Last week, as Nvidia's market cap flirted with $3.5 trillion, a parallel universe unfolded: a cohort of AI startups saw their secondary market valuations slashed by 40%. The divergence isn't random noise. It's the signature of a rolling bubble—a phenomenon that Dhaval Joshi, chief strategist at BCA Research, has framed with uncomfortable precision. In a recent report, Joshi warns that the AI sector isn't heading toward a single, cataclysmic bust. Instead, it's undergoing a sequence of localized overheatings, each cooling before the next ignites. This is not a market that will break in one blow; it will fracture in pieces, and the pieces will keep moving.
Joshi's framework challenges the binary narratives dominating crypto and tech circles. The popular consensus oscillates between 'AI will conquer all' and 'the bubble is about to pop.' Joshi offers a third path: a structural, cyclical decay that spreads risk across time and layers. For crypto analysts like myself, who have spent years dissecting narrative cycles in DeFi and NFTs, this resonates deeply. The AI bubble is not a single supernova; it's a series of smaller flares, each consuming its own fuel before passing the torch.
Context: The Architect of the Rolling Bubble
Dhaval Joshi is not a fringe voice. BCA Research has advised institutional investors for over five decades, and Joshi's track record includes prescient calls on macro trends. His current thesis rests on a simple observation: AI capital expenditure is concentrated in discrete layers of the technology stack. Infrastructure (GPUs, data centers) saw the first wave of euphoria, followed by foundation models (OpenAI, Anthropic), then tooling (development frameworks), and finally applications (enterprise solutions). Each layer attracts capital until it reaches a saturation point—where marginal returns on investment collapse—and then the narrative shifts to the next layer.
This is not a new pattern. The internet bubble of the late 1990s unfolded in a similar cascading sequence: semiconductors, then portals, then e-commerce, then fiber optics. Each sub-bubble inflated and deflated in turn, but the aggregate market didn't crash until the final layer—telecom infrastructure—collapsed under the weight of its own hype. Joshi argues that AI is following the same playbook, but with a compressed timeline due to faster capital cycles.
Core: The Mechanics of Rotating Overvaluation
To understand the rolling bubble, we must dissect the four layers of the AI stack and the capital misallocation within each.
Layer 1: Infrastructure – The GPU and data center frenzy. Nvidia's market cap alone exceeds the combined GDP of many nations. Cloud providers (Microsoft, Google, Amazon, Meta) have committed over $200 billion in cumulative CAPEX for AI compute. The problem? Utilization rates remain opaque. Industry sources suggest that a significant portion of rented GPU capacity goes unused, especially after the initial training phases. Based on my audit experience in blockchain infrastructure, I've seen similar patterns: projects overspend on hardware to signal credibility, leading to ghost capacity. The 'capital misallocation' Joshi warns about is most visible here. The ROI of a GPU cluster is not guaranteed; it depends on sustained demand for inference, which is still nascent.
Layer 2: Foundation Models – The race to be the 'best' LLM has created a winner-take-most dynamic, but with diminishing returns. OpenAI, Anthropic, and Google DeepMind are burning billions on training runs, while smaller players struggle to differentiate. The narrative has shifted from 'model size' to 'model efficiency,' but the capital locked in training runs is enormous. When the hype cycle moves to Layer 3, these model companies will face a funding crunch. Trust is a variable, not a constant—especially when investors start asking about unit economics.
Layer 3: Tooling & Middleware – Frameworks like LangChain, vector databases, and MLOps platforms saw a surge in 2023-2024. But many of these tools are wrappers around existing technologies, with low switching costs. The capital influx inflated valuations without durable moats. As the hype rotates to applications, tooling will be the first to deflate, as it lacks the narrative 'stickiness' of end-user products.
Layer 4: Applications – This is the current frontier. Companies like Palantir, Salesforce Einstein, and a swarm of AI writing assistants are now the darlings of the market. Yet, the revenue figures remain modest compared to the infrastructure spend. The 'capital misallocation' here is the gap between promise and execution. I recall a conversation with a startup founder who admitted that 80% of their AI-powered customer service bot's calls were still handled by humans. The code whispers truths only the silent can hear—and the silence is the absence of real ROI.
Joshi's framework implies that the bubble will not collapse uniformly. Instead, it will deflate layer by layer, with each layer's decline cushioned by the next layer's ascent. This creates a 'rolling floor' that prevents a total market meltdown but extends the period of overvaluation. The hidden risk is that the rolling process masks the accumulation of systemic fragility. When the final layer—applications—fails to deliver the promised returns, there will be no next layer to catch the fall. That is when the rolling bubble turns into a single, synchronized crash.
Contrarian: The Delusion of Managed Decline
The conventional wisdom is that a rolling bubble is safer than a single bubble. It allows for gradual price discovery, and the infrastructure built during the boom might have lasting value—like the fiber optic cables from the 2000s, which eventually carried the internet traffic of the 2010s. This is a comforting narrative, but it's also a trap.
First, the 'residual value' argument for AI infrastructure is weaker than it appears. GPUs depreciate rapidly, and data centers have high energy costs. If the application layer fails to generate enough demand, the idle GPUs will become stranded assets, not productive capital. The railroad bubble of the 19th century left behind useful tracks, but it also wiped out generations of investor wealth. The same could happen here.
Second, the rolling bubble creates a 'liquidity mirage.' Because each layer attracts new capital, the overall market remains liquid, preventing a panic. But this liquidity is shallow—it's rotating from one sector to another, not expanding the total pool. When the rotation stops, the entire market will feel the vacuum. In the red, I found the quiet signal: the VIX has remained low, but cross-asset correlations are rising. That's a warning that the rolling machine is about to run out of gears.
Third, the crypto market's relationship with AI adds another layer of complexity. Crypto native investors are increasingly drawn to AI tokens (e.g., Render, Fetch.ai, Bittensor). These tokens are a bet on the application layer or the compute layer. If the AI bubble rolls from infrastructure to applications, these tokens might see a temporary boost. But if the underlying infrastructure value collapses, the entire crypto-AI ecosystem could suffer a contagion. The narrative that 'AI will flow into crypto' is a double-edged sword: it might attract capital, but it also ties crypto's fate to a broader market dysfunction.
Takeaway: Navigating the Next Wave
For the disciplined investor, the rolling bubble offers both opportunity and peril. The key is to track the rotation signals, not the absolute valuation. Watch for the inflection points: when Nvidia's data center revenue growth decelerates, capital will flow to models; when model funding rounds tighten, it will flow to applications. Each transition is a chance to rebalance.
But the ultimate question remains: what happens when the last layer—applications—fails to deliver the 'killer use case' that justifies the entire pyramid? History suggests that the rolling bubble will eventually become a standing wave that crashes. The only defense is to maintain liquidity and avoid the trap of anchoring to a single layer. To hold firm is to understand the void—the space between narratives where value is measured not by hype, but by the quiet ledger of real utility.
The AI bubble is not a binary event. It's a process, a sequence of overconfidence and correction. Joshi's framework gives us the map, but the terrain is shifting. The wise trader will not fight the rolling tide; they will ride its waves, knowing that every wave eventually breaks on the shore.