Intel’s Computex 2026 vision for AI PCs, agentic data centers, and custom silicon

Intel’s Computex 2026 keynote was less about a single chip launch and more about a full-stack AI strategy. From AI PCs and handheld gaming to agentic data centers, rack-scale inference, and custom silicon for medicine and industry, Intel is trying to position itself as the silicon backbone of the AI era.

Intel’s new leadership and a return to engineering roots

Intel’s CEO Lip-Bu Tan opened by tying today’s AI race back to the origins of Silicon Valley and Taiwan’s “Silicon Island.” He emphasized that Intel is refocusing on what it has always done best: engineering.

All core engineering teams now report directly into the CEO, with execution and technical excellence framed as the company’s top priorities. The message was clear: Intel wants to be judged on shipped silicon, not slideware.

AI PCs and the Core Ultra Series 3

The first big focus was the PC, which Intel still sees as the starting point for mass-market AI. The company highlighted its Intel 18A process technology and the Core Ultra Series 3, described as Intel’s first major AI-first PC platform on 18A.

Core Ultra Series 3 combines:

• High-performance CPU cores for responsiveness
• A significantly improved integrated GPU for graphics and light compute
• A low-power NPU (neural processing unit) for on-device AI
• Updated multimedia engines for modern content workloads

According to Intel, more than 300 Core Ultra Series 3 designs are already shipping across consumer and commercial laptops, with a focus on premium performance and battery life.

Bringing AI to mainstream laptops: Intel Core Series 3

To push AI capabilities into more affordable devices, Intel has repackaged the same 18A IP into the new Intel Core Series 3, introduced in April.

Key points:

• Over 70 designs already in development or shipping
• Nearly 400 total designs across Core Ultra and Core Series 3 in just a few months
• Focus on thin, light, mainstream laptops with “premium feel” and all-day battery life

Intel also took a small jab at competitors by stressing that its mainstream systems still offer multiple ports and broad connectivity, rather than relying on a single USB‑C port.

Handheld gaming powered by Arc G3

One of the fastest-growing PC segments is handheld gaming. Intel is targeting this space with the Arc G3 GPU, derived from the Core Ultra Series 3 platform.

According to Intel’s claims, Arc G3 delivers:

• Over 40% higher performance than competing solutions in tested games
• Comparable performance at roughly half the power
• 1080p gaming across AAA titles, with many running above 120 fps

New handheld devices based on Arc G3 are expected to ship starting this month, with more designs coming throughout the year.

From PCs to the edge and “physical AI”

Intel is extending the same 18A-based IP into edge computing, where demand is surging for AI-enabled devices in the physical world. The company is reusing its PC building blocks for edge systems and industrial deployments.

Highlights include:

• Core Series 3 being used in more than 130 edge designs across manufacturing, robotics, retail, and other verticals
• Over 4,000 edge ecosystem partners building on Intel platforms
• A long-term bet on “physical AI” – AI embedded in machines, robots, and devices – which Intel cited as a potential $25 trillion market by 2050

Intel’s strategy is to offer not just chips, but complete reference platforms: silicon, chipsets, hardware designs, and software stacks tailored to specific physical AI use cases.

Hybrid AI with Perplexity: on-device plus cloud

On-device AI was a recurring theme, and Intel showcased a concrete example with Perplexity’s “Computer” system – an AI operating system built around multiple cooperating agents.

In a demo scenario for a private equity analyst working on a confidential deal (“Project Falcon”), Perplexity’s system:

• Runs smaller models locally on an Intel Core Ultra Series 3 machine
• Uses local AI to classify which files are sensitive and must never leave the device
• Orchestrates which parts of the task can safely be sent to large cloud models
• Combines local and cloud outputs into a single coherent result

This “hybrid agentic inference” approach aims to maximize “token value per watt per user” – in other words, getting the most useful AI work done per unit of energy and cost, while preserving privacy. It’s a practical blueprint for how future AI PCs might blend local and cloud intelligence, similar to how other platforms are evolving at events like Google I/O 2026.

x86 and the new Xeon 6 family

Moving from clients to the cloud, Intel doubled down on x86 as the backbone of general-purpose computing. The company noted that IDC expects 8 out of 10 servers installed through 2030 to be x86-based.

Intel’s CPU roadmap now centers on two flagship core types:

• P-cores (performance cores) for maximum single-thread and latency-sensitive workloads
• E-cores (efficiency cores) for dense, power-efficient throughput

These cores power everything from PCs and edge devices to data center platforms. Under Lip-Bu Tan’s leadership, Intel’s stated goal is to build “the best CPU cores in the world,” especially for compute-intensive workloads.

Xeon 6 and Xeon 6 Plus: built for agentic AI at scale

The data center portion of the keynote focused on the new Xeon 6 family, split into two main lines:

• Xeon 6 (P-cores): for high-performance, latency-sensitive workloads
• Xeon 6 Plus (E-cores): for high-density, power-efficient workloads

Xeon 6 Plus, built on Intel 18A, offers:

• Up to 288 E-cores per CPU
• 576 MB of L3 cache
• Up to 576 cores in a dual-socket server
• Over 36,000 cores in a 32U rack configuration

Intel framed this as “leadership compute for the next era of cloud and network infrastructure,” especially as AI inference workloads grow to an estimated 40% of total data center power demand.

Why agentic AI changes data center design

Intel drew a sharp distinction between traditional AI inference and emerging agentic AI workloads.

In classic LLM inference:

• A user sends a single prompt
• The model spends most of its time on GPU-heavy compute
• The workload is dominated by one long forward pass through a large model

In agentic AI:

• Agents are given goals, not just prompts
• They plan, act, check rules, call tools, read and write files, and spawn sub-agents
• Workloads become iterative, branching, and highly CPU-intensive for orchestration

Intel showed a side-by-side comparison:

• Traditional inference: GPU-heavy, with a 7:1 GPU-to-CPU utilization ratio
• Agentic pipeline: near parity or even CPU-heavy, with different stages mapped to different cores (E-cores for linting, P-cores for web fetch and compile, etc.)

This shift drives demand for dense CPU infrastructure to coordinate and run large numbers of agents, while still pairing with GPUs or other accelerators for heavy model math.

Rack Scale Blueprints and Foxconn partnership

To meet these new workloads, Intel is moving beyond individual servers to “Rack Scale Blueprints” – reference designs for entire racks built on open standards.

Two example blueprints were highlighted:

• A high-performance rack based on Xeon 6 with P-cores
• A high-density “agent rack” based on Xeon 6 Plus with E-cores, capable of running up to ~150,000 agents per rack

Intel is working with partners like Foxconn to turn these blueprints into real products. Together, they plan to develop, integrate, and commercialize differentiated rack-scale AI infrastructure tuned for diverse workloads.

Disaggregated AI inference with SambaNova and Vector Core Compute

Beyond CPUs and GPUs, Intel is betting on heterogeneous, disaggregated inference – splitting different parts of the AI pipeline across specialized chips.

With SambaNova, Intel is building rack-scale AI systems that combine:

• Intel Xeon 6 processors for orchestration and tooling
• SambaNova SN50 Reconfigurable Tensor Units (RTUs) for token generation and decode
• Nvidia GPUs for prompt caching and fast prefill

In a live demo, the same model and prompt ran on two stacks:

• Disaggregated stack: CPUs + SambaNova RTUs + GPUs
• GPU-only stack

The disaggregated setup delivered 2–3x lower end-to-end latency than GPUs alone in internal tests, showing how mixing architectures can accelerate agentic AI.

On top of this, Vista Equity Partners and Cambium Capital have launched Vector Core Compute (VC2), which they describe as the first commercially available architecture for disaggregated inference. VC2 is:

• Built on the SambaNova stack with Intel Xeon infrastructure
• Air-cooled and designed for low-latency, low-cost inference at scale
• Already deployed in a Los Angeles data center, with over 50 more deployments planned to convert existing facilities into inference-optimized data centers

Together AI is the first commercial customer, offering this architecture as a service. With over 90 portfolio companies and 750 million users, Vista expects to support more than 10 billion agents on this kind of infrastructure.

Intel enters the custom silicon market

Another major announcement was Intel’s formal entry into the purpose-built (custom) silicon market, leveraging its foundry capabilities and IP portfolio.

Two flagship customers were highlighted:

• Google: Intel is delivering an Infrastructure Processing Unit (IPU) – a critical piece of hyperscale data center infrastructure – already designed and deployed in production.
• Ericsson: Intel is providing next-generation infrastructure silicon for Ericsson’s global telco deployments.

Intel sees custom silicon as a high-growth area and is positioning itself as a partner for both hyperscalers and startups that need chips tailored to their workloads.

Brain-inspired and biomedical AI partnerships

Beyond traditional compute, Intel showcased several partnerships at the intersection of AI and biology.

Echo Neurotechnologies: brain-trained AI

Echo Neurotechnologies is working with Intel on brain-inspired computing. Instead of just being “brain-inspired,” they are training algorithms directly on human brain activity.

Key ideas:

• High-resolution recordings of human cortex activity during real-time language processing
• Algorithms designed around how the brain actually computes – spikes, sparse communication, and tightly coupled memory and compute
• Early focus on streaming speech and restoring communication for people who have lost the ability to speak

The long-term vision is AI that approaches the efficiency of biological computation, potentially informing future neuromorphic hardware.

Greenstone Biosciences: AI and patient-specific organoids

Greenstone Biosciences, founded by Stanford cardiologist Joseph Wu, is partnering with Intel to accelerate drug discovery.

Greenstone has built what it calls the world’s largest biobank of human induced pluripotent stem cells (iPSCs). From a small blood sample, they can generate organoids – miniaturized versions of organs like the heart, brain, liver, kidney, and gut – that are genetically matched to individual patients.

Intel’s role is to provide scalable AI compute for:

• Processing and storing massive imaging and biological datasets
• Running AI models to analyze how different drugs affect patient-specific organoids
• Helping predict safety and efficacy faster and at lower cost

This combination of human biology and AI could significantly speed up and personalize medicine in the coming decade, echoing broader trends in AI-powered healthcare tools similar to those we’ve seen in other domains like discussions around the AI singularity.

Industrial and quantum partnerships: Hitachi and Siemens

Intel is also deepening partnerships in industrial and quantum domains.

With Hitachi, Intel plans to combine advanced computing with Hitachi’s industrial expertise to build “intelligent solutions” that bridge digital and physical systems. This includes work around foundry tools and quantum computing systems.

With Siemens, Intel is expanding collaboration across the entire semiconductor value chain:

• EDA automation and AI-assisted design tools
• Product lifecycle management and factory automation
• Electrification, quality, and sustainability in manufacturing
• Using chips produced in this ecosystem inside Siemens’ own industrial products

The goal is a tightly integrated loop where Intel’s chips power Siemens’ systems, and Siemens’ tools and factories help design and build Intel’s chips.

Intel’s AI-era roadmap: from silicon to systems

To close, Intel framed its opportunity across four major compute ecosystems:

• PCs: AI-first laptops and handhelds powered by Core Ultra and Core Series 3
• Edge and physical AI: robots, retail, manufacturing, and other devices built on the same IP
• Foundational data centers: Xeon 6 and Xeon 6 Plus for traditional and AI workloads
• Emerging “intelligent centers”: agentic AI infrastructure, disaggregated inference, and custom silicon

Over the past year, Intel says it has:

• Ramped Intel 18A to high volume with multiple products
• Hit key milestones in advanced packaging
• Grown its foundry business with new external customers
• Launched new SoCs across mobile, cloud, and 5G
• Rebuilt partnerships across the ecosystem
• Opened new business lines in custom silicon and AI-optimized systems

The company’s message at Computex 2026 was that it is “built different, built together, built on Intel” – aiming to be the engineering engine behind the next generation of AI, from the laptop on your desk to the racks powering billions of agents in the cloud.