AI: Is the Bubble Actually the Flywheel? MIT economist Ricardo Caballero makes a fascinating argument in his recent working paper, Speculative Growth and the AI "Bubble": The real question is not whether AI is a bubble, but whether a bubble itself can create the fundamentals of the future. Traditional finance assumes that valuations are derived from fundamentals. Future cash flows determine today's prices. If prices rise far above expected cash flows, we call it a bubble. This logic underpins value investing, discounted cash flow (DCF) models, and much of the Efficient Market Hypothesis. Caballero extends this causal relationship into a feedback loop. Prices do not merely reflect the future—they help shape it. High valuations increase firms' ability to raise capital. That capital finances investment. Investment builds productive capacity. Higher productivity eventually generates stronger future cash flows. In other words, valuations that initially appear detached from fundamentals can become part of the process that creates those very fundamentals. (This bears some resemblance to George Soros' idea of reflexivity.) The paper argues that whenever market valuations influence investment decisions, rising prices can actively help create future economic fundamentals. The key reason this mechanism may hold for AI is that AI is fundamentally different from traditional capital. Conventional capital is subject to diminishing returns. Build more factories, and eventually demand saturates, excess capacity emerges, and returns on capital decline. Caballero argues that AI is better understood as a form of scalable labor-like capital. GPUs, foundation models, and AI agents do not merely add more machines—they expand the economy's effective labor supply. In his framework, AI capital performs tasks that would otherwise require human labor. As AI capital accumulates, productive labor capacity expands alongside it, substantially weakening the traditional law of diminishing returns. The paper goes even further. AI investment also changes the distribution of income. A growing share of income flows to capital owners, who tend to save a larger fraction of their earnings. Higher savings increase the supply of long-term capital, push down long-term interest rates, and allow the economy to sustain a larger capital stock. Caballero calls this the Funding Feedback: more capital formation lowers future financing costs, and lower financing costs encourage even more capital formation. Instead of the negative feedback embedded in standard growth models, the system begins to exhibit positive feedback. This leads to two fundamentally different long-run equilibria. In one world, AI investment remains insufficient. Capital accumulates slowly, and productivity growth stays persistently weak. In the other, AI continues attracting abundant financing. Massive investments flow into data centers, GPUs, foundation models, and AI agents, ultimately producing a high-capital, high-productivity equilibrium. The intriguing part is that although this superior equilibrium exists, rational markets may never reach it on their own. Caballero shows that starting from today's low-capital equilibrium, even perfectly rational investors may fail to coordinate on the better outcome. The logic is circular: without enough capital today, future productivity cannot accelerate; without higher future productivity, today's valuations remain subdued; without high valuations, firms cannot finance the necessary investment. The economy becomes trapped in a self-reinforcing equilibrium. This is precisely where the bubble matters. Elevated valuations allow firms to raise capital. That capital finances more GPUs, larger models, and more autonomous agents. Those investments eventually raise the economy's productive capacity. The bubble is not the destination. It is the bridge. This is also why the paper repeatedly emphasizes fragility. The real danger is not that the bubble eventually bursts. The danger is that it bursts too early. If financing dries up before sufficient AI infrastructure has been built, investment stalls, AI development slows, and the expected productivity gains never materialize. But if enough data centers, compute infrastructure, models, and AI agents are already in place before valuations normalize, the high-capital equilibrium can sustain itself even after the speculative premium disappears. The timing of the correction matters far more than the correction itself. The Internet provides a classic example. The dot-com bubble collapsed spectacularly in 2000. Yet the fiber-optic networks, servers, software, data centers, and engineering talent remained. The bubble disappeared, but the Internet revolution had only just begun. AI may follow a similar path. The difference is that what survives this time may not simply be digital infrastructure—but intelligence itself. Going One Step Further I believe Caballero's framework can be extended even further. His paper models AI as replicable labor. In reality, AI is increasingly becoming replicable researchers. If AI can not only perform labor but also conduct scientific research, write software, design chips, discover new materials, and invent better AI models, then it changes not merely the production function—it changes the innovation function itself. Historically, innovation has depended on the number of scientists, engineers, and exceptionally talented individuals. As a result, major technological revolutions have typically taken decades to unfold. This is one of the fundamental reasons behind the long duration of the Kondratiev waves. The economy does not naturally produce a technological revolution every fifty or sixty years. Rather, innovation resources themselves have historically expanded very slowly. AI may be the first technology capable of breaking this constraint. Future innovation will no longer depend solely on human intelligence. Instead, it may become the combined output of humans plus millions of AI agents. Eventually, much of it may even be driven primarily by AI itself, powered by ever-expanding compute. As computational capacity continues to grow, so does the economy's ability to innovate. For the first time, innovation itself becomes a production factor that can be capitalized, scaled, and continuously expanded. Now combine this with the rapid progress of coding agents, research agents, autonomous scientific discovery, and Recursive Self-Improvement (RSI). The feedback loop becomes dramatically stronger. More AI accelerates research. Faster research produces better models. Better models further accelerate research. This becomes a genuine Intelligence Flywheel. The rate of innovation itself begins to accelerate—not merely the efficiency of production. "Slowly, Then Suddenly" This is why I have long believed that AI's economic payoff is likely to follow the pattern of "slowly, then suddenly." Today, investors mainly see spending on GPUs, model training, and data centers. The return on investment appears modest, leading many to conclude that AI is simply another bubble. But these investments are not primarily buying today's profits. They are purchasing tomorrow's intelligence capital. Once model capabilities cross certain critical thresholds, AI agents begin operating throughout enterprises, labor substitution accelerates, and productivity could experience a highly nonlinear jump. At that point, valuations that once appeared excessive may suddenly look entirely justified. Caballero's original feedback loop is: Valuation → Investment → Capital Formation → Fundamentals I suspect AI may ultimately evolve into something even more powerful: Valuation → Investment → Compute → Intelligence → Innovation → More Ideas → Higher Productivity → Higher Profits → Higher Valuations The object generating positive feedback is no longer just capital. It is society's entire capacity to innovate. If this process proves correct, AI will represent more than another technological revolution. It will fundamentally change how technological revolutions themselves are generated. Historically, Kondratiev long waves lasted forty to fifty years not because economics demanded such timing, but because innovation resources were scarce: scientists were limited, R&D capacity expanded slowly, and knowledge diffused gradually. AI is changing those assumptions. Instead of progressively shorter technological cycles, we may witness multiple industrial revolutions unfolding simultaneously on top of a common AI platform: AI-driven drug discovery AI-designed materials AI-created semiconductors AI-powered robotics AI-enabled biomanufacturing ...and many more. Innovation itself becomes industrialized. Technological revolutions become continuous rather than episodic. If Schumpeter made innovation the engine of economic growth, and Romer made knowledge the engine of growth, then RSI and Caballero together may be pointing toward the next frontier of growth theory: Schumpeter's economic cycles depended on disruptive innovation, and disruptive innovation depended on human intelligence—and occasionally, on rare geniuses. AI may be the first technology that turns genius itself into a form of capital: something that can be financed, replicated at scale, continuously improved, and ultimately capable of improving itself. Base on this argument, no matter how large today's AI bubble appears, exponential growth in innovation may allow the economy to absorb it far more quickly than most people expect.

SoftBank CEO Masayoshi Son dismissed AI bubble fears, comparing skeptics to early critics of cars and airplanes. He estimates it will take nearly $5 trillion in global investments annually to build out the necessary data centers, chips, and energy infrastructure for AI. Get the latest news on all things AI:

"Situationship" is my $BTC bull porn essay on how the AI bubble will burst, and why the money printer will go hyper brrrr and take us back to a rip roaring bull market. "The question of internal framing is the key variable that determines whether AI is a bubble. But before we understand the bubbliciousness of AI, the better question to ask is in what kind of businesses are we investing, or in relationship terms, “What are we”? The dichotomy, at least to my Luddite brain, is the distinction between what investors believe AI CAPEX represents: is it technology or real estate? The zeitgeist is that this multi-trillion dollar build out is “technology” and must receive insane growth multiples by the market. But I believe AI CAPEX is just another boring real estate play. However, in this instance, what’s inside the data center is compute that creates silicon-based lifeforms that will aid human civilizational development in the most profound way since the railroads. The distinction between real estate and compute is important because the recently post-pubescent hedge fund bro, the bank, the private credit fund, and ultimately the government believes funding the building of a data center and power plant is like lending to Apple, instead of lending to Lehman Brothers. The bursting of the AI bubble will occur because financial intermediaries, tacitly supported by the US and Chinese governments, will over build data centers and everything that goes into providing the substrate to house chips that train AI models and conduct inference. Therefore, the AI bubble is a credit story like 2008 and not an earnings story like 2000.[1]" Link to my substack and the full article on my bio.

🚨 BREAKING: BRIDGEWATER FOUNDER RAY DALIO JUST SAID LIVE: "THE AI BUBBLE WILL BURST EVENTUALLY. THE ONLY QUESTION IS WHEN." THIS GUY PREDICTED CRASH IN 2008 AND MADE $15.8 BILLION HE KNOWS SOMETHING BAD IS COMING...