AI Agent Era, CPU is a Bottleneck Again
Intel vs Qualcomm vs ARM Winner Prediction

The Great Migration of Bottlenecks: From GPU to CPU. According to a joint study by Georgia Tech + Intel (arXiv:2511.00739, 2025.11), when an AI agent performs external API calls, web searches, code execution, and sub-agent orchestration, all of these coordination layers are concentrated on the CPU. The GPU bottleneck of the Training era is shifting to a CPU bottleneck in the Inference + Agent era. FACT

Structural Demand Explosion. ARM estimates that CPU core demand in AI agent data centers will increase fourfold from the existing 300,000 cores/GW to 1.2 million cores/GW. Intel CFO officially confirmed on the Q1 2026 earnings call that unmet demand for Xeon server CPUs is in the billions of dollars and lead times are six months. FACT

Key Framework for Winner Prediction. The three companies each own bottlenecks at different layers: Intel owns the server CPU supply bottleneck (Xeon), Qualcomm owns the edge NPU performance bottleneck (85 TOPS), and ARM owns the architecture IP bottleneck (the foundation of all non-x86 chips). In any scenario, ARM collects royalties, but the short-term risk-reward optimal solution is Qualcomm. INFERENCE

TAM Sizing. The AI agent CPU market (edge + data center) is expected to grow at a CAGR of +32% from approximately $52B in 2026 to $112B in 2028. Edge (AI PC, smartphones, IoT) is expected to account for 60% of the total, and data centers 40%. INFERENCE

Value Capture Structure. IP licensors (ARM) collect 5-15% of revenue as high-margin royalties, chip designers (Qualcomm, Intel) capture 40-60% as added value, foundries (TSMC) take 20-30%, and OEMs take the remainder. ARM's "architecture toll" model is a structure that maximizes volume leverage. NARRATIVE

S-Curve Position. AI agent computing is currently in the early acceleration phase (adoption rate 5-15%). The distribution of agent tools such as Microsoft Copilot+, ChatGPT Desktop, and Claude Code has begun, but the transition of corporate infrastructure is still in its early stages. This section provides the greatest returns to leading investors, but also carries significant timing risk. INFERENCE