CPU demand is driven by four forces over the next five years. First, agentic AI: continuous inference loops favor power-efficient CPUs over GPU clusters — ARM-based chips are structurally positioned to capture this shift. Second, cloud custom silicon: AWS Graviton, Microsoft Cobalt, and Google Axion are displacing x86 CPUs in cloud workloads, expanding total ARM CPU TAM while compressing Intel and AMD cloud pricing power. Third, AI PC refresh: Microsoft Copilot+ PC requirements mandate NPU-enabled processors, pulling forward an enterprise upgrade cycle that was otherwise flat. Fourth, mobile steady state: smartphone CPU demand is stable with modest content growth from on-device AI features. Bear case assumes AI PC refresh disappoints and enterprise PC demand stays soft through 2028, while ARM cloud displacement slows as x86 incumbents respond. Bull case assumes agentic AI drives a step-change in per-device CPU compute requirements — more cores, higher frequencies, and faster upgrade cycles across cloud, PC, and mobile simultaneously.
“Masayoshi Son expects ARM value to increase tenfold as AI shifts toward CPU-centric computing. Agentic AI runs continuous inference loops — favoring power-efficient CPUs over GPU-heavy training rigs. ARM licensing model scales with every chip shipped across cloud, edge, and mobile.”
“AMD EPYC server CPU revenue hit a record quarter with hyperscaler market share gains continuing. AI orchestration workloads drive demand for high core-count CPUs as the coordination layer between GPU accelerators.”
“ARM architecture royalty revenue +40% Y/Y as cloud CPU deployments expand. AWS Graviton4, Microsoft Cobalt, and Google Axion custom ARM CPUs now represent 20%+ of new cloud instance deployments, displacing x86.”
“AI PC CPU attach rate accelerating as Copilot+ requirements mandate NPU-enabled processors. Intel Core Ultra and Qualcomm Snapdragon X Elite driving PC refresh cycle earlier than expected in enterprise.”
“Intel Xeon data center CPU recovery slower than guided. Gaudi AI accelerator ramp not offsetting CPU ASP pressure. Market share losses to AMD EPYC accelerating in hyperscaler deployments; foundry business progress behind schedule.”
“Qualcomm Snapdragon X Elite laptop CPU wins at Dell, HP, Lenovo, and Microsoft Surface — first meaningful ARM CPU penetration in Windows PC market. Battery life and on-device AI performance driving enterprise consideration.”
OpenAI, SoftBank, and Oracle's $500B US AI infrastructure initiative — "Terafab" — is the largest single demand event in semiconductor history. Phase 1 ($100B) is already under construction in Texas. Each AI data center requires dense CPU racks for orchestration, networking, and retrieval-augmented inference coordination: estimate 4–6 high-core-count CPU sockets per GPU node. At full buildout, Stargate alone could add 1–2 years of incremental server CPU demand.
$500B total capex × ~30% servers & networking = $150B hardware
$150B × ~15% CPU content of AI server BOM = $22.5B in CPUs (4-yr total)
Less existing baseline CPU demand in forecast = ~$10–18B net incremental uplift
─── Cross-check (bottom-up) ───────────────────────────────────────
1M GPUs ÷ 8 GPUs/server = 125K compute nodes
125K nodes × 2 sockets × $4K avg EPYC ASP = $1B direct compute CPUs
+ 3× support/orchestration/storage servers = +$3B
× 4-year ramp + networking overhead → $10–18B cumulative
Elon Musk's xAI is expanding Colossus from 200K to 1M+ GPUs across Memphis and new sites. Each Grok training and inference cluster requires substantial CPU infrastructure for data pipelines, retrieval, agent orchestration, and interconnect switching. xAI has signaled it will be among the largest GPU buyers globally through 2027, with CPU demand scaling proportionally.
Expansion: 200K → 1M GPUs = 800K incremental GPUs
800K GPUs ÷ 8 per node = 100K new compute nodes
100K nodes × 2 CPU sockets × $4K EPYC avg = $800M direct compute CPUs
+ 2× support / orchestration / storage servers = +$1.6B
× 18–24 month ramp + overhead → $3–6B cumulative
The Trump administration is using CHIPS Act incentives and tariff pressure on Asian imports to reshore CPU and semiconductor manufacturing. Intel's Ohio and Arizona fabs are under construction; Qualcomm is exploring US production. A domestic CPU supply chain would add a cost premium but insulates against geopolitical disruption. Full reshoring is unlikely — partial reshoring of advanced packaging and final test is more probable.
US domestic CPU production target by 2030 ~20% of global market
$128B CPU TAM (2030 base) × 20% domestic share = $25.6B US-made CPUs
US fab cost premium over Taiwan: 30–50% = $7.7–12.8B cost uplift
Tariff pass-through on remaining imports (25%) = +$3–5B price effect
Probability weight at 58% → $5–12B expected TAM impact
AWS Graviton5, Google Axion, Microsoft Cobalt, and Meta's custom CPU designs are displacing Intel Xeon and AMD EPYC in cloud deployments. ARM-based cloud CPUs offer 40–60% better performance-per-watt at lower ASP, compelling hyperscalers to accelerate in-house silicon. Estimated 30–40% of new cloud CPU sockets will be ARM-based by 2028. This expands total CPU TAM while compressing x86 vendor pricing power.
Hyperscaler server CPU market 2028 estimate ~$30B/year
ARM share shift: 5% → 30–40% (+25–35pp) = +$7.5–10.5B ARM CPU revenue
ARM Holdings royalty uplift (~$0.10/chip×volume) = +$1–2B
x86 ASP compression (partial offset, -10–15%) = -$2–3B
Net TAM expansion vs. ARM-absent baseline → $8–14B
Masayoshi Son's thesis: agentic AI runs continuous inference loops rather than batch GPU jobs, making power-efficient CPUs structurally more important than GPU clusters for most agent tasks. If 10–15% of enterprise software shifts to agent-native architecture by 2029, per-server CPU demand increases 2–3× as agents require persistent memory, fast context switching, and low-latency retrieval — all CPU-bound workloads.
Global servers: ~15M deployed; 10–15% → AI-inference grade
= 1.5–2.25M servers needing high-core-count CPU upgrade
× 2 sockets × $3K incremental ASP over current = $9–13.5B upgrade cycle
+ 500K new AI-inference deployments × $8K CPU = +$4B new builds
+ Enterprise PC: 10% of 500M PCs × $500 NPU prem = +$25B (4-yr)
Probability-weighted at 65% → $15–30B net TAM uplift