2 AI Stocks That Survived the March Sell-Off -- and Look Stronger Because of It

$100 billion by 2031: Arm forecasts the data-center CPU market will reach $100B by 2031 and aims to capture $15B in CPU revenue (and $25B total) from its new Arm-designed chips. AMD generated $16.6B in data-center revenue last year, is rolling out a Venice chiplet architecture to increase core counts for agentic AI workloads, and has two GPU partnerships each projected to be worth over $100B, positioning both firms to capture material share as AI agents drive increased CPU demand.

Analysis

The CPU-for-agents narrative creates a bifurcated market: one camp pays a premium for tight integration and power-efficiency (favours IP-light, software-validated incumbents) while another prizes open-architecture scale (favours chiplet and foundry-flexible designs). Expect the real competitive battleground to be latency-sensitive orchestration stacks and tool invocation throughput, not peak FLOPS — meaning wins will hinge on compiler/runtime superiority, hypervisor hooks, and telemetry rather than raw core counts alone.

Supply-side constraints will be the overlooked governor on revenue ramp: TSMC/ASML lead times, advanced packaging slots, and memory subsystem allocations (HBM/DDR) create multi-quarter lags between design wins and material revenue; a vendor can show a “design win” but still be throughput-capped for 6–18 months. Power/TCO proofs will be the dominant commercial lever — a claimed architectural advantage without 12–24 month field TCO data will not move large hyperscaler purchase schedules.

Catalysts to watch with clear timing: public hyperscaler OR OEM platform validation (quarterly releases), wafer allocation notices and foundry guideposts (TSMC capacity windows announced semi-annually), and first comprehensive third-party benchmarks that measure agentic workloads end-to-end (likely 3–9 months). Tail risks are typical: software lock-in, Nvidia or a major hyperscaler vertically integrating a control-plane that obviates third-party CPUs, or an unexpected foundry shock that compresses available advanced packaging capacity for a year.

The market currently prices an outcomes binary; the prudent stance is to scale exposure into objective validations rather than narratives. Reweight decisions should be made off verifiable signals: (1) multi-customer production shipments, (2) independent agentic workload TCO studies, and (3) sustained fab capacity commitments covering the next 12–24 months.

Key Decisions for Investors

• Long AMD (AMD) — 2–3% nominal NAV. Entry: scale in over 2–6 weeks on any pullback >8%. Horizon: 12 months. R/R: target +40% if AMD gains share and monetizes chiplet roadmap; downside -25% if pricing or memory inflation compresses server ASPs. Key monitor: server CPU mix and gross margin expansion each quarter.
• Pair trade: Long AMD / Short INTC (1:1) — 1–2% NAV. Horizon: 6–12 months. R/R: expect ~+25–35% relative outperformance if AMD continues share gains and Intel faces margin pressure; limited downside if Intel stabilizes via pricing or enterprise bundling. Exit triggers: Intel announces meaningful platform refresh with competitive perf/Watt or near-term foundry re-contracts.
• Directional ARM (ARM) option collar — buy ARM equity sized 1–2% NAV plus buy a 6–12 month 20% OTM put (or buy an 18-month 30% OTM call spread if you prefer limited capital outlay). Entry: initiate on a 10–15% pullback or immediately if your mandate tolerates higher narrative risk. R/R: asymmetric upside if multi-hyperscaler validations arrive within 12–24 months; downside capped via put/call-spread to limit narrative-driven drawdowns.
• Tail hedge using NVDA (NVDA) — buy 3-month 7–10% OTM puts sized 0.3–0.5% NAV to protect against a rapid GPU-led rotation/market shock. Cost is expected low relative to implied protection; this protects against broad AI-led re-rating reversal that would compress high-multiple peers simultaneously.