The History That Suggests an AI Bubble

Nvidia has reached a $5 trillion valuation even as leading AI developer OpenAI remains unprofitable and is burning through billions while venture capital continues to fund OpenAI, Anthropic and other startups. The authors warn that large language models have outpaced existing benchmark-based evaluation methods, creating a 'crisis of evaluation' that makes it unclear whether current investments — many predicated on substantial automation within three to five years — are building durable infrastructure or inflating a bubble that could precipitate another AI downturn.

Analysis

Market structure: The immediate winners are dominant GPU suppliers (NVDA) and data‑center ecosystem firms (chips, HBM memory, fabs, cloud capex) because vertical integration and design wins create durable pricing power; losers are early‑stage AI startups and high‑multiple AI labelers that lack unique data or compute advantages. Expect NVDA to sustain >30–40% gross margin pricing power near term as lead times remain 6–12 months and backlog persists, concentrating share in top‑2 GPU vendors and select foundries. Cross‑asset: a re‑rating or shock would push equities into risk‑off (bonds rally, USD bid) while elevated NVDA options IV (~>80%) keeps skew and commodity sensitivity (copper/energy) elevated.

Risk assessment: Tail risks include a benchmark/evaluation collapse that triggers a 30–60% de‑risking of non‑profitable AI names, an export/regulatory action against advanced chips to China within 3–9 months, or a major NVDA supply glitch. Short term (days–weeks) catalysts are earnings, model releases, and fundraising rounds; medium (3–12 months) are capex cycles and memory supply; long term (1–3 years) is actual enterprise productization and ROI. Hidden dependencies: cloud provider capex, HBM supply, and power constraints could throttle adoption unexpectedly.

Trade implications: Direct play: size NVDA exposure 2–3% of portfolio with a 6–12 month horizon, pairing each long with a 6‑month 10% OTM protective put (cost ≤3–4% premium historically). Pair trade: long NVDA (or AMAT/MU for supply exposure) vs short ARKK or small‑cap AI names (1–2% notional) to capture dispersion if the bubble pricks within 3–9 months. Options: sell 30–60 day covered calls on existing NVDA longs to harvest elevated IV; buy 1–3 month NVDA put spreads if IV>90% as cheaper tail hedge.

Contrarian angles: Consensus fears may underprice structural compute growth — if enterprise adoption converts 10–20% of workflows over 3–5 years, NVDA earnings could justify current multiples; conversely, many public AI names are priced for 3–5 year perfect execution and are vulnerable to 50%+ drawdowns. History (1980s expert systems) warns that hype cycles wipe out marginal players but consolidate incumbents — favor durable moats and avoid funding‑dependent startups. Unintended outcome: a drawdown could create high‑conviction M&A/reimplementation opportunities in 12–24 months.

Key Decisions for Investors

• Establish a 2–3% long position in NVDA (ticker NVDA) for a 6–12 month holding period; simultaneously buy a 6‑month 10% OTM put for each lot to cap downside (acceptable premium target ≤4% of notional).
• Implement a relative‑value pair: long NVDA (or AMAT MU exposure split 60/40) and short ARK Innovation (ARKK) equal‑dollar for 1–2% of portfolio notional to profit from concentration vs speculative AI beta over 3–9 months.
• If NVDA implied vol >90% pre‑earnings, sell 30–60 day covered calls against new or existing NVDA longs to collect premium; if IV spikes and NVDA falls >25% from peak, buy NVDA 3‑month put spreads (e.g., 20%/30% OTM) as a cost‑efficient hedge.
• Reduce private/venture AI allocation by 30–50% over next 90 days; redeploy proceeds to semiconductors (AMAT, MU) and data‑center REITs (EQIX) totaling 3–5% incremental portfolio exposure to capture secular capex while avoiding early‑stage execution risk.
• Set hard stop/triggers: cut public high‑valuation AI longs if their price/sales >20 and revenue miss >10% on next report, or if a major regulatory export control is announced within 30–90 days; consider adding to NVDA/AMAT on any >20% market‑wide drawdown.