The Rolling AI Bubble: A Structural Audit of Capital Misallocation

The ledger does not forgive. Over the past 12 months, the top four cloud providers burned through $200 billion in AI capital expenditure. Yet the revenue generated from AI services barely covers 30% of that outlay. The data shows a pattern that defies the simple narrative of a single AI bubble about to burst. Dhaval Joshi, chief strategist at BCA Research, calls it a "rolling bubble." I call it a structural misalignment of capital flows across the AI tech stack. The market is not pricing a uniform crash. It is pricing a sequence of localized overvaluations that rotate through infrastructure, models, tools, and applications. This is not a story of euphoria. It is a story of deferred reckoning. Joshi’s thesis is grounded in a simple observation: the AI industry is not a monolith. It is a stack. At the bottom, you have the physical layer: GPUs, data centers, networking gear. Above that, the model layer: foundational LLMs, training pipelines. Then the tooling layer: frameworks, middleware, orchestration. Finally, the application layer: chatbots, code assistants, vertical SaaS. Each layer has its own capital cycle, its own valuation dynamics, and its own timeline for ROI. The market does not inflate all layers simultaneously. It rotates. In 2023, the spotlight was on NVIDIA and the infrastructure layer. In 2024, it shifted to model companies like OpenAI and Anthropic. Now, in early 2025, the narrative is pivoting toward applications—Palantir, Salesforce, and a crowd of AI-native startups. This rotation is not organic. It is a symptom of capital chasing the next narrative before the previous one has delivered. Trust nothing. Verify everything. I have spent the past three years auditing smart contracts and scaling protocols. I have seen the same pattern in DeFi: a liquidity cycle that inflates one protocol, then moves to the next, leaving the first with a hangover of unproductive TVL. The AI rolling bubble is no different. The critical metric is not the total market cap of AI stocks. It is the capital efficiency of each layer. Based on my forensic analysis of public filings and private data, the infrastructure layer currently exhibits the widest gap between CAPEX and earned revenue. NVIDIA’s data center revenue hit $47 billion in fiscal 2024, but the actual utilization of rented H100 compute is highly variable. Spot prices for GPU compute have dropped 40% since Q3 2024. That is a classic sign of oversupply. The market is building compute capacity faster than applications can consume it. That is capital misallocation. The model layer is equally fragile. Foundation model companies are spending billions on training runs, but the revenue from API calls and licensing is insufficient to cover the cost. OpenAI’s annualized revenue is estimated at $3.4 billion, yet its operating expenses—including compute, talent, and infrastructure—exceed $7 billion. The gap is closed by venture capital and strategic investments. That is a subsidy, not a business model. The tooling layer, meanwhile, is overcrowded. There are over 200 AI middleware startups, most of them burning cash to acquire users. The application layer is the only one that shows signs of genuine product-market fit, but even there, retention rates are inconsistent. The average 30-day user retention for AI chatbots is below 15%. The data is clear: the rolling bubble is not a sustainable cycle. It is a deferral of reconciliation. Contrary to popular belief, the rolling bubble does not reduce systemic risk. It amplifies it. Each rotation creates a new set of financial commitments that are built on the assumption that the next layer will justify the previous one. This is a recursive dependency. If the application layer fails to generate sufficient revenue to justify the model layer’s valuation, the model layer’s capital base will collapse. That will cascade back to the infrastructure layer, which will then face a sudden drop in demand. The market is currently pricing a smooth transition. The historical record suggests otherwise. The dot-com bubble was not a single event. It was a series of rolling mini-bubbles: semiconductors (1998), portals (1999), e-commerce (2000), and optical networking (2001). Each one collapsed in sequence, but the overall market crash was delayed until the last layer failed. The same dynamics are at play now. Complexity is the enemy of security. The more layers, the more points of failure. What does this mean for investors? The rolling bubble model implies that shorts will be punished until the rotation stops. You cannot short the entire AI sector and expect a linear payoff. The market will always find a new narrative to inflate. But the risk is not evenly distributed. The infrastructure layer is the most vulnerable to a correction because it is the most capital-intensive and the least directly tied to end-user demand. The application layer, on the other hand, may be the most resilient—if it can demonstrate real ROI. The real opportunity is not in timing the crash. It is in identifying which layers have the strongest fundamentals when the rotation pauses. That requires a granular, data-driven approach. I have built a framework that tracks three metrics: capital efficiency (revenue per unit of CAPEX), utilization (real demand vs installed capacity), and retention (user stickiness). The current signal is clear: the infrastructure layer is flashing red. The model layer is yellow. The application layer is green, but only for a handful of companies with proven unit economics. Let me give you a concrete example from my own work. In early 2024, I architected a smart contract system for a DeFi yield aggregator. We designed an oracle aggregation mechanism to prevent flash loan attacks. The key lesson was that layers of abstraction hide risk. The more layers between the capital and the actual use case, the harder it is to audit the true state. The AI rolling bubble is the same. The distance between the GPU sale and the end-user value is three layers deep. Each layer adds latency, cost, and uncertainty. The market is currently ignoring that distance. It is relying on a narrative that AI will eventually solve every problem. That narrative is not backed by data. The ledger does not forgive. Here is the contrarian angle: the rolling bubble might actually be the best possible outcome for the AI industry. A single, catastrophic crash would destroy capital and talent, setting back progress by years. A rolling bubble allows for gradual correction, forcing capital to flow to the most efficient layers. It is a form of Darwinian selection. The survivors will be the companies that can demonstrate real product-market fit, not the ones that ride the narrative wave. The market is already showing signs of this. The AI startup funding in Q1 2025 dropped 25% year-over-year, but the deals that closed were larger and later-stage. The capital is concentrating. The weak are being weeded out. This is healthy. The risk is that the correction does not happen fast enough. If the next rotation fails to materialize—if the application layer disappoints—the entire structure could still collapse in a synchronized manner. The probability of a synchronized crash is low in the short term, but rising over a 12-month horizon. What specific signals should you watch? First, the GPU spot price. If it continues to decline, infrastructure demand is softening. Second, the revenue growth of leading model companies. If OpenAI’s growth slows below 50% year-over-year, the model layer is in trouble. Third, the user retention numbers for AI applications. If the average D30 retention stays below 20%, the application layer is not delivering value. All three signals are currently trending negative. The market is pricing a continuation of the rotation. The data suggests a pause. The question is not whether the bubble will burst. The question is which layer will be the first to crack. My reading of the Joshi thesis is that the market is in a state of strategic denial. The rolling bubble is a narrative that allows investors to stay long without confronting the underlying capital misallocation. It is a comfortable story. But the ledger does not care about comfort. The data will eventually force a reconciliation. The only way to prepare is to audit each layer independently. Trust nothing. Verify everything. Look at the raw numbers. The infrastructure layer is overbuilt. The model layer is overvalued. The tooling layer is overcrowded. The application layer is underdeveloped. The rolling bubble is not a permanent state. It is a transition. The transition will end. When it does, the winners will be the protocols and companies that have built real, provable, and sustainable value. The rest will be written off as a capital misallocation. Complexity is the enemy of security. The AI stack is complex. The bubble is complex. The solution is simple: focus on fundamentals. Revenue per dollar of CAPEX. Utilization rates. Retention. These are the metrics that matter. Ignore the narrative. The data will tell you when the roll ends.