# Speculative growth

The bubble question is usually binary: do high prices reflect fundamentals, or are they a bubble? Speculative growth is a third answer. It takes the fundamentals as *not* given — investment responds to valuation, so the capital installed during a boom changes the economy that later prices it. Once prices shape the fundamentals against which they're judged, "justified" and "bubble" stop being exhaustive: a valuation can be unsustainable and still leave a permanent real legacy.

The name and the modern formalization come from Ricardo Caballero's [[speculative-growth-ai-bubble]] (MIT, July 2026), building on the earlier speculative-growth mechanism of Caballero, Farhi, and Hammour (2006), where asset values, funding conditions, and accumulation reinforce one another.

## The mechanism

Three objects the bubble question runs together, separated:

1. **Is the technology productive?**
2. **Are peak valuations sustained?**
3. **Does the capital financed at those valuations remain?**

Speculative growth is the case where (1) is yes, (2) is no, and (3) is yes — the technology is real, the peak prices are not, and the capital installed at those prices stays.

The engine is a **wealth-saving feedback that creates multiple steady states.** New capital (AI capital that performs tasks previously done by labor) expands capacity and shifts income toward capital owners, who save more. More saving lowers the interest rate consistent with a larger installed capital stock. Past a threshold this produces two equilibria: a low-capital state and a self-sustaining high-capital state with a permanently lower long-run interest rate.

The catch: **rational pricing from the low-capital state stays on the low-capital path.** A high-capital steady state exists, but no rational transition reaches it — the economy won't climb there on fundamentals alone. The missing force is a **temporary belief-supported overvaluation:** investors perceive high returns, valuation rises, investment accelerates, and capital moves toward the region where the high-capital economy sustains itself. (During the transition, investment demand actually *raises* the interest rate; only the destination has a lower one.)

## The fragility

The overvaluation must eventually correct. The whole legacy hinges on timing:

- **Correction late** (enough capital installed): valuation falls back to the *high-capital* rational path, and the capital remains. The boom paid for a permanent upgrade.
- **Correction early** (learning removes the perceived high branch too soon): the perceived high-capital path disappears, valuation crashes to the low arm, and the transition collapses. The capital doesn't stick.

This is the load-bearing assumption, and the one critics single out — "enough capital installed before learning removes the wedge" is doing all the work, with no guarantee the real world hits that window.

## Distribution

Same logic drives who wins and who pays. **Workers gain** through capital deepening — they operate with a larger conventional capital stock, so wages rise *even as the worker share of income falls*. **Capitalists carry the exposure** — they finance the boom at belief-supported prices that later correct, and compress their own consumption to do it; their realized return nets the destination gain against the price correction.

The distributional framing is where the model's assumptions bite hardest: it treats workers as holding no assets and consuming their wage, and as "protected on the downside" — a modeling choice critics note removes the downside risk workers actually face (job loss, no asset cushion), and ignores that historically much US capital was pension assets backing workers.

## The capital-reusability crux

Whether speculative growth applies to a given boom reduces to one empirical question: **does the installed capital retain value after the correction?** The reference case is the dot-com fiber overbuild — dumb glass in the ground, cheap to upgrade by swapping endpoint equipment, eventually used for exactly its imagined purpose; the builders went under and new owners bought the capacity at a profit-supporting price. Long-haul fiber, railroads, and the airline industry are the standard analogies.

The AI-specific worry is that its capital may age like tulips, not fiber: GPUs and current-node compute depreciate fast, cost gigawatts to run, and hit a cost cliff where new hardware beats free old hardware. Data centers and their power/cooling/grid buildout are the durable part; the silicon may not be. If the "higher capital" only has value to AI and crypto workloads and those bust, the legacy evaporates. This is the open question the mechanism can't settle on its own.

## Cross-references

- [[speculative-growth-ai-bubble]] — the paper
- [[ai-bubble-pale-horses]], [[subprime-ai-crisis]], [[ai-too-expensive]] — the pure-correction / doom framing this nuances rather than contradicts
- [[hold-on-to-your-hardware]] — the second-order hardware-supply effect of the same capital boom