The AI Stocks to Sell Immediately… and the Ones to Buy Instead

The author warns that AI data-center builders face significant risk as their businesses are capital-intensive and commoditized, citing Oracle’s $300 billion cloud agreement with OpenAI (4.5 GW between 2027–2032) and estimating roughly $50 billion per gigawatt (implying ~$225 billion in build costs versus $300 billion revenue, with optimistic profit math producing ~$145 billion over eight years). If OpenAI or equivalent customers fail to honor commitments, pricing pressure and vendor financing could erode margins—putting companies like Oracle, Amazon, Equinix and Digital Realty at downside risk—while recommending investors favor “AI Appliers” such as PayPal, which leverages AI for fraud detection (blocking ~$500 million/quarter) and stands to benefit from cheaper compute without funding the infrastructure buildout.

Analysis

Market Structure: Capital-intensive AI data centers (ORCL, EQIX, DLR, AMZN infra) face commodity-like pricing pressure because compute is fungible and customers will shop on price; this compresses gross margins and shifts win probability to software/service “appliers” (PYPL, select healthcare and logistics SaaS) that monetize model outputs. Expect a Darwinian shakeout over 12–36 months: providers with diversified, long-duration revenue (colocation contracts, enterprise lock-ins) will fare better than speculative build-outs tied to one anchor customer.

Risk Assessment: Tail risks include anchor-client default (e.g., OpenAI renegotiating or walking), a hard drop in utilization (30–50% vacancy), or regulatory limits on LLM compute that reduce demand; each could push infrastructure equity down 30–60% and widen credit spreads in 6–18 months. Hidden dependencies include vendor financing, power/real-estate constraints, and Nvidia (NVDA) supply bottlenecks — loss of chip supply or price spikes would flip the calculus and re-empower builders. Key catalysts: OpenAI contract performance updates, quarterly utilization metrics from DLR/EQIX, and NVDA inventory/supply guidance.

Trade Implications: Tactical bias is underweight infrastructure REITs and ORCL, overweight AI Appliers (PYPL) and selective enablers (NVDA on pullbacks). Use relative-value pair trades (long PYPL vs short ORCL/EQIX) and defined-risk options to express convexity; expect 6–12 month windows for material re-rating. Rebalance corporate credit exposure: hedge ORCL/EQIX paper if spreads widen >75bp from current levels.

Contrarian Angles: Consensus underestimates stickiness of high-quality colocation and enterprise cloud contracts — not all infrastructure supply will go unsold; some names (EQIX with diversified tenants) may be oversold by 20–40% on panic. Conversely, the market may be underpricing the chance of a major AI customer default or collective price erosion; historical parallel to Lucent warns that capex-heavy winners can become losers quickly. An unintended consequence: cheaper compute could accelerate software-driven margin expansion for dominant appliers, concentrating profits in a few winners.

Key Decisions for Investors

• Establish a 2–3% long position in PYPL (equities) with a 12-month target of +30–50% and a hard stop-loss at -15%; alternatively buy PYPL 9–12 month call LEAPS (buy the 25–30% OTM call) to capture asymmetric upside if AI-driven margins accelerate.
• Initiate a 1–2% short position in ORCL (equities) paired with a 6–9 month 150/120 put spread (adjust strikes to current levels) to limit capital at risk; target ORCL downside to low $100s within 6–18 months if utilization or OpenAI revenue flags.
• Cut direct exposure to data-center REITs (DLR, EQIX) by 30–50% over the next 30 days and redeploy into AI Appliers and NVDA on pullbacks; if EQIX/DLR CDS or bond spreads widen >100bp, add protection or increase shorts.
• Implement a relative-value pair trade: long PYPL (2%) vs short EQIX (1%) for 6–12 months, and monitor quarterly utilization and customer concentration reports—exit if EQIX reports <5% QoQ utilization decline or PYPL reports material AI revenue re-acceleration (>10% incremental margin).