Intel is betting that the next phase of AI infrastructure won't look like the last one. At Computex 2026, the chipmaker unveiled its Xeon 6+ processor and outlined a sweeping vision for how inference, the process of running trained AI models to generate answers, is reshaping computing from data centers to handheld devices. The company argued that as AI moves from training massive models to actually using them in production, the balance of computing power needs to shift dramatically.

Why Is Inference Suddenly So Important?

For years, the AI industry focused almost entirely on training, the computationally expensive process of teaching models on massive datasets. But Intel's leadership made a clear case that inference, the lighter-weight task of running those trained models to answer questions or complete tasks, is becoming the real bottleneck. The company noted that inference and agentic AI tasks, where AI systems make decisions and take actions autonomously, are increasing the importance of CPUs in coordinating reasoning and orchestration.

This matters because it changes what kind of hardware companies actually need. Intel said concerns about privacy, security, compliance, and cost are pushing customers toward hybrid computing models, where some AI processing happens on local devices and some happens in the cloud. This split approach lets companies keep sensitive data close to home while still drawing on large-scale cloud systems when needed.

The scale of this shift is substantial. Intel's research forecasts suggest that AI inference workloads could account for nearly 40 percent of all data center power demand by 2030, while wider use of agentic AI could sharply increase token consumption, the measure of how much text an AI model processes.

What Hardware Changes Are Coming?

Intel's announcements spanned three major product categories, each designed to handle inference at different scales. The Xeon 6+ processor, built on Intel's 18A process technology, features 288 efficiency cores and 576 megabytes of L3 cache, designed specifically for data center inference workloads.

On the consumer side, Intel introduced the Core Ultra Series 3, which it described as its first product built on 18A technology. The chip is already included in more than 325 consumer and commercial designs. The company also introduced the Arc G3 series, based on the same architecture and aimed at handheld gaming devices, with availability expected later in June 2026.

For edge computing, the category that includes manufacturing equipment, robotics, and retail systems, Intel said 18A already has more than 130 edge designs in development. The company's broader edge network includes more than 4,000 ecosystem partners and more than 100,000 deployments across sectors including manufacturing, robotics, and retail.

How Are Companies Actually Implementing Hybrid AI?

Intel's keynote featured a discussion with Perplexity, an AI search company, about how hybrid computing works in practice. Perplexity has built what it described as a hybrid local server for inference orchestration, designed to move workloads between local and cloud environments depending on device features and available resources.

"Inference workloads are reshaping system design away from the ratios used in large model training," explained Intel executives during the presentation.

Intel Leadership, Computex 2026

This orchestration layer is critical because it decides in real time whether a particular AI task should run locally on a user's device or be sent to the cloud. The decision depends on factors like the complexity of the task, the capabilities of the local hardware, and the sensitivity of the data involved. Intel argued that this approach is becoming increasingly important as companies try to balance performance, privacy, and cost.

Steps to Understanding the Shift From Cloud-Only to Hybrid AI

• Local Processing: Simple, routine AI tasks run directly on your device, keeping data private and reducing latency, the delay between when you ask a question and when you get an answer.
• Cloud Processing: Complex tasks that require more computing power or access to real-time information are sent to cloud servers, where they can leverage larger models and more resources.
• Orchestration Layer: Software decides which tasks go where based on device capabilities, data sensitivity, and available resources, optimizing for speed, privacy, and cost simultaneously.
• Infrastructure Design: Data centers are being redesigned to handle inference-heavy workloads rather than training, shifting from GPU-dominated systems to more balanced CPU and accelerator configurations.

Intel also announced a collaboration with SambaNova, Vista Equity Partners, and Cambium Equity on inference systems intended to reduce cost and energy use. The partners plan to launch a new inference cloud service called Vector Core Compute, using infrastructure from Intel, Nvidia, and SambaNova.

What Does This Mean for Computing Beyond Data Centers?

Intel's broader pitch emphasized that AI infrastructure needs to span multiple form factors and use cases. The company highlighted its work on purpose-built silicon for specific customer workloads, including collaborations with Google on IPUs, specialized processors for inference, and with Ericsson on wireless infrastructure chips.

Intel also outlined industry-specific work with companies including Hitachi, Siemens, Echo Neurotechnologies, and Greenstone Biosciences. These partnerships focus on sectors such as energy, industrial automation, biomedical engineering, and drug development, where customers are seeking chips tailored more closely to their computing requirements.

The overarching message from Intel's Computex presentation was that the next phase of AI infrastructure will be defined not by who can build the biggest training systems, but by who can efficiently run inference across the widest range of devices and environments. As AI models become more prevalent in everyday applications, the ability to run them locally, securely, and cost-effectively is becoming a competitive advantage. Intel's strategy suggests that the future of AI computing is not centralized in the cloud, but distributed across data centers, edge devices, and consumer hardware working in concert.