In a September 2026 whitepaper1, Meketa Investment Group researchers Alison Adams and Frank Benham argue that the AI investment cycle is far more analytically complex than the single bust narrative dominating current commentary. Their paper does not predict an outcome. What it does, with care and rigour, is map the terrain of what could happen next, and it provides investors with a more durable framework for navigating it.
The Stakes Are Enormous
The scale of the AI buildout warrants serious attention. Adams and Benham note that Alphabet, Meta, Amazon, and Microsoft have collectively spent approximately $1.1 trillion on AI-related investments over three and a half years. Looking ahead, they observe that "another $4 trillion to $8 trillion could be invested in AI-related infrastructure over the next five years." Hyperscaler capital expenditure reached $725 billion in 2026, up 77 percent year over year. Public markets absorbed the SpaceX IPO at roughly $75 billion, the largest on record, and both Anthropic and OpenAI filed confidentially with valuations estimated above $1 trillion. The financing structure supporting all of this is, by any measure, historic in scope.
Four Endings, Not One
The authors resist collapsing this complexity into a simple boom-bust scenario. They separate four distinct end states: a financial break, in which the financing structure fails before the assets it funded generate returns; a rotation, in which capital expenditure proves justified in aggregate but value migrates away from currently favoured companies; a transition, in which a financial break clears the way for a longer phase of falling costs and widespread adoption; and a demand disappointment, in which usage simply fails to grow into the capacity being built, without any accompanying credit event.
Adams and Benham are direct on this point: "These outcomes are not mutually exclusive, and they may occur one after another."
The Frameworks That Matter
Each outcome maps to a different body of economic thought, and the paper works through each with discipline. On the financing question, the authors apply Minsky and Kindleberger, noting the emergence of circular financing arrangements where, for example, Nvidia invests in CoreWeave, CoreWeave uses those proceeds to purchase Nvidia chips, and Nvidia recognizes that purchase as revenue. They write that the same dollar "can be recognized as revenue at more than one point in that circuit," implying that reported revenues may overstate actual demand. Special purpose vehicles, multi-year compute commitments sitting off balance sheets, and aggregate hidden debt estimated at $1.65 trillion complicate the picture further.
On creative destruction, the authors invoke Schumpeter. Chegg is the clearest example. A decade of accumulated proprietary homework solutions ceased to be a defensible asset the moment a general-purpose model could reproduce the function at no marginal cost. "No credit event, financing failure, or leverage was involved," Adams and Benham observe. For legacy companies across the economy, the existential pressure is real. Alphabet, once a disruptor, now "views itself as being in an existential race to reinvent and restructure its business model."
On the question of what follows a break, the authors turn to Carlota Perez, whose historical analysis of five technological revolutions finds a consistent pattern: an installation period, a financial collapse as the turning point, and then a deployment phase of falling prices and broadening adoption. The telecom buildout is instructive. Carriers went bankrupt. The fiber stayed in the ground. The consumer internet of the following decade was built on capacity its original owners could not monetize. What failed, Adams and Benham note, was "the claims on the infrastructure, not the infrastructure itself."
On demand, the authors apply the Jevons paradox, observing that as steam engines became more efficient, coal consumption rose rather than fell. Inference margins at Anthropic have climbed from 38 percent to over 70 percent, and demand remains strong enough that scarcity is showing up in prices. The unresolved question is whether AI tokens, unlike coal, face a demand ceiling once current applications are saturated.
Five Key Takeaways for Advisors and Investors
- Concentration is the common risk across all four scenarios. The companies currently priced as AI's primary beneficiaries occupy an unusually large share of public equity benchmarks, and any reshuffling of the industry registers across broad equity allocations regardless of which path materializes.
- A financial break would not settle the question of AI's economic significance. Infrastructure built ahead of demand has historically been completed and eventually used, even where the companies that built it did not survive.
- Circular financing arrangements between hyperscalers, chip suppliers, and AI laboratories warrant close monitoring, as they may cause reported revenues to overstate genuine end demand.
- A slowdown in capital expenditure is an ambiguous indicator. Read against falling prices and rising usage, it signals deepening deployment. Read against deteriorating credit metrics, it signals financing constraints. Context determines meaning.
- Utilization rates against committed capacity are more informative than token volumes alone, particularly given that the most aggressive capital expenditure plans rest on the proposition that machine-initiated AI demand has no obvious ceiling.
Adams and Benham close with a caution that every investor watching for a single dramatic event would do well to absorb. "Investors watching for a single Minsky moment may find that the more consequential developments unfold across several years and are visible only in aggregate."
Footnote:
1 Adams, Alison, and Frank Benham. "The AI Boom: What Could Come Next." Meketa Investment Group, Sept. 2026, https://meketa.com/leadership/ai-boom-what-could-come-next/.
