The latest earnings call from a major cloud provider painted a picture of relentless AI capex—$60 billion in a single quarter, with no sign of slowing. The market yawned. Then, a whisper from BCA Research's Dhaval Joshi: this isn't a bubble about to burst. It's a rolling bubble, migrating across AI's tech stack like a wildfire that refuses to die. For crypto natives, this isn't just a macro signal—it's a mirror. We've seen this movie before. Mapping the chaos to find the signal in the noise.
Context: The Narrative Cycle of Bubbles
In the summer of 2020, I was neck-deep in Compound Finance's yield curves, tracing the eToken interest rate models across five chains simultaneously. The 'money lego' narrative was still a whisper. But I saw the pattern: capital flowing into DeFi, inflating yields, then rotating to the next hot protocol. That was a rolling micro-bubble, contained within crypto. Today, AI is doing the same on a macro scale. The tech stack splits into four layers: infrastructure (GPUs, data centers), models (LLMs like GPT-4), tooling (frameworks, middleware), and applications (agents, SaaS). Capital sloshes from one layer to the next, leaving a trail of inflated valuations and eventual corrections.
Joshi's insight is that AI's bubble isn't a single balloon ready to pop; it's a sequence of smaller bubbles, each rolling forward as the next narrative takes hold. This is precisely the structure we saw in the 1990s internet mania, from semiconductors to portals to e-commerce. In crypto, we saw it with DeFi summer → NFT mania → Layer2 hype. The map is not the territory, but the story is. And the story now is that AI's bubble is alive, just moving.
Core: The Mechanical Dance of Capital Misallocation
From the ashes of Terra, we learned to walk. That collapse taught me that capital misallocation isn't a bug—it's a feature of early-stage narratives. When I reverse-engineered Arbitrum's fraud proof mechanism after the crash, I saw how resilient infrastructure could survive the narrative's death. AI's rolling bubble works the same way. The current capital misallocation is concentrated in infrastructure: NVIDIA's $3 trillion market cap, hyperscalers' $200B+ annual capex. But the ROI on those GPUs? Still murky. H100 spot prices have dropped 30% in Q2 2025, signaling oversupply. Yet the narrative hasn't rotated yet.
Here's the core mechanism: each layer of the AI stack has a 'narrative half-life.' Infrastructure's peak was 2023-2024, driven by the 'scaling law' story. Now, as scaling laws hit diminishing returns, the narrative is shifting to 'agent economies'—AI agents autonomously transacting on blockchains. This is where crypto enters the frame. I've been digging into three protocols: Fetch.ai, a new Tokyo-based startup, and SingularityNET. The 'agent-to-agent' micro-transaction narrative is dry brush waiting for a spark. But the capital hasn't fully rotated yet. Based on my audit of on-chain data, the volume of AI-related token transactions on Ethereum L2s has grown 4x since January, but the real inflow is still in the hands of VCs, not retail. The bubble is still in the 'smart money' phase.
Contrarian: The Crypto Beneficiary Theory
Conventional wisdom says that if AI bubble pops, it drags crypto down with it. I'm skeptical. A rolling bubble doesn't burst all at once—it deflates locally. When the infrastructure layer corrects, capital doesn't vanish; it migrates. Where? To the next story that promises asymmetric returns. Crypto narratives—especially those tied to AI (DePIN, agent platforms, compute markets)—are perfectly positioned to absorb that liquidity. Think of it as a 'narrative spillover.' In 2022, when Terra collapsed, capital didn't leave crypto; it rotated into ETH and L2s. The same pattern repeats.
My contrarian angle: the AI bubble's rolling nature actually creates a 'flight to complexity' effect. As simple narratives (buy NVIDIA) peak, sophisticated investors seek alpha in harder-to-value assets. Crypto's AI tokens, with their messy tokenomics and uncertain roadmaps, are the perfect hunting ground. But there's a trap: most of these projects have zero revenue. The signal is in the code—check if the team has a working product on L2, not just a whitepaper. When the crowd jumps, I look for the net.
Takeaway: The Compass for the Next Rotation
So where are we in the cycle? I track three signals: (1) NVIDIA's data center revenue growth rate—if it drops below 50% YoY, the infrastructure narrative is done; (2) the number of AI agent transactions on L2s—currently 120k/day, needs to hit 1M for 'agent economy' to be real; (3) the spread between AI token prices and their development activity on GitHub. My current read: the infrastructure layer is 80% of the way through its bubble, the model layer is 60%, and the application layer is just 20%. The next spark? Watch for a major AI agent platform launching on a high-throughput L2. That's when the crypto-native AI narrative will catch fire.
Rebuilding the compass after the storm passes. The rolling bubble means we have time—but not for complacency. Stories drive value, not just algorithms. The algorithm today says: position in AI-agent infrastructure on L2s, hedge with short positions on overvalued GPU proxies. The narrative says: the next wave is coming, and it speaks smart contract. Are you listening?