CPU demand set to surge with agentic AI; AMD best positioned

As artificial intelligence moves beyond training large models toward building autonomous agents capable of reasoning and acting independently, the semiconductor architecture powering AI data centers must evolve. The article argues that this shift fundamentally reshapes the role of CPUs relative to GPUs. Graphics processors excel at parallel tasks and raw computing power, making them ideal for model training. Central processors, by contrast, handle sequential reasoning—the step-by-step logic that allows an AI agent to pause, evaluate, and decide before acting. This difference in workload profiles means the balance between the two processor types will shift dramatically.

Today, training workloads run at an 8-to-1 GPU-to-CPU ratio. As AI inference (the production phase where trained models generate answers) becomes more prevalent, that ratio compresses to 4-to-1. For agentic AI, the article projects it will reach 1-to-1—meaning as many CPUs as GPUs in the data center. Nvidia estimates this market transition could reach $200 billion（約32兆円） within the next few years, underscoring the scale of the opportunity.

Among the three CPU vendors—AMD, Arm, and Intel—AMD emerges with the clearest competitive advantage. The company has been steadily taking market share from Intel in data center CPUs, leveraging strong technology and architectural innovation. Its forthcoming Venice architecture represents the culmination of this strategy: it will support up to 256 cores, an unusually high count that mirrors the demands of agentic AI workloads, where more cores function like a larger workforce for autonomous systems. Beyond CPUs, AMD has diversified into inference GPUs, recently acquiring memory optimization company MEXT to strengthen its position. The chiplet design from MEXT allows more memory to be packaged with the GPU, a critical advantage in inference workloads. AMD has already converted this into two massive commercial partnerships: one with OpenAI and another with Meta.

Intel has seen its stock surge roughly 323% over the past year, driven largely by elevated data center CPU demand. However, the article cautions that this gain masks underlying vulnerabilities. While Intel remains a substantial CPU vendor despite having lost technological leadership to AMD, its core PC business has stagnated, and rising component costs—including memory, not just CPUs—could eventually damp demand for personal computers. The company's foundry business, which manufactures chips for external customers, continues to burn losses. Once valued cheaply for its physical assets, Intel no longer trades at a discount following its year-long rally, raising questions about valuation sustainability.

Arm shocked the market by announcing it would abandon its traditional licensing model and begin designing and selling its own CPUs, a dramatic strategic pivot driven by the scale of the data center opportunity. Arm's architecture, which differs from the x86 standard used by Intel and AMD, has become the dominant standard in smartphones and powers custom data center chips from cloud giants including Nvidia, Amazon, and Alphabet. At the time of its CPU announcement, Arm projected the data center CPU market would reach $100 billion（約16兆円） over five years and believed it could capture 15% of that share, translating to $25 billion（約4兆円） in 2031 revenue, of which $15 billion（約2.4兆円） would come from CPUs alone. Yet Arm confronts significant headwinds. Its core smartphone business faces pressure from high memory costs that are expected to raise handset prices and curb demand. Securing adequate manufacturing capacity to support its CPU ambitions remains uncertain. Perhaps more damaging, Arm now risks alienating its own customers—the cloud providers and chip designers who have licensed its technology—by competing directly with them as a merchant CPU vendor. This dynamic could undermine relationships that Arm has relied on for decades.

The article identifies a structural shift in AI infrastructure driven by the emergence of agentic AI—systems that reason sequentially and act autonomously. This shift fundamentally alters the balance between CPUs and GPUs. While GPUs have dominated AI data centers because they excel at parallel processing and raw compute, CPUs are better suited to the sequential reasoning that agentic systems require. Nvidia's projection of a $200 billion（約32兆円） market for this workload signals both the scale of opportunity and the credibility of the thesis.

The three chipmakers face materially different competitive positions. AMD has first-mover advantage and existing relationships (deals with OpenAI and Meta), plus a second revenue stream from inference GPUs. Intel, despite a 323% stock surge over the past year, remains hampered by a weak broader PC business and losses in its foundry division—gains in data center CPUs may not offset these structural headwinds. Arm, the third player, has the most aggressive growth claims (15% market share in a $100 billion（約16兆円） market) but faces real execution risks: manufacturing capacity is unproven for its CPU ambitions, and by building its own chips, it now competes against the customers (Amazon, Alphabet, Nvidia) who have historically licensed its architecture for custom designs. This conflict of interest may limit Arm's upside relative to AMD's straightforward position as a merchant CPU vendor.

No comments yet. Be the first to share your thoughts!