Over the past 12 months, the total value locked in AI-related crypto protocols has surged 300%, yet on-chain activity metrics suggest less than 10% of that capital is being used for actual computation. This is not a crash warning—it is a structural signal. The capital is not evaporating; it is rotating. This pattern is not unique to AI. It is the same rhythm I have observed in Layer 2 scaling since 2022: a rolling bubble that moves from infrastructure to execution to application, leaving a trail of misallocated resources and fragile technical debt.
Context: The Rolling Bubble Framework
The source material for this analysis is a recent report by Dhaval Joshi of BCA Research, which argues that AI is not a single monolithic bubble about to burst, but a sequence of rotating mini-bubbles. Joshi identifies four layers: chip infrastructure, base models, developer tools, and applications. Capital flows into one layer, inflates it, then moves to the next as the first layer’s growth narrative cools. The risk is not a sudden crash but a persistent misallocation—money chasing narratives rather than sustainable value creation.
This framework resonates deeply with my own observations in the Layer 2 space. Since 2023, I have audited seven rollup protocols, and the capital rotation pattern is identical. First, the data availability layer (Celestia, EigenDA) attracted massive inflows, with valuations reaching tens of billions. Then, the execution layer (Arbitrum, Optimism, zkSync) took over, followed by cross-chain messaging and interoperability stacks. Now, the capital is rotating toward AI-agent frameworks on L2s. Each rotation leaves behind a layer of infrastructure that is overcapitalized relative to its actual usage.
Core: Parsing the entropy in Layer 2 state transitions
During my 2024 audit of an Optimistic Rollup’s fraud proof mechanism, I discovered a latency vulnerability that could be exploited during high-volatility events. The challenge period was designed for a single state transition, but in practice, the protocol had to handle simultaneous challenges from multiple users. The gas cost of resolving these disputes scaled non-linearly, making the system economically fragile. This is a direct consequence of capital misallocation: the team had spent 80% of its budget on modular DA integration and only 20% on the core fraud proof logic. The result was a system that looked secure on paper but had hidden structural weak points.
This is not an isolated case. In my 2020 DeFi composability audit, I modeled the liquidation risks of leveraged ETH positions on Aave buying UNI on Uniswap V2. The simulation revealed that the capital flow into liquidity mining created a systemic risk that was invisible to individual protocols. The same pattern repeats in L2s today: capital flows into the DA layer because it is the hot narrative, but the actual execution layers remain underfunded. The result is a stack where the most critical components—state verification, sequencer decentralization, censorship resistance—are the least robust.
I have seen this entropy firsthand. In 2022, I spent four months reverse-engineering Celestia’s Data Availability Sampling mechanism. The cryptographic proofs were elegant, but the economic assumptions were fragile. The model assumed that light nodes would always be online to sample, but in practice, node churn during a market downturn could reduce sampling frequency, creating a window for data withholding attacks. The capital flowing into DA projects was based on a theoretical promise, not a verified operational reality. The rolling bubble makes this worse: by the time the vulnerability is discovered, the capital has already rotated to the next layer, leaving the previous layer under-maintained and over-exposed.
Contrarian: The invisible costs of abstraction layers
The conventional wisdom is that rolling bubbles are dangerous because they delay a correction and accumulate systemic risk. That is true, but there is a more insidious blind spot: the abstraction layers themselves become the risk. When capital rotates, the infrastructure built in the previous wave becomes a legacy system that new projects must integrate with, but the original developers have moved on. This creates what I call “spaghetti code of legacy DeFi”—complex integration points that no one fully understands.
In my 2026 zkML prototype, I spent five months building a neural network verification circuit in Circom. The circuit worked, but the computational cost was prohibitive for mainnet deployment. The capital chasing AI-agent narratives on L2s is now pouring into projects that promise to verify AI outputs on-chain, but they are building on top of the same fragile abstraction layers from the previous bubble. The result is a stack of stacked risks: modular DA with unproven economic security, execution layers with latency vulnerabilities, and now verification layers that are computationally impractical. The market is paying for the promise of composability, but the actual cost of abstraction is rarely visible until the next bubble rotation exposes the cracks.
Takeaway: Vulnerability forecast for the next rotation
The next wave of the rolling bubble in L2s will be AI-agent integration. Capital will flow to projects that claim to bridge AI inference with on-chain settlement. But based on my technical analysis, the projects that will survive are those that focus on verification, not hype. The real vulnerability is not in the AI layer itself—it is in the legacy infrastructure that the AI layer will depend on. The DA layer, the execution layer, and the cross-chain messaging layer were all built for a different capital regime. When the bubble rotates again, the projects that over-invested in modularity without proving real demand will be the first to collapse. The question is not whether the bubble will burst, but which layer will be the first to fall and whether the capital that remains will be enough to rebuild the foundation.