AI Capex: It's Not All Or Nothing

Jeff Blazek of Neuberger Berman counsels that recent AI-driven volatility mixes signal with noise and investors should instead position for a multiyear AI buildout that will accelerate creative destruction at the company and industry level, increasing the need for selective portfolio construction. He argues a durable capital cycle is already reshaping cash flows, balance sheets and competitive dynamics; Neuberger Berman reports $538 billion in assets under management as of June 30, 2025.

Analysis

Market structure: AI is creating a multiyear capex cycle concentrated in cloud providers (MSFT, GOOGL, AMZN), GPU leaders (NVDA, AMD), and data‑center REITs (EQIX, DLR) while compressing margins for labor‑intensive IT/services and legacy software. Expect pricing power for proprietary model owners and chip architects; commodity suppliers (power, copper) see higher secular demand—data‑center power draw could rise 20–30% in the next 3–5 years in hotspots. Cross‑asset: rising tech capex and dispersion should lift IG credit spreads modestly (+10–30bp) for high‑capex names, raise option implied vols episodically, and put modest upward pressure on energy/industrial commodities and a structurally stronger USD if US retains tech leadership.

Risk assessment: Tail risks include export controls or China decoupling (large revenue hit if >10% of sales blocked), heavy regulation of foundation models (data/privacy fines 1–5% revenue), and a GPU supply shock reversal that collapses margins. Immediate (days) = headline‑driven vol; short‑term (weeks‑months) = re‑ratings around earnings/guidance and capex announcements; long‑term (years) = creative destruction shifting cash flows and market share. Hidden dependencies: talent scarcity, energy grids, and proprietary data access; second‑order risk is vertical fragmentation (different stacks per region).

Trade implications: Favor infrastructure and concentrated chip/IP exposure with risk controls: buy NVDA and MSFT exposure for 12–24 months; overweight EQIX/DLR for 2–4 years; underweight/short select labor‑heavy IT services (ACN, CTSH) where automation reduces billable hours by 5–15% over 2 years. Use option spreads to cap cost: calendar/LEAP call buys for convexity and sell short‑dated calls after major product launches. Entry: scale over 3–6 months, add on pullbacks of 10–20%; set target realizations of +30–50% and trailing stops of 15–25%.

Contrarian angles: Consensus hyper‑focus on NVDA upside may underprice normalization of GPU ASPs and competition—if supply grows 30–50% over 12 months, NVDA could reprice lower even as demand rises. Historical parallel: 1999–2003 internet capex saw concentrated winners and many vanishing vendors; here, expect similar consolidation—pick owners of data and distribution (MSFT, GOOGL) rather than pure hardware suppliers. Watch unintended consequences: accelerated automation could shrink aggregate serviceable addressable market for outsourcers and raise regulatory backlash against monopolistic data practices.

Key Decisions for Investors

• Establish a 2–3% portfolio long in NVDA via a 12–18 month call‑spread (buy Jan 2026 1,200C, sell Jan 2026 1,800C) to capture secular AI GPU demand while capping cost; add on any pullback >15% within 6 months.
• Add 2–3% long exposure to cloud/platform leaders: 1% MSFT (stock) and 1% GOOGL (stock) or replace with 18–24 month LEAP calls (delta ~0.6) to play model hosting & monetization over a 12–36 month horizon.
• Initiate 1–2% long in data‑center REITs (EQIX or DLR) funded by reducing 1–2% exposure to Accenture (ACN) or Cognizant (CTSH); thesis: 24–48 month demand for colocation grows faster than billable hours for legacy services.
• Open a 1% notional hedge: short ACN (or buy put spread ACN 6–12 month) sizing to cap portfolio risk from services compression—target 15–25% downside over 12 months if automation reduces fees.
• Monitor regulatory/export catalysts: avoid adding China‑exposed semis/cloud names >5% revenue from China until next 90 days; if export controls tighten (announcement → 10% revenue at risk), reduce related positions by 30% within 5 trading days.