Michael Burry of "The Big Short" Says an AI Slowdown Would Be "Self-Serving" for OpenAI and Anthropic

Michael Burry argues that frontier AI labs' calls for an industry slowdown may be self-serving, masking high compute costs, intensifying open-source competition, slowing growth and IPO-readiness concerns. OpenAI reportedly lost tens of billions of dollars in 2025 and may delay an IPO until at least next year, while Big Tech's planned AI infrastructure spending exceeds $750 billion this year. Anthropic, despite advocating AI safety constraints, is reportedly advancing an IPO that could raise about $100 billion and said it expects a second consecutive quarter of positive adjusted operating income.

Analysis

The investable signal is not a near-term demand collapse but a widening gap between compute infrastructure monetization and frontier-model economics. If model vendors defer training cycles or stretch capacity commitments, the first pressure point is utilization and pricing for leased GPU capacity—not necessarily NVDA unit demand, where hyperscaler capex and sovereign AI orders provide a longer backlog. The more exposed public proxies are AI hosting/power-constrained infrastructure names whose valuations assume sustained scarcity pricing; a shift toward smaller, more efficient or open-weight models would compress their revenue-per-MW assumptions over the next 6-18 months.

A credible IPO window for a frontier lab would likely redirect the narrative from existential scarcity to unit economics: inference gross margin, customer concentration, deferred compute liabilities, and the cash cost of retaining model leadership. That disclosure could reset private-market marks and create sympathy pressure on public AI software multiples, particularly firms priced on broad “AI attach” claims without measurable net-revenue retention or incremental margin. Conversely, an IPO that demonstrates durable enterprise inference demand and positive contribution margins would validate the compute buildout and extend NVDA’s earnings-duration premium.

Consensus appears too binary: either AI safety rhetoric means capex slows, or it is pure promotion. The actionable distinction is whether contracted compute is cancellable, take-or-pay, or merely an aspirational commitment. A slowing training cadence can coexist with rising inference demand, leaving NVDA relatively insulated while weakening marginal data-center economics. The thesis is falsified if disclosed frontier-lab commitments prove firmly take-or-pay and hyperscaler capex guidance remains above current expectations through the next two reporting cycles.

Key Decisions for Investors

• Maintain NVDA as the preferred liquid AI exposure rather than adding speculative infrastructure beta; reassess over the next 1-3 months after hyperscaler capex updates. Reduce if aggregate 2027 capex guidance falls more than 10% or management signals a material shift from accelerated compute to internally optimized ASICs.
• Avoid treating any prospective frontier-lab IPO as a standalone bullish AI catalyst. On filing, focus on revenue growth, inference gross margin, customer concentration, and minimum compute-payment obligations; absent those disclosures, no recommended position.
• Watch for a pair-trade setup: long NVDA versus short an AI infrastructure/hosting basket only after evidence of capacity repricing or reduced contracted MW additions. Target a 10-15% relative move over 3-6 months; stop if utilization remains tight and pricing rises despite a slower training cadence.
• For broader tech exposure, favor companies with independently measurable AI monetization over firms relying on narrative multiple expansion. A deterioration in enterprise AI software net-revenue retention or an increase in GPU-cloud discounting would be the trigger to de-risk high-multiple AI software.