The AI bubble can wait, but your exit plan can't

Forrester says an AI infrastructure bubble is not about to burst, noting that roughly 90% of build-out funding comes from large, financially healthy technology companies, but warns a correction could disrupt suppliers, contracts and prices. It estimates investments break even only if AI adds 5% to US GDP by 2035; projected productivity gains may contribute about two percentage points, leaving roughly $1.3 trillion in value to be found elsewhere. Bain estimates the AI industry would need $6 trillion in annual revenue by 2031 to sustain infrastructure spending, while Forrester advises enterprises to measure returns and prepare for supplier stress.

Analysis

The investable risk is a dispersion trade, not a blanket AI short. Cash-funded hyperscalers can absorb weaker returns longer than leveraged or externally financed compute builders, so a demand miss is more likely to first pressure financing-sensitive capacity providers and contract economics than trigger an immediate industry-wide stop. CoreWeave is the clearest vulnerability in the supplied set; Oracle is more nuanced: financing and customer-prepayment exposure warrant diligence, but its broader business and cloud demand make it a less direct proxy for a standalone compute-builder short. Anthropic’s price cuts illustrate the second-order squeeze: cheaper inference can expand usage while lowering revenue per token and weakening returns on newly built capacity. That benefits enterprise buyers and may accelerate adoption, but makes supplier utilization, recognized revenue versus contracted commitments, and customer concentration more important than headline demand.

Near term (days to weeks), sentiment and funding terms can drive volatility; avoid extrapolating an industry correction from one vendor’s stress. Over 1–3 months, watch utilization, backlog conversion, financing costs, prepayments, and hyperscaler capex commentary. Over 6–18 months, the key test is whether paid workloads broaden beyond productivity enhancements enough to support capacity and pricing. The contrarian risk to bearish positioning is that falling inference prices stimulate workloads and cash-rich platforms sustain investment, delaying any capacity reset. The thesis weakens if providers demonstrate durable utilization and recognized revenue catching up with commitments without worsening funding dependence; it strengthens with capex cuts, refinancing pressure, or service degradation.

Key Decisions for Investors

• Prefer a relative-value expression: underweight CoreWeave versus cash-generative hyperscalers rather than shorting the entire AI infrastructure complex. Size for high volatility; reassess if utilization and recognized revenue consistently validate contracted demand.
• Treat Oracle as a diligence/watch position, not an automatic short. Verify the scale, duration, and recourse of external financing, customer prepayments, and capacity commitments; a worsening funding mix alongside weaker cloud growth would strengthen the downside case.
• For Anthropic, lower token prices are a potential volume catalyst but a monetization risk. Track paid usage and revenue per unit alongside price reductions; avoid treating cheaper tokens alone as evidence of either durable demand or collapsing economics.
• Over the next 1–3 months, monitor vendor disclosures and hyperscaler capex guidance for backlog-to-revenue conversion, utilization, funding costs, and cancellations. A material capex pullback or deterioration in service commitments is a downside catalyst; improving conversion without rising leverage is a thesis-falsifying signal.
• For enterprise-exposed portfolios, favor providers and contracts with portability and multiple sourcing options; avoid assuming that a signed commitment guarantees durable service or attractive pricing. No broad sector short is warranted absent evidence that demand and utilization are actually falling.