Intel at Computex 2026 leans on new client, edge, and rackscale AI push

Intel used its Computex 2026 keynote to connect a familiar narrative, x86 powering computing for decades, to a newer theme, building AI systems from the device to the data center. CEO Lip-Bu Tan took the stage in Taipei and framed Intel’s platform strategy as an effort to support privacy-sensitive workloads at the edge while also scaling to cloud and enterprise deployments where inference needs more context and compute.

Tan credited x86 for its long-running role in data centers and beyond, and he pointed to Taiwan’s manufacturing and systems ecosystem as an important partner for Intel’s products across desktops and notebooks and other form factors. The keynote then shifted to Intel’s client and physical AI roadmap, with new leadership for client computing and physical AI, Alex Katouzian, outlining how Intel wants to extend its newest process technology into a broader set of computing experiences.

Katouzian began with the Core Ultra Series 3, described as Intel’s first product built on Intel 18A. He said the chip line is designed to deliver a “full XPU experience,” combining a CPU for general-purpose tasks, a GPU for graphics and acceleration, and a low-power NPU for AI workloads. Intel also said the Core Ultra Series 3 has been set into “more than 325 designs” already across consumer and commercial products, tying the launch to fast platform adoption rather than just a single reference product.

Next, Katouzian reintroduced the Intel Core Series 3, calling it a mainstream option that uses the same underlying IP as the Ultra line. Intel positioned the Core Series 3 around all-day battery life, connectivity-focused port availability, and performance in “incredibly thin” mainstream PC designs. He then introduced an Intel Arc G3 series processor line, built on the same Intel Core Ultra Series 3 architecture, but optimized for handheld gaming, with availability targeted “later this month,” according to the keynote.

For edge and physical AI, Intel said Intel 18A already has “over 130 edge designs.” The company also cited a large partner and deployment base, with “over 4,000 edge ecosystem partners” and “over 100,000 edge deployments” across areas including manufacturing, robotics, and retail. The implication is that Intel is treating edge compute not as an experimental niche, but as a foundation to expand into new physical AI form factors such as robotics and other AI devices that operate outside traditional cloud data centers.

Tan returned to the stage with Perplexity CEO Aravind Srinivas for a discussion of hybrid AI computing optimized from cloud to edge. They argued that cost, compliance, security, and privacy requirements are pushing inference to be split between local device compute and cloud environments. Srinivas described how Perplexity built a hybrid local server for inference orchestration, enabling workloads to scale dynamically between local and cloud based on device capabilities.

In the data center segment, Kevork Kechichian, executive vice president of Intel’s Data Center Group, announced Intel Xeon 6+ processors. Intel characterized Xeon 6+ as built on Intel 18A technology and highlighted specific design targets, including 288 e-cores and a 576MB L3 cache, with an emphasis on compute density and efficiency for enterprises balancing AI readiness against “mission critical” day-to-day workloads. Intel also raised a forecast claim, stating that recent research expects AI inference to become nearly 40% of data center power demand by 2030, and linked that pressure to the rise of agentic AI workloads that require more iterative CPU-coordinated reasoning.

To address those workload changes, the keynote turned to rackscale systems, with Tan describing the need to reinvent compute beyond the socket for agentic AI. Intel and Foxconn were positioned as partners for rackscale AI infrastructure, with Foxconn described as supporting systems integration for rackscale inference and agentic AI workloads. Intel also highlighted the role of token usage, saying reports indicate an agent can consume up to 1,000 times more tokens than single-turn reasoning, and argued that future systems must be optimized for token consumption and generation alongside CPUs.

Intel further said it is collaborating with SambaNova, Vista Equity Partners, and Cambium Equity on cost- and energy-efficient inference solutions. Tan said SambaNova CEO Rodrigo Liang joined the stage to discuss additional announcements, including a plan for “fully disaggregated inference” via a new inference cloud offering called Vector Core Compute. The keynote closed by broadening the “purpose-built silicon” theme, including work with Google on IPUs for cloud providers’ optimal performance, and a partnership with Ericsson aimed at wireless infrastructure chips. Intel also referenced collaborations with Hitachi, Siemens, Echo Neurotechnologies, and Greenstone Biosciences across energy, industrial automation, biomedical engineering, and drug development, positioning vertical solutions as the next step after silicon and systems.