AI, bubble or flywheel?

MIT economist Ricardo Caballero proposed a very interesting perspective in his latest working paper, *Speculative Growth and the AI "Bubble"*:
The real question is not whether AI is a bubble, but whether the bubble itself can create future fundamentals.
Traditional finance holds that valuation comes from fundamentals. Future cash flows determine today's price. If the price is far higher than cash flows, it's a bubble. This is the logic shared by almost all value investing, DCF models, and efficient market theory.
Caballero adds a circular causality to this logic. Price not only reflects the future but also shapes the future. High valuations bring financing ability, financing ability brings capital formation, capital formation improves productivity, and productivity ultimately improves future cash flows. Thus, a valuation that initially seemed disconnected from fundamentals becomes part of the formation of future fundamentals (somewhat like Soros's reflexivity?).
The paper argues that when valuation can affect investment, price increases themselves can help create future fundamentals.
The key to this logic holding for AI is that AI is not capital in the traditional sense.
Ordinary capital follows diminishing marginal returns. Building more factories eventually leads to insufficient demand and overcapacity, with lower returns on capital.
But Caballero argues that AI is more like a "labor-like capital" capable of continuous expansion. GPUs, models, and agents do not just increase the number of machines; they constantly increase the effective labor in the entire economy. The paper directly models AI as capital that can perform tasks originally done by labor, so as capital increases, labor capacity expands simultaneously, and the diminishing returns on capital are significantly weakened.
If we dig deeper, there is an even more important finding: AI investment changes income distribution.
More income flows to capital owners, who naturally have a higher propensity to save. Increased savings mean increased long-term capital supply, lower long-term interest rates, and a larger capital stock that is more easily borne by the entire economy. The paper calls this the Funding Feedback. The more capital formed, the lower the future financing cost; the lower the financing cost, the more capital formation is supported. The entire system begins to show positive feedback, rather than the negative feedback in traditional growth models.
Thus, the economy begins to exhibit two completely different long-term equilibria.
In one world, AI investment remains insufficient, capital formation is slow, and productivity stays low for a long time.
In the other world, AI receives sustained financing, with large-scale construction of data centers, GPUs, models, and agents, eventually forming a new high-capital, high-productivity equilibrium.
What's truly interesting is that although the high-capital equilibrium exists, it cannot be reached automatically by rational markets alone. The paper proves that starting from today's low-capital state, even if all investors are fully rational, they will not actively jump to that better future. The reason is simple. Without sufficient capital today, there will be no high future growth; without high future growth, there will be no high valuation today; without high valuation, there will be no capital formation. The entire system becomes self-locked.
The bubble precisely breaks this cycle.
High valuations allow companies to raise funds, build more GPUs, train larger models, deploy more agents, and ultimately truly improve the productivity of the entire economy. The bubble is not the long-term equilibrium but the bridge to it.
This is why the paper repeatedly emphasizes Fragility. The real question is never whether the bubble will burst, but whether it bursts too early. If capital formation has not yet occurred and financing stops, then the entire AI buildout is interrupted, and future growth disappears. If enough data centers, models, agents, and infrastructure have been completed before the bubble bursts, then even if valuations eventually return to normal, the high-capital equilibrium can still be maintained. The paper explicitly states that the key is not whether a correction occurs, but whether the correction occurs too early.
The internet is a typical example. The dot-com bubble burst completely in 2000, but the fiber optics, servers, software, data centers, and internet talent all remained. The bubble disappeared, but the internet revolution truly began. AI is likely to go through a similar process, except what remains is not just the network but intelligence itself.
However, I think Caballero's framework can be pushed one step further.
The paper models AI as "replicable labor," but in reality, AI is increasingly approaching "replicable researchers." If AI can not only replace labor but also participate in research, write code, design chips, discover new materials, and develop new models, then it changes not just the production function but the innovation function.
In the past, innovation capacity mainly depended on the number of scientists, engineers, and talented individuals. Therefore, major technological revolutions usually required decades of accumulation, which is why Kondratiev waves have persisted for so long. It's not that the economy naturally experiences a revolution every sixty years, but that innovation resources themselves grow too slowly.
AI is breaking this constraint for the first time.
Future innovation capacity will no longer rely solely on the human brain; it may be Human + Millions of AI Agents. Going further, innovation capacity might even rely solely on AI (compute power).
As compute power continues to grow, innovation capacity also continues to grow. Innovation has become a capitalizable, scalable production factor for the first time.
If we combine this with today's rapidly developing Coding Agents, Research Agents, automated research, and recursive self-improvement (RSI), the feedback becomes even stronger. More AI brings faster research, faster research produces better models, and better models continue to improve research efficiency, forming a true Intelligence Flywheel. The speed of innovation itself begins to accelerate, not just production efficiency.
This is why I have always believed that the economic returns from AI are likely to follow a "Slowly, Then Suddenly" pattern.
What everyone sees today is GPU investments, model training, and data center construction, with ROI not looking very high, leading many to question whether AI is a bubble. But what these investments are truly buying is not today's profits but future intelligent capital. When model capabilities cross a certain threshold, large-scale agents enter enterprises, labor substitution begins, and productivity may experience a nonlinear leap. The seemingly excessive valuations of the past few years will then begin to materialize.
This means that the feedback loop proposed by Caballero:
Valuation → Investment → Capital Formation → Fundamentals
May evolve further into:
Valuation → Investment → Compute → Intelligence → Innovation → More Ideas → Higher Productivity → Higher Profits → Higher Valuation
What truly forms positive feedback here is not just capital, but the entire society's innovation capacity.
If this process holds, then the change brought by AI may not be just another technological revolution, but a change in the mechanism by which technological revolutions themselves arise.
Historically, the Kondratiev long waves lasted forty to fifty years, largely not because of economic laws, but because innovation resources were always scarce: limited scientists, limited R&D capacity, slow knowledge diffusion. AI is changing this premise.
In the future, we may not see increasingly shorter Kondratiev waves, but rather multiple industrial revolutions emerging continuously on the same AI platform: AI drugs, AI materials, AI chips, AI robots, AI biomanufacturing... Innovation becomes industrialized, and technological revolutions begin to occur consecutively.
If Schumpeter made innovation the core of growth, and Romer made knowledge the core of growth, then RSI and Caballero together point to the core proposition of the next stage of growth theory:
Schumpeter's earlier business cycle theory relied on creative destruction, which depended on human brains and occasional geniuses. AI, for the first time, makes such genius itself a form of capital that can be invested in, mass-produced, continuously enhanced, and self-reinforcing.
From this perspective, no matter how large the current bubble, it may be quickly absorbed in the face of exponential innovation growth.