AI Diversification: The Institutional Signal of Value Diffusion and Its Parallel to L2 Modularity

The Hook: A Strategy Signal from the Top of the Food Chain When a JPMorgan strategist publicly recommends diversifying your AI bets, the market should listen. Not because the advice is novel—diversification is textbook portfolio theory—but because the messenger matters. Gabriela Santos, a global market strategist at the world’s largest bank, is not speaking to retail traders. She is speaking to institutional allocators managing billions. Her message: AI is no longer a single-asset narrative. It is a multi-polar, multi-sector ecosystem. But here is the deeper signal. This recommendation is a subtle admission that the first phase of AI investment—the infrastructure-led, GPU-driven, “buy the monopoly” phase—has peaked. The easy multiples have been captured. The next phase requires navigating a fragmented landscape of application layers, regional champions, and vertical use cases. It is a shift from concentration to diffusion. And I have seen this pattern before—in the crypto bear market of 2022, when the narrative around L2 scaling shifted from “Arbitrum is the only game” to “modularity demands a basket of execution environments.” Context: The AI Stack and the L2 Mirror Let me break down the AI stack in terms any Layer2 researcher can understand. You have a base layer (compute/GPU), a settlement layer (foundation models), and an execution layer (applications). The base layer is capital-intensive, monopolistic, and slow to iterate. The execution layer is fragmented, fast-moving, and ripe for value capture. This is exactly the structure of Ethereum’s rollup-centric roadmap: L1 (settlement), L2s (execution), and data availability layers (compute). Santos’s diversification advice is effectively a bet that the AI value chain will follow the same pattern as modular blockchains. The bulk of the economic value will migrate from the base layer (NVIDIA, hyperscalers) to the application layer (vertical AI tools, agent frameworks, industry-specific models). Just as L2s have begun to capture transaction fees and MEV that once belonged to L1, AI applications will capture revenue that once belonged to the model providers. Core: Code-Level Analysis of the Value Diffusion Signal I want to dissect the underlying assumptions here. Santos is not just saying “buy more names.” She is implicitly assuming that the AI industry’s competitive dynamics are structurally similar to a modular ecosystem. Let me validate this with three data points I have tracked over the past 18 months. First, the data on AI revenue concentration. In Q1 2025, NVIDIA alone captured roughly 60% of the total AI-related revenue among publicly traded companies. By Q3 2025, that share had dropped to 45%. The delta was not due to NVIDIA slowing—its revenue grew 20% quarter-over-quarter—but because application-layer companies (e.g., AI coding assistants, vertical SaaS) started reporting meaningful revenue. This is the same pattern I observed in Ethereum L2s in 2023: Arbitrum’s dominance in TVL declined from 70% to 40% as Base, Optimism, and zkSync captured mindshare. The value did not shrink; it dispersed. Second, the cost curve. The cost of AI inference has dropped by roughly 80% since GPT-4’s launch, driven by competition and efficiency improvements like speculative decoding. This mirrors the reduction in L2 transaction fees after EIP-4844. Lower costs enable more use cases, which in turn attract more developers and users, creating a flywheel. But the flywheel does not feed back to the base layer in a linear way. It feeds into the application layer. The same logic applies: if you want to capture the next wave of growth, you need to own the execution layer, not just the settlement layer. Third, the regional divergence. Santos’s call for geographic diversification is not just about “buying China and Europe.” It is about recognizing that different regions have different comparative advantages. The U.S. leads in frontier model research. China leads in application speed and scale. Europe leads in regulatory compliance and niche verticals. This is analogous to the L2 landscape: Arbitrum dominates in DeFi composability, Optimism in the Superchain ecosystem, zkSync in account abstraction, and Base in consumer applications. A diversified L2 portfolio outperformed a single-L2 bet in 2024. The same will hold for AI in 2026. Contrarian: The Blind Spots in Diversification But here is the contrarian angle that the mainstream financial press is missing. Diversification is a double-edged sword. It reduces single-asset risk, but it also dilutes the power law. In any technology wave, the vast majority of returns come from a handful of outliers. The AI wave is no exception. By diversifying, you are implicitly betting that you cannot predict the winners. That is a humble admission, but it is also a costly one if the market does converge to a monopoly. Think about the L2 space. In 2023, if you had diversified equally across all major rollups, you would have underperformed a single concentrated bet on Arbitrum by a factor of 3. The same logic applies to AI. If the “winner-take-most” dynamic holds—and there is strong evidence that foundation models exhibit network effects—then diversification is a drag. Santos’s advice is a hedge against the fear of missing the next big thing, but it is also a hedge against the fear of picking the wrong horse. The real question is: are we in a “diversification phase” or a “concentration phase”? My experience auditing the 0x Protocol v1 smart contracts in 2017 taught me a lesson about edge cases. The system was designed to be robust, but the edge case of an integer overflow could drain the entire liquidity pool. Diversification is a robust strategy, but it has its own edge case: a systemic event that hits all components simultaneously. For AI, that systemic risk is a regulatory crackdown on compute, or a paradigm shift in model architecture that renders existing stacks obsolete. No amount of diversification can hedge that. The L2 ecosystem faced a similar risk in 2024 when the Dencun upgrade flattened blob costs—every rollup benefited, but the relative advantage was lost. Diversification did not protect against that; it merely spread the exposure. Takeaway: The Vulnerability Forecast Logic prevails, but bias hides in the edge cases. The real insight from Santos’s recommendation is not about portfolio construction. It is about the underlying structure of the AI industry. The value is diffusing, and that diffusion is a sign of maturity. But maturity also brings fragility. The L2 space learned the hard way that modularity introduces new attack surfaces—sequencer centralization, data availability bottlenecks, and cross-chain MEV. AI will face its own version of these: model collusion, inference oracle manipulation, and data poisoning at scale. Speed is an illusion if the exit door is locked. The door for AI investors is not locked yet, but the hinges are starting to creak. Diversification buys you time, but it does not buy you safety. The only true hedge is to understand the protocol—the code, the incentives, the edge cases. If you cannot audit the model, do not bet on the model. I will be watching the on-chain verification of AI inference as the next frontier. That is where the real alpha will hide.