As Agentic AI moves into real-world workflows, Intel believes the AI era will not have GPU as the sole winner, with CPU demand poised to return to a growth trajectory.

During a keynote address at Computex 2026 on Tuesday, Intel CEO Lip-Bu Tan outlined a clear industry vision: future AI infrastructure will not be dominated solely by GPUs but will instead enter an era of heterogeneous computing where CPUs, GPUs, ASICs, and custom chips develop synergistically. With the rapid adoption of Agentic AI, data center demand for CPUs is expected to re-enter a growth cycle.

Based on this assessment, Intel comprehensively showcased progress in its 18A process node volume production, its Xeon 6 product portfolio, and the expansion of its custom chip business. In collaboration with partners including Perplexity, SambaNova, and Foxconn, the company demonstrated new computing architectures designed for Agentic AI. Tan stated in his speech:

"We will not rest on past glory; we are building a new Intel."

This strategic repositioning comes as Intel's fundamentals show signs of recovery. Two years ago, the company faced significant transformation pressure. Under Tan's leadership, its Q1 2026 revenue, profit, and Q2 guidance all exceeded market expectations. Tigress Financial Partners analyst Ivan Feinseth remarked that Tan is "exactly the CEO Intel needs," significantly improving the company's execution while continuing its established strategy. Notably, over the past year, Intel's stock price has surged 454%, at one point reaching a new all-time high.

For the capital markets, the core signal from this presentation was not merely a single product update but Intel's attempt to re-establish its positioning within the AI infrastructure supply chain: transitioning from a traditional CPU manufacturer to a full-stack computing platform covering PCs, edge computing, data centers, AI racks, and custom silicon.

Agentic AI Alters Compute Demand, Intel Bets on CPU Revival

Over the past two years, the large language model wave fostered a market consensus: AI demand equates to GPU demand. However, Intel explicitly proposed in this presentation that this logic is being reshaped by the rise of Agentic AI.

Intel posits that compute consumption for traditional generative AI is heavily concentrated on GPUs, whereas Agentic AI has evolved from single-turn Q&A to autonomous workflows executing complex tasks involving planning, tool calling, database access, code execution, and multi-agent collaboration—tasks that rely heavily on CPUs.

Data presented on-site showed that in traditional AI inference, the resource demand ratio between GPU and CPU is approximately 7:1. In Agentic AI workflows, CPU demand increases significantly, even surpassing GPU demand in some scenarios.

Based on this, Intel predicts future AI systems will evolve from singular model inference platforms into complex workflow systems with autonomous decision-making capabilities. In this architecture, the CPU will assume the role of orchestrating and coordinating the entire system. As stated in the presentation: "CPU orchestrates the show"—the CPU is once again becoming a critical component in AI infrastructure.

Xeon 6 Plus Debuts, Targeting the Agent Era Data Center

Based on the above rationale, Intel officially launched the Xeon 6 Plus processor built on the 18A process. This product features 288 E-Cores and 576MB of L3 cache, targeting cloud computing, network infrastructure, and AI inference scenarios.

Intel believes future data centers will simultaneously host traditional enterprise applications and AI workloads. Enterprises will not only need GPUs for model inference but also a significant number of CPUs for agent orchestration, tool calling, and workflow management. On-site data indicated that a dual-socket Xeon 6 Plus server provides 576 CPU cores, with a single rack capable of deploying over 36,000 cores, supporting up to approximately 150,000 concurrent agents.

This signifies a shift in the key metrics for evaluating AI infrastructure value: from pure GPU count to the scale of supported agents and inference efficiency. The launch of Xeon 6 Plus is not just a technological iteration but a crucial strategic move for Intel from "CPU-centric general-purpose computing" to "CPU+GPU collaborative agentic computing."

From Selling CPUs to Selling Racks, Intel Enters the AI Rack Market

Beyond the chips themselves, Intel is extending into system-level solutions. CEO Lip-Bu Tan announced the Rack Scale Blueprint initiative, collaborating with partners through open standards to develop rack-scale AI infrastructure solutions. Initial partners include Foxconn and SambaNova.

According to the presented plans, different rack architectures can be optimized for high-performance computing or high-density agent deployment, helping enterprises rapidly build AI infrastructure without needing to integrate complex hardware and software systems themselves.

This approach aligns with current trends in the AI industry. As AI system scale grows, enterprise customers increasingly prefer purchasing validated full-rack solutions rather than buying chips separately and integrating them.

For Intel, this is not just a product extension but a reshaping of its ecosystem advantages. Leveraging its longstanding presence in the server market and ODM partner network, Intel aims to rebuild a competitive edge in the system-level AI infrastructure domain.

Disaggregated Inference: Intel Demonstrates New Heterogeneous Computing Paradigm

The "disaggregated inference" architecture jointly showcased by Intel and SambaNova was another technical highlight. This solution decomposes AI inference tasks: GPUs handle prompt preprocessing, SambaNova RTUs handle token generation, and Xeon processors handle agent orchestration and tool execution.

On-site data was compelling: for the same model and task, this CPU+GPU+RTU collaborative approach achieved 2 to 3 times the performance improvement compared to a pure GPU architecture.

Intel noted this result validates its core thesis—future AI infrastructure will enter an era of heterogeneous computing, with different chips optimized for different stages, rather than a single processor handling everything. This also explains why Intel continues to invest in its diversified portfolio of CPUs, GPUs, and custom ASICs.

From Google to Ericsson, Intel Doubles Down on Custom Chips

Beyond its standard product lines, Intel formally announced its entry into the Purpose-Built Silicon market. Srini, head of the business, noted that over the past decade, hyperscale cloud providers have validated the value of the custom chip model—and Intel possesses full-stack capabilities from design and manufacturing to packaging, enabling it to help customers develop proprietary compute architectures.

Currently, Intel has disclosed two key collaborations: with Google to provide Infrastructure Processing Units (IPUs), and with Ericsson to jointly develop next-generation communications infrastructure chips. CEO Lip-Bu Tan stated that enterprises will increasingly prefer building proprietary computing platforms around their specific workloads rather than adopting a "one-size-fits-all" unified architecture. Intel aims to precisely target this fast-growing market through its custom chip business.