Intel Computex 2026 Keynote. Key Takeaways & Full Transcript
Before we get to what Intel had to say at Computex 2026, I thought it might be fun to look at what they said exactly one decade ago, back in 2016. This was actually the first time I wrote about Intel/Computex, and one of the first articles I ever posted online.
As an Intel employee based in Asia, I attended Computex in person on a number of occasions in the 2005-2014 timeframe. Back then, there was a huge focus on client computing and each year Intel and its partner companies would come on stage showing off yet another laptop and extolling its virtues. I found this really tedious and boring, as you can likely tell from the article. Thank goodness the Computex speakers have something more exciting to talk about nowadays.
Side Notes:
It’s really interesting to see how Intel’s executive lineup as completely changed from a decade ago. Not a single one of the executives who spoke at Computex 2016 are still with the company. In fact, they’ve all been gone for around five years or more already.
I thought this 2026 Intel keynote was pretty ho-hum. LBT brought on so many speakers that the flow was constantly being interrupted, and he was basically just mostly choreographing the event. I also don’t find him a compelling keynoter. He constantly has to look to his monitor for his lines making his delivery seem awkward and artificial. He’s certainly no Jensen when it comes to keynotes.
On a more positive note, Alex, Kevork and Srini seem like really smart & experienced executives. I expect good things from them.
Corporate Execution and Heritage
Intel CEO Lip-Bu Tan opened the address by emphasizing a rigorous return to Intel’s core engineering roots over his 14-month tenure. Highlighting Intel’s deep, 40-year historic tie to Taiwan’s semiconductor ecosystem, Tan underscored that corporate focus has been shifted entirely toward execution, performance, and cross-ecosystem collaboration.
The Four Compute Ecosystems & Client Innovations
Intel is targeting trillion-dollar opportunities across four structural domains: personal computers, the edge/physical AI, foundational data centers, and emerging digital intelligence centers.
Alex Katouzian (recently hired, Qualcomm veteran), lead for Client Compute and Physical AI, announced the full-scale volume ramping of Intel’s 18A process technology. Key product expansions include:
Core Ultra Series 3: Built on 18A, securing over 300 premium mobile and commercial design wins.
Core Series 3: Launched in April for mainstream, thin form-factor PCs, scaling rapidly to over 70 designs.
Arc G3: A specialized 18A handheld gaming GPU delivering AAA gaming at 1080p, running 40% faster and at half the power of competitors.
Edge & Physical AI: Deploying Series 3 silicon across 4,000 edge ecosystem partners, positioning Intel to capture a projected $25 trillion physical AI market by 2050.
Hybrid Agentic Inference
Perplexity CEO Aravind Srinivas joined Tan to demonstrate the Perplexity Computer AI operating system running locally on Core Ultra Series 3 hardware. The architecture executes local, hybrid agentic inference, keeping highly confidential enterprise data secure on-device while delegating non-sensitive research tasks to cloud servers. This framework optimizes token value per watt by balancing on-device privacy with data center capacity.
Data Center: Xeon 6 Plus & The Architecture Shift
Kevork Kechichian (recently hired, previously also a Qualcomm veteran) introduced the Intel Xeon 6 Plus, built on the 18A process. Engineered for foundational cloud and network infrastructure, it packs 288 E-cores and 576MB of L3 cache to deliver unprecedented rack density (up to 150,000 agents per 32U rack).
Crucially, Intel demonstrated that agentic AI workflows alter infrastructure demands. Unlike traditional, GPU-heavy LLM inference, autonomous agents iteratively think, plan, and use tools. Because these processes rely heavily on the CPU, agentic workloads bring the ratio of CPUs to GPUs closer to parity. To rapidly scale these architectures, Intel launched its Rack Scale Blueprints initiative, partnering closely with Foxconn to commercialize system-level AI solutions.
Heterogeneous Compute & Purpose-Built Silicon
SambaNova CEO Rodrigo Liang showcased the SM50 Samba Rack, demonstrating the world’s first heterogeneous disaggregated inference utilizing Intel Xeon 6 CPUs, SambaNova RDUs, and NVIDIA GPUs together. This combined stack runs agentic AI workloads two to three times faster than standalone GPU systems by offloading tooling to the CPU and token decoding to the RDU. Vista Equity Partners confirmed a major deployment of this air-cooled architecture to scale low-cost enterprise inference.
Finally, Srini announced Intel’s official entry into the custom, purpose-built silicon market. Key custom deployments include delivering vital Infrastructure Processing Units (IPUs) for Google and next-generation telco silicon for Ericsson. Intel also highlighted strategic vertical collaborations in biomedical engineering (Echo Neurotechnologies and Greenstone Biosciences) and industrial automation (Hitachi and Siemens) to co-develop custom, brain-inspired, and AI-driven hardware.
Full transcript below the fold.
