In agentic AI workloads, CPU tool execution accounts for 50-90% of total latency. The bottleneck shifts from GPU excess → CPU shortage, and the CPU:GPU ratio regresses from 1:8 to 1:1. The x86 vs ARM architecture war has entered a new phase.
The bottleneck in the AI industry is structurally shifting from GPU to CPU. FACT According to a joint study by Georgia Tech + Intel (arXiv:2511.00739), CPU-based tool execution (tool-calling, web scraping, code execution) accounts for 50-90% of end-to-end latency in agentic AI workloads. GPUs are only involved in token generation, but the actual value creation work is performed on the CPU.
INFERENCE The CPU:GPU ratio is regressing from the training era (1:7~8) to the agentic era (1:1), and ARM estimates that CPU core demand will increase by 4x. This fundamentally changes the value distribution structure of the semiconductor value chain. The GPU bias in data center CAPEX (2024-2025) will shift to balanced CPU/memory investment (2026-2027).
NARRATIVE The x86 (Intel) vs ARM (Qualcomm/ARM Holdings) architecture war has reignited in a new battlefield called agentic AI. Intel has vertical integration strengths with Xeon 6-month lead time supply constraints + 18A self-fab strategy, while the ARM camp claims power efficiency advantages with 136-core AGI CPU + 85 TOPS NPU.
Conclusion: Bottleneck owner = Intel (data center) + Qualcomm (edge). Intel's DCAI sales of $5.1B (+22% YoY) and Xeon supply constraints prove pricing power. Qualcomm is preempting the edge AI agent market with a 3x increase in OEM design wins (20→58). ARM is an indirect beneficiary of TAM expansion due to the nature of its licensing model, but the valuation (P/E 110-121x) is burdensome.
The AI agent CPU compute TAM is estimated to grow from $52B in 2026 to $112B in 2028. INFERENCE This estimate is the sum of (1) the existing server CPU market ($38B, 2025) + (2) additional demand for agentic workloads (incremental increase due to changes in the CPU:GPU ratio) + (3) the edge AI PC/device AI processor market.
FACT Data Center Segment: Intel DCAI sales are $5.1B per quarter (annual rate of $20.4B), occupying ~55% of the server CPU market. If the incremental increase in CPU demand for agentic AI (ARM estimate of 4x) is realized, the server CPU TAM can expand to $72B by 2028. This includes existing cloud server replacement demand + new agent-specific infrastructure investment.
FACT Edge Segment: The AI PC market is currently dominated by x86 with 71% and ARM with 24%. Qualcomm's OEM design wins increasing 3x from 20 to 58 in 6 months suggests that demand for on-device AI agents is growing explosively. Intel Panther Lake (180 TOPS platform total) and Qualcomm X2 Elite Extreme (85 TOPS NPU) will aggressively target this market in the second half of 2026.
INFERENCE Sensitivity of Key Assumptions: Even if the CPU:GPU ratio changes to only 1:4 (conservative), the CPU TAM expands by +50% compared to the existing one. If the 1:1 scenario (ARM estimate) is realized, it expands by +115%. It is estimated that the CPU share relative to the current data center GPU CAPEX will increase from 15% to 35-40%.