Late-Cycle Investment Theory

Late-cycle investment theory is the body of practitioner and academic thinking about how investors should allocate, size positions, and manage risk when markets are demonstrably late in an economic expansion, a credit cycle, or an asset-class bubble. The term draws on work by Howard Marks, Seth Klarman, Stanley Druckenmiller, and the post-2008 macro-hedge-fund literature, together with the Austrian-school tradition on malinvestment and credit cycles. The core insight is that late-cycle environments are characterized by stretched valuations, widespread risk-taking, narrow spreads, abundant liquidity, and widespread consensus that 'this time is different' — conditions in which conventional diversification and passive exposure fail to protect against the eventual reversal. AI has become the central late-cycle story of the 2020s.

Mechanically, late-cycle theory identifies a recurring set of features and prescribes behaviors for each. Features include: speculative excess concentrated in a narrative sector (today: AI infrastructure, foundation models, and adjacent picks-and-shovels); credit expansion to lower-quality borrowers; tightening labor markets; rising inflation pressure that eventually forces central bank action; and increasingly asymmetric return distributions where upside is muted but downside is large. Prescribed behaviors include: reducing net exposure, increasing the margin of safety on individual positions, hedging tail risk, rotating toward cash-generative assets with less narrative dependence, and shortening duration in credit. The framework is explicit that these are about asymmetric payoff protection, not market timing — no one reliably knows when the cycle ends.

The advantage of late-cycle theory is that it converts a vague sense of 'we are late' into a specific operational checklist. It also forces the analyst to distinguish between late-cycle and post-peak, which have different optimal behaviors (defensive vs. opportunistic). The cost is that late-cycle theory is repeatedly early — cycles can run long, the marginal upside from being positioned for reversal can be small for years before the reversal arrives, and the framework is easy to weaponize into perpetual bearishness. There is also a survivorship bias: the practitioners most associated with late-cycle theory got famous because their calls eventually worked, not because they were correct in real time. The framework is also under-specified for AI specifically: the AI investment cycle has unusual features (capex intensity concentrated in a handful of firms, hyperscaler balance sheets absorbing the spending, revenue growth that may or may not justify the spend) that don't fit cleanly into either traditional cyclical or bubble frameworks.

Open questions include how to operationalize late-cycle signals without relying on market timing, how to identify whether the current cycle is the late stage of the AI investment boom or a sustained new plateau, and whether the conventional late-cycle playbook (cash, defensives, hedges) is the right one when central banks are explicitly backstopping asset prices. The deeper question — whether AI capex is the productive infrastructure buildout that the late-1990s telecom buildout turned out to be (eventually justified by the resulting network value) or the speculative excess of a dot-com-style bubble — is the dominant macro question of the decade. Late-cycle theory provides the conceptual vocabulary for the debate, but the answer is empirical and will not be known for several years.