The Rolling AI Bubble: A Structural Audit of Capital Flows and the Coming Rotation

I trace the shadow before it casts. Over the past seven days, the AI narrative has shifted again. Nvidia's stock dips 3% on a whisper of slower GPU demand. Meanwhile, a new crop of AI-agent tokens on Solana surge 40%. The market is not crashing. It is rotating. And beneath the surface, a pattern emerges that every DeFi auditor knows by heart: when liquidity migrates, vulnerabilities follow. Dhaval Joshi, chief strategist at BCA Research, recently warned that the AI bubble is not a single supernova about to collapse. It is a series of rolling bubbles, each one inflating, then deflating, as capital moves from one layer of the tech stack to the next. This thesis, reported by Crypto Briefing, resonates with a specific truth I encountered during the 2017 ICO audit of Ethlance. Back then, I spent six weeks line-by-line reviewing a token distribution contract. The integer overflow I found would have drained the treasury. The fix was elegant, but the lesson was structural: value migration is never smooth. It leaves traces. Joshi's rolling bubble model maps neatly onto the AI stack: infrastructure (GPUs, data centers), base models (LLMs), tooling frameworks, and application layers. Each layer acts like a liquidity pool in a DeFi protocol. When one pool's yield drops, capital flows to the next. But unlike a well-designed AMM, the AI market lacks transparent invariants. The result is capital misallocation—a term Joshi uses, but one that carries a specific weight for anyone who has watched a stablecoin depeg. Before I dive into the core analysis, let me ground this in context. The AI industry's capital expenditure has reached staggering levels. In 2024, Microsoft, Google, Amazon, and Meta collectively spent over $200 billion on AI infrastructure. That number is projected to grow. Yet, the revenue from AI products—while growing—remains a fraction of that spend. This is not necessarily a problem. In the early days of the internet, fiber-optic cable was laid far ahead of demand. But the key difference is that fiber was a passive asset. GPUs are active. They consume energy, require cooling, and depreciate. The capital misallocation is not just about dollars; it is about the ongoing cost of maintaining a machine that may not yet have found its purpose. My 2020 deep dive into Curve Finance's stableswap invariant taught me that even the most elegant system can hide fragility. I wrote a Python script to simulate 10,000 arbitrage attacks. The invariant held, but only because the design assumed a certain liquidity distribution. The AI bubble's rolling structure assumes a similar equilibrium: that capital will always find the next hot layer. But what if the next layer fails to materialize? What if the yield on base models collapses before the application layer is ready to absorb it? Here is the core technical observation. The rolling bubble thesis implies a sequence of valuations that are not independent. Each layer's valuation is a function of the previous layer's hype. Infrastructure is priced on the assumption that base models will need exponentially more compute. Base models are priced on the assumption that applications will generate billions of users. Applications are priced on the assumption that the infrastructure is already built. This is a circular dependency. In smart contract security, we call this a reentrancy lock—a call that depends on the state of another contract before it updates its own. If one contract fails, the entire chain reverts. I have seen this pattern before. In 2021, I reviewed the random seed entropy for an Art Blocks generative art project. The block hash dependency had a predictability flaw. The artist was grateful I found it privately. That experience taught me that the most beautiful systems often conceal the simplest vulnerabilities. The AI bubble's beauty is its narrative of continuous progress. But the vulnerability is that each layer's valuation is a call on the next layer's success. When the call fails, the bubble doesn't pop everywhere at once. It deflates layer by layer, like a series of interconnected price oracles feeding the same protocol. Let me be precise. The capital misallocation risk is not evenly distributed. Infrastructure—GPUs, data centers—has a tangible asset backing. Even if demand softens, the chips can be repurposed. The 2022 Terra Luna collapse taught me that the most dangerous assets are those that cannot be unwound. UST's luna minting mechanism created a phantom liquidity that evaporated when the peg broke. I built a simulation model of the de-pegging. It showed that the fragility was structural, not market sentiment. Similarly, the AI bubble's most fragile layer is likely the one with the least tangible value: the pure API model providers. They compete on margin, have no switching costs, and depend on continuous capital infusion. When the bubble rolls away from them, they will not have a GPU to sell. They will have a pricing model that no longer works. This is the contrarian angle. The common narrative is that the AI bubble is a single entity, and its collapse will take everything down with it. Joshi's rolling model suggests a different outcome: a gradual, sector-by-sector recalibration. But the hidden risk is that the rolling process itself accelerates. In DeFi, we see liquidity crunches when multiple pools are drained simultaneously. The same can happen here. If the infrastructure layer's growth slows, it sends a signal to the model layer, which then triggers a reassessment of the application layer. The roll becomes a cascade. My 2025 work on AI-agent security frameworks gave me a front-row seat to this interconnection. We designed a code-stasis verification layer for autonomous on-chain transactions. The key insight was that AI hallucinations could lead to unintended interactions. The bubble's rolling is itself a form of collective hallucination—each layer convincing itself the next layer will justify the current valuation. When the hallucination stops, the stasis breaks. What does this mean for the crypto market? The Crypto Briefing article that reported Joshi's views is itself a signal. Crypto media covers AI bubble narratives because their audiences are attuned to risk and reward in speculative assets. The rolling bubble thesis offers a nuanced framework: AI is not a bubble to short, but a series of opportunities to time. For crypto investors, this implies that capital may rotate from AI to crypto when the AI bubble cools. We saw a hint of this in early 2025 when AI token narratives surged as GPU stocks wobbled. The rotation is already happening. But there is a deeper structural implication. The AI bubble's rolling nature mirrors the fragmentation of liquidity in cross-chain protocols. Every new chain worsens the problem. Every new AI layer inflates the bubble. The solution is not to build more layers, but to create a unified invariant—a way to measure the true value of each layer relative to the others. In DeFi, we have AMMs and lending protocols. In AI, we have no such mechanism. The capital allocation is driven by narrative, not by a transparent pricing oracle. I listen to what the compiler ignores. The compiler ignores the emotional state of the developer. The market ignores the fragility of the rolling structure. But the signs are there. The GPU rental spot price has been declining for three months. The number of new AI startups has plateaued. The funding rounds for base model companies are getting harder to close. These are the static signals. The pulse is in the static. What is the takeaway? The rolling AI bubble is not a call to panic. It is a call to position. The coming rotation will create winners and losers. The winners will be the assets that have real utility independent of the next layer's hype. The losers will be the ones that depend on constant capital inflow to sustain their valuation. For crypto, this means that AI-linked tokens tied to actual infrastructure (GPU compute, decentralized storage) may survive the rotation better than pure AI agent tokens. The latter are more like the model layer—priced on promise, not on proof. Vulnerability is just a question unasked. The question we should ask is: what happens when the roll stops? The answer is not a crash, but a quiet redistribution. The capital will find a new home. The question is whether that home is prepared for the arrival. Finding the pulse in the static. The static is the noise of billions in capital chasing the same narrative. The pulse is the underlying data—the real utilization rates, the actual revenue per GPU, the churn of AI app users. These metrics are the invariants. They are the audit trail. I trace the shadow before it casts. The shadow is the next layer's valuation. The body is the capital that will move there. In the void, the bytes whisper truth. The truth is that the rolling bubble is a natural market mechanism. It is not a bug; it is a feature of a capital-intensive, fast-evolving technology. But features have vulnerabilities. And vulnerabilities need auditors. The market is the protocol. The narrative is the smart contract. The capital is the transaction. We are all just watching it execute. The question is: will the code hold? Security is the shape of freedom. The freedom to rotate capital without losing value. The shape of that freedom is transparency. The AI industry needs its own version of a formal verification—a way to prove that the next layer's valuation is grounded in the previous layer's reality. Without that, the rolling bubble will eventually become a crashing wave. Not today. Not tomorrow. But when the macro tide turns, the roll will become a fall. I trace the shadow before it casts. The shadow is already here. The next rotation will be from AI to crypto. The only question is whether the crypto market is ready to absorb the liquidity without breaking its own invariants. The answer, as always, lies in the code.