Speculative Growth and the AI "Bubble"

Working paper by Ricardo J. Caballero (MIT and NBER), dated July 15, 2026 (first draft December 2025). It develops a formal macro-finance model for a claim the AI-bubble debate keeps missing: a valuation can be both unsustainable and leave a permanent real legacy. The full mechanism is on speculative-growth; this page is the paper's argument, framing, and reception.

The thesis

High AI valuations are usually read in binary terms — fundamentals or bubble. Caballero's third possibility: because investment responds to valuation, the capital installed during a boom changes the economy that later prices it, so "justified" and "bubble" stop being exhaustive. A temporary overvaluation raises investment; if enough capital is installed before the valuation corrects, the economy is left permanently with a larger capital stock, higher wages, and a lower interest rate. The technology can be real, peak valuations can be unsustained, and the capital installed during the boom can remain.

He opens with Keynes (1931) on 1925–1929: "there can, I think, be no doubt that the world was enormously enriched by the constructions of the quinquennium from 1925 to 1929" — the boom's physical legacy outlasting the boom.

The model

Three blocks:

• q-theory investment — asset prices drive real accumulation, so a valuation wedge moves capital.
• Wealth-in-utility capitalists — a rising propensity to save with wealth (the non-homothetic structure from Straub 2019), so the interest cost of capital falls as capitalist wealth rises. This is the feedback that creates a low-interest-rate high-capital steady state.
• Bayesian beliefs — investors estimate a persistent excess return from a noisy signal, price capital off the posterior mean, and revise as evidence arrives. The baseline path is an optimistic prior that later data don't confirm (equivalently, an adverse ex-post realization under genuine uncertainty about a new technology). The temporary belief wedge supplies the transition force that fades when the data don't confirm it.

The task-based technology (AI capital substitutes for labor tasks, shifting income to capital owners) follows Acemoglu–Restrepo (2018) and Moll–Rachel–Restrepo (2022). The current cycle motivates it: market cap concentrated in AI firms, announced data-center/power/compute investment pointing to large real accumulation.

What's fragile, who pays

The legacy survives only if the correction comes after enough capital is installed — land on the high rational arm and the capital stays; correct too early and it collapses back to the low arm. Distribution follows the same logic: workers gain higher wages at the high-capital destination even as the worker share falls; capitalists finance the belief-supported valuation and eat the correction. The broader lesson Caballero draws: judge an asset-price boom not only by whether current prices match current fundamentals, but by whether the boom finances the capital accumulation that makes the high-valuation outcome self-sustaining.

Where it sits in the wiki

This is a deliberate counterweight to the pure-correction narrative in subprime-ai-crisis and ai-too-expensive — not a denial that a bubble exists or that it pops, but an argument that popping isn't the whole story. It shares the ai-bubble-pale-horses premise (overvaluation, inevitable correction) and diverges on the aftermath. It's also the theoretical companion to hold-on-to-your-hardware: the same capital surge that eats consumer DRAM/NAND is the accumulation that might leave the permanent legacy.

Interesting from the discussion

The HN thread (55 points) split between people engaging the economics and people rejecting the frame:

• The capital-reusability crux dominated. The dot-com fiber analogy (overbuild → crash → capacity eventually used, new owners buy cheap) vs the counter that AI capital is compute for current nodes that depreciates fast, unlike dumb fiber in the ground. Sharpest version (Retric): "20-year-old fiber can have another 10-20 years... most of the AI investment is in compute/manufacturing capacity for current nodes which doesn't age nearly as well." The tulip-mania counter: a glut of post-crash GPUs may find no productive use. This is now the speculative-growth page's "capital-reusability crux" section.
• Energy legacy is contested too. Long sub-thread on whether data-center power buildout leaves useful grid capacity or a stranded glut — one commenter with grid-interconnect detail arguing data centers mostly connect to the grid rather than dedicated plants, so the "new energy infrastructure" legacy is thinner than assumed.
• Animats' historical frame (well-received): three prior speculative overbuilds with good legacies — the dot-com web (overpromotion created the B2C web that wasn't pre-ordained), long-haul fiber (Global Crossing), and by extension the current one. Flagged the paper's line "Workers supply labor, hold no assets, and consume their wage" against the era when most US capital was pension assets backing workers.
• The labor-incidence objection. bluefirebrand: "I don't give a single damn if 'the economy' grows if it means my skills become worthless." dankai: the paper calls workers "protected on the downside" while its model removed the downside risk workers actually face; no mention of taxes.
• The "propaganda" reading. A long thread accusing the paper of dressing up a K-shaped economy (asset owners up, worker share down) as "rational" via math, countered by others explaining multiple-equilibria models properly (multiple states each rational to stay in but impossible to rationally move between) and pointing to Bret Devereaux's high/low-equilibrium ancient-economy essays as the intuition.
• LLM-writing accusations. Several commenters flagged the prose as AI-written — "wedge," "load-bearing," "crux," "spike" as tells — and mocked the pivotal sentence: "If enough capital has been installed before learning removes the wedge... I'm gonna need an honest caveat on the load-bearing assumption here."
• AloysB's readthrough: assumes real ROI on AI labor (still debated), assumes falling rates, and "correction" can be a euphemism for a disastrous financial crisis that takes years — "Do we really need to engineer a financial crisis to build energy facilities?"