Move Over, Nvidia GPUs. The AI CPU Era Is Here

Arm launched its first in-house AI CPU and said agentic AI will drive continuous inference demand, sending Arm shares up 16%. Intel and AMD told investors they will raise CPU prices by up to 15% starting this month, and each stock rose about 7% on the report; Nikkei warned of upcoming CPU shortages. Data-center CPU capacity per GW is forecast to grow more than fourfold, Nvidia is entering the CPU market, and Arm expects its AGI CPU in H2 with Meta as a flagship partner, implying a broad sector tailwind for CPU makers.

Analysis

The move to continuously running, agentic AI is not just a demand shock for more silicon — it rewires where value accrues across the stack. Firms selling licensed ISA royalties (lower cash conversion) versus those selling physical CPUs (higher gross margin but capex and inventory risk) will see diverging P&L trajectories over 12–24 months; expect reported revenue growth to lag gross-margin expansion for product sellers as they absorb NRE and inventory build.

On the supply side, CPU demand competes directly with GPUs and HBM for advanced-node wafer capacity and substrate/test services, creating a multi-quarter bottleneck in advanced packaging rather than pure logic masks. That bottleneck amplifies pricing power for vendors who control integration pathways (IP + design ecosystem + customer commitments) and favors players that can lean on multi-sourced packaging and software optimization partners to shorten time-to-deploy.

Key downside catalysts are software portability and algorithmic efficiency: meaningful progress in quantization/compilation or a surge in LLM sparsity techniques could cut token compute per inference by 30–50%, collapsing near-term CPU demand. Geopolitical export controls and a rapid fab-capacity response (TSMC/Intel foundry ramps) are second-order risks that can normalize pricing within 12–24 months.

Net: prefer exposures that capture (a) continued memory scarcity and (b) vendors able to monetize both IP and product without bearing sole capex risk. Avoid single-architecture bets that assume immediate, ubiquitous Arm-software parity; the adoption curve will be multi-year and lumpy, creating option-like outcomes for equities.

Key Decisions for Investors

• Long ARM (ticker: ARM) — buy 12–24 month call LEAP or a call spread sizing 1–2% AUM. Rationale: captures licensing → product optionality and data-center design wins while capping premium outlay. Target: 2.5x on realized server wins in 12–24 months; Stop: -40% of premium if no reported design wins/bookings in next two quarters.
• Long Micron (ticker: MU) — 6–12 month call or 3–4% net-long position in stock. Rationale: memory tightness (HBM/GDDR) supports >30–50% upside if HBM allocations persist through next capacity cycle. Risk: cyclicality; tighten stop to -20% if end-market inventory destocking indicators rise.
• Pairs trade — long AMD (ticker: AMD) / short Intel (ticker: INTC) 6–12 months, net delta ~0: overweight AMD by 1.2x size vs short INTC. Rationale: AMD best positioned to capture share in server CPU demand with fewer legacy-capex drag; Intel more exposed to execution and margin pressure despite price hikes. Target: relative outperformance of AMD vs INTC of +25–40%; adverse case: relative underperformance >20% triggers unwind.
• Hedge/Options on NVDA (ticker: NVDA) — buy 9–12 month protective puts sized to 0.5–1% AUM or sell covered calls against existing NVDA exposure. Rationale: market is repricing inference winners; NVDA faces incremental CPU competition and is the clearest single-stock tail risk if agents favor CPU economics. Objective: limit portfolio gamma exposure while keeping upside participation.