Red Hat Summit 2026 | Bill Pearson, Intel & Taneem Ibrahim, Red Hat

Clips
More from Red Hat Summit 2026

Bill Pearson

VP, Data Center

Intel

Taneem Ibrahim

Director of Engineering, AI Inference

Red Hat

play_circle_outline Inference-Dominant AI: Scaling Open-Weight LLMs on RHEL/OpenShift with vLLM/LMD and Cost‑per‑Token Optimization

play_circle_outline Native Intel Xeon Support in Red Hat AI Enterprise 3.4: Balancing CPU‑GPU Ratios and Workload Orchestration

play_circle_outline CPUs Lead Inference for Agentic AI: Cost, Power, and Space Advantages for Tool-Calling and Orchestration

play_circle_outline Day‑Zero Model Support: Validated, Reproducible Model Catalog and Quantized Models on Hugging Face

Info
Transcript

Bill Pearson, Intel & Taneem Ibrahim, Red Hat

Bill Pearson

VP, Data Center Intel

Taneem Ibrahim

Director of Engineering, AI Inference Red Hat

In this interview from Red Hat Summit 2026, Taneem Ibrahim, director of engineering for AI inference at Red Hat, joins Bill Pearson, vice president of data center and AI at Intel, to talk with theCUBE's Rob Strechay and Rebecca Knight about how a balanced CPU-GPU infrastructure strategy is reshaping the economics of AI inference at enterprise scale. Ibrahim explains how demand for enterprise-grade inference exploded following the open-weights model wave, driving Red Hat to become the largest commercial contributor to vLLM and to build Red Hat AI Enterprise as... Read more

explore Keep Exploring

When did inference become the dominant complexity and cost in AI/LLM deployment, and what are customers reporting about their needs and challenges? add

How are you collaborating with Red Hat and the open-source community to enable vLLM and AI inference on Xeon CPUs, and what role do CPUs play (versus GPUs) in current AI workloads? add

How should organizations balance the use of CPUs and GPUs for AI inference workloads, and what advantages do CPUs offer? add

How does Red Hat handle model optionality and support for models that aren't yet in its catalog, and what should organizations focus on in that situation? add

bolt Powered by CUBE AI

Bill Pearson, Intel & Taneem Ibrahim, Red Hat

search

Rebecca Knight

>> Hello everyone and welcome back to Atlanta, where we are theCUBE and we are live at the Red Hot Summit at the Georgia World Congress. I'm your host, Rebecca Knight, alongside Rob Strechay. We're talking AI inference.

Rob Strechay

>> We are, that's such a shock though. But I think what we're going to be talking about is really how you get to scale with it and how you get to understanding that infrastructure is not all as attainable as we would want it to be kind of thing.

Rebecca Knight

>> Indeed. Indeed. But it is achievable if we-

Rob Strechay

>> It is achievable.

Rebecca Knight

>> Exactly. So, with that, I would like to welcome our next guest to the show, Taneem Ibrahim, director of Engineering AI inference at Red Hat. Welcome, Taneem.

Taneem Ibrahim

>> Thank you.

Rebecca Knight

>> And I'd like to welcome back to the show, Bill Pearson, VP Data Center and AI at Intel.

Bill Pearson

>> Thanks. Good to be back.

Rebecca Knight

>> Yeah. So, Taneem, inference is really where the real complexity and cost is right now. From your seat as the person who leads this at Red Hat, when did this become the dominant problem and what are you hearing from customers on this?

Taneem Ibrahim

>> Absolutely. That's a great question. So, when we started with the inference story, the first story was all around how can you get the largest model to deploy on the stack of GPU clusters, right? So, you had ChatGPT moment three years ago, we all talked about it, then all kinds of open weights models start to show up. And that's when we saw the customers coming to Red Hat, like, "How can I make this open weights model scale on Red Hat platform? How can I make them scale on RHEL? How can I make them scale on OpenShift?" And that's when the friction moment that came in for us is like, how do I take this project called vLLM, where we're the largest commercial contributor to, and work it at scale with a project like LMD? So, that's the main benefit of like how do you drive the cost per token down so that you can operationalize your AI, you can govern your AI, you can deploy it at scale.

Rob Strechay

>> Yeah. I mean, I love how you frame that because I look at it as there's a split in the road here. You're either going to be a token consumer or a token producer. Bill, help us understand how inference is really driving Intel's roadmap and what does that mean for the silicon that's coming out?

Bill Pearson

>> Yeah, it's quite interesting. And the first thing is we have great partners. We've been working with the open source community and with Red Hat for a long time on this. Some of it is enabling upstream vLLM, so that it just runs out of the box on our platforms, our CPUs. And some of it is partnering with Red Hat to make sure it shows up in their products as well. So, we're excited about the Red Hat AI 3.4 release, which has full support for Xeon and vLLM. With that, what we're seeing is AI is the workload and AI inference in particular. And now, with agentic coming into play, we're seeing that the demand for CPUs with AI is even more important than it was before when it was just a GPU-CPU combination. People are realizing that they need to have the right ratios, that we need to have the right performance and that it isn't a one-size-fits-all approach, but rather, what's my workload? What's the outcome I'm looking for? And how do I put together the right combination of hardware and software to go and deliver that outcome?

Rebecca Knight

>> Taneem, so Red Hat AI enterprise is distinct from OpenShift. In one sentence, can you describe what it is and what it does?

Taneem Ibrahim

>> Enterprise is the platform that gives you consistency when you're deploying models across your hybrid cloud, brings you the hybrid AI platform and it provides you a way to scale your model very efficiently across any hardware, any models and you can also govern them because that's another important part. All these agentic tools that Bill was talking about, they generate a lot of tokens. These reasoning models generate a lot of tokens. So, how do you govern them? How do you comply with whatever industry you're in with that policy? That's what Red AI enterprise provides.

Rob Strechay

>> So, let's play off of both of these because I think one of the things and we were talking about it leading up to this is that if AI was a nail, GPUs were the hammer. And everybody, "GPU this, GPU that. You get a GPU. You get a GPU," up until you couldn't. But to me, that doesn't make sense in normal data centers. It doesn't make sense in how clouds are built. Talk, Bill, about how the CPU plays a major role in this whole inference as we go forward because people can't retrofit. I was telling the story about a CO, you can't go retrofit water cooling into the CO of a telco.

Bill Pearson

>> Yes.

Rob Strechay

>> Digging up the CO is not going to be very popular with the telco, so help us understand.

Bill Pearson

>> Yeah. I think the shift to inference is really when this started. People realizing that sure you can do inference on a GPU, but it's expensive, it takes a lot of power and it's hard to fit into small spaces. What we've realized, of course, is that CPUs are good for a lot of models up to certain sizes. They can deliver good enough performance for a number of concurrent users. And so, depending on your workload, depending on your constraints, whether they're power, space, cooling, you can go and fit your workload into a CPU and have the outcome that you're looking for with the right KPIs and such. And as we've gone through this with our customers in the industry, we've seen that people often have just assumed I've got a hammer. I need the nail to hit it with. And once they take a step back to saying, "Well, wait a minute, I have these CPUs in my data center," or, "I need to figure out how to balance the right number of CPUs with the right number of GPUs to achieve that outcome I'm looking for," they're actually going to get better results at a better price point for delivering lower cost tokens like you were just talking about.

Rob Strechay

>> Absolutely.

Bill Pearson

>> Yeah.

Taneem Ibrahim

>> And then, the other advantage with CPU is that you already have them in your data center. They're already sitting there, right? And there's a lot of workload and inference that can run on CPU today, like your agentic workload, your tool calling, your data orchestration, that's a big part of making your agents work really well. So, all of that workload does not require a GPU and what it does is actually complimentary. It actually opens up more GPU workload driven. So, areas where you actually need to do high concurrent, high really major inference workload with large models, you have more GPU capacity available for yourself to deploy those models on top of it.

Rebecca Knight

>> The word that is synonymous with Red Hat is open source. And so, at a time when a lot of vendors are pushing for proprietary AI stacks, how do you talk about Red Hat's open source approach and why it matters so specifically for AI?

Taneem Ibrahim

>> Yeah. So, our entire Red Hat AI enterprise is built around open source components, starting with obviously vLLM and then LMD, which is a distributed vLLM platform. And then, we also have projects like OGX and OpenShell that we are heavily investing in. So, one of the projects that I want to talk a lot about is the OGX project, which is Open GenAI Stack. This is actually a project that unifies the API layer. So, if you think of different model providers, they give you different types of APIs. OpenAI has chat completions and responses, Anthropic has their own messages API, Gemini has their own... Every model provider has a specific API that you must conform to. So, how do you unify that, right? Because you are a developer who is having to make your workload run across open weights model, as well as proprietary models. You need a layer that unifies that story around. And then, let's say you also need to do a RAG chatbot and you need a way to talk to a vector database. How do you make that storage-retrieval process, that embedding process unified under a single layer of API? That's what OGX, or Open GenAI Stack provides you. It provides you that one single layer. So, regardless of which model you switch to, which hardware Intel Xeon or NVIDIA GPU or whatever you switch to, it doesn't really matter. API stays the same so your workload does not have to change. It gives you the consistency. So, by staying with all these open source components, it's allowing us to provide that consistency in our Red Hat Enterprise product. And also, most importantly, transparency is extremely critical in this AI workload. If you don't have that transparency, you don't have the safety guardrails around it. So, this also allows you to have all the safety guards because you know what's going into your product as a consumer.

Rob Strechay

>> And just building on that, Bill, I mean, I've known Intel for a long time and always been active in open source for a long time and various different projects from PyTorch, and Hugging Face, and now vLLM. How does that really help when you're looking at this about helping customers really focus on the use cases and what is right for-

Bill Pearson

>> Sure. I mean, open source is critical in so many ways. One of them is that it really allows us to innovate together as a community. So, many of the features, the capabilities, the performance that we're driving through open source is something that we're doing together as a community. So, attention mechanisms, for example, as they evolve, we all benefit and win. Another great advantage for customers is transparency, which we were just talking about, right? Particularly in sovereign where you want to make sure that you know what code is running your application in your data center. And for Intel, what we like about open source also is that it's something that our customers are used to. So, they have familiar interfaces and APIs and ways of working and they don't have to change that just because they're on our hardware. They can continue with the same software stack, the same tooling, same implementations across our hardware and other hardware. By doing that, it makes it much easier on them to get their job done and again, get the KPIs that they're looking for in their organization.

Rebecca Knight

>> Well, speaking of partnerships and collaboration, Jensen has called the Intel NVIDIA collaboration historic. Most people, of course, in and out of the tech industry think of you as competitors. So, how do you describe and explain the partnership?

Bill Pearson

>> Yeah. I mean, it's part of the new Intel, just approaching the industry with a little bit of humility, being humble, and really listening to what customers want. And we have heard that customers really love NVIDIA GPUs, they also love Intel CPUs. So, putting those two things together is a natural fit.

Rebecca Knight

>> It's like peanut butter and jelly.

Bill Pearson

>> Peanut butter and jelly. So, anywhere we can help a customer win, that's the approach that we're taking.

Rob Strechay

>> So, what gets you excited about where Xeon-based AI is going over the next 12 to 18 months?

Bill Pearson

>> Yeah, there's two things. One is the possibilities around inference. Just starting there, being able to use hardware you already have, evolve it to which workloads and which inference use cases are going to work on Xeon, which I'm going to need to compliment with GPUs and figure out how to build that infrastructure. The second piece is around agentic AI. It's really changing the game and when we look at that and orchestrating all of the workloads, making tool calls and doing things that have traditionally been done on the CPU, but with an agentic lens, there's a much, much larger role for CPU in that agentic AI than there was say in training where it was all about the GPU. And so, you see the ratio of CPU to GPU changing. You also see people innovating around, "Gee, how do I orchestrate these things? And at what level do I orchestrate, so I can send the right workload to the right silicon in the right way, again, to drive the outcome that our customers are looking for?"

Rebecca Knight

>> Taneem, let's talk a little bit about operationalized AI because that is a term that Red Hat uses very prominently. For a CIO or a CISO in a regulated industry, financial services, healthcare, what is the business case specifically for using AI in the governance infrastructure?

Taneem Ibrahim

>> Absolutely. So, Red Hat Enterprise provides the consistency. So, for example, operational perspective, you may need to deploy your model on a cloud platform, you may need to deploy it on on-prem, you may have it on the edge, but for your administrative suite who is managing that platform, they're using the same platform to get the same metrics. So, the metrics is what drives your compliance, so you need to comply to a certain level of metrics. You get those metrics from the single place. Red Hat Enterprise provides the entire dashboard to provide you all the metrics that you need to get access to. Again, you also need to see your token budget. Token budget is on top of the CTO and especially the CFO's mind today. So, how do you manage your token budget across the cloud, hybrid cloud, data center, and edge? Red Hat Enterprise provides that. So, maybe it's deployed in Intel Xeon CPU on a data center, maybe it's deployed on another provider on a cloud. You can see the token for each model, which tenant within your organization is using which token, where it's going. All of that operational consistency is provided on Red Hat Enterprise, which is why Red Hat Enterprise is the platform to choose when you were looking at doing hybrid AI at scale for your own next agentic workflow.

Rob Strechay

>> I mean model optionality is one of the big things. When you look at the whole idea of model optionality, you guys have some validated models. What if one of the models is not in the catalog or something? What should organizations really focus on with that?

Bill Pearson

>> I mean, we've been partnering to put more and more models in the catalog, building model cards and also showing steps how to reproduce the results that we get with those models. So, what our customers can do is come to the catalog, see what's there, see how to deliver that value that we're describing there. So, all the steps, all the approach. And if something's not there, we can always add it, take requests and go look at it together. We continue to evolve this in partnership to make sure that we get everything that our customers are looking for there.

Taneem Ibrahim

>> Yeah. And speaking of partnership, obviously working with Intel, we're also working with a lot of the model providers. So, one of the things at Red Hat AI Enterprise we really focus on is providing day-zero support for models. So, like let's say when Gemma 4 came out, we had it on day zero on Red Hat Enterprise available for customers to use. And then, we are also working together with Intel, we do a lot of evaluation on those models to ensure that they're within compliance, the scores are good, they actually perform well for different workloads. Which one is good for agentic workload? Which one is good for code gen? Which one is good for just a truthful chatbot? So, on and so forth. So, all of those, we try to get them uploaded as fast as possible to the catalog. And on top of that, we also optimize those models. So, we have quantization algorithms that we have created that allows you to get more out of your Intel Xeon servers or NVIDIA GPU servers. That allows you to get more memory that's available to load more models into it. And there are lots of advanced model optimization techniques. And the most important thing, being an open source company, we upload all those optimized model that we train on our huggingface.com/redhat repository for customers to deploy on any hardware anywhere.

Rebecca Knight

>> So, as we're wrapping up here, I want to hear from both of you about your best advice, best practice for a CIO who is watching this and wants to move from AI evaluation to execution. Taneem?

Taneem Ibrahim

>> Sure. So, I feel with Red Hat AI Enterprise 3.4, the platform is already here. You have a platform that's a metal-to-agents platform that will allow you to deploy any model anywhere on any hardware. And with Intel Xeon now being one of our first class hardware provider that we actually provide GA support for, the options are just increasing. And I want to see the CIO and the CTOs to look at Red Hat AI Enterprise 3.4 as the hybrid AI platform that allows them to scale for the future of AI agentic workflow.

Rebecca Knight

>> Bill, last word?

Bill Pearson

>> How do I add onto that, right? Red Hat AI Enterprise 3.4 with Intel Xeon, you can't go wrong.

Rebecca Knight

>> Okay. All right. Words to live by. Great. Bill and Taneem, thank you both so much for coming on the show.

Taneem Ibrahim

>> Thank you.

Rebecca Knight

>> Great.

Bill Pearson

>> Thanks. It's great to be here.

Rebecca Knight

>> 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.