, but it's the biggest build out we've seen in the industry ever, and you're starting to see it happen, enabling all kinds of use cases. We have a returning CUBE alumni, Alexis Crowell, CMO and President of the Americas Alex Zahra. Welcome back to the CUBE. Thank you, thank you so much for having me. Thanks for coming back on. We had you guys on in Mobile World Congress, which was smashing because everyone was talking about Edge, which fits in nicely with what you're doing. and GTC, obviously that's the mega event. And then two weeks ago at the RACE Summit, you guys were on again as robotics and agents kick in, which highlights a lot what you're doing. It does. Give us an update on the latest and greatest.

Alexis Crowell

>> Happy to. So as you know, right, we started intentionally at the edge so that we built really efficient performant compute. So we've got customers that are putting us in drones to go search and find people and help prevent against wildfires. So there's some really interesting, fun stuff happening there. But the product line scales into servers and it scales into on -prem enterprise multi -user LLMs that are really solving a lot of what people are concerned about, which is what happens when I throw all of this compute into the cloud and then I lose a little bit of control or I don't know what my token costs are going to be. So we're helping bring some of that back on -prem with the server solutions that really really maximize infrastructure capacity without having to completely rebuild your data center.

John Furrier

>> Talk about that token cost equation, how that relates to enterprise, because this is the biggest conversation happening, because compute and GPUs and the role of pre -filled decode, disaggregated infrastructure, disaggregated inference is coming up too, so starting to see smaller footprints, different use cases, what's the impact to the enterprise?

Alexis Crowell

>> Well, it really comes down to what are you trying to solve for? My favorite example is if you're a legal aid and you're trying to help build a contract, you don't necessarily need to have a language model that had French poetry trained as well, right? Like that didn't need to be part of the source material. So you can use a much smaller model and get really accurate, really important productivity out of that scenario. And that's where solutions like Accelera really play in because we can do all of that at the edge without having to leverage big frontier models. Those have their place. I'm not trying to knock them, they're amazing. I use them. But they're not necessary all the time. You can use an open source model and really solve for what you need.

John Furrier

>> Talk about where you guys have the momentum right now because as the market develops, we see the AI factors in the service for sure. Rack scale systems, those are the monster racks. but they're starting to see different levels of clusters, of systems, the edge is out there. We're seeing, you mentioned drones. Yes. Sovereignty is about control. Yeah. Enterprises love control and they have their deterministic workloads and the agents, but talk about where you guys are winning your deals, where momentum is.

Alexis Crowell

>> Absolutely, so there's a few different places. When we were founded, Vision was the entirety of the market, so we are still really winning in Vision for sure, but now it's not just Vision. It's vision, search, and summarization. So companies that are trying to understand what's happening in their retail chains, right? We have a huge convenience store in South America that's deploying us to be able to not only monitor inventory, but then to real -time update and understand what's happening and then use reasoning to actually go action on that. And you can do that both out at the edge within the store with a single chip, or you can do that back in a centralized location on a few chips across the entirety of your stores.

John Furrier

>> So people want to bring their LLMs to their own LLMs for the company. They also use the general purpose, large scale. So you've got general intelligence, that's the internet, open AI, cloud, that's great. But then you start to see the specialty AI. How do you guys talk about those two forces? They're not necessarily mutually exclusive in some ways.

Alexis Crowell

>> No, they're not. We very much see them as complementary. It's very similar to tools. If you're going to go build a house, you need a hammer and a screwdriver and a saw and all these things. right? Each one of them has their purpose. Where we find ourselves is not so much just a hammer in that, you know, you can hit anything with it, but it's more of, as you know what your use case is, you're going to want hardware that is really purpose -built and optimized around that. Now, the good news is, is we're completely programmable. So we can, if it's inference and it's matrix math, I can run it. But more importantly, it's highly, highly optimized for power, meaning energy costs, and performance from the edge all the way through to where it allows customers to say, okay, maybe I don't need that much performance today, but it gives me some headroom to move into what's coming next.

John Furrier

>> You know, the inference conversation has shifted a lot. I want to get your thoughts. You had mentioned earlier at the top of the interview about cloud and relationship with the drones. Pulling stuff from the cloud to the device is something that's happening. But AI inference has many different use cases. Where do you guys see that, and how do companies that have a lot of cloud pull that down and bring that capability, that intelligence to the device?

Alexis Crowell

>> You know, there's a number of different ways, right? There's the, I have a local model that's only trained on my data, mine is a Lexus, like it's got my bank accounts, my things. I don't necessarily want one specific company to have access to all of that in one place. You're starting to see more of an enterprise grade solution actually end up in homes for those types of things, right? On the enterprise side, you see companies companies that have trained entirely on their data. In our case, as an example, we built a tool called Wingman. And Wingman is an entire LLM. It uses the backbones of some of the major models, but it's built on all of our documentation, all of our GitHub repo, everything to make programming really, really simple. And now in minutes, you can get something up and running. So the reason I say all of this is it really depends on what problem you're trying to solve, right? And that's always the first question is, what problem are we trying to solve? and then let's go figure out what tech stack makes the most sense.

John Furrier

>> You guys have the AI accelerator, you've got the systems and the software. That's right. It's interesting, I was having another conversation with another company in a different area and they make silicon, but they said, no, no, we don't want to be known as the silicon player, but they also have cybersecurity in silicon. Yeah. I don't want to be known as a cybersecurity company. He's actually a controller, the software is driving it. You guys have a very similar pattern. How do you tell that story? Because you don't want to get pigeonholed into one bucket. Right. Because you're doing all, but the software tooling is critical for developers because they have to do the work. Exactly. I want to get my stuff in there, get intelligence injected in. Talk about that dynamic, because you have a lot of piece parts, but they work together with the software. Why is that important?

Alexis Crowell

>> They're designed together very intentionally. I mean, we have as many folks working on the software side of our business as we do on the hardware side. And it's not that I'm trying to monetize the software. I'm trying to, to your point, make sure that developers have a frictionless experience. variants, right? Because if I go in and I try and disrupt what a developer is doing, that doesn't work. You and I have been in this industry a really long time. That doesn't work, right? So our goal and the reason we co -design our software is so that as somebody goes and buys one of our cards or one of the systems, they can just start, right? It's all PyTorch. It's all Onyx -based. It's all the same stuff they're doing today. It's just landing on a much more performant per watt, performant per dollar solution as a result.

John Furrier

>> One of the things you guys, I'm impressed by, and you guys talked about it in some of the interviews, is the ease of use. The ability to stand up something so compelling, so fast. With the coding assistant agents out there. So all this is coming to the table. So if I'm a developer, explain to me, I already kind of know the answer, but explain to me if I didn't know, what's in it for me? Because right now I'm coding maybe on a box that's Nvidia, and then okay, I write software, but I want to move it into say a robotic system. Okay, does the right chip work there? So now I got to port that code, cause I'm developing here. This has become kind of, I won't say, it's kind of a friction point. For sure. How do you guys address that? Cause the developers just want to, I don't mind coding on a super device or the box itself, but I got to get it into smaller form factors.

Alexis Crowell

>> Yeah, the good news is, is if you've coded it on any platform, you are likely already using PyTorch, right? From an inference perspective, that is by far the dominant platform. That is what we support. We support Onyx as well, but we support the major solutions. And then our tool chain allows that port really fast, right? You literally have a chatbot that is hyper -optimized on our system and our documentation to where you can have 20 % to 30 % better accuracy than just running Cloud Code alone, right? Cloud Code is amazing, but it's not hyper -tuned on what we do. So we built a tool to make that port really easy and not make it hard for developers.

John Furrier

>> And they get what they want.

Alexis Crowell

>> They get what they want. They can test it. We've got cloud deployment so they can test it. Try and make it as easy as possible.

John Furrier

>> Okay, so a lot's changed since we last talked when you were on. I'm expecting it to be just as crazy speed -wise. What's your focus now as you go to market? The market's hungry for solutions. It is. What's your strategy? What's the plan?

Alexis Crowell

>> Well, the good news is our strategy hasn't changed. So that is fabulous. We already have products in the embedded space, the edge space, into the server space, and we have deployments with customers with that. But our focus, candidly, right now is just making sure we continue to deploy those and support the customers and feed the demand as best as we can. As you know, we announced Europa last year. It's on track. It's coming to market this year. We've got good news coming soon. So all of that. We're executing. We are. We're absolutely executing.

John Furrier

>> What's next? What can we expect without giving away the farm? Yeah. Sharing too much information. But you're a private company. You're not public yet. Yet.

Alexis Crowell

>> Yet. That's right. Yet. Exactly. So we have always had the strategy and the vision from before we were funded to go edge all the way into cloud. Right. And we we've announced that we are part of a program called DARE, which is the digital autonomy for RISC -V in Europe. We are part of that in so much as it's helping us architect what a cloud solution would be like based on the hyper efficient edge architecture that we started with. Right now, obviously there are constraints, you can relax, you can get more power, et cetera, in the cloud, but it's that same architecture that will flow through, and that's what we're working on.

John Furrier

>> All right, well, put a plug in for what's going on in the Americas. Yeah. You can see I'm also in charge of the Americas. What's new, what are you looking to do? Hire, the goals, people watching, customers.

Alexis Crowell

>> We're constantly looking for really good talent. If there's folks that are interested, please call me. We've got a couple of really big shows coming up where you'll see even more of what we're doing, both at AI Infra Summit in September. Please come join us there. Yeah, exactly, Santa Clara Convention Center, and then Super Compute in November. We will be there in full force.

John Furrier

>> Yeah, we'll be there as well. Alexis, thanks for coming on. Appreciate you coming on.

Alexis Crowell

>> Yeah, thanks for having me.

John Furrier

>> We'll see you at Super Compute, and maybe OCP, maybe in Santa Clara.

>> Absolutely, that sounds great.

John Furrier

>> Thanks for coming on for today.

>> Thank you.

John Furrier

>> I'm sure for a third year in a row, doing the AI leaders coming together here for the AI Summit here, the CUBE and the NYSEY, of course the Pool Party. Be right back with more after this short break. Thank you.