Shift Toward Agentic AI Reshaping Hardware Ratios

At the OCP APAC 2026 conference, ADVANCED MICRO DEVICES (AMD) indicated that the evolution of artificial intelligence from simple chatbots to “agentic AI” is expected to significantly alter the demand for central processing units (CPUs). According to reports from Digitimes, the transition toward autonomous AI agents is projected to shift the traditional CPU-to-GPU ratio from approximately 1:4 toward 1:2 or even 1:1.

Industry analysts note that while standard chatbots are primarily GPU-bound during the inference stage, agentic AI requires substantial computational cycles for orchestration, memory management, tool calls, and control logic. These specific tasks lean more heavily on CPU performance, potentially driving a resurgence in demand for high-performance processors alongside the continued expansion of GPU infrastructure.

Architectural Developments and Memory Management

The shift toward agentic AI is driving discussions regarding how hardware architectures handle large-scale context windows. While unverified discussions on social media platforms like Reddit suggest that NVIDIA’s Rubin architecture may already utilize a hierarchical memory system to manage context, AMD is expected to develop competing solutions to address similar bottlenecks.

Current AI models often rely on “context compaction” or Retrieval-Augmented Generation (RAG) to manage large datasets, which can lead to performance degradation as context windows expand. Speculative technical theories suggest that a tiered memory approach—utilizing quick memory for immediate context and “cold storage” for long-term data—could allow for context windows reaching tens of millions of tokens without the performance loss associated with current methods. Such a system would theoretically allow an agent to pull data from various tiers of memory to fulfill a query, reducing the need for external text files to maintain task continuity.

Supply Chain Constraints and Material Bottlenecks

As the demand for AI-capable hardware increases, the industry faces potential bottlenecks in specialized materials required for chip manufacturing. One critical component is T-glass, a high-temperature-stable glass fiber fabric used as a substrate to provide structural stability to chips. This material prevents the warping or deformation of the substrate caused by the extreme heat generated by modern AI processors.

NITTO BOSEKI CO LTD currently holds a significant market position, reportedly controlling approximately 90% of the T-glass market. While competitors such as Taiwan Glass are reportedly still in the research phase, NITTO BOSEKI CO LTD is expected to face capacity constraints, with new production capacity not anticipated until mid-2027 according to Trendforce.

Other specialized materials identified as potential supply chain focal points include high-performance copper foil from MITSUI KINZOKU CO LTD for high-speed signal transmission, and ABF insulating film from Ajinomoto. These materials are essential for maintaining the integrity of the increasingly dense and hot-running circuitry found in advanced AI servers.