1 Agentic AI Chip Stock to Buy and 1 to Sell

The article favors AMD over Arm as agentic AI is expected to increase server-CPU demand, with the GPU-to-CPU ratio in AI data centers projected to fall from 8:1 for LLM training to 1:1 for AI-agent workloads. AMD estimates the data-center CPU market could reach $220 billion and believes it can capture 50% share, while its inference-GPU opportunity is supported by reported large deals with OpenAI, Meta and Anthropic. Arm is viewed negatively despite forecasting 15% share of a potential $100 billion server-CPU market by 2031, due to its nearly 140x forward P/E, anticipated smartphone-volume declines, and potential gross-margin pressure from shifting into chipmaking.

Analysis

The investable issue is not whether agent workloads raise CPU intensity, but whether that demand accrues to merchant silicon or hyperscaler-designed Arm-based CPUs. AMD gains most if enterprises standardize on general-purpose x86 clusters and inference deployments require frequent refreshes; its CPU attach can also improve GPU platform economics versus NVDA. The risk is that the largest buyers—AMZN, GOOG and META—continue shifting incremental capacity to internally designed Arm silicon, limiting the addressable merchant CPU pool even as total compute spending rises.

AMD's reported customer commitments should not be capitalized without verified purchase obligations, delivery schedules and pricing: AI infrastructure announcements commonly represent multi-year capacity intentions rather than near-term revenue. Over the next 1-3 months, the relevant catalyst is AMD's data-center revenue guide, MI-series supply commentary and EPYC share data versus INTC, not generalized agentic-AI adoption metrics. A weaker-than-expected cloud capex read or evidence that inference is becoming materially more GPU/NPU-efficient would compress the AI-premium multiple before it changes AMD's earnings base.

ARM is more nuanced than a simple valuation short. Moving into chips could create channel conflict and lower consolidated gross margin, but higher-value compute subsystems and royalty participation in custom server CPUs can raise revenue per deployed core without ARM carrying full manufacturing economics. The near-term asymmetry remains unfavorable if handset royalties soften while data-center design wins take years to monetize; however, a disclosed major hyperscaler chip customer, stronger royalty rate, or sustained licensing growth would make a high-short-interest valuation short vulnerable.

Contrarian view: agentic AI may initially be CPU-intensive at the orchestration layer, but optimization pushes work into accelerators, networking and purpose-built inference ASICs over 6-18 months. The cleaner relative expression is therefore AMD share gains versus ARM's expectation premium, rather than an outright bet on a mechanically higher CPU-to-GPU ratio.

Key Decisions for Investors

• Initiate a 3-6 month relative-value position: long AMD / short ARM, dollar-neutral, only after confirming AMD's next data-center guide implies continued sequential growth. Target 15-20% relative outperformance; stop if ARM reports licensing/royalty growth that materially outpaces handset exposure or announces a named hyperscaler server-chip program.
• Use AMD upside selectively via 6-month call spreads rather than outright shares ahead of earnings, sized to premium at risk. The thesis requires EPYC share gains and credible MI-series conversion; exit on a data-center revenue-guide miss or a material reduction in AI accelerator supply commentary.
• Maintain INTC as the key falsification monitor rather than a default short: evidence of sustained Xeon stabilization, improved gross-margin trajectory, or major cloud design wins would weaken the AMD CPU-share leg. Conversely, further share losses support adding to AMD after earnings confirmation.
• Do not establish a standalone ARM short solely on headline valuation. Set an alert around the next earnings release for smartphone royalty trends, gross-margin mix and disclosed server-chip economics; absent those data, the risk of a narrative-driven squeeze outweighs the near-term catalyst quality.