LIVE: Intel CEO Lip-Bu Tan Delivers Major COMPUTEX 2026 Keynote in Taiwan | APT

Thank you, Lip-bu. It's so great to be here, specifically at this point in our history, global history, collective history, and be at Computex with the blue badge. I'm very happy and humbled to be here to share with you some of the innovations that we have. Let's talk a bit about what this AI thing is about. When we say foundational, we mean the workloads that keep the world running. Currently, we have data centers and there's a number of items and workloads and entities that run on these data centers. For example, we have 5G networks that keep us connected, databases that keep our data safe, and cloud services that power our daily lives. We expect demand for these workloads to grow in size and capacity between now and 2030 from 80 gigawatts to about 100 gigawatts. Most of you involved in this domain understand the extent of this type of expansion. These workloads are broad, they are mission-critical, so special attention has to be taken when running them. But also, they require performance, efficiency, security, and resiliency, and we can't emphasize enough all these four factors. That is why we are excited to have Intel Xeon 6 Plus introduced at Computex this week. It has 288 E-cores, a massive 576 megabytes of L3 cache, built with our Intel 18A technology. We can't emphasize enough the value that Intel technology brings to data center products. But most importantly, it delivers efficiency and density, which enables our partners to save very precious real estate, have more compact servers and racks. This is leadership compute for the next era of cloud and network infrastructure. Xeon 6 Plus launches with the strength of our ecosystem that's been built over decades of data center development, both from a hardware but also from a software and infrastructure perspective. Moreover, our ODM partners are bringing Xeon 6 Plus solutions to the market today. These range from full rack-scale deployments to server-level designs. Xeon 6 Plus joins our lineup of data center processors next to our already launched Xeon 6 based on P-cores. Both of these categories and classes of solutions deliver new performance and choice for all the enterprises whose infrastructure backbone is built on x86 and Xeon. This is critical for enterprises that need to increasingly balance preparing for AI workloads but at the same time running their day-to-day mission-critical applications. So let's switch gears and talk about how Intel is certifying the deployment of intelligence at scale. It's undeniable that enterprise infrastructure today will have to evolve to keep up with the AI demand. Recent research forecasts that AI inference workloads are expected to become 40% of all data center power demand and much more than they are today. We have these two paradigms where we have the foundational data centers keeping on running their traditional workloads, but at the same time they have to figure out ways of building their infrastructures to serve intelligence at scale. This is where Intel and Xeon 6 Plus come in. Up to now, training split the data center into two. On one hand, we have CPU-led enterprise infrastructure, and on the other hand, we have GPU-heavy AI factories. That was a very clear divide for a while, and we've all been accustomed to that reality. But as AI moves into real workflows, data tools, governance, the needs change. The next wave is not just about training models; it is about putting AI to work. So let's look at why Agentic AI changes the infrastructure equation. The way AI inference works is straightforward. We take a prompt, it gets fed into an LLM where it spends most time reasoning about the prompt. We've all seen this, we've done this thousands of times, and out comes an answer. In this case, a lot of time is spent computing the large language model, which is mostly GPU and compute-intensive. Now, the way agentic AI works is radically different. It's given goals rather than prompts. We've all seen the different types of loops that people are running on this agentic AI. It's also very iterative in nature but also prompted by automation, and thinking, planning, acting, and reflecting are a natural way of these agents interacting with us. As it works, it uses tools, reads and writes files, checks rules, and other aspects that were in the traditional realm of CPUs and x86. For each step, the type of underlying compute needs is very different, and we'll show that in a bit. This is particularly important as agents scale up their work, spawning new agents that work concurrently. The category and the complexity of agents are going to be very different depending on the complexity of the work. That's the main reason that there's such a rapid increase in CPU demand for agentic AI. The CPU orchestrates the show. Now, what we're seeing is the balance and the ratio of one CPU to eight GPUs and more is coming much closer to par. So let's take a look at a real example. John.

Thanks, Kavouk. You talked about how Agentic AI is changing the compute requirements. Let's take a look at a real example. I have a traditional AI inference setup on the left-hand side of the screen. Let's send a request: write a Python function that calls an OpenAI-compatible chat completions API. The model gets the response, generates code, and sends the request back. Take a look at the slider on the top of the screen. GPU dominates, nearly 7 to 1, GPU-heavy. In contrast, let's take a look at an agentic AI system. Across the top, look at the pipeline stages. Green is GPU work, blue is CPU work. Linting is happening on our Xeon 6 Plus processor with efficiency cores. Web fetch and compile is happening on our Xeon 6 performance cores, and unit testing is coming back and running on our Xeon 6 Plus efficiency cores. The right class CPU for each stage of the pipeline. Take a look at the slider across the top again. We're near par, but CPU-heavy this time. What's this look like when we multiply that by millions of queries a day? As you mentioned, each Xeon 6 Plus processor has up to 288 cores. That's 576 cores per two-socket server. When we look at that from a rack-scale perspective, that gives us over 36,000 cores per 32U of compute space.

Thank you, John. Wow, this is pretty amazing and some data to ponder on. By far, the density of CPU we showed has the highest density per rack ever. But also looking at the number of agents, and these are the new metrics that are emerging, we can safely say that that particular rack can run up to 150,000 agents. So good news to all the CIOs in the audience: now your very expensive GPUs can see more utilization because of our solutions. Now, both Xeon 6 with P-cores and E-cores are built for intelligence at scale. There are different cores, of course, but we've seen the workloads that require very high-performance cores pushing the frequencies, but also there's a need for very high-density, power-efficient cores. We've seen all the workloads, we've run all the analysis, and we are delivering these solutions to all of you now. Having said that, we are working with our customers and partners to make sure that each solution is tailored to your needs. So, I'd like to welcome Lip-bu back on stage to talk about the server and rack-scale solutions that our partners are working on. Thank you.

Thank you, Lip-bu. I'm so excited to be here today. Wonderful product and amazing event.

Yeah, that's right. I'm proud of the work we have done together, from AI servers to data centers and to edge computing. Today, we're excited to announce the next step in our partnership. Intel and Foxconn are working together to develop rack-scale products built upon Intel Xeon processors. Together, we will focus on exploring the development, integration, and commercialization of differentiated rack-scale AI infrastructure solutions, leveraging complementary architectures to address diverse AI workload requirements. Together, we will continue to deepen and expand our partnership, unlocking new opportunities ahead. Through this collaboration, we will deliver system-level AI solutions to our joint customers, enabling more integrated and scalable computing environments. This marks an important step ahead, and we look forward to unveiling more in the near future. Thank you, Lip-bu.

Fantastic. Thank you for having me.

Thank you, Lip-bu.

Absolutely. We've been busy. Earlier this year, we announced a multi-year collaboration to deliver high-performance, cost-efficient AI infrastructure solutions based on Xeon infrastructure. We've been building something really special, excited to show you today. This is the SN50 rack we announced earlier this year. Rack-scale AI infrastructure built for agentic workloads. It uses Intel Xeon 6 processors with SambaNova SN50 RDUs and is shipping to customers later this year. Today, we're also excited to demonstrate the world's first heterogeneous disaggregated inference using SambaNova's RDU with Intel's CPU and NVIDIA GPUs. What you're about to see is the same prompt, the same model, running side by side, two different stacks. The one on the left is GPUs, RDUs, and CPUs disaggregated inference. And this one on my right is GPUs on their own. They both get fed the same prompt and the same model, just different stacks. The disaggregated inference stack is taking off. What's happening here is you have the Xeon 6 processors doing all the tooling execution, you have SambaNova RDUs doing the decode and generating all of the tokens, and then you got the GPUs performing the prompt caching and the faster prefill, reducing overall time. When all three chips are working together, you dramatically reduce the end-to-end latency and the agents for the fastest need for agentic AI. On the other side, the GPU stack is still working away. So the initial result of our work in disaggregated inference is the GPUs, the RDUs, the CPUs. That's the fastest, and artificial analysis and our tests found it to be two to three times faster than just the GPUs alone. This gives us an early look at how fast this can be. The most exciting part about all this is that we have tremendous customer interest in these solutions.

Thank you so much.

Great. Thank you. Lip-bu, good to see you.

Yes.