Key Takeaways
The focus of the AI arms race is shifting from GPUs to the long-overlooked CPU, as the rise of complex agentic AI workloads threatens to change the fundamental architecture of data centers. The required CPU-to-GPU ratio is now expected to shift from a range of 1:4 to 1:8 to as tight as 1:1, creating a severe production bottleneck for a component that was previously an afterthought.
"CPU is facing an extremely serious capacity shortage," Dylan Patel, chief analyst at SemiAnalysis, said in an April interview. He noted that the paradigm for AI workloads is evolving from simple text generation to complex, multi-step tasks coordinated by AI agents, a process that is intensely CPU-dependent.
The market research firm TrendForce affirmed this judgment in a recent report, projecting the CPU-to-GPU ratio would narrow to between 1:1 and 1:2 in the era of agentic AI. In these new workloads, the CPU handles the "orchestration layer"—planning tasks, calling tools, and managing data flow between models. A 2025 academic paper, "A CPU-Centric Perspective on Agentic AI," found that CPU-based tool processing can account for up to 90.6% of total latency in agentic tasks. Arm calculates that this translates to a four-fold increase in demand, from 30 million CPU cores per gigawatt in traditional AI data centers to 120 million cores for agentic AI.
This structural demand shock is reshaping the competitive landscape, putting immense pressure on Intel's historical dominance while creating massive growth opportunities for AMD and new entrants Nvidia and Arm. For investors, this opens up new vectors for capitalizing on the AI infrastructure buildout beyond the well-established GPU trade.
The shift first destabilized the traditional x86 market. Intel, whose Xeon processors held over 95% of the server market, saw its position erode after its 7nm process yield issues delayed the Sapphire Rapids chip by nearly two years, opening the door for AMD's EPYC Milan. Intel's 2026 roadmap, including the 288-core Xeon 6+ and 256-core Xeon 7, relies on its still-unproven 18A process node. TrendForce reports that yield issues may delay mass production of these chips into 2027, likely allowing AMD to continue gaining market share with its 256-core/512-thread EPYC Venice, built on TSMC's N2 process.
The more significant change is the entry of non-traditional players. In March 2026, GPU-giant Nvidia announced it would sell its Vera CPU as a standalone product. The chip, based on TSMC's N3 process, features 88 cores and can be directly linked with Nvidia's GPUs via its NVLink-C2C interconnect. That same month, Arm ended its 35-year history as a pure IP licensor by announcing its own CPU, the Arm AGI. The 136-core chip, also on TSMC's N3 process, has already secured design wins with Meta, OpenAI, and Microsoft. Cloud providers are also accelerating their in-house CPU designs, with AWS's Graviton5, Microsoft's Cobalt 200, and Google's Axion all targeting cost reduction for AI workloads.
The surge in CPU demand creates a new, under-appreciated investment vector beyond GPUs. While Intel (INTC) faces significant execution risk on its 18A process, AMD (AMD) is positioned to continue its market share gains. The entry of Nvidia (NVDA) and Arm (ARM) opens new, multi-billion dollar revenue streams for both companies. This expansion of non-traditional CPU designers also directly benefits IC backend service firms like Global Unichip Corp. (GUC), which handles design for Google and Microsoft, and the broader TSMC advanced packaging ecosystem.
This article is for informational purposes only and does not constitute investment advice.
