The AI Bubble is Not a Bubble. It's a Rotating Carousel of Capital Misallocation.
The market is screaming for a simple narrative. AI is a bubble. It will pop. The end. But the market is wrong. It's also lazy. The reality is far more complex, and far more dangerous for the passive investor who thinks they can just short the QQQ and call it a day.
Based on a recent report from BCA Research strategist Dhaval Joshi, which I've been dissecting all week, the real pattern isn't a single, fragile balloon. It's a rotating carousel of capital misallocation. A rolling bubble. And this distinction, between a bubble and a rolling bubble, is the single most important concept for navigating the next 12 to 18 months.
s fragmented logic.
Context: The Narrative Hunter's Framework
We've been here before. The dot-com era wasn't a single bubble. It was a sequence. Chips (Intel, Cisco) โ Portals (Yahoo, AOL) โ E-commerce (Amazon, eBay) โ Pets.com (the absurdity endpoint). Each phase had its own narrative, its own heroes, and its own brutal crash. The capital didn't leave the tech sector; it rotated. It flowed from one exhausted narrative to the next, inflating a new, equally fragile balloon, until the music stopped.
Joshi's framework applies this same historical lens to the current AI mania. The key insight is that the AI ecosystem is not a monolith. It's a multi-layered stack. The capital is not leaving AI. It's migrating, layer by layer, leaving a trail of overvalued corpses in its wake.
The technical stack is clear: Infrastructure (Chips/Compute) โ Model Layer (Foundation Models) โ Tooling/Middleware โ Application Layer (Vertical Solutions). The capital is currently in motion. The question is not if the bubble will pop, but which layer is currently in the crosshairs, and which layer is about to get pumped.
Core Analysis: The Mechanism of the Rotational Misallocation
Let's get concrete. The "Rolling Bubble" thesis is a structural critique of capital efficiency. It's not a prediction of a market crash tomorrow. It's a diagnosis of a systemic flaw. The flaw is that capital is being allocated based on narrative momentum, not on a sober, risk-adjusted assessment of return on investment (ROI).
Phase 1: The Infrastructure Overhang (2023-2024). This is the most obvious phase. The market fell in love with the "picks and shovels" narrative. Nvidia became a $3 trillion+ company. The hyperscalers (Microsoft, Google, Amazon) committed to a combined CAPEX of over $200 billion. This is the classic "capital misallocation" Joshi warns about. The market priced in 10 years of demand for AI compute in under 18 months.
The unspoken truth here, based on my own audit experience in Prague, is that a significant portion of this compute capacity is currently sitting idle or underutilized. The H100 spot prices have been declining. The ROI on these massive GPU clusters is not yet proven. The capital was deployed based on a narrative of infinite demand, not on a current, verified revenue stream. This is not a collapse. It's a hangover.
Phase 2: The Base Model Squeeze (2024-2025). The capital is now rotating to the model layer. OpenAI, Anthropic, and a dozen other labs are in a desperate race to raise capital at ever-larger valuations. The narrative is "the OS of the future" or "the intelligence layer." But the business model is fragile. The cost of frontier model training is astronomically high. The competitive moat is questionable when the underlying technology (the Transformer architecture) is open-source. The capital is flowing to these entities because the infrastructure narrative is tapping out. The infrastructure story is now "old news". The market needs a new narrative. The model layer is it. But the unit economics are brutal. The capital misallocation is now shifting from compute to the cost of training the next frontier model.
Phase 3: The Coming Application Layer Saturation (2025-2026). This is where the rolling bubble gets interesting. Once the model layer narrative becomes "priced in" and the marginal returns on training bigger models diminish, the capital will rotate to the application layer. We'll see a flood of AI-powered SaaS companies, AI copilots, and vertical solutions. The market will be flooded with "AI-powered" everything. The narrative will be "the revenue is coming." But the reality is that most of these applications will be commoditized wrappers around the same base models. The capital misallocation will shift to customer acquisition costs (CAC) and the race to capture market share in a market that is not yet large enough to support all the players.
This is the core of the Rolling Bubble thesis. It's a series of local, sequential over-investments. The crash is not a single event. It's a slow bleed across different layers of the tech stack.
Contrarian View: The Bubble as a Value Creator
The narrative is always "The bubble will pop and destroy everything." This is too simplistic. The contrarian view, which Joshi's framework implicitly supports, is that a rolling bubble is a value-creating mechanism for the long-term investor, even if it's a wealth-destroying mechanism for the short-term momentum trader.
The capital misallocation in Phase 1 (infrastructure) is creating a massive, over-provisioned physical asset base. This is not a pile of useless Pets.com inventory. It's a fleet of GPUs that will be used for decades. The bubble in compute is effectively subsidizing the creation of the world's most powerful computing network. The cost was inflated, but the asset is real.
The capital misallocation in Phase 2 (models) is funding the frontier of AI research. The billions of dollars being poured into OpenAI and Anthropic are paying for the talent that is pushing the boundaries of what's possible. The bubble inflates the cost of talent, but it also accelerates the R&D timeline.
The real risk is not the individual bubble in a single layer. The real risk is the accumulated systemic fragility. If the capital rotation slows down, or if the macro environment (interest rates, geopolitical risk) hardens, the entire carousel could stop spinning. The local bubbles could de-synchronize and deflate simultaneously. That's the 2000-level crash scenario. But it's not the base case. The base case is a slow, grinding, multi-year rotation of capital from one overvalued sector to the next.
Takeaway: The Call for a New Kind of Investor
The analysis is clear. The investor who buys "AI" as a monolith is a fool. The investor who shorts "AI" is fighting a multi-headed hydra. The only winning strategy is an active, multi-layer, structural approach.
The market is not asking for a bet on AI. It's asking for a bet on the timing of the capital rotation. It's asking for a bet on the ROI of the current layer. The game is no longer about picking the right technology. It's about predicting the narrative cycle.
The next big move? The capital is migrating from the model layer back to the application layer. The narrative is shifting from "build the intelligence" to "package the intelligence." The next 6 months will be about identifying the application layer projects that have real, defensible revenue. Not just a ChatGPT wrapper. The application layer companies that survive the next rotation will be the ones with strong unit economics. The rest will be relics of the carousel.
The market is not irrational. It's just rotating. The only sin is not knowing which way the carousel is turning.
