Red Hat Summit 2026 | Chris Wright, Red Hat

Clips
More from Red Hat Summit 2026

Chris Wright

Chief Technology Officer and Senior Vice President, Global Engineering

Red Hat

play_circle_outline Engineers First: Platform Continuity on Linux and Kubernetes — AI as Just Another Application

play_circle_outline Human-Centered Trust and Sovereignty in AI: Identity, Least-Privilege, and Audit Trails for Responsible Agents

play_circle_outline Performance and cost focus: right-size models, token economics, consumption-based pricing

play_circle_outline Safe Agent Governance: Guardrails, I/O Filtering, Access Controls, and Safety via Evaluations, Unit Testing, and Red Teaming

Info

Chris Wright, Red Hat

Chris Wright

Chief Technology Officer and Senior Vice President, Global Engineering Red Hat

search

Rebecca Knight

>> Good morning everyone and welcome back to theCUBE's live coverage of the Red Hat Summit here in Atlanta, Georgia. I'm your host, Rebecca Knight alongside of Rob Strechay, chief analyst. Rob, we are doing a veritable who's who of the Red Hat leadership team this morning.

Rob Strechay

>> It keeps on rolling and I love where this is going because we get to get progressively even more technical and bring in customers and partners as the morning goes on as well. So this has been fantastic.

Rebecca Knight

>> Indeed. Well with that, I would like to welcome to the show Chris Wright, Chief Technology Officer and Senior Vice President of Global Engineering at Red Hat and host of the podcast, the very popular podcast, Technically Speaking. Welcome, Chris.

Chris Wright

>> Thank you. Great to be here.

Rebecca Knight

>> Welcome back to the show.

Chris Wright

>> Yeah, yeah.

Rebecca Knight

>> So you have been at Red Hat through some of the biggest shifts in enterprise techs. Linux, enterprise, I'm sorry, Kubernetes, now AI. One of the things you said on stage this morning was you really paid tribute to the people who work behind the scenes, the operators who are figuring out problems at two o'clock in the morning, the architects, the engineers. And you said the reason we at Red Hat can stand up and say AI is just an application is because of the humans that have built those foundations.

Chris Wright

>> Yeah. I think it's really important to acknowledge a couple of things, the real people doing the real work, but also the platform continuity. We're building on Linux, we're building on Kubernetes. Yeah, the workload is different and the problems we can solve with this new technology, those problems are different and it's super exciting, but it's built on things that we already know and it's like standing on the shoulders of giants. You can't be there without all this infrastructure support. And it's not just us, it's a whole broad community. Plus it's all of the people putting this into production environments to run their business today with applications and tomorrow with AI and AI agents. So it's cool. I like it.

Rob Strechay

>> Yeah. I mean, I look at it and go, there's this tension, I guess you could say, between the economics of AI and kind of trust. And I think one of the things that you talked about in your section this morning was really how do you get to trust? And what do you see that Red Hat is really leading in the open source communities to kind of bring these technologies that help enable trust, enable sovereignty, enable a number of these different things that people find critical?

Chris Wright

>> Yeah. I mean, trust is, go back to that human comment, trust is a very human thing. So we just start in communities. Communities are built from people and people develop trust relationships, and that's how the open source communities work. In a business, of course, trust is critical. Trust is part of being reliable and safe for your business. Trust has an aspect of guardrailing that you have the inputs and outputs of what you're asking the model to do as feeling safe for your business. But all the way into as you're building agents that can write code and do things, make real actions within your real business, how do you trust that? You got to give it the right sandboxing. You got to put protections around the agent, give it least privileges so it doesn't... Think about read versus read, write, very big difference. And so reading data, great. Being able to write back to it or delete it could be a real problem. So how do you manage that in scale with potentially hundreds or thousands of agents? So I think building trust is critical. And also who is the agent? What's the agent identity? Who is it working on behalf of? How do you build this audit trail of activity? So there's all these different aspects of trust that I think are critical to making it safe and trustworthy for the enterprise to put this into production.

Rebecca Knight

>> You have described what Red Hat is building as the standard foundation for the AI era, the standard execution layer. Why does the industry need that right now?

Chris Wright

>> I think the industry always benefits from having these sort of technology pillars of standardization. Why? Well, for one thing, you just build industry-wide efficiency. We have a focal point. Linux is a great example. There's so many ways that Linux can be used. And then as a result, there's all these different applications and ecosystems that develop around Linux, and then distribute that with Kubernetes, a similar story, building consistency. We're really focused especially on inferencing with a industry standard tool like vLLM creates again, a focal point. So the model providers, they build to vLLM before they ever release the model. So that's creating this efficiency at scale. And then when you bring it into the business, you've got the same ability to create operational efficiency. You know what you're targeting as a builder. So I think it's really important to have these standardized building blocks and help move the industry forward quickly.

Rob Strechay

>> Yeah. I couldn't agree more. And I think when you start to look at it, it's always interesting because there's a big thing around convergence and we kind of talked a little bit with Ashesh about this. What do you see as kind of that converged reality and how does that aid people making things more simple, like you said, to reduce that cognitive flow?

Chris Wright

>> Yeah. Well, I think a big question is start like, where do I even get started? So you got to understand what is your use case. And then as you pick through the use case, you got to understand within the enterprise, what context do I need access to? It could be data, could be tools that you're calling. And as you step through each of these pieces, you start to see that having a converged platform that can run say applications... Again, an agent in one sense, it's just an application wrapped around a model. Having an application platform that can support lifecycle management, the security boundaries, the trust boundaries of that application, the connections to that model, bringing that converged view I think is extremely powerful and it helps you move quickly from what is my use case to how do I put that into production.

Rebecca Knight

>> How are you thinking as CTO about performance and cost? Because there is a lot of panic about how expensive AI is getting to run. How do you think about it for Red Hat and how do you also think about approaching customers with the challenges that they're having with this?

Chris Wright

>> Yeah, tokens for everybody.

Rebecca Knight

>> Tokens for all. You get a token, you get a token.

Rob Strechay

>> Well, if you're Jensen, it's $250,000 a tokens for a $250,000 engineer. I'm like, I don't know how the CFO actually plans for that, but yeah.

Chris Wright

>> Well, it's a real... So there's always this tension. And I know internally, we use AI a lot from any aspect of the business, but think of from my point of view within the engineering, with developers, what are you going to do? You're going to pick the best model you can find and pin to it. It doesn't matter what task you're trying to achieve. Well, it turns out it's great in certain scenarios. It's still good in other scenarios, but now you're just basically wasting tokens. It's like you wouldn't take your F1 car to the grocery store. So pick the right tool for the right job. It's a lot of what we're focused on, enabling that choice, that flexibility that allows you to pick the most performant per cost, even per power ratio for the job that you're trying to do. And I think that's critical and we'll see already shifts in how consumption looks moving from per seat pricing to more consumption based pricing, which exposes more and more of the raw costs of producing tokens. You're going to have to shift from being just a consumer of tokens to being able to produce your own tokens and then balance how do you manage your token consumption. And I think that's going to be a really important part of the conversation throughout this year.

Rob Strechay

>> And I mean, you guys made the acquisition of Neural Magic actually a couple years back now almost, like two and a half years or so ago.

Chris Wright

>> Just over a year.

Rob Strechay

>> Oh, just over a year.

Chris Wright

>> Yeah.

Rob Strechay

>> Oh, wow.

Chris Wright

>> Time flies. AI years.

Rob Strechay

>> We were in Denver, so you're right. I was like, okay, I'm trying to figure out where I was when that happened. But it does seem... I mean, it is. It's like dog years, but accelerated so much faster but when you look at AI. So when you look at like vLLM, LLMD, and kind of the underlying infrastructure, I mean, to me that seems to be one of the things that you guys are leaning in on because it helps get to those token economics to that efficiency. Is that where you see a lot of that happening?

Chris Wright

>> Yeah, absolutely. So I mentioned that the model providers build to vLLM. They're also trying to improve their own model performance. And that could be architectural. How is the model built? So we reflect those architectural shifts in the underlying infrastructure and then we distribute that across whatever infrastructure you're using to serve these models. And in that combination of the models, architectural refinements, plus the distribution of how we manage all the queries hitting this hardware, we can really reduce the cost. So we've seen things like 3X token output improvements just with all of the work that we're doing. That's software improving the efficiency of hardware. We can also reduce the time to first token and stabilize the overall tail latencies so it becomes very reliable in these mission-critical contexts. So it's performance, it's scalability, it's reducing cost, it's making it consumable for the enterprise, and all of that happens right at the inference layer.

Rebecca Knight

>> I want to go back to trust and governance and safety because people, boards are rightly concerned about AI agents making decisions on their own. How are some of the products that you're building in terms of the agent ops identity framework, how is that keeping these agents in check and creating more accountability?

Chris Wright

>> Yeah, it's a couple of different pieces. So one of it is just basic identity. Who is this agent? And there's also a notion of delegation. So I might kick off a job. You need to establish the relationship between who started it and who's doing the work and even an agent can have a sub agent so it can get pretty interesting and complicated managing with that identity, authorized access so that you're never surprised at who's doing what, and then bringing in notions of the guardrailing of the activities. So you can filter all of the inputs coming into the model. You can filter the outputs coming out of the model to make sure you're staying in compliance, but also verifying that. I think there's a couple of different tools that are really important. One is evaluations. It doesn't get talked about as much. Maybe it's not as exciting from a pure technology point of view. Fundamental to understanding does this work or does this not. So understanding what you expect and then verifying that you're getting what you expect. We do this all the time in applications. It's called unit testing and system testing and integration testing. Evaluations are a critical part of this. And then you can also do red teaming and literally attack the system. You can even use known bad examples and magnify those with synthetic data generation to make this swarm of attackers on a system to really give you the confidence that it's doing exactly what it should be doing, it's authorized and it's clear identity. So there's a lot happening in building trust to make this accessible for enterprises.

Rob Strechay

>> So there's a lot of talk because the agents have to be able to talk to either other agents or talk to tools and be able to leverage that. What do you see as kind of the path forward? Because there's a lot of, I would say there's some great stuff going on in the open source community and about how you interact with that or have skills and MCP and things. How do you see that all coming together as you try to enable trust because there's the whole identity thing as well that we've talked about briefly?

Chris Wright

>> An easy example, MTP I think is a really important tool in the toolkit. An easy example would be the world is built around identity as you and I, and we feel pretty confident in trusting ourselves. So if you have the authorized access to your bank account, it's just all in. You may want to give an agent read only access to your bank account. And that isn't even necessarily a credentialed option with your service provider. The MCP tool layer can help gate how you even get access to the APIs, which may have pretty broad permissions. You can scope those permissions down through the ways you expose tools. So it's just an easy example. I'd love to know where my monthly credit card bills are going to platforms I'm not even looking at anymore. I sure don't want that agent to go buy something or maybe donate my money to this new Bitcoin wallet that it found. So that I think is an easy example. Skills are just a nice way to explain to the system how to do something and that can include things like similar to tool calling and it kind of brings in this fun resurgence of the CLI. I'm a nerdy guy. I love Linux. I love the terminal. CLI wasn't what we were talking about just even a year ago. But building these tools that are aimed at agents being able to make calls to an API, leveraging a CLI defined by a skillset is a great example of bringing... The CLI is cool again. I love it. I love it.

Rebecca Knight

>> Chris, hardware partnerships. We have all over the floor here, we have AMD and Nvidia. You're co-engineering with these organizations. What do these partnerships look like behind the scenes?

Chris Wright

>> So hardware is pretty part and parcel to what we do. Our whole history includes this very important part of being an operating system company. So we're always there with our hardware partners. It's enabling hardware. It's exposing this consistent set of APIs to applications. What's shifted so dramatically, I mean, it's unbelievable how big the change has been, is the speed. Used to be design cycles and tape out cycles and these things happen over years, two, three years. Now our partners are on this annual cadence of pushing out new hardware and we've had to fundamentally change how we build and deliver software to keep up with hardware, which you never would have imagined just even a few years ago. And so behind the scenes, it's deep co-engineering. It's being sure that when the hardware comes out, the software's there ready on day zero to support that hardware. And it puts us all in this perpetual foot race of just full speed ahead, bringing great new technology at the hardware level, exposed through software. It's all about performance, bringing down cost of token production, doing it within a power envelope, and that's what we're driving towards.

Rebecca Knight

>> Excellent.

Rob Strechay

>> Yeah. I mean, I love that because I think when you start to look at it, I think GPUs are great, but there's still a lot of CPUs in the world and there's a lot of people who want to go out to the edge, to the radio towers, to the shop floor where they're building things and have robotics. Do you see that as also one of the challenges that you guys just love that is going after, to your point about efficiencies and being able to working with all of the people who are here, the AMDs, the Intels, the Nvidias and all around that?

Chris Wright

>> Yeah. I mean, vLLM is this inference engine for us. Initially, it was focused predominantly on just a few accelerators and over time we've expanded what that looks like, including the CPU as a way to do serving of models. So you think about the smaller models and maybe industrial automation right on the shop floor and a very small form factor at a base station, you might not need this gigawatt scaled powered type of device to do a more function specific job. So we're optimizing for that full breadth. Heterogeneity absolutely is the future. Building heterogeneity, not just in hardware, but in the workloads and the kind of models that you use to support your workloads, the bigger ones, the smaller ones, tuned for a specific task. That's exactly what we're focused on and it makes it really fun because it's super dynamic. It's changing all the time. And old is new and new is old and all the players are converging in different ways. It makes it a really exciting space.

Rebecca Knight

>> A great note to end on. Chris, thank you so much for coming on the show.

Chris Wright

>> Yeah, thanks for having me.

Rebecca Knight

>> Podcast is Technically Speaking for all of our viewers out there. I'm Rebecca Knight for Rob Strechay. Stay tuned for more of theCUBE's live coverage of the Red Hat Summit. You're watching theCUBE, the leader in enterprise tech news and analysis.

Chris Wright, Red Hat

search

Rebecca Knight

>> Good morning everyone and welcome back to theCUBE's live coverage of the Red Hat Summit here in Atlanta, Georgia. I'm your host, Rebecca Knight alongside of Rob Strechay, chief analyst. Rob, we are doing a veritable who's who of the Red Hat leadership team this morning.

Rob Strechay

>> It keeps on rolling and I love where this is going because we get to get progressively even more technical and bring in customers and partners as the morning goes on as well. So this has been fantastic.

Rebecca Knight

>> Indeed. Well with that, I would like to welcome to the show Chris Wright, Chief Technology Officer and Senior Vice President of Global Engineering at Red Hat and host of the podcast, the very popular podcast, Technically Speaking. Welcome, Chris.

Chris Wright

>> Thank you. Great to be here.

Rebecca Knight

>> Welcome back to the show.

Chris Wright

>> Yeah, yeah.

Rebecca Knight

>> So you have been at Red Hat through some of the biggest shifts in enterprise techs. Linux, enterprise, I'm sorry, Kubernetes, now AI. One of the things you said on stage this morning was you really paid tribute to the people who work behind the scenes, the operators who are figuring out problems at two o'clock in the morning, the architects, the engineers. And you said the reason we at Red Hat can stand up and say AI is just an application is because of the humans that have built those foundations.

Chris Wright

>> Yeah. I think it's really important to acknowledge a couple of things, the real people doing the real work, but also the platform continuity. We're building on Linux, we're building on Kubernetes. Yeah, the workload is different and the problems we can solve with this new technology, those problems are different and it's super exciting, but it's built on things that we already know and it's like standing on the shoulders of giants. You can't be there without all this infrastructure support. And it's not just us, it's a whole broad community. Plus it's all of the people putting this into production environments to run their business today with applications and tomorrow with AI and AI agents. So it's cool. I like it.

Rob Strechay

>> Yeah. I mean, I look at it and go, there's this tension, I guess you could say, between the economics of AI and kind of trust. And I think one of the things that you talked about in your section this morning was really how do you get to trust? And what do you see that Red Hat is really leading in the open source communities to kind of bring these technologies that help enable trust, enable sovereignty, enable a number of these different things that people find critical?

Chris Wright

>> Yeah. I mean, trust is, go back to that human comment, trust is a very human thing. So we just start in communities. Communities are built from people and people develop trust relationships, and that's how the open source communities work. In a business, of course, trust is critical. Trust is part of being reliable and safe for your business. Trust has an aspect of guardrailing that you have the inputs and outputs of what you're asking the model to do as feeling safe for your business. But all the way into as you're building agents that can write code and do things, make real actions within your real business, how do you trust that? You got to give it the right sandboxing. You got to put protections around the agent, give it least privileges so it doesn't... Think about read versus read, write, very big difference. And so reading data, great. Being able to write back to it or delete it could be a real problem. So how do you manage that in scale with potentially hundreds or thousands of agents? So I think building trust is critical. And also who is the agent? What's the agent identity? Who is it working on behalf of? How do you build this audit trail of activity? So there's all these different aspects of trust that I think are critical to making it safe and trustworthy for the enterprise to put this into production.

Rebecca Knight

>> You have described what Red Hat is building as the standard foundation for the AI era, the standard execution layer. Why does the industry need that right now?

Chris Wright

>> I think the industry always benefits from having these sort of technology pillars of standardization. Why? Well, for one thing, you just build industry-wide efficiency. We have a focal point. Linux is a great example. There's so many ways that Linux can be used. And then as a result, there's all these different applications and ecosystems that develop around Linux, and then distribute that with Kubernetes, a similar story, building consistency. We're really focused especially on inferencing with a industry standard tool like vLLM creates again, a focal point. So the model providers, they build to vLLM before they ever release the model. So that's creating this efficiency at scale. And then when you bring it into the business, you've got the same ability to create operational efficiency. You know what you're targeting as a builder. So I think it's really important to have these standardized building blocks and help move the industry forward quickly.

Rob Strechay

>> Yeah. I couldn't agree more. And I think when you start to look at it, it's always interesting because there's a big thing around convergence and we kind of talked a little bit with Ashesh about this. What do you see as kind of that converged reality and how does that aid people making things more simple, like you said, to reduce that cognitive flow?

Chris Wright

>> Yeah. Well, I think a big question is start like, where do I even get started? So you got to understand what is your use case. And then as you pick through the use case, you got to understand within the enterprise, what context do I need access to? It could be data, could be tools that you're calling. And as you step through each of these pieces, you start to see that having a converged platform that can run say applications... Again, an agent in one sense, it's just an application wrapped around a model. Having an application platform that can support lifecycle management, the security boundaries, the trust boundaries of that application, the connections to that model, bringing that converged view I think is extremely powerful and it helps you move quickly from what is my use case to how do I put that into production.

Rebecca Knight

>> How are you thinking as CTO about performance and cost? Because there is a lot of panic about how expensive AI is getting to run. How do you think about it for Red Hat and how do you also think about approaching customers with the challenges that they're having with this?

Chris Wright

>> Yeah, tokens for everybody.

Rebecca Knight

>> Tokens for all. You get a token, you get a token.

Rob Strechay

>> Well, if you're Jensen, it's $250,000 a tokens for a $250,000 engineer. I'm like, I don't know how the CFO actually plans for that, but yeah.

Chris Wright

>> Well, it's a real... So there's always this tension. And I know internally, we use AI a lot from any aspect of the business, but think of from my point of view within the engineering, with developers, what are you going to do? You're going to pick the best model you can find and pin to it. It doesn't matter what task you're trying to achieve. Well, it turns out it's great in certain scenarios. It's still good in other scenarios, but now you're just basically wasting tokens. It's like you wouldn't take your F1 car to the grocery store. So pick the right tool for the right job. It's a lot of what we're focused on, enabling that choice, that flexibility that allows you to pick the most performant per cost, even per power ratio for the job that you're trying to do. And I think that's critical and we'll see already shifts in how consumption looks moving from per seat pricing to more consumption based pricing, which exposes more and more of the raw costs of producing tokens. You're going to have to shift from being just a consumer of tokens to being able to produce your own tokens and then balance how do you manage your token consumption. And I think that's going to be a really important part of the conversation throughout this year.

Rob Strechay

>> And I mean, you guys made the acquisition of Neural Magic actually a couple years back now almost, like two and a half years or so ago.

Chris Wright

>> Just over a year.

Rob Strechay

>> Oh, just over a year.

Chris Wright

>> Yeah.

Rob Strechay

>> Oh, wow.

Chris Wright

>> Time flies. AI years.

Rob Strechay

>> We were in Denver, so you're right. I was like, okay, I'm trying to figure out where I was when that happened. But it does seem... I mean, it is. It's like dog years, but accelerated so much faster but when you look at AI. So when you look at like vLLM, LLMD, and kind of the underlying infrastructure, I mean, to me that seems to be one of the things that you guys are leaning in on because it helps get to those token economics to that efficiency. Is that where you see a lot of that happening?

Chris Wright

>> Yeah, absolutely. So I mentioned that the model providers build to vLLM. They're also trying to improve their own model performance. And that could be architectural. How is the model built? So we reflect those architectural shifts in the underlying infrastructure and then we distribute that across whatever infrastructure you're using to serve these models. And in that combination of the models, architectural refinements, plus the distribution of how we manage all the queries hitting this hardware, we can really reduce the cost. So we've seen things like 3X token output improvements just with all of the work that we're doing. That's software improving the efficiency of hardware. We can also reduce the time to first token and stabilize the overall tail latencies so it becomes very reliable in these mission-critical contexts. So it's performance, it's scalability, it's reducing cost, it's making it consumable for the enterprise, and all of that happens right at the inference layer.

Rebecca Knight

>> I want to go back to trust and governance and safety because people, boards are rightly concerned about AI agents making decisions on their own. How are some of the products that you're building in terms of the agent ops identity framework, how is that keeping these agents in check and creating more accountability?

Chris Wright

>> Yeah, it's a couple of different pieces. So one of it is just basic identity. Who is this agent? And there's also a notion of delegation. So I might kick off a job. You need to establish the relationship between who started it and who's doing the work and even an agent can have a sub agent so it can get pretty interesting and complicated managing with that identity, authorized access so that you're never surprised at who's doing what, and then bringing in notions of the guardrailing of the activities. So you can filter all of the inputs coming into the model. You can filter the outputs coming out of the model to make sure you're staying in compliance, but also verifying that. I think there's a couple of different tools that are really important. One is evaluations. It doesn't get talked about as much. Maybe it's not as exciting from a pure technology point of view. Fundamental to understanding does this work or does this not. So understanding what you expect and then verifying that you're getting what you expect. We do this all the time in applications. It's called unit testing and system testing and integration testing. Evaluations are a critical part of this. And then you can also do red teaming and literally attack the system. You can even use known bad examples and magnify those with synthetic data generation to make this swarm of attackers on a system to really give you the confidence that it's doing exactly what it should be doing, it's authorized and it's clear identity. So there's a lot happening in building trust to make this accessible for enterprises.

Rob Strechay

>> So there's a lot of talk because the agents have to be able to talk to either other agents or talk to tools and be able to leverage that. What do you see as kind of the path forward? Because there's a lot of, I would say there's some great stuff going on in the open source community and about how you interact with that or have skills and MCP and things. How do you see that all coming together as you try to enable trust because there's the whole identity thing as well that we've talked about briefly?

Chris Wright

>> An easy example, MTP I think is a really important tool in the toolkit. An easy example would be the world is built around identity as you and I, and we feel pretty confident in trusting ourselves. So if you have the authorized access to your bank account, it's just all in. You may want to give an agent read only access to your bank account. And that isn't even necessarily a credentialed option with your service provider. The MCP tool layer can help gate how you even get access to the APIs, which may have pretty broad permissions. You can scope those permissions down through the ways you expose tools. So it's just an easy example. I'd love to know where my monthly credit card bills are going to platforms I'm not even looking at anymore. I sure don't want that agent to go buy something or maybe donate my money to this new Bitcoin wallet that it found. So that I think is an easy example. Skills are just a nice way to explain to the system how to do something and that can include things like similar to tool calling and it kind of brings in this fun resurgence of the CLI. I'm a nerdy guy. I love Linux. I love the terminal. CLI wasn't what we were talking about just even a year ago. But building these tools that are aimed at agents being able to make calls to an API, leveraging a CLI defined by a skillset is a great example of bringing... The CLI is cool again. I love it. I love it.

Rebecca Knight

>> Chris, hardware partnerships. We have all over the floor here, we have AMD and Nvidia. You're co-engineering with these organizations. What do these partnerships look like behind the scenes?

Chris Wright

>> So hardware is pretty part and parcel to what we do. Our whole history includes this very important part of being an operating system company. So we're always there with our hardware partners. It's enabling hardware. It's exposing this consistent set of APIs to applications. What's shifted so dramatically, I mean, it's unbelievable how big the change has been, is the speed. Used to be design cycles and tape out cycles and these things happen over years, two, three years. Now our partners are on this annual cadence of pushing out new hardware and we've had to fundamentally change how we build and deliver software to keep up with hardware, which you never would have imagined just even a few years ago. And so behind the scenes, it's deep co-engineering. It's being sure that when the hardware comes out, the software's there ready on day zero to support that hardware. And it puts us all in this perpetual foot race of just full speed ahead, bringing great new technology at the hardware level, exposed through software. It's all about performance, bringing down cost of token production, doing it within a power envelope, and that's what we're driving towards.

Rebecca Knight

>> Excellent.

Rob Strechay

>> Yeah. I mean, I love that because I think when you start to look at it, I think GPUs are great, but there's still a lot of CPUs in the world and there's a lot of people who want to go out to the edge, to the radio towers, to the shop floor where they're building things and have robotics. Do you see that as also one of the challenges that you guys just love that is going after, to your point about efficiencies and being able to working with all of the people who are here, the AMDs, the Intels, the Nvidias and all around that?

Chris Wright

>> Yeah. I mean, vLLM is this inference engine for us. Initially, it was focused predominantly on just a few accelerators and over time we've expanded what that looks like, including the CPU as a way to do serving of models. So you think about the smaller models and maybe industrial automation right on the shop floor and a very small form factor at a base station, you might not need this gigawatt scaled powered type of device to do a more function specific job. So we're optimizing for that full breadth. Heterogeneity absolutely is the future. Building heterogeneity, not just in hardware, but in the workloads and the kind of models that you use to support your workloads, the bigger ones, the smaller ones, tuned for a specific task. That's exactly what we're focused on and it makes it really fun because it's super dynamic. It's changing all the time. And old is new and new is old and all the players are converging in different ways. It makes it a really exciting space.

Rebecca Knight

>> A great note to end on. Chris, thank you so much for coming on the show.

Chris Wright

>> Yeah, thanks for having me.

Rebecca Knight

>> Podcast is Technically Speaking for all of our viewers out there. I'm Rebecca Knight for Rob Strechay. Stay tuned for more of theCUBE's live coverage of the Red Hat Summit. You're watching theCUBE, the leader in enterprise tech news and analysis.