Michael Burry's next 'Big Short': An inside look at his analysis showing AI is a bubble

Michael Burry and former Scion associate Phil Clifton argue that the current generative AI infrastructure buildout is economically overextended: hyperscalers have roughly quadrupled capex to about $400 billion annually with Man Group estimating $3 trillion over five years, while OpenAI revenues are expected to top ~$20 billion — a gap Clifton says may not justify the investment. Scion draws a parallel to the early-2000s telecom overbuild (U.S. capacity utilization ~5%, wholesale pricing collapsed ~70%) and flags risks including longer book depreciation vs. rapid Nvidia product cycles and early project cancellations (Microsoft trimming ~2 GW of planned data-center projects), positioning Burry and Clifton on the cautious/bearish side of AI investments.

Analysis

Market structure: Short-term winners are GPU vendors (NVDA) and power/grid suppliers during the build phase; longer-term losers are hyperscalers and legacy datacenter capex-heavy contractors if utilization stays below ~30–40%. The telecom-2000 parallel implies a credible risk of a rapid price collapse in wholesale compute capacity that would compress gross margins for MSFT and BABA cloud segments and force write-downs across hardware vendors.

Risk assessment: Tail risks include a rapid markdown cycle where GPU resale prices drop >40% within 12 months, accelerated obsolescence forcing immediate impairment charges, or regulatory restrictions on AI compute exports; any of these would widen IG/ HY spreads for tech issuers in 3–12 months. Hidden dependencies: depreciation policies (6-year server life vs ~12-month effective GPU cycle), secondary markets for used GPUs, and grid constraints that can create localized oversupply/undersupply dynamics. Key catalysts are NVDA quarterly guide, hyperscaler capex cuts (watch MSFT/AMZN/GOOGL Qs), and Alibaba management statements in the next 30–90 days.

Trade implications: Tactical short NVDA bias via 3–6 month put spreads sized 2–3% of portfolio (target -25% move) and a pair trade long MSFT (2–3%) vs short NVDA (2–3%) to express asymmetric risk—MSFT is more diversified and less binary. Consider short copper exposure (ETFs/futures) 1–2% if multiple hyperscalers announce >15% capex cuts within 6 months; buy VIX 1–3 month call calendar as crash insurance for tech correlation spikes.

Contrarian angles: Consensus underweights software-led monetization and Nvidia CUDA lock-in, which could sustain GPU economics longer than hardware depreciation models imply; if used-GPU markets tighten and NVDA guidance holds, shorts will get squeezed. Monitor three thresholds as decision triggers: NVDA revenue guide miss >5%, MSFT capex cut disclosure >15%, or secondary GPU price collapse >30% — any one should materially shift positioning within 30–90 days.

Key Decisions for Investors

• Establish a tactical 2.5% portfolio short on NVDA via a 3–6 month 25–35% OTM put spread to limit capital at risk; exit if NVDA guidance beats and forward revenue growth implied in guide >15% for next two quarters.
• Implement a 2.5% pair trade: long MSFT (2.5%) funded by short NVDA (2.5%) equity exposure to play diversification/visibility advantages; trim the long if MSFT gross margin falls >150bps or cloud growth slows below 20% YoY on two consecutive quarters.
• Deploy a 1–2% short position in copper futures/ETFs if two or more hyperscalers announce cumulative capex reductions ≥15% within 90 days, targeting a 10–20% downside in industrial metals demand over 6–12 months.
• Buy a 1–2% portfolio hedge: VIX 1–3 month call calendar (staggered expiries) sizing to cover 40–60% of tech book delta; liquidate if realized correlation between NVDA and S&P tech basket drops below 0.6 for 30 days.
• Monitor leading indicators daily for 30–90 days: NVDA guidance/ASP language, MSFT/BABA capex line items, and used datacenter GPU auction prices (if used-GPU prices drop >30% in 60 days, increase short NVDA to 4–6%).