The market is wrong about the AI bubble.
It is not a single, monolithic explosion waiting to vaporize portfolios. It is a sequence of controlled detonations—a rolling bubble that migrates across the AI tech stack, leaving local wreckage while the broader narrative stays alive. This is the core thesis of Dhaval Joshi, chief strategist at BCA Research, and it deserves far more attention than the lazy 'AI crash imminent' headlines.
Let me be clear: I am not here to defend inflated valuations. I have spent the last 18 years dissecting crypto bubbles—2017 ICOs, 2020 DeFi yield farms, 2021 NFT manias. I know the smell of capital misallocation. But Joshi's framework is intellectually honest. It forces us to abandon binary thinking and embrace a more surgical approach to risk.
Context: The Single-Bubble Fallacy
The prevailing narrative treats AI as a single asset class—buy NVDA, short everything else. But the AI industry is not monolithic. It is a four-layer stack: infrastructure (chips, data centers), models (foundation LLMs), tooling (frameworks, middleware), and applications (enterprise SaaS, agents). Each layer has its own capital cycle, its own hype curve, and its own liquidity dynamics.
From 2023 to 2024, we saw a textbook rotation: first, Nvidia and GPU infrastructure absorbed the bulk of capital. Then, as OpenAI and Anthropic raised billions, the model layer became the hot narrative. By late 2024, the spotlight shifted to application-layer plays like Palantir and AI-native startups. The market did not crash; it shuffled.
This is precisely what Joshi means by a 'rolling bubble.' Capital does not exit the AI theme entirely. It migrates from one layer to the next, leaving a trail of overvalued assets in the first layer while inflating the next. The result is a prolonged period of systemic overvaluation—but not a simultaneous collapse.
Core: The Mechanics of a Rolling Bubble
To understand why this matters, you must map the capital flows. Based on my experience auditing DeFi protocols and institutional portfolios, I have developed a simple framework for tracking bubble rotation: follow the marginal dollar.
In 2023, the marginal dollar went into GPU procurement. Nvidia's data center revenue grew 400% year-over-year. Cloud hyperscalers—Microsoft, Google, Amazon, Meta—collectively spent over $200 billion in capex, most of it on AI infrastructure. The capital was chasing hardware scarcity.
By mid-2024, the marginal dollar began shifting to model training. OpenAI raised $6.6 billion at a $157 billion valuation. Anthropic secured $4 billion from Amazon. The narrative was 'foundation models are the new operating system.' Capital flowed into compute-intensive training runs, not into revenue-generating products.
Today, the marginal dollar is rotating toward applications. AI-native SaaS companies are raising rounds at 50x ARR. Enterprise AI adoption is accelerating, but the ROI is still unproven. The capital is chasing a promise of future cash flows, not current profitability.
Yields are taxes on risk you don't see. In a rolling bubble, the yield on each layer compresses as capital floods in. The risk is not that the bubble pops—it is that the capital misallocation becomes so severe that the next layer cannot absorb the inflow, and the entire system enters a state of 'froth saturation.'
I have built a quantitative model to track this. The key metric is the 'capital efficiency ratio'—the ratio of revenue generated per dollar of capital invested in each AI layer. In the infrastructure layer, this ratio is currently ~0.3x (for every $1 of capex, only $0.30 of revenue is realized). In the model layer, it is even worse: ~0.1x. In the application layer, it is ~0.6x, but falling as competition intensifies.
A rolling bubble continues as long as there is a new layer to absorb the hot money. Once all layers are saturated, the system either enters a 'super-cycle' of compounding speculation or suffers a synchronized correction. The tipping point is when the capital efficiency ratio across all layers drops below the cost of capital. Based on current interest rates, that threshold is approximately 0.25x. We are dangerously close.
Contrarian: The Decoupling Thesis
The conventional wisdom says: 'AI is a bubble, short it all.' But a rolling bubble creates a decoupling effect. One layer can crash while another stagnates, and the overall index may not reflect the pain. This is the contrarian angle that most investors miss.
If the infrastructure layer deflates—say, GPU prices drop due to oversupply—the model layer may benefit from cheaper compute. The model layer's collapse could then lower the cost of building applications, boosting the application layer. The bubble does not burst; it transfers value.
This is not theoretical. In the dot-com era, the semiconductor bubble of 1995-1996 deflated first, then the portal bubble (Yahoo, AOL) burst in 1999, and finally the e-commerce and fiber optic bubbles popped in 2000-2001. Each phase had its own winners and losers. The Nasdaq Composite crashed 78% from peak to trough, but the drawdown was spread over five years of rolling corrections.
Utility is dead. Long live speculation. The AI industry's utility is not in question. The speculation is. The rolling bubble structure means that the market can sustain overvaluation for years, provided that each new 'hot layer' offers a compelling narrative. The true risk is not a sudden crash, but a long, grinding period of low returns for late-stage capital.
For crypto investors, this has direct implications. If AI capital rotates out of infrastructure and models, it may eventually spill into adjacent speculative assets—including crypto. We saw this in 2021 when institutional capital rotated from DeFi to NFTs to gaming. The same pattern could emerge: AI overvaluation leads to a 'search for yield' in crypto markets, especially in tokenized AI compute markets or decentralized GPU networks.
But do not mistake this for a bull signal. It is a liquidity cascade. Eventually, all bubbles converge to the same endpoint: mean reversion. The question is timing, and the rolling bubble framework suggests we have more runway than the permabears assume.
Takeaway: Position for Rotation, Not Collapse
Stop trying to short the entire AI sector. It is a fool's game. Instead, build a rotation map. Track which layer is absorbing the marginal dollar. When the current layer (applications) shows signs of saturation—declining revenue multiples, rising churn, falling capital efficiency—it is time to rotate out and wait for the next bubble to form.
I am not predicting a crash. I am predicting a slow bleed of overvalued layers, each one deflating in turn, while the AI narrative evolves to keep the speculative engine running. The market will not give you a clean exit. It will offer a series of small fires, not a single inferno.
The real question is not whether the AI bubble will burst. It is whether you have the discipline to rotate before the capital does.
The answer will determine whether you survive this cycle, or become another statistic in the liquidity graveyard.