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

Arm launched its first in-house AI CPU and said agentic AI will shift compute toward continuous inference; Arm stock jumped 16% on the announcement. Intel and AMD told investors they will raise CPU prices by up to 15% starting this month, and both stocks rose about 7%; data centers are forecast to grow CPU capacity per GW by more than four times, implying tighter supply and higher pricing. Nvidia is also entering the CPU market and Arm expects its AGI CPU in H2 with Meta as a flagship partner, indicating broad sector upside for CPU makers.

Analysis

The market is re-pricing compute along an orthogonal axis to GPUs: latency-aware, always-on inference that favors power, density and software-stack integration over raw floating‑point throughput. That shift amplifies value to firms that can monetize architecture-level advantages (power per throughput, software hooks, and royalty capture) and to supply-chain nodes exposed to mature-node wafer demand and packaging — a different set of winners than the leading-edge GPU ecosystem.

Second-order winners include OS/cloud orchestration vendors that sell stack optimizations (quantization, batching, model sharding) and colocation providers that can re‑architect racks for higher CPU thermal density; conversely, hyperscalers may face margin pressure and push back via bulk procurement contracts or in‑house silicon to blunt vendor price power. A credible, durable cycle requires multi-quarter capacity tightness and sticky ASPs; if either collapses (fast fab ramp, aggressive hyperscaler hedging) the current re-rating could unwind quickly.

Key risks: rapid adoption of task‑specific accelerators or algorithmic efficiency (sparser models, aggressive quantization) that reduce CPU token cost per inference, and policy/export controls that bifurcate supply chains. Near term (days–months) catalysts are pricing announcements and Q2 guidance from major CPU vendors; medium term (6–24 months) is hinge‑on product ramp execution and foundry capacity allocation. The consensus captures the headline CPU demand rotation but underweights the competitive response from hyperscalers and the margin‑compression risk if multiple suppliers flood the market simultaneously.

Key Decisions for Investors

• Pair trade (3–9 months): Long ARM Jan-2027 50–80% notional in-call LEAP (buy call spread) to capture execution optionality and royalty upside, funded by short NVDA Jan-2027 5–10% notional covered calls — R/R ~2:1 if ARM proves platform traction; stop-loss: cut ARM exposure on failed partner wins or missed roadmap by next two quarters.
• Diversified CPU exposure (6–12 months): Buy AMD and INTC near-term (equal weight) to play ASP reset across vendors — use a 60/40 AMD/INTC mix, hedge macro with a small short position in a hyperscaler ETF (IYW or direct MSFT/GOOGL) to protect versus hyperscaler pushback; target 20–30% upside vs 10–15% downside in stress.
• Convex hedge (months): Buy out‑of‑the‑money put spreads on Micron (MU) and memory ETF sized ~5% portfolio to protect vs rapid memory normalization that would damp GPU-to-CPU substitution economics; cut if memory spreads widen >25% QoQ.
• Event-driven short (days–weeks): If public filings show large multi-year supply agreements between hyperscalers and fabs (TSMC/Intel Foundry), short ARM/AMD 2–4 week straddles to collect premium — thesis: long‑dated supply commitments compress near-term volatility.